EPISODE 2026-06-25

AI:AM LIVE — June 25, 2026 — Surviving on Top of the Frontier Labs: Eric Olson & Tricia Martinez

The opening tracked a frontier-economics news cycle: the first legal test of the export-control blockade on Anthropic's Fable 5 and Mythos models, with legal-tech firm Legion suing the Trump administration over its authority to force the models offline; Tom Brown replacing Dario Amodei as Anthropic's lead voice in Washington; Snowflake CEO Sridhar Ramaswamy's GLM 5.2-versus-Opus 4.7 benchmark thread and what a near-frontier open-source model means for SaaS margins; Anthropic's accusation that Alibaba used nearly 25,000 fraudulent accounts to extract 28.8 million Claude exchanges to train Qwen, and the deeper question of whether distillation spreads Anthropic's alignment work as a side effect; and a wave of Google DeepMind departures alongside ban-superintelligence calls from MIRI's Nate Soares and Francis Fukuyama. Eric Olson — CEO and co-founder of Consensus, the AI research engine over 200M+ peer-reviewed papers used by 2.5M+ people a month — then joined to locate where 'AI for science' actually stands and to answer the question every AI investor keeps asking: do application-layer companies built on top of the labs get steamrolled, or compound? He detailed Consensus's 'recipes' architecture, its routing of narrow tasks to sub-billion-parameter BERT-series models, a roughly days-to-hours speed-up for researchers, and why a price-discrimination gap between API and consumer pricing is navigable but not unlimited. Tricia Martinez — founder and CEO of Dapple — followed on sovereign, single-tenant, in-country AI infrastructure; the company says it booked over $100M in enterprise contracts in its first five months on a $30M seed, and Tricia made the case for an asset-light, software-first 'Enterprise OS Cloud' against pointed questions on its 91–94% utilization claim, 6-to-9-month deployment timelines, and Azure-native dependency. The hosts closed by debating whether a company like Dapple is a market maker or a defensible control plane, what software is worth when agents are the primary buyers, and the value of human time and attention in a hyper-deflationary software world.

▶ Full show on YouTube𝕏 Live broadcast

The June 25 show opened on a frontier-economics news cycle. The export-control blockade on Anthropic's Fable 5 and Mythos models reached its first legal test, with legal-tech firm Legion suing the Trump administration over whether it ever had the statutory authority to force the models offline — and Nathan reading the move through the administration's pattern of fast unilateral action followed by back-channel negotiation. A personnel shift framed the conversation: Tom Brown, Anthropic's chief compute officer, has replaced Dario Amodei as the company's lead interlocutor in Washington, with early feedback reportedly far warmer to his build-everything framing. The hosts then turned to GLM 5.2, the Chinese open-source model that drew a detailed Opus 4.7 comparison thread from Snowflake CEO Sridhar Ramaswamy, and to Anthropic's accusation that Alibaba used nearly 25,000 fraudulent accounts to extract 28.8 million Claude exchanges to train Qwen — opening a debate over whether large-scale distillation might, paradoxically, spread Anthropic's alignment characteristics across the field. They closed the opening on a wave of Google DeepMind departures and the rising 'ban superintelligence' chorus from MIRI's Nate Soares and Francis Fukuyama.

Two application-layer founders followed. Eric Olson — CEO and co-founder of Consensus, an AI research engine over a corpus of 200M+ peer-reviewed papers used by more than 2.5 million people a month — joined to locate where AI for science really stands and to take on the platform-risk question head-on: do companies built on top of the frontier labs get steamrolled, or compound? Then Tricia Martinez — founder and CEO of Dapple — made the case for sovereign, single-tenant, in-country AI infrastructure, defending a company that says it booked over $100 million in enterprise contracts in its first five months on a $30 million seed. The hosts stayed on after each guest to pressure-test the claims — on routing and price discrimination after Eric, and on whether Dapple is a market maker or a durable control plane after Tricia — before previewing the next day's guests, Cameron Berg on AI welfare research and Forum AI's Robbie Goldfarb on bringing human taste into AI systems.

The rundown

  1. 9:14Opening29 min
    Opening: Export Controls in Court, GLM 5.2, and the Distillation DebateThe first legal test of the Fable 5 / Mythos export-control blockade, with Legion suing the Trump administration; Tom Brown replacing Dario Amodei as Anthropic's lead in Washington; Snowflake CEO Sridhar Ramaswamy's GLM 5.2-versus-Opus 4.7 benchmark and the SaaS margin squeeze; Anthropic's accusation that Alibaba extracted 28.8M Claude exchanges via ~25,000 fraudulent accounts to train Qwen, and whether distillation spreads alignment as a side effect; plus Google DeepMind departures and ban-superintelligence calls.

    Nathan and Prakash opened mid-thought on the economics of large-scale AI infrastructure buildout, noting that as long as inference demand holds each successive generation of investment pays back, even if per-generation margins narrow slightly.

    The dominant story of the morning was the Fable 5 / Mythos export-control blockade entering its first legal test. A legal technology company called Legion filed suit against the Trump administration, arguing the government lacked statutory authority to force Anthropic to withdraw its most capable models. Prakash walked through overnight developments: a brief reappearance of the Fable 5 endpoint in Amazon Bedrock's model listing had fueled speculation before being quickly dampened — Anthropic's head of growth confirmed the models remain unavailable to all customers. Nathan catalogued additional signals pointing toward an eventual return — a model-name string quietly reintroduced in a Claude Code update, rising prediction markets — and offered a structural read on the Trump administration's governing pattern: rapid unilateral moves followed by back-channel negotiation, leveraging what he called 'Schrödinger's authority' to extract concessions before courts ever weigh in. On the legal merits, Nathan cited commentary from the Scaling Laws podcast suggesting a strong statutory argument that current export-control law simply does not reach API-delivered AI services.

    A notable personnel development framed the conversation: Anthropic co-founder Tom Brown has replaced Dario Amodei as the company's lead interlocutor in Washington, with early administration feedback reportedly warm — Brown's infrastructure-and-compute framing aligning naturally with the Trump team's build-everything posture. Prakash captured the moment with a joke circulating in DC circles: it was as if someone gestured at Dario and said, 'that's just my quant — leave him in the background.' Nathan reasoned that whatever path Anthropic took through the 'penalty box,' it has ended up exactly where it needed to be: in the room as the rules governing future model releases are being written.

    The hosts turned next to GLM 5.2, a Chinese open-source model that had circulated widely following a detailed benchmarking thread from Snowflake CEO Sridhar Ramaswamy comparing it against Opus 4.7. Prakash framed the stakes: SaaS companies are under severe margin pressure as AI spending competes within fixed IT budgets, and a model merely 'in the ballpark' of Anthropic's best represents genuine relief. Nathan praised the thread's intellectual honesty — it catalogued real limitations, including a tendency for GLM 5.2 to spin out into excessive tool calls on failing agentic runs, narrowing the apparent cost advantage — while stressing that the threshold effect of being competitive matters far more than precise performance gaps. The Zapier founder added context: GLM 5.2's 23% ARC-AGI score sits right at the inflection point where frontier closed models previously saw agentic capability take off, and no open-source model has yet crossed 25% — putting the open frontier roughly six to nine months behind.

    That naturally led into the distillation debate. Anthropic formally accused Alibaba of using nearly 25,000 fraudulent accounts to extract 28.8 million Claude exchanges for training Qwen, and Prakash described a rumored gray-market ecosystem in China where resellers maximize tokens per account and wholesale the outputs at roughly 70% discounts off US pricing. Nathan walked through his evolving view: the raw-data problem is tractable with money, but — as Zvi Mowshowitz has pushed him on — distillation is doubly efficient because it also transfers Anthropic's hard-won taste and alignment characteristics, not just raw capability. Prakash offered a provocative reframe: for a safety-first organization built on a 'race to the top,' widespread distillation might actually advance the mission, spreading Anthropic's alignment characteristics to Chinese labs that, in practice, behave as pragmatic utilitarian models — 'with some Chinese characteristics,' as Nathan quipped. Nathan suggested Anthropic could formalize this dynamic by releasing public alignment datasets, a public-good move he found surprisingly absent.

    The segment closed on mounting talent attrition at Google DeepMind — Arthur Kami and Rishab Jain the latest names to depart — alongside calls for a superintelligence ban from MIRI's Nate Suarez and, notably, political scientist Francis Fukuyama. Nathan acknowledged the signal is strengthening while maintaining confidence in top leadership — Jeff Dean, Shane Legg, and Demis Hassabis. Prakash transitioned directly into the introduction of the day's first guest, Eric Olson, CEO of Consensus, an AI-powered scientific research platform serving over 2.5 million users.

    This is definitely a pattern from the Trump administration. We see a very quick move followed by rapid negotiation or some sort of walkback. Whether it was actually a legally defensible maneuver in the end doesn't really matter, because they're able to use the Schrödinger's authority that they have to compel some action, get somebody to the table, extract some concession, and a deal gets worked out before the legal process ever has time to run its course.

    The feedback on the Washington side has been: he's so much better. He's not a freak. We can talk to him and he listens. The joke going around is that someone just did that to Dario — 'that's just my quant, leave him in the background.'

    With some Chinese characteristics.

    Legion sues the Trump administration to restore Fable 5 — the first legal test of the export-control blockade. Legal-tech firm Legion filed suit arguing the government lacked statutory authority to force Anthropic to withdraw Fable 5 and Mythos. A brief reappearance of the Fable 5 endpoint in Amazon Bedrock's listing stirred return speculation before Anthropic's head of growth confirmed the models remain unavailable to all customers. Nathan read the administration's pattern as fast unilateral moves leveraging 'Schrödinger's authority' to extract concessions before courts weigh in.

    Tom Brown replaces Dario Amodei as Anthropic's lead voice in Washington. Co-founder and chief compute officer Tom Brown has taken over the Washington conversation, with reportedly far warmer feedback from the administration — his infrastructure-and-compute, build-everything framing aligning naturally with the Trump team's posture. Nathan argued Anthropic ended up where it needed to be regardless of the path: in the room as the rules for future model releases are written.

    GLM 5.2 vs Opus 4.7: a near-frontier open-source model and the SaaS margin squeeze. Snowflake CEO Sridhar Ramaswamy posted a balanced GLM 5.2-versus-Opus 4.7 thread; Nathan praised its honesty (flagging GLM 5.2 spinning out into excessive tool calls on failing runs) while stressing the threshold effect of being competitive. The Zapier founder noted GLM 5.2's 23% ARC-AGI score sits near the agentic-takeoff inflection point, putting the open frontier roughly six to nine months behind.

    Anthropic accuses Alibaba of distilling Claude at scale to train Qwen. Anthropic alleged Alibaba used nearly 25,000 fraudulent accounts to generate 28.8 million Claude exchanges between April and June. Citing Zvi Mowshowitz, Nathan noted distillation is doubly efficient — transferring Anthropic's taste and alignment, not just raw capability — prompting Prakash's reframe that for a 'race to the top' safety org, widespread distillation might actually spread alignment characteristics to Chinese labs.

    Google DeepMind attrition and the rising 'ban superintelligence' chorus. Researchers Arthur Kami and Rishab Jain are the latest to depart Google DeepMind; Nathan acknowledged a strengthening signal while keeping confidence in Demis Hassabis, Jeff Dean, and Shane Legg. MIRI's Nate Soares and 'End of History' author Francis Fukuyama both called for banning superintelligence.

    Lightly edited · timestamps jump to YouTube
    9:14

    Nathan Labenz: Demand is there — there's not that much slack. They're making money on each phase of the buildout, and each subsequent phase gets bigger. Maybe the economics get a little less favorable from one generation to the next, but the scale gets bigger too. The question of where it stops is a good one. But if it were to stop, as long as inference demand is still there, they'll be fine. It all pays back as long as inference demand holds.

    9:51

    Prakash: Inference demand certainly seems to be there. There's a shortage of chips in the market, a shortage of people who are already set up, and a shortage of qualified operators — as we've seen when trying to use inference providers that don't adhere to certain standards. Let me segue quickly to what happened overnight. We are now seeing the Fable 5 / Mythos export-control saga in court, finally. A company called Legion, a legal tech company, is suing the Trump administration, arguing the government lacked the legal authority to force Anthropic to shut off Fable and Mythos 5. We had expected to see this. There were some green shoots — on Amazon Bedrock there was a sign that the Fable 5 endpoint had been reopened and was getting listed again in available models, though it wasn't available for access. Then Anthropic's head of growth, Amol Asavari, stepped in a few hours ago to say that no, Fable is not available for anyone. So it remains blocked. We have the first legal challenge at this point.

    11:30

    Nathan Labenz: There was a discovery in a Claude Code point update — a string reintroduced referencing Fable using more credits, some kind of operational message to users — which had people chattering that it might be coming back. The prediction markets went up. The reporting is that Dario has stepped back from this, and Tom Brown, Anthropic's chief compute officer, is leading the Washington conversation. Multiple signals suggest it might soon return.

    On the case itself: it's first of all great marketing for a legal tech company to jump in and sue. Questions over whether they have standing — I'm not qualified to answer. But it sounds like they've made a decent argument on standing, and a pretty strong argument that the law as written just doesn't apply to technologies of this shape. The argument is that you can't do this to an API under current law. From what I've seen from the Scaling Laws podcast hosts, Alan Rosenstein and Kevin Fraser, they think it's a pretty compelling case.

    I'm reminded of what Dean told me in our conversation last week: don't read this as a principled move or a declaration of long-term policy from the administration — similar to Judd's comment about having some cognitive empathy for them. He basically said: I think they freaked out based on perhaps an imperfect understanding of what was happening. But given all the warning messages they'd received over time, you can't entirely fault them for getting alarmed. They reached for the tool that would get the thing off the market regardless of whether they really believed it would be long-term defensible policy.

    This is definitely a pattern from the Trump administration, time and again. We see a very quick move followed by rapid negotiation or some sort of walkback. Whether it was actually a legally defensible maneuver in the end doesn't really matter, because they're able to use the Schrödinger's authority that they have to compel some action, get somebody to the table, extract some concession, and a deal gets worked out before the legal process ever has time to run its course. That still seems like the safe bet here.

    14:53

    Prakash: We'll see. I remember we were giving over/under odds on end of July, beginning of July for Fable coming back. Barely five or six days left now, so I'm not that hopeful. I'm more of the mind that it will be another model that gets released — not the Fable model. They'll dress it up as something else.

    One note from yesterday: the Washington lead for Anthropic is no longer Dario. Tom Brown, co-founder of Anthropic, has taken over as the lead for speaking to Washington. And the feedback from the Washington side has been: he's so much better. He's not a freak. We can talk to him and he listens. The joke going around is: you know how sometimes you go in as a two-person team to a VC firm and one person says something odd or disconcerting, and the CEO turns around and says, 'Oh yeah, that's just my quant'? The joke is that someone just did that to Dario — 'that's just my quant, leave him in the background.' That's the joke going around at this point.

    16:28

    Nathan Labenz: It's the smartest of times, it's the stupidest of times. It's hard to read. There's still a story you could tell where this is all a galaxy-brained maneuver from Anthropic that adds up to them getting what they want. I highly doubt this is how they wanted it to go down — it really doesn't seem that way. But as you noted, they are now in the room as much as anyone could ever be in the room as the details of how new models will be allowed to come online are being hashed out.

    Maybe they had to get there through the penalty box, but maybe that's where they wanted to be regardless. Tom Brown doesn't have much of a public presence, but being the chief compute officer presumably has him talking a language the Trump administration is inclined to hear. He's the buildout guy, the we-gotta-solve-the-energy-bottleneck guy, and they're very onboard with those things. Relatively subtle reframings that aren't even substantively that different could just get people vibing a little more with what Anthropic's trying to do. Probably a wise move to make a substitution.

    18:15

    Prakash: If you go in saying 'we need to build faster' rather than 'we need to slow this thing down,' that automatically flips the conversation. You're talking about doing the same thing — making sure the models are safe — but the context you inject it into and the priorities of the administration do matter. Washington is what it is.

    Let me segue. GLM 5.2, a model from Zhipu AI, has been making the rounds. Over the last 24 to 48 hours we've seen the Snowflake CEO Sridhar Ramaswamy post a GLM 5.2 versus Opus 4.7 breakdown. Worth noting: Ramaswamy was formerly the technical lead on Google Search for many years. About a week ago he gave an interview saying SaaS firms are in survival mode — if they don't figure it out, they will die, including his own firm. Most leaders paper over that. He just said it. And you can see that even he is now posting about how an open-source model compares to Opus 4.7, because the absolute margins being taken by the model firms are very significant and very painful to other SaaS providers. Enterprises budget AI out of the IT bucket, so more spend on AI means less spend on everything else.

    20:41

    Nathan Labenz: I read that whole thread — it was very well done and very balanced. This was not a hype thread. Broadly, Opus 4.7 comes out better, all things considered. There are details around how much you're really saving, because GLM 5.2 is actually a little more token-hungry and makes a lot more tool calls in cases where it's not on the path to success — it seems to spin out and fire off tons of tool calls and burn a bunch of tokens that way. Those subtleties suggest the headline cost advantage isn't quite as compelling as it initially looks.

    But that's also kind of burying the lede. The high-level summary is: it's not far behind, they consider it a viable alternative, and the threshold effect of being in the game versus not being in the game is probably a lot more important than any of these caveats. The caveats actually boost the credibility of the report. When I see someone say 'oh, it's just as good,' I think: was it really? Probably not. When I see these guys say 'it's not just as good — here are the subtle ways we measured it — but we still think it's basically viable and we're going to use it,' that is a very credible report from a company I think is shooting straight. It is a big deal.

    22:49

    Prakash: It is. We've had other comments worth flagging. One is from the founder of Zapier, who is also a major donor of the ARC-AGI project. He noted that GLM 5.2's 23% ARC-AGI score is right on the border of the agentic takeoff we saw with Opus 4.5 and GPT in Q4 2025. Crossing 25% was pivotal for other frontier closed models, and to date no open-source model has crossed it. So he's saying we're somewhere between six and nine months off the frontier time frame.

    23:56

    Nathan Labenz: This really begs the question of how much they're keeping up due to distillation. I don't think anybody outside of Anthropic has a great sense of that. The fact that GLM 5.2 seems to think it's Claude in some way is an interesting indicator — there's some amount of riding the wake of the American frontier companies going on. But how much? And at what point do they have a good enough model to run their own recursive self-improvement loop? Is there a threshold beyond which it becomes less important to have access to the latest Claude to distill from — and you can do a scalable oversight, constitutional recursive self-improvement feedback loop on your own, as Anthropic is already doing? It seems like they're probably getting close to that point.

    25:19

    Prakash: Let me drop in exactly that. Anthropic yesterday accused Alibaba of continuing to distill Claude on a large scale to train Qwen — specifically, nearly 25,000 fraudulent accounts generating 28.8 million Claude exchanges between April and June. Billions of tokens distilled out in total. Anthropic says this is part of a broader pattern of adversarial distillation where Chinese labs harvest outputs from US frontier models to train rival systems at a fraction of the cost.

    Right in that distillation zone. There are also rumors: Claude accounts in China are offered at a steep discount compared to abroad. The reason is that Chinese vendors set up systems that continually maximize every token in the context window per account, then resell that token data to companies like Alibaba. They maximize tokens per account and resell to multiple buyers — allowing them to offer Claude access at something like a 70% discount to US pricing. A $200 Claude account there would go for $50. Which is insane when you think about it.

    27:20

    Nathan Labenz: The value of the data is pretty high. The alternative is hiring humans to generate it, which is a lot more expensive. I've gone back and forth on this. In some moments I feel like the data problem is the easiest to solve — if you said 'Nathan, here's a gigawatt data center, go train a frontier model,' I'd have a lot of work ahead of me. But if you said 'here's a billion dollars, go get 28 million traces of decent reasoning,' that seems a lot more tractable. Just a matter of money, not anything too galaxy-brained.

    But Zvi pushes back on me on this. He makes the point that distillation is doubly efficient: you're not just taking a shortcut on the data, you're taking a big shortcut on taste — all the subtle aspects of what makes Claude good. Someone has already done all the work to zero in on the desired behavior, and you get that for free. That is pretty compelling. Would Fable make a big difference if you could distill from it? I'm sure it would. But it does seem like they have enough — or if not yet, they're very close to a point where they can start their own synthetic data flywheel that depends much less on Claude going forward. I'd still expect there's enough capability already captured in the GLM stack that cutting them off would not stop their progress.

    30:16

    Prakash: It strikes me that Anthropic, as a safety organization, should almost want other people to distill off their models. The entire idea behind Anthropic was this race to the top Dario talks about — you introduce safety features that other firms are forced to adopt because it's the prosocial thing to do, and all the researchers will want to work on this more prosocial method of creating AI. If you let people shortcut and distill off your models, you're aligning the other models by default — they pick up all of your taste and value characteristics. And that is actually what we see in practice when we test Chinese models: none of them are communist-aligned. They're all pragmatic utilitarian models. So I wondered—

    31:24

    Nathan Labenz: With some Chinese characteristics.

    31:26

    Prakash: With some Chinese characteristics. So I wondered to what extent it really matters to anyone in safety, beyond the commercial competitive aspects. I don't know.

    31:44

    Nathan Labenz: I've often thought that a public good Anthropic might create — and they've done a little of this, but far less than I expected — would be big public datasets for alignment of various kinds. You just put those out there and say: if you want to take advantage of the alignment and character work we've done, here's what we think good looks like across a range of dimensions, and you can train on that. It wouldn't need to include frontier research capabilities or kernel optimizations, but it could help people refine their models. I imagine a lot of companies would want to use it. I'm surprised there hasn't been more of that. It wouldn't necessarily change the core problem — there'd still be strong incentive to extract coding capability that Anthropic wouldn't release publicly — but why haven't there been more public-good alignment datasets? Seems like a gap that wouldn't be too hard to fill and would be pretty valuable.

    33:26

    Prakash: Speaking of aligning models — we have a wave of researchers leaving primarily Google DeepMind. Arthur Kami has just decided to depart. Rishab Jain leaving DeepMind as well. And Nate Suarez from MIRI is, as usual, saying we should ban superintelligence — it'll be the end of history. Francis Fukuyama, the original author of 'The End of History,' is also saying we should ban superintelligence. The question that strikes me: why are people leaving Google DeepMind?

    34:13

    Nathan Labenz: I don't know — very similar to what I asked Swix on Monday about what's up at Google. At some point it does become a worrying signal. I still have a lot of confidence in top leadership there. I think one should never bet against Demis, never bet against Jeff Dean, never bet against Shane Legg — those people have been incredibly prescient, and I can't believe they're totally dropping the ball. But the signal is getting stronger. It really started with Carlini — that was a while back now — and it seems like that kicked off a notable sequence of these moves. I don't know if it's a crisis yet, but it won't be too much longer, if this trend continues, before it really does become a crisis for Google.

    35:22

    Prakash: So speaking of researchers and science — let me bring in our first guest for today. It is Eric Olson. Eric Olson is the CEO and co-founder of Consensus, an AI-powered operating system for scientific research currently utilized by over 2.5 million researchers, clinicians, and students every month. He's a former Division I football player at Northwestern University who transitioned into predictive analytics and data science. He launched Consensus in 2021 to solve a fundamental bottleneck in human progress: the sheer impossibility of keeping up with the millions of scientific papers published every year.

    Rather than building a general chatbot that risks inventing fake citations, Eric and his team architected a different approach. Consensus searches a verified database of over 220 million peer-reviewed papers, utilizing advanced multi-agent AI to read the literature, extract methodologies, and synthesize the findings. Backed by a recent $30 million Series B, the company is pushing beyond simple search to build tools that actively assist in complex literature reviews and gap analysis.

    Eric brings a distinctly pragmatic worldview to AI. He believes AI should not replace the human scientist, but rather give the brightest minds superhuman leverage by eliminating the tedious manual work of evidence retrieval. As nations increasingly view their scientific data and AI infrastructure as vital sovereign assets, Eric brings a vital perspective. Eric, welcome to the show.

    37:24

    Nathan Labenz: Thank you so much. That was a wonderful intro. I gotta bring you to pitches with me.

    You're not the first to say that, actually. We've got some very bright Silicon Minds helping us do deep research on every guest.

    Amazing. I'm really, really happy to be here and to touch on all the topics Prakash just mentioned.

  2. 37:48Interview57 min
    Interview: Eric Olson — AI for Science and Surviving as an Application-Layer CompanyEric OlsonEric Olson — CEO and co-founder of Consensus, an AI research engine over 200M+ peer-reviewed papers used by 2.5M+ people a month — on where AI for science stands and whether application-layer companies built on the frontier labs get steamrolled or compound. He detailed Consensus's 'recipes' architecture, exponential growth in query complexity, routing narrow tasks to sub-billion-parameter BERT-series models (recovering ~95% of frontier performance down to ~1B parameters), a days-to-hours research speed-up, three drivers of stickiness, and why the API-versus-consumer price gap is navigable but not unlimited.

    Nathan opened by asking Eric Olson, CEO and co-founder of Consensus (AI-powered scientific search), to set the level on where AI for science actually stands. Eric positioned Consensus as deliberately building for the leverage available today rather than betting on a 'push a button, get a scientific discovery' future — a future that is uncertain, and that, if it arrived, would change everything for everyone, including Consensus.

    Prakash asked about the next layer of search — whether passive, intent-anticipating search would supplant active querying, and how Consensus uses behavioral metadata to inform it. Eric agreed that proactive search is part of the evolution, pointing to PubMed-style alerts as an early version. But he pointed to a more striking trend already underway: the explosive growth in query complexity. Over just the past year, average query length on Consensus has grown exponentially. Users now submit multi-step research briefs — search this, cross-reference that, synthesize a gap-analysis report, save the ten most relevant papers. He attributed the shift primarily to expectations raised by tools like Claude Code.

    Nathan pressed on how Consensus balances language-model flexibility with systematic, citation-grounded guarantees, and whether companies in the space are converging or diverging. Eric described a 'recipes' architecture that classifies user requests into guardrailed agentic paths and always runs searches under the hood to ground every response in literature — even if that occasionally means telling a user 'we couldn't find papers on this.' Better models are causing some convergence across competitors, since all can now delegate planning to the model rather than hard-coding it, but domain-specific trade-offs will continue to differentiate products.

    Prakash fired a three-part question: user breakdown by domain, token costs and open-source adoption, and observed speed-up for users. Eric said about 70% of Consensus users come from academia across all research disciplines, and 30% from professional contexts (roughly half clinicians, half R&D, with bio heavily dominant in both). Token costs are rising steeply as agentic task lengths grow even as per-token prices have dipped; open-source models are now used in the product more than six months ago, as the capability gap to frontier has narrowed. Speed-up in science: roughly days to hours — two to three days of work is now four to five hours — real, but less steep than fully verifiable domains like coding.

    Nathan offered a theory: until enough models cross a capability threshold, science products should default to the best available frontier model since scientists demand high-quality output; after that crossing, intelligent routing becomes the real strategic moat. Eric largely agreed and gave a detailed picture of how Consensus already routes: BERT-series models (sub-billion parameters, including domain-specialized variants like BioBERT and PubMedBERT) handle field-of-study classification and search-relevance weighting in sub-0.1 seconds, while frontier models handle complex synthesis. For a narrow classification task, he put the recoverable performance at roughly 95% of frontier capability down to about a 1B-parameter model with a good fine-tuning set.

    Nathan asked how users decide between Consensus and competing tools and how sticky the product is. Eric named three decision factors: trust and accuracy (a single wrong citation or missed seminal paper can permanently lose a researcher, who often runs personal verification tests); feature depth (Consensus's reference-management workflow — save, organize, analyze, and share papers — has been a key differentiator); and UX simplicity (the single most common thing heard from users who have tried all the tools). He acknowledged switching costs are low and framed earning user loyalty as a daily challenge.

    With a few extra minutes available, Nathan raised the price-discrimination question: per-token API pricing appears to be more than 10× more expensive per token than heavily subsidized consumer subscriptions, creating a structural headwind for application-layer companies. Eric was measured: some gap is fine and navigable — the app layer's job is to allocate tokens far more efficiently than individual end users — and the real risk is if the gap becomes 'gigantic.' Open-source progress is the natural competitive check on that, which is why application-layer founders should root for open source.

    After Eric signed off, Nathan and Prakash extended the discussion into the broader platform-power question. Prakash drew an airline analogy: frontier AI companies have high fixed GPU costs and are running classic price-discrimination strategies — premium real-time capacity versus economy/batch tiers — to maximize GPU fleet revenue. He also noted the underappreciated competitive threat: Anthropic is not just a model provider but an application competitor led by Mike Krieger, co-founder of Instagram, whose product organization is setting the pace the rest of the industry imitates. Nathan flagged the newly announced Mirandil — elvish for 'guarded secret,' founded by ex-Anthropic researchers — as an explicit counter-movement aimed at making recursive self-improvement tools broadly available. The pair also discussed whether frontier labs might acquire scientific publishers or health-record systems like Epic (estimated $30–60B valuation, controlling an estimated 70–80% of US electronic health records) as ways to entrench data moats.

    If we move into a world that is truly push a button, get science out of it, I think a lot of things are gonna be different. A lot of companies are gonna be screwed.

    The difference in complexity of queries has just grown. Exponential doesn't even capture the extent that is coming.

    You can get about 95% of the performance from a frontier model for a very small classification task going all the way down to about a one-billion-parameter model, if you put in the work and give it a good fine-tuning set.

    37:48Where are we right now in terms of AI for science — how would you locate us on the road to actually AI-powered discovery?
    AI for science is behind leading domains like coding but appears to be on a similar curve. There is evidence of some fundamental differences between science and inherently verifiable domains. Consensus's bet is on building for the leverage available today rather than waiting for push-button scientific discovery — if that future arrives, great for the world, but it's not a bet worth making as a company.
    41:35What does the next layer of search look like — passive, intent-anticipating search — and how does user metadata help inform it?
    Proactive search is part of the evolution; PubMed-style topic alerts are a simple early version that users already love. But the bigger shift Consensus is observing is an explosion in query complexity: average query length has grown exponentially over just the past year. Users now submit multi-step research briefs expecting the product to search, cross-reference, synthesize gap analyses, and save relevant papers — all in one prompt. Expectations set by tools like Claude Code are driving this.
    45:12How do you balance language model flexibility with systematic, gap-free guarantees around literature? And are companies in the space converging or diverging on strategy?
    This is the core product tension. Consensus uses a 'recipes' architecture that classifies user requests into guardrailed agentic paths and always runs searches under the hood to ground responses in literature — even if that occasionally means telling a user 'we couldn't find papers on this.' Better models are causing more convergence across competitors, since all can now delegate planning to the model rather than hard-coding it, but domain-specific trade-offs will continue to differentiate products.
    49:10What's the breakdown of users by domain? Are token costs driving open-source adoption? And what kind of speed-up are your users actually seeing?
    About 70% of users come from academia across all research disciplines; 30% from professional contexts roughly split between clinicians and R&D (with bio heavily dominant in both). Token costs are rising steeply as agentic task lengths grow; open-source models are now used in the product more than they were six months ago as the capability gap to frontier has narrowed. Speed-up: roughly days to hours — two to three days of work is now four to five hours — real, but less steep than fully verifiable domains like coding.
    55:31My theory: science products should default to the best frontier model until a capability threshold is crossed, then the strategic value shifts to intelligent routing. Where do you think we are on that curve?
    Routing has always been part of the picture, even when the capability gap was large. Consensus already routes 20-odd small classification tasks — field-of-study classification, search-ranking signal weighting — to sub-billion-parameter BERT-series models (BioBERT, PubMedBERT) returning results in sub-0.1 seconds, while frontier models handle complex synthesis. The frontier-to-open-source gap has narrowed in the last few months, enabling more open-source use in complex tasks too. Build your product for optionality so you can always swap things out.
    57:56What specific tasks are routed to small models, and what base models do you use? How much frontier performance can you recover on those narrow tasks?
    Field-of-study classification is a representative example: knowing whether a query is biomedical (where sample size and study design drive ranking) versus computer science (where recency and citation velocity matter more) lets Consensus weight search-ranking variables appropriately. Primarily BERT-series models are used, fine-tuned with human labels or model-generated labels. For a narrow classification task with a good fine-tuning set, you can recover roughly 95% of frontier performance down to about a 1B-parameter model.
    1:01:55How do users decide between Consensus and alternatives, and how much loyalty does your product create given low switching costs?
    Three factors: (1) Trust and accuracy — a single wrong citation or missed seminal paper can permanently lose a researcher who ran a verification test; (2) Feature depth — Consensus's reference-management workflow (save, organize, analyze, share papers) has been a major differentiator; (3) UX simplicity — the single most common thing heard from users who've tried all the tools is 'yours is the easiest to use.' Switching costs are low; you earn the right to keep users back every day.
    1:05:31Should policy address the price discrimination between per-token API pricing and consumer subscription plans, where consumers appear to get more than ten times as many tokens per dollar?
    Some gap is fine and navigable — the app layer's job is to allocate tokens far more efficiently to solve specific problems than a consumer using a chat interface. A gigantic, persistent gap is the real risk. Antitrust treatment would more likely push consumer pricing toward usage-based models than lower API prices. More plausibly: large volume commitments will unlock big discounts, similar to GCP. Open-source progress is also a natural competitive check on frontier pricing power.
    Lightly edited · timestamps jump to YouTube
    37:48

    Nathan Labenz: Where are we right now in terms of AI for science? There's this constant disagreement — AI seems to have made a discovery here or there, and on the other hand there's a kind of denialism, which is sometimes easy to caricature. How would you locate us on the road to actually AI-powered discovery?

    38:27

    Eric Olson: Compared to the domains that have jumped out in front — chief among them coding — AI for science is definitely behind, but I think it's still very much on a similar-ish curve. The question is whether there's something fundamentally different about science versus something that is so inherently verifiable, like coding. Nobody really knows, but there is evidence that there are some fundamental differences. Our view is: there's so much leverage to be had today, so many problems that still need to be solved, that we want to build for the problems we see today and in the next couple of years — not necessarily bet on the 'push a button, get a scientific discovery out the other side' that some more research-focused companies are going after. If they solve that, great — the world will be a much better place. I don't think anybody knows if it's actually going to be possible, and there's so much room and so much love to be had at what's available today, and that's what we're attacking.

    39:39

    Nathan Labenz: That's a pretty interesting perspective. You don't often hear CEOs frame it quite that candidly. I think what you're saying is: that future might work, and if it does, we're going to be in a tough spot as a business, but the world will be in a great spot. You seem willing to bite a bullet that most people try to talk around.

    40:03

    Eric Olson: Number one, you have to make bets as a company, and there's gonna be risk on the table no matter where you place them. If we move into a world that is truly push a button, get science out of it, I think a lot of things are gonna be different — a lot of companies are gonna be screwed. It's gonna change the whole world. So it's quite a big bet to bet on that specifically. And I think there's so much room for gray, too: versions where Consensus is still a huge part of the process, where we're building the best possible tools involved in the scientific process — or we're APIs that plug into those harnesses that run the whole science process at the push of a button. We're already used by a number of organizations today that power internal agentic systems with our search API. I can completely see a role, as we build more of this tool suite, where our tools power those bigger, broader systems. Anything short of the very far end of the spectrum — full automation requiring no third parties — there's going to be places for tools and companies doing things like us. And I think nobody knows if we can actually get to that very, very far end.

    41:35

    Prakash: What does the next layer of search look like? The first layer is reactive — you ask a question, you get an answer. The next layer might be passive search: you don't have to ask the question; you get prompted the answer upfront, preempting the process based on what the system knows you're interested in. To what extent do you see the metadata you're collecting on what people want to know right now, in their scientific search, helping you form that next layer?

    42:26

    Eric Olson: It's a neat way to phrase it. Proactive searching and predicting intent is absolutely part of the evolution — there were even very simple versions of this a decade ago, like PubMed alerts that run saved searches and email you a report every morning. People love those features, and we can obviously do that far more intelligently with AI. But where I see the bigger shift in how people interact with products like ours is something slightly different: the complexity and length of what they're asking the product to do. If you look at the average length of our queries today versus even just a year ago, it has grown exponentially. When you go look at what those queries are, it's typically: look up this thing, then go look up this other thing, synthesize it back to me in a research-gaps analysis report, and save the ten most pertinent papers. It's this multistep expectation of much more complex work from the product. Part of that is our product getting better so people know they can just ask that. Way more of it is the expectation from using other tools — knowing they can type in a very messy, very long multistep process and expect some usable output. Search in science is a fundamental starting point — you do it in service of some further downstream task. What we're seeing is not only the complexity of the search task growing, but people also putting complex search plus task execution all into one prompt. The difference in complexity of queries has just grown — exponential doesn't even capture the extent that is coming.

    45:12

    Nathan Labenz: How do you think about getting the best of language model flexibility while also making certain guarantees to users around systematic, thorough, gap-free processing of the literature? And are companies in your space converging on similar strategies or diverging?

    45:56

    Eric Olson: This tension is just the core tension of building our product. We don't have all the liberties of a general-purpose product that can throw a computer at the problem with no guardrails. The negative is we can't just let things run wild; the positive is the space of what users are asking about is smaller for a verticalized product — it's an achievable trade-off. What we basically do is give guardrailed paths, because it's more predictable what paths users want to go down. We have what we call 'recipes' in our agentic product: given what a user asks, we classify it into a recipe — a systematized step process we know users are likely to want for that type of request. We also have pure flexible routing for the rest. But we always run searches under the hood to ground all responses in literature, and we put those constraints on ourselves — sometimes in ways that make the experience worse in edge cases, like when a user asks for flexible advice on something and we say, 'Hey, we couldn't find any papers on this, sorry.' That is a worthwhile trade-off for grounding everyone else's responses correctly. On convergence: about a year ago, products in our space were diverging more; in the last couple of months we're starting to converge. The better models get, the less hard the choices — you can let the model decide the steps to take rather than betting on them yourself. But given the guardrails we all still have to put in place in this domain, there are still going to be decisions and trade-offs that give products different shapes and things they lean into. I have seen more convergence recently.

    49:10

    Prakash: Among your users, what percentage are bio versus other scientific verticals? Second: how are you dealing with token costs — are they significant enough that you're constantly evaluating open-source models to replace frontier models? And third: what kind of speed-up are your users getting? In coding, people can get a week's work done in three or four hours if they organize well. What kind of speed-up are you seeing in science?

    50:03

    Eric Olson: On the user breakdown: compared to some contemporaries, we lean more academically general. About 70% of our users come from an academic context — that's across all domains of science, roughly the general distribution of what people study in school with some modifier for how research-intensive the subject is. We have art students, history students, sociology researchers. About 30% come from a professional context, pretty evenly split between clinicians and R&D. In that R&D bucket, it's primarily bio, with some materials science and engineering. So bio is by far the biggest bucket — but we've done really well in the academic world because of our initial deep focus on search, which is so fundamental to academic research. We lean a little less pure-bio than some other products in this space.

    On token costs: they've significantly gone up. Even as average model costs have come down somewhat in the last year, the length of the tasks is just so much more — we're giving so much more to the model to do. Costs are increasing for every agentic product, not just ours. It's always on our mind. One thing that helps: because we have those specialized guardrailed recipes, we aren't just saying 'model, just figure this out and iterate a million times.' We're more efficient than the average product in this broad space. But that doesn't mean it isn't increasing. In the past couple of months, a bunch of new open-source models have gotten closer to frontier, and we're using them in spots in the product. I'd be surprised if it doesn't keep following that pattern: when there's a capability jump, frontier leads and it'd be a mistake not to lean in; then the gap closes. Build your product so you can quickly swap things out and always be weighing the cost, speed, and capability trade-offs.

    On speed-up: it's a similar-shaped answer to where science fits alongside other domains. In science it's days to hours — what might have been two to three days of work is now four to five hours. Is it as steep as fully verifiable domains where end-to-end execution can mostly be done? No. There's still so much manual intervention. But we are going up that curve.

    55:31

    Nathan Labenz: Let me throw a theory at you. For the application layer, I think there may be a phase-change moment, and when it kicks in will vary by domain — science will probably be one of the higher bars. Until you reach that bar, a strong default would be to use the very best model, because you're dealing with scientists who want good output. My guess is there will be strong customer demand for Fable once we get our access back, and you'll be willing to pay up for it because the results are just that much better. Then at some point, as enough models cross a threshold, your strategic position is really about routing — figuring out which model is best for which task, most cost-effective for which thing. And that's something the frontier models themselves will never be motivated to optimize for your customers the way you can. How do you feel about that framing, and where do you think we are with respect to 'just use the best model for everything' versus 'you're creating real strategic value as a routing layer'?

    57:26

    Eric Olson: Even six months ago, when we were more 'use the frontier to solve all the problems,' there was still lots of routing going on — lots of small tasks offloaded to very small specialized models, sometimes even sub-billion-parameter. So even when the capability gap was large, we were still doing routing.

    57:56

    Nathan Labenz: Can you talk a bit about what those tasks are — like, is it relevance classification?

    58:01

    Eric Olson: That can be one. Something as simple as: classify the field of study of the query so we know which search-ranking variables to care about. In biomedical research, experimental design matters enormously — sample size, study duration, where the study took place. In computer science, recency and citation velocity of the paper matter more — and who the researchers were. To know how to weight each of those variables, we have a small model that classifies the field of study. That does not need to be jammed into a giant prompt. It does not need to be a one-second latency API call to a frontier model. It should be a self-hosted 800M-parameter model fine-tuned on a few examples — returning results in sub-0.1 seconds. There are probably 20 different versions of those small classification routers informing downstream things on query time. All of those should always use a really small model — not even just for cost, but purely on latency.

    59:14

    Nathan Labenz: One more double-click: what base models are you using for that? Are you distilling from something like Claude into a smaller model? And how much frontier performance can you recover — can you get 90%, 95% of the way back in that narrow domain?

    59:55

    Eric Olson: A lot of BERT-series models — BERT variants of different sizes and domain specializations. There's BioBERT, PubMedBERT, and very small BERT iterations that are very good at classification tasks because that's what they were designed for. BERT was the leading architecture five years ago and is still popular for small classification and relevance tasks. We use human labels for some of these fine-tuning sets, and model-generated labels for others — you could call that distillation. But the objective isn't to replicate the full frontier model; it's to train something to do one very small specialized thing. Because of that narrow scope, you can retain a ton of performance. I'd say you can get about 95% of frontier performance for a very small classification task going all the way down to about a one-billion-parameter model, if you put in the work and give it a good fine-tuning set. That number changes as you add more classes or complexity — but you can also close the gap by making the model bigger.

    1:01:48

    Nathan Labenz: Very helpful — I appreciate the specificity.

    1:01:55

    Nathan Labenz: Competitively: how do people decide what to use? When choosing between Consensus and alternatives, how much loyalty does your product create, versus users doing their own routing between tools? What are the decision criteria that matter most?

    1:02:29

    Eric Olson: I'd put them in three buckets. First, trust and accuracy — it's just so important in our space. The second you see a citation attributed incorrectly, or a paper summarized in a way that isn't actually represented in the paper, or someone types in a question they know deeply and we don't return the seminal paper — that's a huge trust-losing event. Researchers love to run their own little verification tests: they'll type in their own name and expect their best papers to come up instantly. The second you fall short of that standard, you may never win those users back. In churned-user interviews, we hear that: 'You fell short in this way, and I immediately moved off your tool.' Second is feature set and capability. We have to make trade-offs about what paths to go down as a product. One differentiator: we went deep into reference management because of our search focus — saving papers, organizing them into folders, analyzing and sharing them. A lot of people use us specifically because we have that, while other tools will build other features users gravitate toward. Third is taste and UX. Our product is very user-friendly — a simple, easy-to-use interface — and we've emphasized that more than any other tool in our space. The single most common thing we hear from users who've tried all the tools: 'I like your interface, yours is the easiest to use.' That still matters and always will. On switching: the switching costs are just low. You have to earn the right to keep winning those users back every single day, and if you don't adopt that mindset, you're just not gonna win.

    1:05:17

    Nathan Labenz: We're just about at our planned time, but we have an open slot for the next 30 minutes. Do you have time for a couple more?

    1:05:27

    Eric Olson: I can do one or two more.

    1:05:31

    Nathan Labenz: How about your thoughts on price discrimination? You're paying a lot more at the API price than I'm paying with my Claude Max plan if I'm maxing it out — and reports suggest it's more than ten times as many tokens per dollar on the consumer side. Do you think there's a role for policy to try to even that out? I'm a little worried this gap is going to make it extremely difficult at the app layer long-term. Whether that's antitrust or some other policy, there's a question of whether AI providers should have more balance between first-party customers and API customers serving end users.

    1:06:36

    Eric Olson: You're basically saying: because we're paying per token, we're getting charged much more than somebody who can just buy the $200-a-month plan and there's no equivalent deal structure — that's basically what you're saying?

    1:06:48

    Nathan Labenz: Yeah. Reports suggest it's more than ten times as many tokens per dollar on the consumer plan.

    1:06:59

    Eric Olson: If there were ever antitrust treatment of models as commodities — which is definitely possible — I'd imagine the pressure would actually go the opposite direction: consumer end users get priced more on usage too, trending toward usage-based pricing overall. If it went the other way, giving companies like us better rates, that'd be wonderful. I could also see a world more like cloud services, where you make big volume commitments and get large discounts — commit to ten million dollars next year and get a much lower cost per token. I'd be surprised if we don't move more toward that. But I think it's okay if there is a gap. There being a gap is not fundamentally a problem. It is our job to allocate those tokens much more efficiently to solve the specific problem for our users than somebody just talking on their $200-a-month plan to a base model in a chat interface. We should be able to do that better. The gap getting large is a problem. There being a gap, I don't think is fundamentally a problem.

    1:08:52

    Nathan Labenz: Thank you. I don't know exactly where I land on this either. I've been feeling like AI is pretty cheap — at the same time I hear founder friends saying costs are getting out of control. And then I realize the ratio I'm personally enjoying right now is huge. Some normalization feels probably healthy for the overall ecosystem, just to avoid everything collapsing into the foundation model companies.

    1:09:29

    Eric Olson: This is also probably why founders have to root for open source to keep pushing frontier — it means they don't have the ability to make that gap gigantic. If the gap between frontier and open source is huge, frontier labs have all the power in the world to keep that gap wherever they want. If open source gets close to frontier, they have much less pricing power. But application-layer companies focused on specific problems have so many levers to fight against this. This is simply the core task we're all doing: solve the specific problem for our users in a really high-leverage way. If we can really do that, we will find ways around different pricing structures.

    1:10:16

    Nathan Labenz: No excuses. Play like a champion.

    1:10:18

    Eric Olson: Football player by trade. Not a Notre Dame one, though.

    1:10:23

    Nathan Labenz: Go Cats.

    1:10:26

    Eric Olson: Go Cats. One and O against Notre Dame in my career, I will say.

    1:10:30

    Nathan Labenz: Fantastic. Prakash, anything else you want to jump in with? Eric, any concluding thoughts?

    1:10:38

    Eric Olson: This has been great, Nathan. Really appreciate having me on.

    1:10:40

    Nathan Labenz: Thanks for joining us. Great to meet you.

    1:10:43

    Eric Olson: Have a great rest of day, guys.

    1:10:47

    Nathan Labenz: I'm so torn on this question — it just keeps coming up over and over again. Concentration of power on one hand versus, in this case, concentration of power on the other hand, because you do want a healthy app ecosystem. And I'm wary of what happens if this greater-than-ten-to-one cost advantage persists indefinitely. For the things I'm doing day to day, I don't find myself reaching to other apps all that much, because I just have a lot of tokens on Claude Max. I'm an early adopter and more comfortable doing my own scaffolding, but it's getting easier and easier over time.

    The flip side: if I, as an individual user, lose my subsidy — and it is quite subsidized — the cost per token on my Claude Max is cheaper than running Llama open-source at inference, so if I'm maxing it out they're definitely subsidizing me. If I lose that subsidy, the concentration of buying power potentially shifts to enterprise, and you have a different problem: already-profitable enterprises are far better positioned in a pure auction market for tokens. So the anti-concentration argument actually works both ways. Right now, the app layer is getting squeezed and the individual layer is being empowered. Flip it — empower the app layer more — and you disempower the individual and potentially empower enterprise more. I'm not sure that's better. It's just hard to evaluate.

    1:13:40

    Prakash: I think you can look at other high-fixed-cost, variable-revenue businesses. The classical example is airlines: you've already paid for the plane upfront, and then you have to maximize revenue on that fixed asset. You end up with first class, business class, economy — multiple pricing strategies to survive. Business class: you can change flights easily, arrive without a booking and board, they maintain more capacity for you. Economy: they overbook you and bump you. You can see the same concepts in AI — batch processing versus real-time processing, dedicated capacity with latency guarantees versus low-cost providers without guarantees.

    The premium product is the frontier lab — state-of-the-art intelligence for some window of a few days to a few months — and within that window they have a GPU fleet to maximize revenue off of. So OpenAI still serves GPT models at something like 25 cents per million tokens — very competitive with open source — because those are older models on high-margin capacity. What they're doing is GPU-fleet revenue maximization, very similar to airlines. And you're going to see those same strategies: premium branding, feature differentiation — some firms saying 'our math model is better,' some saying 'our coding is better.' I think a lot of people haven't figured out that this game yet. They're still thinking model versus model, not: how do you maximize revenue on a fixed GPU fleet?

    1:18:43

    Nathan Labenz: I have no beef with price discrimination in the airline industry. And I've been broadly not worried about net neutrality — I remember when that was the sacred cow, and the first Trump administration weakened those rules. At the time: 'This is going to be the end of the internet.' I predicted it wouldn't happen. History vindicated that — I've never heard one story of an ISP blocking content unless you pay extra. The internet continued to improve. I say all this as a lifelong techno-optimist libertarian. And still, somehow, this feels like it might be a little different. Maybe AI is just a paradigm breaker — not obviously, but plausibly — where the tokens themselves are so much of the value that we could see a real loss of diversity at the app layer if a ten-to-one-plus cost ratio persists indefinitely. Business class tickets are certainly often a multiple of economy, so it's not qualitatively different — but it is tough at the app layer. Andrew from Tasklet said that point blank: the number one place people go when they churn from Tasklet is Claude Code. So for him, it's an absolute strategic imperative to become platform-neutral — to be a layer that helps consumers play the foundation model companies off against one another. In that, there's some defensibility. Ten to one, though — that's a big ratio.

    1:21:53

    Prakash: I think perhaps the problem is different, because it's not just that the foundation model companies have the model — they also have the app. It's not just Claude the model; it's Claude the model with a bunch of apps built on top of it. And those apps are built by Mike Krieger, who was co-founder and CTO of Instagram — arguably the all-time best consumer app in history. He became head of product at Anthropic a couple of years ago, recruited the entire product team, and ran that organization. And he didn't make stupid mistakes: Claude Code was created by Boris, and rather than let him walk to Cursor, they pulled him back and elevated Claude Code into Claude Code on desktop, Claude Max, and now more. So you have to take a step back and say: Tasklet isn't competing with foundation model researchers. They're competing with arguably the leading product mind of the last fifteen to twenty years, with experience taking an app from eleven employees to trillion-dollar scale. The innovations coming out of Anthropic's product organization are being imitated across the entire space — Codex command line and others are all imitating what Anthropic built. That integration between product and model teams is what app-layer companies are really up against. That's a tough part.

    1:25:13

    Nathan Labenz: There's another level too that they haven't done much at all yet, but could easily do: only deploy their latest and greatest models into their own first-party products. Right now with Fable, OpenAI supposedly will have an answer soon-ish, but Google does not have a model that can really compete with Fable on the API today. If Anthropic decided they wanted to make Fable only available to first-party users — tough to compete with that. If that much raw intelligence just isn't available via the API, what are the app-layer companies going to do?

    1:26:46

    Prakash: I'll also say the moat is perhaps even less than we think, because there was a release of an ad product — a Chrome extension that would sit on top of Claude or ChatGPT web pages and inject a little ad while you were waiting for the response, and pay you for watching it. And when you have that extension, it's also reading tokens in and out. You could imagine an opt-in version: here's your $200 Claude subscription, put in the extension and watch ads and we'll pay you $50 a month back, opt into the data sharing and we'll pay you another $50. Consumers are free to do this — by US rules, it's their data. You'd effectively have near-immediate distillation happening on the other end. People don't do it because there's not enough money to make it worth their while right now. But if the consumer is annoyed enough, you'll get very fast distillation out of it. So there's always this middle ground where firms clamp down on what they don't like, but not so hard that extremely motivated actors will go out of their way to work around it.

    1:29:04

    Prakash: I honestly think it's very difficult for app-layer companies in scientific search because there are strategies the frontier labs haven't done yet that seem fairly obvious — like buying sources of data. Scientific papers are gated behind journal subscriptions; getting all the subscriptions you'd need would probably cost two to twenty million dollars. Researchers end up going to Sci-Hub, which is based in Russia and pirating hundreds of millions of journal articles, under federal investigation, with the founders facing drug-dealer-level extradition scrutiny if they leave Russia. And all US researchers openly use these sites. I don't understand why the model companies don't just buy access and plug it in directly. All scientific publishers combined are probably worth around $20B total and make less than $1B in revenue per year in aggregate. Roll them all in, plug the data in, and you have the publishers alongside you. I know it's not their core business, and I know that once you do that, app-layer scientific-search companies really have nothing to do — if the papers are already in Claude and ChatGPT with a direct front end, what's the point? So I do wonder about that.

    1:31:27

    Nathan Labenz: The money's going to be there. I was looking up Epic Health — estimated valuation somewhere in the $30 to $60 billion range, which tops out at Cursor scale. If you wanted to get your hands on the system that controls something like 70 to 80% of all American electronic health records, that's a pretty direct path to an enormous amount of data, and it's very affordable for these companies at this point.

    This connects to something I wanted to flag: yesterday, Mirandil launched — or rather announced their funding without putting forward their product in detail yet. The name means something like 'guarded or kept secret' in Elvish. They're former Anthropic people who were starting this company before the Fable launch became public — presumably seeing the writing on the wall around recursive self-improvement tools being restricted. They're making an explicit appeal on an anti-concentration-of-power basis: 'We were at Anthropic. They're doing recursive self-improvement and don't want to let you have it. So we're bringing that to an independent company and making it available to all.' I'm not sure how well that's going to work — some very smart people involved, but keeping up with Anthropic's resources is hard. Maybe they don't have to keep up if Anthropic doesn't want to compete in that space. Seems like they'll be behind, but maybe being the best open option in the market is enough. Very interesting development.

  3. 1:34:22Interview39 min
    Interview: Tricia Martinez — Sovereign, In-Country AI InfrastructureTricia MartinezTricia Martinez — founder and CEO of Dapple — on dedicated, single-tenant, in-country AI clouds for regulated enterprises, the 'Enterprise OS Cloud,' and the company's claim of over $100M in contracts in its first five months on a $30M seed. She fielded pointed questions on the 91–94% utilization claim, 6-to-9-month deployment timelines, Azure-native dependency, financing structure, and long-term defensibility (framed as enterprise switching costs, a self-improving operating platform, and a hard-to-recreate ecosystem). Her contract, utilization, and deployment figures are company claims.

    Nathan and Prakash briefly closed out a discussion of a newly announced frontier AI lab — characterized by ideologically motivated founders, $200M from a16z, and a reported large but undisclosed NVIDIA investment — before pivoting to the Tricia Martinez interview.

    Prakash introduced Tricia Martinez, founder and CEO of Dapple, as an entrepreneur offering a 'third path' between renting shared public-cloud capacity and building private data centers from scratch. Dapple deploys dedicated, single-tenant, in-country AI clouds — isolated physical hardware governed by local law, operated with modern cloud-software interfaces. Tricia previously helped launch one of Africa's first digital banks and served as a White House Fellow across both the first Trump and Biden administrations at the Department of Energy, where she built an AI-and-energy strategy across seventeen national labs. Her co-founder and COO, Salam Al Musawi, brings an engineering background with claimed deployment of more than 300,000 AI accelerators globally. Prakash noted Dapple's announced $30M seed round backed by Raptor Group and ION Pacific, and the company's claim of over $100M in secured enterprise contracts within its first five months of operation.

    After a brief audio-quality fix at the top of the segment, Nathan opened with a direct question: who is so eager to spend this money that Dapple claimed $100M in signed contracts within five months? Tricia argued that enterprise and government demand has been bottled up for years — public clouds lack capacity for large enterprise workloads, NeoCloud bare-metal options require enterprises to build their own operational stack, and on-premises buildouts are too costly and complex to staff. Dapple's 'Enterprise OS Cloud,' she said, collapses that complexity into a single operating layer for deploying and running production AI across dedicated infrastructure. She teased that upcoming growth announcements would be even more dramatic.

    Prakash pressed on Dapple's heavy Azure-native marketing language — integration with Azure Machine Learning, Azure Arc, and Azure Monitor. Tricia confirmed that unifying public and private cloud into a single frictionless experience is Dapple's core IP, but declined to name the strategic cloud partners involved, saying official announcements would come within months.

    Nathan questioned the 91–94% GPU utilization figure on Dapple's website, finding it surprisingly high for a single-tenant infrastructure play. Tricia explained that Dapple's deployments are backed by long-term enterprise contracts — typically three, five, or seven years — making the model about committed capacity and customer outcomes rather than spot-market utilization.

    On what workloads sustain that utilization, Tricia identified three categories: large-scale model training for AI-native companies building domain-specific foundation models; production-scale inference requiring predictable performance, low latency, and strong governance; and AI agents embedded in internal operations and customer-facing products. She noted Dapple started with mission-critical verticals — finance, cybersecurity, government, defense — but said Fortune 1000 demand had turned out far larger than anticipated, with most customers simply desperate to access any capacity so they can begin building.

    Prakash asked how Dapple achieves 6-to-9-month deployment timelines when even record-setting data-center builds take roughly fifteen months. Tricia described a network-orchestration model: Dapple partners with existing data-center operators and infrastructure providers, then contributes capital, GPU infrastructure, enterprise demand, and its software operating layer. In some deals Dapple helps finance the infrastructure directly; in others it deploys on partner facilities.

    Nathan raised a macro risk question about the AI infrastructure finance stack — whether long-term commitments paired with short-term revenue flows could create cascading fragility under an external shock. Tricia acknowledged the market has grown more disciplined, with financing groups increasingly requiring demand-backed contracts rather than speculative buildout. She argued Dapple's model is more de-risked because enterprise customers make large upfront down payments and commit to multi-year relationships. She also cited growing regulatory risk around compute export restrictions as a mounting threat to competitors relying on certain geopolitical off-takers.

    Prakash characterized Dapple as more asset-light than the typical NeoCloud. Tricia agreed, framing the company as software-first: it partners for data-center operations and GPU ownership where possible rather than trying to own everything, and believes that positioning is where Dapple will ultimately win.

    Nathan explored Dapple's long-term defensibility in a world where AI coding agents could eventually simplify even hyperscaler cloud complexity. Tricia said there is no single moat; defensibility comes from three reinforcing layers: enterprise relationships and the high switching costs of moving mission-critical AI infrastructure (governance, networking, compliance, security, and operational workflows all move together); an operating platform that gets smarter with each deployment, improving placement, capacity planning, performance, cost, and policy management; and a broad ecosystem of infrastructure providers, capital partners, silicon partners, and enterprise customers that is difficult to recreate simultaneously. She drew an analogy to AWS abstracting on-premises complexity for the enterprise cloud era, calling Dapple the same abstraction layer for the AI infrastructure era.

    Prakash asked what numerical KPIs enterprise customers track most closely. Tricia listed five: time to deployment; availability and reliability for mission-critical and in-country isolated workloads; scalability as AI usage grows; governance and compliance — which she called 'probably one of the most important things that no one can support them on in the market right now'; and predictable economics over the full deployment lifetime.

    Nathan asked, drawing on Tricia's Africa background, how accessible AI infrastructure is to African enterprises. Tricia said Dapple's demand is concentrated in APAC, the US, and Europe, with minimal signal from South America, Africa, or parts of the Middle East. She described Africa as falling behind on AI production capacity, with cost as the primary barrier — citing new hardware like Nvidia's Vera Rubin as making deployment even more prohibitive. She called it a genuine problem that governments are trying to address but struggling to solve.

    Prakash asked what mistakes enterprise AI buyers most commonly make. Tricia said most are making operating-model mistakes rather than technology mistakes: treating AI like another software project rather than recognizing it as a fundamentally infrastructure problem that requires rethinking compute, data, governance, security, and operations. A second common error is optimizing for today's proof-of-concept scale without planning for the architecture required when every business unit wants AI.

    Nathan asked whether Dapple is best understood as a market maker or as an agent representing enterprise customers. Tricia rejected both frames, calling Dapple an operating platform: it orchestrates the right infrastructure for each customer's AI strategy and requirements, with Dapple's software sitting across every deployment — regardless of whether the underlying data center is owned by Dapple or a partner.

    Prakash closed with a question on pricing dynamics mid-deal. Tricia described a 'take it or leave it' forcing function driven by capacity scarcity — sales cycles simply cannot stretch to six months because the capacity is gone. She said Dapple honors a quoted price once offered to a customer, even if market pricing shifts during deployment.

    The reality is the enterprise and the government and some digital natives for that matter have been sitting on the sidelines. Why? Because the public cloud doesn't have capacity for them. They're growing faster than ever.

    Our moat isn't owning the most GPUs. It's becoming that trusted operating layer that enterprises build their AI strategy around.

    This is the price right now. This is the location. You want it, take it, or don't — because tomorrow it's probably gone.

    1:38:41Who's so eager to spend this money with you that you've gotten to $100 million in five months?
    Tricia claimed that enterprise and government demand has been bottled up for years because all three existing options are inadequate: public clouds lack capacity for large enterprise workloads; NeoCloud bare-metal options require enterprises to build their own operational stack and software; and on-premises buildouts are too costly and complex to staff. She and co-founder Salam Al Musawi saw no one properly servicing the enterprise gap and built Dapple's Enterprise OS Cloud as the single operating layer to fill it. She said that in their first few months live, they secured over $100 million in customer contracts — and teased that upcoming growth announcements would be even more dramatic.
    1:42:41Your marketing emphasizes you are heavily Azure-native with integration into Azure ML, Azure Arc, and Azure Monitor. Are you transitioning customers from Azure to your cloud, or does your cloud sit on top of Azure?
    Tricia declined to name strategic cloud partners, saying official announcements would come within months. She explained that all three major public clouds are struggling to deliver capacity to their strategic customers because they're consuming that compute themselves. Dapple's core IP is pulling public and private cloud together into a unified, frictionless experience — because no enterprise can build an entirely new software engineering team to manage a separate cloud.
    1:44:29Your website shows 91 to 94% GPU utilization for single-tenant infrastructure. That's surprisingly high — how does that work, and if they're at 94% capacity, where do enterprises go to explore new use cases?
    Tricia explained that enterprise AI differs fundamentally from traditional cloud models. A meaningful portion of Dapple's deployments are backed by long-term enterprise contracts — three, five, or seven years — so the model is about committed capacity and customer outcomes rather than spot utilization. She described it as a take-it-or-leave-it committed capacity contract, noting enterprise customers are sticky and unlikely to leave once mission-critical infrastructure, governance, networking, and compliance workflows are embedded.
    1:46:27What kinds of workloads are running at that utilization level, and are you building ramp capacity into contracts for experimentation?
    Tricia identified three primary workload categories: large-scale model training for AI-native companies building domain-specific foundation models; production-scale inference requiring predictable performance, low latency, and strong governance; and AI agents embedded in internal operations and customer-facing products. She said Dapple started with mission-critical verticals — finance, cybersecurity, government, defense — but found that the broader Fortune 1000 demand was far larger than anticipated. Most customers, she said, are simply desperate to access any capacity to begin deploying AI.
    1:49:42How does Dapple achieve a 6-to-9-month deployment timeline? That doesn't seem long enough to build a physical data center. Are you leasing capacity from existing NeoCloud operators and putting your software stack on top?
    Tricia confirmed the network-orchestration model: Dapple partners with world-class data center operators and infrastructure providers, then contributes capital, GPU infrastructure, enterprise demand, and its software operating layer to turn those facilities into AI-ready infrastructure. In some deals Dapple helps finance the infrastructure directly; in others it deploys on partner facilities. She added that Dapple is also building its own campus infrastructure for ahead-of-demand capacity, and that the Dapple software layer sits across all deployments regardless of ownership structure.
    1:51:51How much do you see across the AI infrastructure space the pattern of long-term commitments generating short-term revenue, and could an external shock cascade through that structure?
    Tricia said the market has grown significantly more disciplined over the past twelve to eighteen months, shifting from speculative 'build it and they'll come' infrastructure to demand-backed contracts. She argued Dapple's model is more de-risked than competitors because enterprise customers make large upfront down payments and commit to multi-year relationships, enabling Dapple to finance GPUs and data centers. She also flagged that CSPs and NeoClouds relying on certain geopolitical off-takers face growing risk from compute export restrictions by the current administration.
    1:55:34It strikes me that the Dapple model is really much more asset-light than the existing NeoCloud model. Is that correct, and is that an advantage?
    Tricia agreed. She said most NeoClouds and CSPs are trying to own everything — data centers, GPUs, and demand simultaneously. Dapple came in as a software-first company that partners with data center operators, debt providers, and demand sources rather than trying to own the full stack. She noted there are scenarios where Dapple does own GPUs or build campuses, but those exist alongside a broader partner network. She said being software-first is where Dapple will ultimately win.
    1:57:27In a world where AI coding agents increasingly simplify cloud complexity, what is Dapple's defensible form factor eighteen-plus months from now — what remains a step above what someone could get directly from the hyperscalers?
    Tricia said there is no single moat; defensibility comes from three reinforcing layers. First, enterprise relationships and switching costs: once mission-critical AI infrastructure is deployed, the governance, networking, compliance, security, and operational workflows all move together — enterprise won't want to touch that. Second, an operating platform that gets smarter with each deployment, improving placement, capacity planning, performance, and cost management. Third, a broad ecosystem of infrastructure providers, capital partners, silicon partners, and enterprise customers that is hard to recreate simultaneously. She drew an analogy to AWS abstracting on-premises complexity, calling Dapple the same abstraction layer for the AI infrastructure era.
    2:01:58What are the two or three most important KPIs enterprise customers track when evaluating their relationship with you?
    Tricia listed five: time to deployment, given the urgency across the market; availability and reliability for mission-critical and in-country isolated workloads; scalability as AI usage grows from hundreds to tens of thousands of GPUs; governance and compliance — which she described as probably the most important thing no one in the market can currently support; and predictable economics over the full deployment lifetime, which every CFO prioritizes above per-GPU pricing.
    2:03:51Based on your history in Africa, how accessible is AI infrastructure to African companies — are they able to buy in, or are they priced out?
    Tricia said Dapple's demand comes largely from APAC, the US, and Europe, with minimal signal from South America, Africa, or parts of the Middle East. She described Africa as a continent falling behind on AI production capacity, with cost as the primary barrier — citing new hardware like Nvidia's Vera Rubin as making deployment increasingly prohibitive for companies and nations that aren't among the largest industrialized economies. She called it a genuine problem analogous to prior technology gaps and said governments are trying to step up but struggling with the economics.
    2:06:39What are the major mistakes enterprise AI buyers are making right now in their purchasing decisions?
    Tricia said most enterprises are making operating-model mistakes rather than technology mistakes. The biggest is treating AI like another software project rather than recognizing it as fundamentally an infrastructure problem that requires rethinking compute, data, governance, security, and operations. A second common error is optimizing for today's proof-of-concept scale — a hundred users — without planning for the architecture required when every business unit wants AI. She noted enterprises also move far slower than AI-native companies, and in a capacity-constrained market, slowness has real consequences.
    2:08:54How would you characterize Dapple's activity — market maker or agent representing your enterprise customers in the market?
    Tricia rejected both frames. She called Dapple an operating platform: it works with enterprises to understand their AI strategy and operational requirements and then orchestrates the right infrastructure to deliver that outcome. She emphasized that the Dapple software sits across every single deployment and transaction, regardless of whether the underlying data center is owned by Dapple or a third-party partner — that consistent software layer is what makes Dapple a platform, not a broker or marketplace.
    2:10:18How often does pricing move while a deal is still in flux?
    Tricia described a forcing function driven by capacity scarcity: sales cycles literally cannot stretch to six months because the capacity disappears. She said the pitch to customers is effectively 'this is the price right now, this is the location — take it or don't, because tomorrow it's probably gone.' Pricing does change in the market, but she said Dapple honors the price it quoted to a customer, even if market rates shift during the deployment period.
    Lightly edited · timestamps jump to YouTube
    1:34:22

    Nathan Labenz: The risk is that if you squeeze too hard, the key things slip through your fingers. Their launch statements read as ideologically motivated — when you invoke concentration of power in a launch statement, I think that's fair to say. They seem to have started this company because they didn't like the direction Anthropic was taking, trying to hold things back. For now it remains a free country and they're going to do it. They've got $200 million from a16z and other VCs, and they say a large investment from NVIDIA on top of that — exact dollar amount not disclosed. It's also an interesting move to see NVIDIA come in and say, we don't want this kind of frontier R&D monopolized, so we'll support a company trying to democratize that capability. The strategic dynamics here are going to continue to fascinate.

    1:35:41

    Prakash: Speaking of decentralizing — our next guest is Tricia Martinez, founder and CEO of Dapple, an AI infrastructure company that is fundamentally rethinking where and how enterprises run their most critical models. Historically, if a company wanted to run heavy AI workloads, they faced a difficult binary choice: either rent shared space from massive public clouds like AWS or Azure, or spend years and billions of dollars building their own private data centers. Tricia and her team are introducing a third path.

    Dapple builds dedicated, single-tenant, in-country AI clouds. That means a bank, a hospital, or a government agency gets its own isolated physical hardware governed strictly by local laws, but operated with the ease and interface of modern cloud software. Tricia previously helped launch one of Africa's first digital banks and mentored startups at Techstars, giving her deep firsthand understanding of what heavily regulated industries actually require to adopt new technology. She is joined at Dapple by COO Salam Al Musawi, whose teams have historically deployed over 300,000 AI accelerators globally.

    Just this month, Dapple announced a $30 million seed round backed by Raptor Group and ION Pacific, and revealed they have secured over $100 million in enterprise contracts in just five months of operation. At a moment when global regulators are demanding sovereign AI, Tricia is building the exact physical and software infrastructure to make it a reality. Tricia, welcome to the show.

    1:37:44

    Tricia Martinez: Hey, Prakash. Hi, Nathan. Thanks so much for having me — and thanks for the intro.

    1:37:50

    Nathan Labenz: Welcome, great to meet you. Can we try to improve your sound a little bit before we go into conversation? There's a lot of background noise — sounds like you're boiling tea right next to your microphone.

    1:38:03

    Tricia Martinez: Oh, really? Maybe I'll put my phone on vibrate — it's blowing up, so that might be why. I can take the AirPods out and go direct if that's easier.

    1:38:16

    Nathan Labenz: See if the input is actually on the AirPods — there's just a lot of background noise.

    1:38:27

    Tricia Martinez: There we go. Is that better? Oh good, my god — sorry, guys.

    1:38:31

    Nathan Labenz: Fantastic. Now we're here. Welcome.

    1:38:36

    Tricia Martinez: Thank you for having me. What do you guys want to talk about?

    1:38:41

    Nathan Labenz: Who's so eager to spend this money with you that you've gotten to a hundred million in five months?

    1:38:45

    Tricia Martinez: I'm not kidding you — it's madness. Where to begin? I like to start by addressing what everyone asks: is there an AI bubble? Is the enterprise really jumping in? Is this just digital natives and hyperscalers? Is this a bubble where we're over-investing in infrastructure and GPUs? I will tell you that framing is wrong. The reality is the enterprise, the government, and even some digital natives have been sitting on the sidelines. Why? Because the public cloud doesn't have capacity for them. They're growing faster than ever.

    They are the largest consumers of compute and need it for themselves. The only options enterprises have are: go to a public cloud they already operate in — but those clouds don't have capacity for them; go to a NeoCloud offering bare metal — but that requires the enterprise to build out its own systems, its own software, and operate a separate cloud alongside the public cloud it already lives in; or build on-premises — which nobody wants to do. It's costly, it's complex, you have to hire a massive team of infrastructure engineers. It doesn't make sense.

    So my co-founder Salam and I came together. He has deep expertise in the GPU space — him and his team have built neo clouds and deployed over 300,000 GPUs. I came from a background of startup investing, and I also served as a White House Fellow under both the first Trump and Biden administrations at the Department of Energy, where I worked across our seventeen national labs building an AI and energy strategy for the nation. We both said: the enterprise is being sidelined. No one is servicing them right now. That's the opportunity.

    We believe at Dapple that AI has fundamentally changed infrastructure. For the last twenty years, enterprises could run applications almost anywhere. AI has changed that. Today, where your compute sits, who controls it, how it's governed, and how quickly it can scale all matter. That's why we built what we call the Enterprise OS Cloud — a new operating layer that helps enterprises deploy and operate production AI across dedicated infrastructure. The simplest way to think about Dapple is as the operating system for production AI. We're not building foundation models. We're not building GPUs. We just make it possible for the enterprise to deploy AI reliably, securely, and at global scale.

    AI infrastructure is becoming increasingly complex, and our job is to make that complexity disappear for the enterprise. Through our network and strategic partners, we've pulled in incredible demand — enterprises have been sitting waiting for years for capacity, literally years. In our first few months live, we ended up securing over $100 million in customer contracts. And the reality is you'll have to have me back, because I'll have some very, very big updates on our growth in just a few months that will wow you.

    1:42:41

    Prakash: Let me take a step back. A lot of your technical marketing emphasizes that you are heavily Azure-native and integrate seamlessly into the Azure Machine Learning control plane, Azure Arc, and Azure Monitor. How does this work? Are you transitioning customers from Azure to your cloud? Does your cloud sit on top of Azure? How does the interaction between the two work?

    1:43:12

    Tricia Martinez: Good question — and you'll have me back in a few months once we've made some official announcements about our strategic partners. But basically: the enterprise lives in the public cloud, and all three public clouds are struggling to deliver capacity to their strategic customers, because they're consuming that compute for themselves. So we saw a unique opportunity. Enterprises are sitting on the sideline. They need capacity and a unified experience — no enterprise wants to build separate clouds the way a NeoCloud offering would require. They can't go build an entirely new software engineering team to manage yet another cloud.

    Pulling together the public and private cloud into a seamless, frictionless experience — that's the IP we've been building. We're doing this with strategic partners who are enabling our rapid growth, and those announcements will come shortly.

    1:44:29

    Nathan Labenz: Another number that jumped out from your website is 91 to 94% GPU utilization on the infrastructure page. For single-tenant dedicated hardware, that's a surprisingly high number. Even at enterprise scale, with all the coordination challenges, you'd expect more headroom. Am I understanding that claim correctly? And does this mean enterprises are running a lot of batch jobs overnight, with you helping orchestrate that? Because if they're already at 94% capacity, where do they go to explore new ad hoc use cases?

    1:45:30

    Tricia Martinez: Enterprise AI is very different from the traditional cloud model. A meaningful portion of our deployments are backed by long-term enterprise contracts — we're optimizing around committed capacity and customer outcomes rather than chasing spot utilization. It's a much more predictable business model. The contract is: this is the capacity, take it or leave it. We're building partnerships with enterprises who are committing to three, five, seven-year deployments with us. Once an enterprise commits to mission-critical AI infrastructure, they're unlikely to want to leave. We're not just moving GPUs — we're moving the governance, the networking, the compliance, the security, the operational workflows.

    1:46:27

    Nathan Labenz: What kinds of workloads are those, though? So much of what I do day to day in AI is spot and unpredictable. Enterprise workloads would have to be much better characterized than mine — that makes sense. But I'm wondering what companies are actually running at that utilization level. And especially: if you have that kind of high utilization from the beginning, are you baking in ramp capacity? Because if they're already at 94%, where do they go to experiment with new use cases?

    1:47:12

    Tricia Martinez: It's about the outcome, not GPU per hour — that's how the market thinks now, but for us it's about what the customer needs delivered. We're seeing three primary categories. First, large-scale model training: AI-native companies need dedicated GPU infrastructure to train their next-generation foundation or domain-specific models. Second, inference at production scale: enterprises are deploying AI applications that need predictable performance, low latency, and strong governance. Third, AI agents and enterprise automation: organizations are embedding AI into internal operations and customer-facing products.

    We originally started focusing on mission-critical workloads — finance, cybersecurity, government, defense-related operations — which require private, sovereign, fully isolated deployments. But in reality, the market is far larger than we anticipated. The entire Fortune 1000 is in serious need of capacity. While not all of them require sovereign-grade isolation, they are all in serious demand, largely because they've been sitting on the sidelines. There's real time urgency. Most of these enterprises do care about compliance and sovereignty — but for many, they're simply desperate to access any capacity so they can start building or deploying AI.

    1:49:42

    Prakash: I've seen that you can get a deployment up and running in six to nine months — which is very fast. But six to nine months doesn't seem long enough to actually build a physical data center; even Elon's all-time record is something like fifteen months from groundbreaking to operation. So in that six to nine months, what exactly are you doing? Are you leasing or renting capacity from existing neo clouds, putting your software stack on top, and then plugging that into customers who are already Azure-native and moving them over? Is that the frame?

    1:50:38

    Tricia Martinez: Yes — and that's a great question. We can't do everything. This market is enormous and the demand we're getting is massive. So we're creating and orchestrating a network. Our model is to partner with world-class data center operators and infrastructure providers, then bring together the capital, the GPU infrastructure, the enterprise customers, and the software operating layer that turns those facilities into AI-ready infrastructure. In some cases we're deploying infrastructure that we help finance. In others, we deploy on partner infrastructure.

    Part of our strategy is building out our own campuses — we need capacity available before the demand hits. But whether we own the GPUs on our balance sheet, finance the data center ourselves, or rely on third-party partners for data center operations or GPU deployment, the Dapple software sits across all of it.

    1:51:51

    Nathan Labenz: I have a question on the overall financial structure we're building in the AI space. It sounds like your customers are making long-term commitments to you, which presumably insulates you from riskier dynamics. But how much do you see across the space as a whole that pattern of long-term commitments generating short-term revenue — creating the potential for a cascading problem if someone has a bad quarter? We've seen that pattern play out in previous financial crisis moments. I'm not sure how much of it is present today or how likely an external shock could be to really throw things off their axis.

    1:53:04

    Tricia Martinez: The market has changed drastically since we jumped in. We're seeing enormous amounts of capital still flowing into AI infrastructure, but investors have become much more disciplined over the last twelve to eighteen months. The market is definitely moving away from 'build it and they'll come' toward infrastructure that has to be backed by real demand and long-term customer contracts. The model we're building is significantly more de-risked than what exists today.

    The CSPs and NeoClouds are servicing huge hyperscaler deals, but there aren't that many of those to go around, and some carry real geopolitical risk. Now with regulatory restrictions increasing, financing groups are worried that the current administration will crack down on which countries or customers can consume compute — and that puts a lot of those CSP operators at risk.

    Our approach is attractive to financing partners precisely because of its repeatability. These are enterprise customers making large upfront down payments, enabling us to finance GPUs and data centers. They're sticky — they repeat with you rather than go find a new provider. Yes, the financial market is becoming more risk-averse because so many NeoClouds and CSPs have taken bad deals, and financing partners are bearing that burden. But our model is a different approach, and I believe it's more attractive to financial markets.

    1:55:34

    Prakash: It strikes me that the Dapple model is really much more asset-light than the existing NeoCloud model. Would you say that's correct? And is that an advantage?

    1:55:50

    Tricia Martinez: Absolutely. Our approach has been: we cannot possibly do everything in this market, and I think most NeoClouds and CSPs are trying to do exactly that — own all the data centers, all the GPUs, all the demand. We came in saying we need strategic partners: data center operators, debt providers, demand aligned through strategic relationships. We can't own everything. Where do we thrive, and where can we work with partners to help us build this greater ecosystem?

    There are scenarios in which we do want to build our own DC campuses. There are scenarios where we own GPUs on our balance sheet. But there are also scenarios where we work with partners to deliver value to customers. At the end of the day, what we care about is solving the customer's problem and delivering value in a timely manner. So yes, it is less asset-intensive compared to other players in the market. Being a software-first company is where we'll ultimately win.

    1:57:27

    Nathan Labenz: Something that was trending yesterday was this notion of the 'software factory' — from Warp, and popping up elsewhere. The CEO publishes a memo saying our job now is to build the factory that builds the product. I'm looking at that concept and wondering: is that one too many levels of abstraction, or is it exactly what the doctor ordered?

    When you think about creating a defensible position over years, one of the challenges is that someone could just go to coding agents and say, 'set me something up on Azure.' If Azure is too complicated today, eventually coding agents will simplify it substantially. What's your form factor long term? In an 18-months-plus world where my coding agent can navigate even the Byzantine complexity of cloud as it exists today — barriers to entry presumably lower, friction reduced — what remains a step above what I can get directly from the hyperscalers?

    1:59:16

    Tricia Martinez: Thinking about it in terms of defensibility — I don't think there's a single moat in AI infrastructure. Defensibility is going to come from a combination of several things that reinforce each other.

    First, enterprise relationships. Having demand in the market is key — if you don't have demand, you have nothing to sell. We've built strategic partnerships and doubled down on enterprise relationships. Once an enterprise deploys mission-critical AI infrastructure, switching becomes extremely difficult. They're not just moving GPUs — they're moving governance, networking, compliance, security, and operational workflows. Enterprise doesn't want to move all of that.

    Second, the operating platform. Every deployment makes our software smarter. We learn how to optimize placement, capacity planning, performance, cost, and policy across all these different environments. Over time, that becomes increasingly valuable.

    Third, the ecosystem — probably the most important. We're building relationships across infrastructure providers, capital providers, silicon partners, and enterprise customers simultaneously, and that is very difficult to recreate. Our moat isn't owning the most GPUs. It's becoming that trusted operating layer that enterprises build their AI strategy around.

    AWS created the cloud for the enterprise by abstracting everything. Every enterprise was building on-premises with clunky, complex ecosystems, and AWS made it frictionless. This is now the new wave. AI is demanding something entirely new from an infrastructure perspective, and once again the enterprise doesn't want to deal with financing, GPU ownership and depreciation, building and managing data centers. We're abstracting everything, similar to how AWS did — but in this new wave of AI infrastructure. The defensibility is much larger than it might appear, because there are so many pieces you have to move for this to work exceptionally well.

    2:01:58

    Prakash: When your customers evaluate their relationship with you, what are the two or three most important numerical KPIs they're looking at? Is it performance per watt? What are they tracking on a weekly or monthly basis?

    2:02:27

    Tricia Martinez: There are definitely technical components, but what matters most to the enterprises we work with comes down to a few things. One: time to deployment. Everyone needs something live, and time is really critical right now. Two: availability and reliability — can they rely on the infrastructure for mission-critical workloads? If they're doing drug discovery, can it be in-country isolated with consistent performance? Three: scalability — as their AI usage grows, can we scale with them? Maybe they start with a thousand GPUs, but in two years they want ten thousand. Four: governance and compliance — probably one of the most important things that no one in the market can currently support them on: meeting regulatory, security, and data residency requirements while operating AI. And five: predictable economics over the life of the deployment. Every CFO cares deeply about that. It's not just the per-GPU price; it's the outcome economics over the full term.

    2:03:51

    Nathan Labenz: Last one for me. Based on your history in Africa, I wonder if you have a take on how broadly this technology is diffusing and how accessible it is to companies there. Are they able to buy in? Are they priced out at this point? What are we going to see in terms of the financial companies, the telcos of Africa, and their ability to keep up with what deeper-pocketed companies in richer countries are doing?

    2:04:38

    Tricia Martinez: That's something I care deeply about. I care about building technology that solves actual problems in our world. Most enterprises have by now proven that AI works — we know it works. The challenge is moving from experimentation into production. But what we see in our demand is: largely APAC and the US, followed by Europe. We're not seeing that from South America, from Africa, from much of the Middle East.

    The more industrialized countries are investing aggressively in AI, which means Africa as a continent — all fifty-four countries — is probably falling behind on the AI production side. The cost is prohibitive. With new hardware like Nvidia's Vera Rubins coming out, costs are increasingly out of reach for many companies and nations. Just like prior technology jumps — blockchain, for example — this is a real barrier. Governments are trying to step up, but these things are genuinely expensive to deploy. That's something we all need to examine seriously.

    2:06:39

    Prakash: When you talk to enterprise AI buyers, what do you think are the major mistakes they're making right now in their purchasing decisions?

    2:06:55

    Tricia Martinez: Most enterprises aren't making technology mistakes — they're making operating-model mistakes. The biggest is assuming AI can be deployed like traditional software. AI is fundamentally an infrastructure problem. It changes how you think about compute, data, governance, security, and operations. In certain scenarios, we do a lot of hand-holding to really ensure the enterprise is making the right decisions.

    A second mistake is optimizing for today's use case instead of tomorrow's scale. Many organizations build a proof of concept that works for one team, but they haven't thought about what happens when every business unit wants AI. The architecture that works for a hundred users often won't work for a hundred thousand. And enterprises simply don't move as fast as AI-native companies. In this market, if you don't move fast, the capacity is gone — and that creates real risk. It takes longer to sell, and there's more thought required about what's being purchased and why.

    2:08:50

    Nathan Labenz: Maybe one more, if you have an extra second.

    2:08:53

    Prakash: Yeah.

    2:08:54

    Nathan Labenz: How would you characterize your activity — market maker versus agent representing your enterprise customers in the market?

    2:09:08

    Tricia Martinez: I'd say neither. We're an operating platform — we're not a market maker, and we don't think of ourselves as a broker or intermediary at all. We work with enterprises to understand their AI strategy and operational requirements and then orchestrate the right infrastructure to deliver that outcome. What's really important to emphasize: no matter whether a third-party partner is running the data center and the deployment, or we are ourselves, the Dapple software sits across every single deployment and transaction. That's the constant. We're building a unified public-private offering that's frictionless and seamless for the enterprise, taking into account their compliance and regulatory needs. They're not buying access to a marketplace — they're buying a production AI platform.

    2:10:15

    Nathan Labenz: Thank you.

    2:10:18

    Prakash: I think that's probably our last question — we don't want to keep you too long. This has been interesting because you are deeply embedded in the sales process happening in the industry right now, and it's very difficult for us as external observers to understand what buyers and the new clouds are actually doing and how deals are moving. One last question: how often do you see pricing move while a deal is still in flux?

    2:11:03

    Tricia Martinez: The upside of this market is that sales cycles literally cannot be six months with an enterprise, because the capacity is gone. So we have a forcing function: this is the price right now, this is the location, you want it — take it, or don't — because tomorrow it's probably gone. We have this constant movement with customers, which is at times difficult for the enterprise, but the reality is everyone wants capacity right now and they're willing to move at a much faster pace. Pricing does change, but once we deliver an offering to a customer, we generally honor that price even if market pricing shifts. If we're committed to a deployment, we're buying at a certain price and we honor that with our customers.

    2:11:52

    Prakash: Thank you, Tricia. Thank you for your time, and we hope to have you back again as the company expands and takes over the control plane of AI infrastructure.

    2:12:06

    Tricia Martinez: Wonderful. Thank you so much for having me. As a takeaway: any enterprise or digital native companies in need of compliant private workloads and looking for the next generation of AI infrastructure, please reach out. I'll be back soon with even greater updates.

    2:12:31

    Nathan Labenz: Indeed. Looking forward to it.

    2:12:33

    Tricia Martinez: Awesome. Thanks, guys. Have a good one.

    2:12:36

    Prakash: You too.

  4. 2:12:52Closing31 min
    Close: Market Maker or Control Plane, and What Software Is Worth When Agents Are the BuyersThe hosts debate whether Dapple is a GPU market maker or a defensible control plane — Prakash on the artisanal reality of bare-metal deployment and lock-in, Nathan skeptical and citing Snowflake/Databricks valuations — then widen into what human-interface software is worth when agents become the primary buyers, a hyper-deflation scenario for software margins, the durable value of human time and attention, and a market check before previewing Cameron Berg and Forum AI's Robbie Goldfarb.

    The hosts picked up a thread on Dapple, an AI infrastructure marketplace, debating whether the company is fundamentally a GPU market maker or something with more defensible technical depth. Prakash explained that deploying AI workloads on bare-metal neo-cloud hardware is genuinely artisanal work: each data center has idiosyncratic GPU behavior driven by wiring, power stability, and failure-rate quirks, requiring a specialized infra team to tune. His case for Dapple's value was that a portable control plane — one that behaves consistently across data centers — plus a shared infra team that regulated-industry CTOs can hand liability off to, amounts to real lock-in and margin opportunity. Nathan remained skeptical that Dapple could solve those problems faster than the neo-clouds themselves, and pressed on margins.

    Prakash noted that the chip generation cycle (H100 → B200 → B300) continuously resets the stability problem: just as one generation gets stable, the next leading-edge generation arrives with fresh quirks, making the reliability gap persistent rather than a one-time fix.

    On the regulated-industry angle, Prakash outlined Dapple's margin thesis: go in at cost for the first enterprise customer, amortize the same infra team across many customers, and capture the cost-improvement tailwind as GPU prices fall while customer contract prices stay fixed — a classic financial-IT lock-in play.

    Nathan widened the lens to Snowflake ($75B) and Databricks ($100B+) as examples of how much the enterprise pays for friction-reducing software layers, expressing honest bafflement at those valuations while acknowledging he may be underestimating switching costs and the value of liability handoff.

    They explored the disintermediation thesis: Nathan argued that AI agents — already giving him precise deep-links into Google Cloud config panels — should eventually reduce the need for middleware vendors. He offered a counter-story, though: if AI is moving so fast and hitting the core of every enterprise's business, no one has bandwidth to rationalize the app layer, and marginal savings on middleware become irrelevant.

    Prakash sharpened the question: what is the value of human-interface software in a world where agents are the primary software consumers? Agents are indifferent to UX polish, will happily navigate ugly APIs, and will drive near-perfect price discovery across the market. He sketched a future where 90–99% of software is consumed by agents, making them the real buyers and the real arbiters of build-versus-buy decisions.

    Nathan agreed margins across the app layer look structurally challenged — lower search costs and lower switching costs push toward perfect-market conditions with thin profits. Prakash extended this into a hyper-deflation scenario where software production costs approach zero, revenues concentrate in zero-sum advertising auctions (the Red Queen races on Google and YouTube), and most producers are squeezed between declining margins and rising customer-acquisition costs.

    Nathan weighed in on advertising's future: human time and attention remains scarce and valuable, so the cost of reaching humans probably stays high. But advertising to agents might open new models — subsidized inference, paid product consideration — where advertisers pay for an agent's compute to evaluate their product, and the human makes a lighter final call. He remained broadly bullish on the value of human time.

    Prakash made an extended aside on vacation planning: most of the time people spend 'planning' a trip is really daydreaming and fantasy, not execution. He argued that AI companies pitching 'AI will book your vacation' are solving the wrong problem — the real opportunity is helping people daydream better, possibly through AI-generated VR highlight reels of trips-not-yet-taken. Nathan agreed, noting that his own AI-assisted family road trip worked because it was highly interactive and iterative, not fully delegated, and that welfare economics research supports the idea that much of a vacation's value lives in the anticipation.

    Quick market check: hyperscalers broadly in the red, semiconductors doing well. Apple down 4.5% after raising device prices. NVIDIA below $200. Markets are speculating on how long the hyperscalers will absorb chip-supplier pricing pressure before slowing data-center build-out — with Microsoft seen as most likely to blink first, while AWS's Jassy is doubling down.

    Nathan previewed tomorrow's show: Cameron Berg joins to discuss AI welfare research — functional emotions and now functional welfare, what happens when a model receives reward and how it may 'feel' about that. Forum AI's Robbie Goldfarb will discuss bringing the highest levels of human taste and judgment into AI systems. Both conversations promise to be thought-provoking.

    No bank CEO wants to deal with it. I don't want to talk about sunspots — I want to talk about interest rates.

    A lot of the vacation planning process is actually just daydreaming and fantasy. You're not actually trying to plan a vacation — you're trying to dream a little bit.

    I'm still generally bullish on the value of human time. Feels like that's going to go up for a while at least.

    Lightly edited · timestamps jump to YouTube
    2:12:53

    Nathan Labenz: You know more about this kind of stuff and all the financing that goes on than I do. But what do you make of her answer to the market-maker-versus-agent question? It seemed to be in some tension with the subsequent comments around facilitating these deals — price is here today, it's gone tomorrow. I don't know what a market maker is if that's not a market maker. It sounded fairly market-maker-like to me. How would you parse that mix of answers?

    2:13:43

    Prakash: There are two ways to look at it. One is, yeah, it's pure market making. The second way is that it's actually quite difficult to deploy the control plane on bare metal on the new clouds right now. Yitay — who was at Google, then left, founded his own company, wrote a fantastic blog post, and then went back to Google — found that every single data center has its own specific quirks on how the GPUs are run. It can be as simple as what kind of wiring they used, whether the power is steady or there's some fluctuation, whether you get more frequent GPU failures. Even the same GPU in a different data center behaves differently. Different data centers have completely different quality, and you end up needing a team like Yitay's to manage the infra in order to use the data center effectively.

    You really only have two choices: use the hyperscaler's control plane or the new cloud's control plane, and you're basically enslaved to it — because that workload doesn't work the same way in another data center. If you move, you have to deploy again and figure out all the infra details: which GPUs work well, how to manage power fluctuations. It becomes a very artisanal deployment process.

    But if you can build a control plane that manages consistently no matter where you deploy, you have the flexibility to move workloads when needed. Maybe it's 5% less efficient than the best case, but you have portability. If you can say 'sign with us for a thousand GPUs, and as you scale we can add locations with the same control plane,' that's valuable. It's what Nebius is trying to build, and what CoreWeave is halfway toward. If they're doing both the market making and managing the infra underneath — so enterprises don't need their own infra team — that's actually quite valuable. It's basically lock-in. Once you have everything running, moving to another control plane means hiring the infra team again, spending a year fine-tuning again, figuring out when the solar flares occur, when the bits flip. No bank CEO wants to deal with that. I don't want to talk about sunspots — I want to talk about interest rates.

    2:18:18

    Nathan Labenz: I remember that blog post, and one anecdote that stood out was that he said there was a time he left a training job running at Google, kind of forgot about it for a couple of weeks, came back, and it was still running. The contrast between that and what you could do on any of the neo-clouds was night and day — something's always going to go wrong on the neo-clouds. That was 2024, though. So an obvious question is: shouldn't the neo-clouds have shored things up quite a bit by now? Maybe some of them can get by without solving it because demand is so high they can sell even if their product is rough around the edges. But it sure seems like they've had plenty of time to solve a lot of those problems.

    2:19:16

    Prakash: What happens is new chips come out. They stabilized on H100s, and H100s are very stable now. But then B200s came out, and you have to go through optimization all over again. On the leading edge, everything is on the leading edge. The chips are less reliable at launch because a lot of this hasn't been figured out. People plug them in, ask why they're failing, go back to NVIDIA, NVIDIA does a fix overnight and deploys to all the GPUs. There's this group activity of discovering the capabilities and pitfalls of each chip generation. It stabilizes over two or three years. H100 eventually became really stable — but then B200s came out and people complained badly about them, and some of those were hardware problems. Now B300s are coming out, and the hope is they fix those hardware issues, not just the software issues. The frontier keeps moving and you keep having to do the work. Maybe you get to auto-fixing all these chip-level problems in the next 18 months.

    2:20:46

    Nathan Labenz: It's a little hard for me to wrap my head around how a company like Dapple fixes that faster or better than the neo-clouds do for themselves, which leaves me back in the spot of thinking it feels more like a market at core. Speed to launch seemed like the number one KPI she discussed — a lot of the KPIs she mentioned were the kind that would go into your overall analysis of AI, but in terms of why buy from her specifically, speed to live was a huge one. And that is a valuable thing to provide today. What would you guess are margins for a company like Dapple? How would you compare it to the 6% real estate agent take on a traditional residential transaction?

    2:21:59

    Prakash: It's really a regulatory issue. She's going to all regulated industries — banking, et cetera. In regulated industries, you often hire an external vendor because you can hand off liability. They have an infra team that gets updates from the new clouds or chip providers on what needs to deploy. Every time you deploy, the CTO of the bank might have to sign off — but the CTO doesn't have an infra team to evaluate this. If you can share one infra team across multiple banks, everyone saves because there's one team evaluating all the fixes and updates, and all the CTOs can sign off together. You're sharing the cost across multiple entities, and that's valuable.

    On margins: in finance, IT providers often go in at cost and fix the price at that cost level. But over time the technology gets cheaper. So you charge a hundred dollars an hour, but in two years the GPU price is 30% of that, and you capture that cost improvement. You lock in regulated firms at basically cost now, and the costing gets optimized over time. Your first customer costs you whatever the infra team costs — but your next customer uses the same infra team, so essentially all of that is margin. It could come down to how they structure contracts.

    2:24:47

    Nathan Labenz: I wouldn't bet on those GPU prices falling, but they can at least hope to amortize their platform development costs at a minimum. I'm always confused by companies like this in general. Snowflake's market cap is currently $75 billion, Databricks is private and over $100 billion. I squint at that and think — really? For a software layer that sits on top of these clouds and manages data that ultimately still sits on AWS hardware somewhere? But I guess this goes to show my naivete about how much friction exists in the enterprise, how much switching cost there is, how much they're willing to pay for predictability or to hand off liability, as you said.

    Those could be interesting canary-in-the-coal-mine companies. If this all goes as I think it will, you're just going to get really good cloud management from your core intelligence provider. I already see this in setting up my own personal AI infrastructure — I'll hit a moment where I think, 'I need to enable what in my Google Cloud project?' And Claude gives me a deep link to exactly the place I need to click and hit enable. How long would it have taken me to find that otherwise? Sometimes, a long time. Disintermediation really does seem possible.

    But there's also a counter-story, which might be the possible salvation for the app layer: everybody's so overwhelmed with everything happening and so focused on applying AI to their core business that they just don't bother swapping vendors on anything. If you're a big bank or pharma company, sure, you could probably save money on your Snowflake or Dapple contract and get the ROI fairly quickly if you put some brains and some AI agents on it. But the AIs are coming right for the core of your business, so you don't have the luxury of rationalizing the margins — every bit of brainpower needs to go to defending the core.

    I can always talk myself into a lot of different stories. I usually think that the faster intelligence progress goes, the longer the tasks, the more reliable the autonomous execution — the worse it's going to be for the app layer. But there's this other story: if it goes fast enough and goes to the heart of people's businesses, there's just no time to even think about rationalizing the app layer, and all those marginal savings opportunities don't matter.

    2:29:40

    Prakash: The question I have is: is there a difference between software that has an interface to humans versus software that only interfaces with agents? Does Snowflake or Databricks have value right now because it provides an interface for humans to interact with cloud systems? If agents were directly interfacing — if no human ever touched Snowflake or Databricks at all, if those names were not even known to humans — what margins could they charge, and what value would they actually provide? An agent will go through the tedium of signing up for four different things, reading ten different lists, reading the docs and the API, doing whatever is necessary to get it done. It's not going to be particularly bothered by usability. It's more than happy to use relatively non-user-friendly code that's more efficient. What value do these things actually have?

    I kind of imagine a future where 90 or 99% of software is used by agents. The agents themselves become the major buyers and will have to do the price discrimination — the buy-versus-build decisions on all of this software. What are the agents willing or able to pay for?

    2:31:50

    Nathan Labenz: It would seem like it's going to be tough for margins in general. The drop in search cost, the drop in switching costs — it all would seem to lead us much closer to perfect market conditions. And in a perfect market environment, not a lot of profits.

    2:32:17

    Prakash: And if you see this kind of hyper-deflation — where the price of all the software stuff starts to plummet — take a business like Google, offering an interface to humans watching ads. If the cost of producing those products, like Intuit's products, is asymptotically approaching zero, Intuit has a problem: to maintain margins, it has to advertise. And you're in this Red Queen race between multiple advertisers on YouTube. I wonder if all the profits end up going to the people who are offering these zero-sum Red Queen races that you have to play in order to access customers and the end buyers who can make decisions.

    2:33:23

    Nathan Labenz: That's been a reality on Google and Facebook for years. In some niches it's been quite consuming — I've seen examples in the digital marketing space where people selling digital marketing services are barely making money once you account for customer acquisition costs on the same platform they're layering services on top of. Advertising in general could go multiple ways, but there's a pretty good case that it will continue to be expensive and valuable because human time and attention remains the most scarce thing.

    Advertising to agents might have somewhat different dynamics, though. We may start to see different ways of cutting up the money. You might be happy for your agent to consume all kinds of promotional materials or product specs, especially if the advertiser is willing to pay for your inference cost to consider their product. I could see different models opening up there. But for a lot of products, people are going to want to make the final call. There does seem to be a market for getting in front of people, consuming a small amount of their time and attention — and the cost of that probably goes up until maybe some point where you trust your agent enough to just let it make calls for you. Then it becomes a very different equilibrium: subsidized inference, paid consideration, some new mode of doing things. That's hard to predict. I don't think people are going to be that quick to let the AI decide where they go on vacation. But there'll be hybrids too — your agent does everything, it's all subsidized, you get a report and make a quick decision. That may feel less like advertising to the human who's the final decision-maker. I'm still generally bullish on the value of human time. Feels like that's going to go up for a while at least.

    2:36:26

    Prakash: I've often seen these ideas about AI helping you book a vacation, and it's always struck me that I've spent a lot more time considering vacation options than I've actually gone on vacation. A lot of the vacation planning process is actually just daydreaming and fantasy. You're not actually trying to plan a vacation — you're trying to dream a little bit. If you had the time, the money, if you could take your kids out of school and go trekking in Machu Picchu — you kind of look at it. Out of every nine vacations I've daydreamed about, maybe one becomes real.

    I think the people focusing on AI booking your vacations are focusing on the wrong thing — the idea that AI is going to help you do this more efficiently. What they really need to figure out is how AI is going to help you daydream better, because the daydreaming is what drives vacation planning, Zillow, even Tinder. All of these things where you spend so much more time searching than executing are really forms of daydreaming. Maybe the solution actually ends up being VR daydreams — vacation sites where the AI generates the entire highlight reel of your trip with you in it. That's super fun. It's not really going to be an AI that plans a vacation, because that's not what you're actually looking for. The intent is not to book a vacation.

    2:38:39

    Nathan Labenz: Yeah, I think that's apt. There's even welfare economics research suggesting that the utility of a vacation is substantially in the anticipation and in the memory of it, not the trip itself. That rings true. In my recent road trip with the kids, which I used AI heavily to plan, it had a lot of that character. The mode that worked for me was very interactive — going off and saying 'plan everything' would not have worked. Give me a ton of options, let me pick some highlights, then find more options based on what I liked. That worked well and led me to something I was really excited about going in, and it actually turned out great. So there's definitely value there, but you can't fully externalize that and capture all of it. There may be a form factor mismatch in some of the ways people talk about this.

    2:40:04

    Prakash: Nathan, another interesting morning on AI:AM. I'm eager to see how the markets have performed. Just to give you a preview — all the hyperscalers are in the red, badly in the red. Semiconductors are doing great.

    2:40:28

    Nathan Labenz: Apple just raised their prices on their devices too, so the effect of the memory crunch is going to be felt in many, many places.

    2:40:40

    Prakash: Apple's down 4.5% for the day, which is a massive move. Even NVIDIA's down below $200 now. I think the effects of Broadcom, OpenAI, the Jalapeno chip are being speculated upon. There is a real question in the market about how long the hyperscalers — really the best-run companies on the face of the earth — are going to be willing to bear this kind of milking by NVIDIA and Micron and the memory producers. The expectation is that Microsoft bows first: at some point Satya says, look, we're going to slow down the data-center build-out. But it's also hard because Jassy at AWS is saying everything is in the black, we're going to make so much money, we're going to deploy more capital, and we are the data-center leader and will outinvest everyone else. We will see.

    2:41:54

    Nathan Labenz: Never a dull moment. We'll see you back here tomorrow morning for another fun edition. Two really good segments coming: Cameron Berg joins us to talk about the latest in AI welfare research — he's really a pioneer of this field. There have been some fascinating papers extending this line of research around functional emotions to now include functional welfare: what can we understand about what goes on when a model gets reward, and how it may feel about that? I think it's almost certain to remain mysterious after tomorrow, but it will be a mind-expanding and potentially hair-raising conversation. And then we have Forum AI — Robbie Goldfarb is really trying to bring the highest level of taste into AI systems. You go to their website and it's just: here are all the experts who we think you would respect as having excellent taste and judgment, and these are the people we're trying to get into the AI. I think that'll be a really interesting conversation as well.

    2:43:36

    Prakash: Absolutely. Looking forward to it.

    2:43:39

    Nathan Labenz: Lots more to come.

    2:43:41

    Prakash: Bye bye.

    2:43:42

    Nathan Labenz: Bye for now.

Opening: Export Controls in Court, GLM 5.2, and the Distillation Debate

The hosts opened on the economics of the AI infrastructure buildout — each generation of investment paying back as long as inference demand holds — before turning to the morning's dominant story: the Fable 5 / Mythos export-control blockade entering its first legal test. Legal-tech firm Legion sued the Trump administration, arguing it lacked statutory authority to force Anthropic to withdraw its most capable models, while a brief reappearance of the Fable 5 endpoint in Amazon Bedrock's listing — quickly dampened by Anthropic's head of growth confirming the models remain unavailable to all customers — fueled return speculation. Nathan catalogued additional signals (a model-name string quietly reintroduced in a Claude Code update, rising prediction markets) and offered a structural read on the administration's pattern: rapid unilateral moves leveraging what he called 'Schrödinger's authority' to extract concessions before courts weigh in. A notable personnel development framed the moment: co-founder Tom Brown has replaced Dario Amodei as Anthropic's lead interlocutor in Washington, his infrastructure-and-compute framing aligning with the administration's build-everything posture.

The conversation moved to GLM 5.2, a Chinese open-source model that circulated widely after a balanced benchmarking thread from Snowflake CEO Sridhar Ramaswamy comparing it to Opus 4.7. Nathan praised the thread's intellectual honesty — it flagged real limitations, including GLM 5.2 spinning out into excessive tool calls on failing agentic runs — while stressing that the threshold effect of being competitive matters far more than precise gaps; the Zapier founder added that GLM 5.2's 23% ARC-AGI score sits near the inflection point where frontier models previously saw agentic capability take off, putting the open frontier roughly six to nine months behind. That led into the distillation debate: Anthropic accused Alibaba of using nearly 25,000 fraudulent accounts to generate 28.8 million Claude exchanges to train Qwen, and Prakash described a rumored gray-market reseller ecosystem offering Claude access at roughly 70% discounts. Drawing on Zvi Mowshowitz's pushback, Nathan noted distillation is doubly efficient because it transfers Anthropic's hard-won taste and alignment, not just raw capability — prompting Prakash's provocative reframe that, for a 'race to the top' safety org, widespread distillation might actually spread alignment characteristics to Chinese labs. The segment closed on mounting Google DeepMind attrition (Arthur Kami and Rishab Jain the latest) and ban-superintelligence calls from MIRI's Nate Soares and Francis Fukuyama.

Interview: Eric Olson — AI for Science and Surviving as an Application-Layer Company

Eric Olson, CEO and co-founder of Consensus, located AI for science as behind leading domains like coding but on a similar-ish curve, and laid out a deliberate bet: build for the leverage available today rather than wager the company on a 'push a button, get a scientific discovery' future that, if it arrives, would reshape everything for everyone — including Consensus. He described a striking trend from his usage data: average query length has grown exponentially over just the past year, with users now submitting multi-step research briefs (search this, cross-reference that, synthesize a gap analysis, save the ten most relevant papers), a shift he attributed largely to expectations set by tools like Claude Code. On balancing model flexibility with citation-grounded guarantees, he detailed a 'recipes' architecture that classifies requests into guardrailed agentic paths and always runs searches under the hood — even when that means telling a user 'we couldn't find papers on this' — and noted that better models are driving some convergence across competitors. On the numbers: roughly 70% of users come from academia and 30% from professional contexts (split between clinicians and R&D, with bio dominant); token costs are rising as agentic task lengths grow even as per-token prices dip; and the speed-up in science runs days-to-hours rather than the steeper gains seen in fully verifiable domains.

Responding to Nathan's theory that science products should default to the best frontier model until a capability threshold makes routing the real moat, Eric agreed routing has always been part of the picture — Consensus already routes some twenty narrow classification tasks (field-of-study classification, search-ranking signal weighting) to sub-billion-parameter BERT-series models like BioBERT and PubMedBERT, returning results in sub-0.1 seconds, while frontier models handle complex synthesis. For a narrow classification task with a good fine-tuning set, he put recoverable performance at roughly 95% of frontier capability down to about a one-billion-parameter model. On competition and stickiness, he named three decision factors — trust and accuracy (a single wrong citation or missed seminal paper can permanently lose a researcher), feature depth (Consensus's reference-management workflow), and UX simplicity — while acknowledging switching costs are low and loyalty must be re-earned daily. On price discrimination, Eric was measured: some gap between per-token API and subsidized consumer pricing is fine and navigable, since the app layer's job is to allocate tokens far more efficiently than an end user in a chat interface; the real risk is the gap becoming gigantic, which is why application-layer founders should root for open source as a competitive check. After Eric signed off, the hosts extended the discussion into platform power — Prakash's airline-style price-discrimination analogy for GPU fleets, the underappreciated threat of Anthropic's Mike Krieger-led product organization, the newly announced Mirandil from ex-Anthropic researchers, and whether labs might acquire scientific publishers or systems like Epic (estimated $30–60B, controlling an estimated 70–80% of US electronic health records) to entrench data moats.

Interview: Tricia Martinez — Sovereign, In-Country AI Infrastructure

Tricia Martinez, founder and CEO of Dapple, pitched a 'third path' between renting shared public-cloud capacity and building private data centers from scratch: dedicated, single-tenant, in-country AI clouds — isolated physical hardware governed by local law, operated through a modern cloud-software layer she calls the 'Enterprise OS Cloud.' Asked who is so eager to spend that the company claims over $100 million in signed contracts within five months, she argued enterprise and government demand has been bottled up for years because all three existing options fall short — public clouds lack capacity, NeoCloud bare-metal forces enterprises to build their own operational stack, and on-premises buildouts are too costly and complex. (Her contract, utilization, and deployment figures are company claims, not independently verified.) Pressed on Dapple's Azure-native marketing, she confirmed that unifying public and private cloud into a single frictionless experience is the core IP but declined to name strategic cloud partners, promising announcements within months. On the 91–94% utilization figure on Dapple's site, she said deployments are backed by long-term contracts (three, five, or seven years), making the model about committed capacity and customer outcomes rather than spot utilization, sustained by three workload categories: large-scale training for AI-native companies, production-scale inference, and AI agents embedded in operations.

On the 6-to-9-month deployment timeline — far faster than record data-center builds — Tricia described a network-orchestration model: Dapple partners with existing data-center operators and infrastructure providers, then contributes capital, GPU infrastructure, enterprise demand, and its software layer, financing infrastructure directly in some deals and deploying on partner facilities in others. Asked about cascading fragility from long-term commitments paired with short-term revenue, she said the market has grown more disciplined, with financing groups increasingly requiring demand-backed contracts, and argued Dapple is more de-risked because enterprise customers make large upfront down payments and commit to multi-year relationships; she also flagged growing compute-export-restriction risk for competitors relying on certain geopolitical off-takers. She agreed Dapple is more asset-light than the typical NeoCloud, framed defensibility as three reinforcing layers (enterprise relationships and switching costs, an operating platform that gets smarter with each deployment, and a hard-to-recreate ecosystem), and drew an AWS analogy — abstracting on-premises complexity for the cloud era, now repeated for the AI-infrastructure era. Drawing on her Africa background, she described the continent as falling behind on AI production capacity, with cost the primary barrier, and called Dapple an operating platform rather than a market maker or broker, with its software sitting across every deployment regardless of who owns the underlying data center.

Close: Market Maker or Control Plane, and What Software Is Worth When Agents Are the Buyers

After Tricia signed off, the hosts debated whether Dapple is fundamentally a GPU market maker or something with defensible technical depth. Prakash made the case that deploying workloads on bare-metal NeoCloud hardware is genuinely artisanal — each data center has idiosyncratic GPU behavior from wiring, power stability, and failure rates — so a portable control plane that behaves consistently across facilities, plus a shared infra team regulated-industry CTOs can hand liability off to, amounts to real lock-in; he noted the chip generation cycle (H100 → B200 → B300) continuously resets the stability problem. Nathan remained skeptical Dapple could solve those problems faster than the NeoClouds themselves, citing Snowflake ($75B) and Databricks ($100B+) as evidence of how much the enterprise pays for friction-reducing layers while admitting he may underestimate switching costs and the value of liability handoff. He offered a counter-story for the app layer: if AI is moving fast enough to hit the core of every business, no one has bandwidth to rationalize middleware spend, making marginal savings irrelevant.

Prakash sharpened the deeper question — what is human-interface software worth when agents are the primary consumers, indifferent to UX polish and driving near-perfect price discovery — and sketched a future where 90–99% of software is bought and judged by agents. Nathan agreed app-layer margins look structurally challenged as lower search and switching costs push toward perfect-market conditions, while Prakash extended it into a hyper-deflation scenario where software production costs approach zero and revenue concentrates in zero-sum advertising auctions. On advertising, Nathan argued human time and attention stays scarce and valuable, even as 'advertising to agents' (subsidized inference, paid product consideration) could open new models. Prakash's extended aside reframed AI vacation-planning startups as solving the wrong problem — people mostly want to daydream better, not book more efficiently — which Nathan tied to welfare-economics research that much of a vacation's value lives in anticipation. A quick market check (hyperscalers in the red, semis up, Apple down 4.5%, NVIDIA below $200) preceded a preview of the next day: Cameron Berg on AI welfare research and Forum AI's Robbie Goldfarb on bringing human taste into AI systems.