$100 Million Researchers and a $200 Billion xAI: An AI Roundtable Does the Math

Unsupervised Learning, the Redpoint Ventures show hosted by Jacob Effron, is trying a news roundtable on top of its usual one-on-ones. The first guests are two people who already argue off-mic. Ari Morcos is CEO and co-founder of Datology AI, a data-curation company; before that he spent about a decade as an ML researcher, including two years at DeepMind and five at FAIR, Meta’s AI lab. His PhD was in neuroscience, in a monkey lab — a detail that will matter once the conversation reaches reinforcement learning.

Rob Toews is a partner at Radical Ventures and runs the firm’s Bay Area office. Radical is an AI-only firm with close ties to researchers such as Geoffrey Hinton and Fei-Fei Li. Toews has led checks into Datology, Writer, and Hebbia; before venture he spent years in autonomous vehicles, and he writes a Forbes column. Effron wants them on the week’s loudest topics: whether model progress is slowing, what RL environments are worth, how to think about lab valuations, and a talent war that now prices individual researchers in nine figures.

Effron tees up the split on Twitter: one camp says GPT-5 is GPT-4.5 with a router, and the era of large general jumps is over; the other points at gold medals in math and coding. Morcos does not think progress is slowing in any significant way. It is harder to measure. For most consumer tasks the models are already good enough, so the average ChatGPT user may not feel the next increment. The gains have moved onto other axes: reasoning, cost, and smaller models matching last year’s giants — which matters a great deal if you actually want to deploy them.

Pre-training hitting diminishing returns is, in his view, the boring fulfillment of a paper. Jared Kaplan’s scaling laws already said that each 10x of data or compute buys a smaller step. GPT-4.5, Llama 4, and Meta’s Behemoth just showed the curve doing what it said it would do. Other axes of scaling still work. The timbre of progress changed. That is not the same as a stall.

Toews takes the other side on purpose. No exponential lasts; every exponential in nature becomes an S-curve, and that is what this era of language models looks like to him. GPT-2 to GPT-3, then GPT-3 to GPT-4, were astounding. The plateau is real, and last year’s story — that post-training compute, test-time compute, and RL would restart the 2020-to-2024 march — is not enough. Reasoning models are exciting, especially in coding. He does not think they restore the old slope. RL shines where answers are verifiable. Most of the world is not math homework. Getting to the next level of intelligence, he argues, will take more fundamental breakthroughs than the public can currently see.

Morcos grants the S-curve and then relocates it. Pre-training of ever-larger, billion-dollar runs has plateaued; the RL gains exist *because* the base models got good enough to allow them. For the way your mom uses ChatGPT, the missing piece is reliability, not a new sense of wonder — GPT-5 is probably a lot cheaper to deploy than GPT-4o, and as a user you may not care. The live questions are whether RL works outside easy-to-verify domains, and whether it generalizes or merely hill-climbs one dataset at a time and then forgets the last one.

Morcos spent much of his research career on out-of-distribution generalization. The modern self-supervised trick, he says, is to train on so much of the web that almost nothing is out of distribution. RL is different. An environment is specific, full of reward hacks, and models are extremely good at finding any gap between the simulation and the world. In the monkey lab, animals learned eye-movement tasks for juice. If there was a way to cheat, they found it. One animal’s pattern of answers made no sense if it had learned the task. Correct trials turned out to be about half a percent brighter. That is what the monkey learned.

He uses the story as a warning label. If the environment is not perfect, the model will overfit it and never generalize. The unsolved problem is building environments accurate enough that the only way to win is to learn the general principle. Code is one of the few domains where that is plausible. Most of the real world is not. That has not stopped a summer of dozens of RL-environment startups, which Effron sorts into two customer stories: sell to the handful of labs doing this at scale, or help enterprises RL their own models because the tasks will not transfer.

Toews has no doubt about the near-term cash. OpenAI, xAI, and the other giant labs have set aside huge budgets, are building internally, and will still buy whatever is good. There is, he says, a tidal wave of cash about to hit anyone writing these environments, so top line will scale fast. The venture question is durability. Environment generation will get more automated; lab needs will shift; five years from now he does not expect massive budgets for this exact product. Act two and act three matter, which is why the Scale AI analogy is both tempting and a founder bet. Scale drew bounding boxes for autonomous vehicles, then sold hard into the U.S. government, then rode RLHF. That path required Alexandr Wang seeing around corners. You are underwriting the weave, not the widget.

Morcos is colder on the lab-facing version. Getting better at building the hundredth environment may transfer in a way that expert human data collection does not, which is an argument for bringing the work in-house — and many labs are already trying to automate it. Scale and Surge built *bespoke* datasets for each customer. RL-environment companies more often build one environment and sell it to everybody, which erases differentiation. If environments are truly the thing that separates frontier labs, those labs have a strong incentive to buy them out or build ones nobody else can use. He expects a lot of this to be internalized, and he would not count on the boom lasting.

The second business — helping companies apply RL on their own data — is the one both guests actually like. Consumer and enterprise needs, Morcos insists, are not the same product with a different sales deck. A consumer wants a model a mile wide and an inch deep: an answer about anything, good enough, not perfect. An enterprise wants the opposite, an inch wide and a mile deep: a narrow set of tasks at five-nines reliability and the lowest possible cost. The barriers to building that, infrastructure and expertise, are eroding fast. Training infrastructure got easier after MosaicML, then SageMaker, Together, and a pile of other platforms. Data is still hard. That is Datology’s mission.

The folklore that customization costs hundreds of millions is, he says, a myth that many people are financially incentivized to spread. There is now plenty of proof you can build an excellent custom model for on the order of single-digit millions of dollars, and that number is falling. Once that is true, custom RL on those models becomes attractive. Proprietary data that exactly matches the task is valuable in a way the same data is not if you are merely feeding a general web model. For the first time, a company can turn its own corpus into capability instead of renting it to someone else’s frontier run.

The week’s consumer headline is Gemini at the top of the App Store for the first time, largely on the back of Nano Banana. Morcos, who was at DeepMind from 2016 to 2018, has been bullish on Google for a long time. The talent density was never the issue. The issue was launching products and wanting to win. That, he thinks, changed on the Gemini team over the last year. Sergey Brin is apparently spending a lot of time with them, which changes the energy. Pair that with Google’s distribution and you get a behemoth.

Toews is on the same page, and more specific. After ChatGPT, the story was that Google got complacent and OpenAI out-executed them; there was truth in it. The structural advantages — resources, compute, bench of talent — did not go away. Rumors say Gemini 3 drops in the next couple of months. His expectation is that it will be state of the art and better than GPT-5. If he had to pick one major lab to keep advancing the frontier, he would pick Google. The financial footnote is less cheerful. Winning puts Google on a platform that is harder to monetize and more expensive to run, a squeeze on revenue and profit until it figures out ads, costs, and its in-house chips. Then, he guesses, the line goes up again.

That is a real threat to the last two funding stories. OpenAI’s round is predominantly a ChatGPT story, and Google can compete for consumers. Anthropic’s is predominantly enterprise API, which is really coding; if Gemini 3 is strong at code, developers switch. Morcos’s own staff, Effron needles, already do: every new model, they move their coding workflows, with zero loyalty. ChatGPT has brand escape velocity. Defaults still matter. When Apple eventually ships a good on-device model, Android’s default will be Gemini and Apple’s default will be whatever Cupertino picks — maybe Gemini, if regulators allow a sequel to the search deal.

Would they buy OpenAI at $500 billion and Anthropic at $180 billion? Effron’s own frame is growth rates and diffusion: OpenAI has about 700 million users, so the question is monetization — ads, commerce, completing end-to-end actions in a way search never did. Anthropic is how much runway is left in coding and the enterprise API. Even if model progress stopped tomorrow, he believes there would be decades of work disseminating what already exists. Morcos would probably buy OpenAI; the upside is smaller at these numbers, the revenue multiple is not crazy, and the real doubts are unit economics and whether non-developers will pay. Everyone he knows in AI pays for ChatGPT. Everyone he knows who is not an AI developer uses it and does not. Monetizing an answer engine is a different problem than a sponsored link above ten blue links. Bias the single answer for a commercial reason and you undermine trust in the answer itself. Search had twenty years to negotiate that truce. Answer engines are at ground zero.

The same day, xAI announced a raise at $200 billion. Toews’s explanation for the number is mostly a mantra: never bet against Elon, a thesis that would have served investors better than almost any other over the last fifteen to twenty years. Set Musk aside, and $200 billion makes less sense to him than either Anthropic’s or OpenAI’s price. xAI has much less of a business, is still playing catch-up — very quickly, to its credit — and is not doing anything fundamentally different in technology or product. Someone’s recent analysis, which he found interesting, is that xAI is Musk’s first me-too company: not starting an EV startup in 2003, not rockets, not Neuralink, just another lab in the mold of the others. Top researchers have been leaving. Culture concerns are the usual Musk-company soundtrack. He will not say anything definitive. He also does not think xAI is as strongly positioned as the other labs.

Morcos says almost exactly the same thing. The numbers do not justify it; the company is valued above Anthropic. It is also Elon, and betting against Elon has been a poor decision for one’s pocketbook. That is the entire analysis, and they let it sit.

Toews entered AI at NeurIPS in 2015, when the conference was still called NIPS and nobody was handing out nine-figure packages. Every year since, researchers have told him this must be the top. In 2017, Intel threw a NeurIPS party in Long Beach with Flo Rida — the *Silicon Valley* gag made flesh — and the room decided the peak had arrived. Friends came out of PhDs with $500,000 offers, which felt insane; some went back to academia rather than ride something that could not last. They missed what came next. The economic-value story, he says, is not a replay of the dot-com bubble. Demand for talent will keep growing faster than supply. Averages will keep rising. The $100 million package may be a moment in time rather than the new mean. Whether those extreme packages continue depends in part on Meta’s superintelligence effort, MSL: if Meta ships a really great model, you will see more of them; if culture blocks that, you will see fewer.

Morcos does the arithmetic that makes the sticker less crazy. What you care about, at a frontier lab or an enterprise, is dollars in and performance out — task quality, inference cost, some mix. Anything that raises that ratio is a compute multiplier. Datology’s core bet is that high-quality data is one: they have seen 10x performance per dollar, which means a much smaller compute budget. Talent is another. If you know how to make the GPUs sing, the increment can easily exceed the cost of the person. In a world of gigantic CapEx, underpricing the human multiplier was the old mistake.

The risk is who the number attracts. Do you get people aligned with the mission, or people whose main project is not getting fired, or someone who will take $10 million in a few months and leave? He has watched the second-order damage at Meta: researchers from his FAIR years who were, in his view, equally worth the new packages, did not get them, were already well paid, and left bitter anyway. If MSL fails, he thinks it will be those cultural effects, not the idea that talent is not a multiplier. He would also not bet against Mark Zuckerberg. One of his more controversial takes is that metaverse spending will look pretty good in the long run. FAIR sat under Reality Labs for about two of his years there, which meant it counted as metaverse spend; he thinks it is plausible Llama’s budget, traced carefully, came from that line. Zuck’s strategic moves, he notes, tend to be lambasted in the moment and called genius later.

The same talent gravity is distorting startup prices. After the Character.AI deal, the floor under a famous team got very high. Thinking Machines priced at $10 billion on day one; SSI went from $5 billion to over $30 billion. Investors treat that as a free call option: big tech has already tried to pay many billions, been rebuffed, and will try again. Downside is covered. Upside is uncapped. That logic is how you get those numbers before there is a product.

Character.AI, Adept, Inflection, and most recently Scale AI and Meta share a structure: an effective acquisition of talent that is not, legally, M&A, twisted so as to duck FTC and Justice Department scrutiny. The new fact, Toews says, is that nobody can predict how the dollars flow. An ordinary acquisition pays the cap table, preferences and all. Because these transactions are not literal mergers, the parties can redefine who gets paid. In some deals investors get their money back; in some they get less; in some they get a better multiple. Founders almost always get paid lavishly. In most deals, employees do not get paid well, if at all. He calls that new in the history of startup finance, harmful, and a transparent loophole. His hope is that it is a blip, and that antitrust authorities adjust.

Morcos is harsher. Founders make a social contract with early employees: you bet your most valuable commodity, time and energy, on this vision, and you share in the upside. Acquire-not-acquire deals break that contract. Over time, people will be less likely to join startups, more likely to take safer jobs, and the ecosystem will starve. Regulation would help. Founders still have a choice at the table. You can make pro-rata payouts a condition. Some do. Some do not. A shell company left behind with worthless equity is, in his view, a choice someone made.

Speed is the part regulators have not solved. Buy an RL-environment company today, wait a year in review, and the asset may no longer matter; the whole point of the structure is that the person can walk in on Monday. Would regulators allow Thinking Machines, SSI, or Mistral to be bought? Toews went to law school and would rather people forgot. By conventional analysis, companies with no revenue and no real market share should be acquirable without changing consumer prices. A creative, proactive regulator in the Lina Khan mold is another story. If she were still running the FTC, he is sure she would come down hard on Google trying to buy SSI. A more laissez-faire shop, by the book, should not have a problem.

On BG2, Sherwin Wu, who runs a lot of engineering for OpenAI’s platform, said he was short all AI infrastructure because the scaffolding changes too fast. Morcos will not sign the blanket. A company like Fireworks, selling fast inference, does not suddenly become pointless; inference is required no matter which model wins. Wrappers that put a lot of scaffolding around today’s GPT-5 so it can do a general-purpose task are exactly the things the next model overwrites. His hotter take is that data labeling is less safe than it looks. You do not need infinite human data. Synthetic data is powerful. The human data you still need gets more narrowly specified, which makes the market smaller. Whether that is two years or five, he thinks it happens.

The durable layer, in his view, is the boring plumbing: fast inference, easy training, the ability to scale, and — soon — infrastructure for constantly updating models. Ranking models at big tech already refresh every couple of hours or they lose revenue; that is where the rest of the industry is going. Toews agrees it is hard to build a durable infra business, and a lot of what is being built will be obsolete. He disagrees that the conclusion is to be bearish on the whole category. Every major technology shift needs a new suite of tools. The job is to find the services whose *need* stays the same while the implementation changes. Inference is a good bet. Human labeling, until recently, was one too.

Datology’s version of that stake in the ground is almost tautological on purpose. No matter what you do with AI, the quality of the data you show the model is critical. Morcos’s claimed unfair advantage is being the best in the world at valuing data for a given downstream use case: you want a model to do X, here is the data, by filtering or by synthesis. The form may shift toward RL in two years. The core stays. Without that, an infra company is playing whack-a-mole with whatever is fashionable this quarter. The only truism he trusts is that what we do today is not what we will do in six months, so the business has to be adaptable around something that remains true.

Will Meta’s superintelligence lab work? Morcos thinks it might, and repeats the Elon rule for Zuckerberg. Hardware is the other half of his optimism: the week’s AR-glasses demos looked like the metaverse thesis in a form factor people might actually wear. Toews has been deep-diving brain-computer interfaces for months and thinks the field is near an inflection, invasive and non-invasive, with different customers. Invasive BCI still mostly means cutting open a skull; Neuralink is the famous case, not the only one. Non-invasive sensors can live in headphones, caps, even on a wrist. Better sensors help. Better AI, extracting signal from the electrical and magnetic noise of a brain, helps more. He expects compelling non-invasive products in the next twelve to twenty-four months, a new way to interface with computers. Meta’s band is an EMG cousin of that idea. Glasses still make more sense to Morcos than the pocket-camera gadgets of *Her* or Humane; Google Glass was early, not wrong. Integrating BCI into glasses over time is the stack he wants.

Asked what they changed their minds about this year, Morcos names RL on language models. It did not work for a long time, then it did, and in hindsight the reason is that the base models got good enough. Take DeepSeek R1’s techniques and put them on models from three years ago, or even one, and they never work. Qwen reinforces well because it was trained on a lot of synthetic reasoning traces; Llama was not, and does not. Toews’s open question is when robotics gets its GPT-3 moment. Almost everything they have discussed is knowledge work in bits. The physical world is still mostly unautomated, or automated with last-generation robots. Physical Intelligence and Skild are building hardware-agnostic foundation models; Tesla, 1X, and Figure are building general-purpose machines that can do what a human can do. Progress is fast. Atoms are exponentially harder than bits. He cannot tell whether that is one or two years away, five, or more — and that timing, he thinks, determines how quickly AI transforms the rest of life.

Effron has changed his own application-layer advice. A year ago he would have said do not spend time on models, go deep on workflow. Domain RL is starting to matter for reliability and quality, and it all comes down to cost: if it is a million dollars instead of a fortune, every enterprise will do it. Toews’s under-discussed theme is recursive AI, systems that autonomously develop stronger AI, the conceptual engine in most superintelligence roadmaps since the 1960s. In the past year it has become more real; more than one high-profile stealth team is building toward it. How well someone gets that to work, he thinks, will set the slope of the next few years. Effron’s gloss is that if you can hill-climb on coding until you have really good AI researchers, questions about generalization outside code start to look secondary.

Morcos closes on the unfashionable input. Data is still the most under-invested area of AI research relative to its impact; evals, which are data in another key, are second. Everyone talks about a data wall. Synthetic data is a large part of the way through it, with real collapse risks if the snake eats its tail, and real methods for not doing that. It was called out as a major part of GPT-5, and of the Chinese models’ gains. Curation plus the right synthetic pipeline is how you get more juice from the datasets you already have, and it becomes more critical as use cases get more domain-specific. People, he thinks, are sleeping on how large that impact will be if you do it correctly.