Bills Doubled Overnight. Next, the APIs Might Vanish.

Six months is an eternity in AI, and Jacob Efron opened the latest Unsupervised Learning roundup by asking Ari Morcos and Rob Toews for the single biggest shift since they last sat down after NeurIPS. Morcos, the Datology founder and a former DeepMind and Meta researcher, did not reach for a new model name. He pointed at coding agents that can finally run on long enough time horizons to be useful. Over Christmas break, he said, people went home, played with the tools, and came back with the same reaction: these things actually work.

That matters because capability has thresholds. Once agents clear one, the work becomes a lot more valuable, which is also why token spending has exploded and “token maxing” became a meme. Inside Datology, Morcos has watched engineers move from individual contributors to managers of agents, context-switching across several runs instead of grinding on one task. The shift only became possible once the agents could stay up long enough to be useful.

He immediately poured cold water on the productivity story. The tools make teams faster in some ways, but they also make it easy to dump a mountain of working code into a repo and open a massive understanding gap. Slop gets cheaper to introduce, review becomes the bottleneck, and nobody wants to live in a world where “my agent reviews your agent’s output.” The slowest step still sets the pace of the whole pipeline, so a local speedup does not automatically lift the system.

Toews went for the jugular. For years his working assumption was that closed models would push the frontier and open weights would trail by only a few months. Over the past half year he has started to doubt whether near-frontier open-weight AI remains a meaningful force at all. Meta, the Western champion of open weights, looks likely to pull back from that strategy. The Chinese labs that then drove state-of-the-art open research are sending a similar signal: Qwen, DeepSeek, and others now keep their strongest models proprietary behind an API and open-weight only smaller, weaker versions.

Morcos agreed on the economics even as he disputed a capabilities cliff. He does not think the trend line of open-weight quality has broken; if anything, the models have continued to converge. What changed is the decision to *release* them. Once a lab has spent huge sums to earn credibility, giving away hosted inference “just fully undermines your business.” He is now much more bearish on how many open models the world will see. 2025 looked like a cornucopia that would only grow; he thinks that was the peak, and that other Chinese labs will likely open a big model for press and then close up.

Asked whether an open-source model company even has a business model, Morcos was blunt: he does not think so. A Red Hat-style premium enterprise subscription does not fit a field where the upfront cost of getting near the frontier is so large. The one workaround he takes seriously is open-sourcing the weights while keeping the harness and scaffolding private, then charging for the full system through an API. Moonshot’s Kimi, he noted, is already doing a version of that and is at a couple hundred million through API and chat. That is also why Nvidia has been pushing so hard, including through Reflection, to keep Western open-model providers alive: more model vendors means more chips sold.

The cost story arrived at the same moment open weights started to look scarcer. Efron noted that for a long time almost all token demand sat at the frontier; only recently have companies looked at the invoices and asked for something cheaper, faster, and smaller. Morcos has seen that shift up close. Extreme compute environments, stronger coding models, and changing output-token efficiency from release to release have combined to shock buyers. Going from Opus 4.6 to 4.7, he said, produced a big difference in token efficiency, and “a lot of people’s bills just doubled overnight.” A year ago enterprises barely cared. Now they call him wanting to cut costs, because long-thinking models can consume a budget almost instantly.

Open models can still help on that bill, especially once you stop treating a model as a naked checkpoint. A model is now the model plus the harness and scaffolding, and a lot of the real innovation is happening in that layer. Morcos pointed to the Mythos work on vulnerability and zero-day finding: plenty of people reproduced similar results with open models and a good scaffold. That is another reason the open-the-weights, keep-the-harness pattern might work. It also pushes threatened companies toward an uncomfortable choice: share data with a frontier lab and eventually get outcompeted, or compete in a niche they actually own. Cursor and others, he said, are already making that move, because the lab will always give itself better rates than it gives the app on top.

Toews’s read on the “apps are cooked” and “SaaS apocalypse” narrative was that market stories swing until the nuance dies. Two things can be true at once. Many traditional software companies, even whole categories, are at real existential risk from OpenAI and Anthropic’s product roadmaps, so a lot of the rerating was rational. The same story was also an overreaction painted with too broad a brush. In venture, the consensus has flipped toward deep tech and hardware and away from apps. Toews still likes a lot of that hard-tech work, but he does not believe one, two, or three companies will win every important market in the world.

Efron worried that VCs are painting themselves into a corner: hide in a pet-store vertical small enough for the labs to ignore, and the category itself may not be interesting; go build data centers, and you still have to beat ten other teams with aggressive financing for the same site. Even the labs cannot do everything. Morcos was surprised OpenAI shuttered or suspended its video effort despite “effectively infinite capital” and a strong team. Video training is expensive relative to text, so compute forced a focus choice. The punchline, delivered as a joke, was that they stripped down in order to acquire TBPN.

That brought the conversation to Toews’s December Forbes prediction that Sam Altman would be out in 2026, a call the others had treated as a joke. In mid-June he said it looked a lot more likely. The vibes have shifted against OpenAI, elements of Altman’s leadership are under question, and the Elon Musk trial, even though Musk lost, damaged Altman’s reputation for trustworthiness. The succession map also changed: Fidji Simo, whom Toews had seen as the obvious groomed successor, stepped back over health concerns. The theory now making the rounds is that OpenAI acquires Sierra and makes board chair Bret Taylor CEO. Toews called that a move in the best interest of shareholders, because people trust Taylor. Efron added a milder variant: an Alphabet-like holding company, with Altman staying atop the parent and ChatGPT becoming its own product line.

The dominant vibe of the past months, Efron said, has been Anthropic’s unprecedented run. The company sucked up the industry’s oxygen, becoming the most consensus name in the field, largely at OpenAI’s expense. Fable’s release was the first sign of light cracks. Morcos argued there is always backlash to whoever is winning, but Anthropic is also making moves that will alienate people. Limiting Fable for anything to do with AI development was the flashpoint. Users were not mainly furious at a limitation. They were furious that it was silent: the model does not refuse and does not say it will not help; it just does a poor job without telling you.

Anthropic can call that safety. Morcos called the claim tenuous and said it looks like competitive positioning. People reproduced many of the same zero-days Mythos found with open models and good harnesses, so he does not see a unique safety risk. He had never seen so many previously bullish Anthropic fans “truly pissed off.” He also thinks the “everyone distills Anthropic” story is overblown. Datology curated the data for Reflection’s Trinity Large without using any closed-source APIs, only public data. Microsoft’s MAI model made a point of using no synthetic data at all, to avoid even the appearance of distillation. As long as an API exists you cannot fully stop distillation, but “the only way any open model can catch up is by stealing” reads to him like copium.

On capabilities, the two guests split. Morcos had only played with Fable for a night and did not personally see massive differences from 4.8; the early takes he heard were all over the map. Toews called it both a big deal and not: it is the latest state of the art, and in a few weeks another lab will have another one. Benchmarks, imperfect as they are, still look like a meaningful step relative to the prior frontier. That matters because it undermines last year’s story that pre-training had hit a wall and only RL and test-time compute could carry things forward. Morcos agreed pre-training did not hit a massive wall; the field was waiting for new chips. Naive scaling of the same recipe, the kind of bet behind 4.5 or Llama 3 405B, did not work well, but people overgeneralized. In deep learning, he said, getting 95 percent of the details right often rectifies to “just not working,” which makes negative results almost impossible to interpret.

Efron pressed Toews on Google: talent and compute looked unbeatable in December, yet the last six months have not felt great, especially in coding. Toews pushed back on the idea that Google has fallen behind. The three labs, in his view, leapfrog one another, and any given month one of them holds the “state of the model.” He conceded Google is behind on coding, but treated that as prioritization rather than decay. Anthropic made coding its north star for years and got catapulted; OpenAI has since doubled and tripled down, and Codex is getting love. Google simply did not put the same weight on it.

The structural case, Toews said, has not changed. Google has a deeper talent bench than OpenAI or Anthropic, a massive cash machine, and a full-stack compute advantage: it designs its own chips and runs its own cloud. Nobody has infinite compute, but Google is better positioned than either rival. Morcos largely shared that view, though he was surprised Google had not improved more since 3.1 and had expected a bigger launch at I/O than just Flash. He also argued that Google may not even be playing the same consumer game. For ordinary questions, tutoring, and phone use, models will be commoditized. Most people will use whatever is on the device, and Google is already the default on Android and iOS unless Apple ships its own models. In that market, the best model does not necessarily win.

That sat awkwardly next to the coding market, which Efron still thinks of as the most meritocratic buyer. Codex looks like an amazing product and still has not made a huge dent; Claude Code remains dominant. Morcos thinks developers remain mercurial, but Claude and Codex have stayed close enough that there was no compelling reason to switch. Fable’s limits, he predicted, will push a lot of AI developers toward Codex in the near term—a “pretty nice gift” for OpenAI among the loudest voices on Twitter. Then he closed the loop back to compute. If models stay close, access becomes the differentiator. It is not hard to imagine Anthropic becoming so constrained that it cuts off the API, because Claude Code makes more money per unit of compute than the API does.

OpenAI is already selling futures that guarantee access to inference tokens, which Morcos called a huge existential threat to anyone building on top of the models. Six months ago, the idea that APIs go away was not really plausible. Now it feels plausible, “not as a business decision, but just purely because of compute constraints.” Toews called a full cutoff extreme but feasible, and sketched an even sharper version: labs might not even put their most powerful models on a public API, reserving them for internal use. How long the bottleneck stays this acute, he said, is the variable that decides whether Morcos’s sketch becomes real.

Relieving that bottleneck is not a two- or three-year story. The market structure is already strange: one company knows how to make the cutting-edge chips, and the most important machine in that process is made by one other company. Challengers are appearing. ASML’s EUV is hitting physical limits on how small a transistor it can print, so startups are raising large rounds around atom lithography—using a beam of atoms instead of light—and X-ray lithography further out on the electromagnetic spectrum. If atom lithography works, Toews said, the machine can be simpler, cheaper, smaller, and higher resolution. It still looks at least five years away. Elon Musk’s “Terafab” idea sits in the same bucket of long-shot unlocks. Toews would be surprised if anything material happens before 2030; the several big unblocks probably arrive together in the early 2030s.

A tighter crunch has second-order effects. Frontier labs with “infinite capital” have not cared much about efficiency; physical limits will force them to. Morcos has long argued you do not need trillion-parameter models to get today’s capabilities, and smaller and smaller models already match the giants of one or two years ago. Distillation and inference-cost work should attract more investment, even if usage still outruns the savings. Alternative chips—AMD, Amazon Trainium, Cerebras, D-Matrix—get a tailwind because there are not enough Nvidia GPUs to go around. They do not, however, solve the shortage. If those vendors did not exist, Nvidia would simply eat the TSMC capacity; the total number of chips stays roughly constant. The same logic, Efron noted, applies to neo-clouds: chips get shuffled among owners, and as long as capacity is tight, every holder has a solid business. The leading indicator is already visible: H100 rental prices bottomed around the previous recording in December and have risen dramatically since.

SpaceX is the strangest character in this plot. It signed huge deals to rent compute to Anthropic and Google, which eased some of Anthropic’s hunger. Toews reads that two ways: padding revenue ahead of an IPO, and an admission that frontier model research is not the company’s foremost priority. A lab that truly needed every GPU would not sell them. Elon companies, he argued, excel at operationally intense atoms-not-bits work, so the durable wedge is standing up giant clusters fast—possibly into the world’s biggest cloud, especially if more compute goes to orbit. He is not bullish that xAI cracks back into the top echelon with Google, OpenAI, and Anthropic, given the past year’s attrition. The rumored $60 billion Cursor deal is, in Morcos’s view, a bid for coding traces and a hedge on a weak coding model. Sixty billion is probably high, but the structure looks like an option: keep optionality until they learn whether traces are what make a great coding model. SpaceX’s S1 TAM chart put enterprise AI around $20 trillion and all of space at five or six hundred billion, for a $28 trillion total. Before Cursor, Toews noted, they had almost no application surface area at all.

Efron noted that they had somehow gone most of an episode without Andrej Karpathy, the show’s usual unifying theme: the man who complains that everything is slop, then joins a lab’s recursive self-improvement team. Both labs have been vocal about automated R&D by 2028. Morcos said his estimate has moved. He is much closer than he was six months ago, and more bullish on the direction, but the bottleneck is still compute. Ideas are not the scarce input, and neither is execution; you still have to run the experiments. Datology has been letting agents do data curation with various amounts of guidance and has seen far more promising results than he expected. Models can already improve models. What he still does not buy is the runaway story in which one player goes super-duper fast and nobody can compete. At least ten companies have the funding, talent, and know-how, and an army of junior-researcher-level systems only gets you so far. Anthropic slowing junior hiring is a confidence signal, he said, but the whole thing will be a lot slower than people claim.

In quickfire, Toews’s disagreement with the current discourse was about energy. In five to ten years, he thinks, we will look back and find today’s AI laughably resource-inefficient: two-gigawatt-scale data centers, about twice the power of San Francisco, versus a human brain that runs the intelligence we are trying to build on 20 watts. There will be no single silver bullet, but hardware bets such as Naveen Rao’s analog-computing company, plus algorithmic and optimization gains, should crush energy use. If that happens, the leading labs’ moat—capital and scale—gets less compelling, because ideas diffuse. OpenAI did not invent test-time compute; o1 just shipped first. Jerry Tworek leaving OpenAI to chase a more efficient paradigm is, in that light, a preview. Morcos expects any real breakthrough to show up as multiple independent discoveries. The space of ideas is not that large, and when an idea is ready, several groups hit it at once.

Morcos’s own disagreement is sociological. He does not buy the “permanent underclass” story, or the claim that AI takes all human jobs in a decade, even as Anthropic says versions of it very visibly. Humans are slow at dissipating tools through the economy, and much of business still runs on human-to-human interaction and trust—barriers he thinks technocrats underweight. He admitted the caveat: maybe he just wants humans to matter for longer. On what else has changed, Toews has pulled in his robotics timelines hard. Six months ago he would have put a GPT-3 moment for general-purpose robot models anywhere from 18 months to five years. Now he thinks robotic foundation models have crossed into commercial viability across many use cases, and once they are deployed the data flywheel will accelerate.

Asked for a spicy bet for the back half of the year, Morcos took the API call. He feels more confident saying it happens by the end of 2027 than in 2026, but he thinks there is a very reasonable chance Anthropic—or OpenAI—suspends or heavily limits API access for a stretch, as a precursor to it happening more often. Toews, declining to top his Altman ouster, predicted that by year-end it will be obvious Anthropic’s next chapter after coding is life sciences and biology. Rumors already have the company standing up wet labs; over a longer horizon he thinks it becomes one of the most important life sciences companies in the world. He revised that, after Efron needled him about Demis Hassabis, to Anthropic *and* Isomorphic Labs. Dario Amodei is a neuroscience PhD who has been passionate about biology for years, but Morcos warned that domain expertise matters more than AI people think. Hassabis paired biologists with AI researchers on AlphaFold and Isomorphic; labs that just throw AI researchers at biology have stumbled. Right now, Morcos would bet more on Isomorphic.