GPT-4-Class Models Will Be Cheap. The Frontier Will Still Need a Patron.

Elad Gil opens a host-only No Priors by listing the model dump of the past week or two: Claude, Grok, Databricks, a pile of launches. Sarah Guo, founder of Conviction and a former Greylock partner, says most people’s priors from a year ago are now wrong. It is very likely, she argues, that the year ends with a handful of GPT-4-level models, and that some of those are open source. Mistral first, then Databricks with DBRX, changed the view of what you can do with a relatively small amount of compute — tens of millions of dollars, not a national-lab budget.

The Databricks team went further and named the curve Mosaic’s Law, after the MosaicML group they acquired: a model of a given capability should require a quarter of the dollar capital every year, from hardware and algorithmic gains stacked together. Guo does not know whether that is grounded in a particular technical belief, but she thinks the landscape has already shifted. A year ago people expected the model layer to be monopolistic or at least oligopolistic. At GPT-4 level, it looks much more open. One notch up in capital, if compute scaling is still the dominant factor, the state-of-the-art question is unanswered.

Gil supplies the other half. For the next couple of years the most cutting-edge models still look like an oligopoly, both because of the capital required and because a strong model helps build the next one — data labeling, reinforcement learning from AI feedback, a bootstrap that pulls the leader further ahead. Under that peak sits a broader swath of models, and that, he argues, quietly moves value capture to the clouds. Whether you host Llama or Claude or some other entrant, inference runs on someone else’s GPUs. Who makes money besides the model providers is, in his view, under-discussed. For applications and enterprises, Guo’s net is still good: a GPT-4-level model already does an awful lot.

Gil asks about the Inflection–Microsoft deal. Guo’s first read is that Microsoft are true believers and that Mustafa Suleyman is a live player. The company sees a product opportunity that needs AI-aware product leadership and research leadership across Microsoft properties. Copilot has real traction in code, she says, but that is still far short of the revenue Microsoft expects from the productivity suite and from search. They are ambitious, and this is a leadership change that supports going after it.

Both companies say the OpenAI partnership continues, and Stargate, the data-center effort, is still on. From the outside it is hard not to see a hedge. Guo is not criticizing OpenAI. She is saying that if you believe this is the most important technical driver for the company, being reliant on an outside player is not a position a trillion-dollar firm likes. Gil adds that Suleyman has had more capital and compute than the vast majority of entrepreneurs and research teams, and Microsoft now has direct access to that. Next-generation spend, perhaps even this generation, needs a level of sponsorship most independent players cannot match — the opposite of Databricks’ cheaper-every-year story.

Guo turns to OpenAI’s voice-cloning announcement. ElevenLabs has real traction; other competitors are pushing features such as latency. Assume, she says, that OpenAI, DeepMind, and maybe others already have very good voice, video, image, and song models. The question is whether they will release them, and what that does to a market beyond text APIs. Gil’s read is that the hesitation is regulatory and societal. Voice deepfakes are harder for the public to catch. Technical workarounds exist: attestation when a voice is first uploaded, other forms of verification. His sense is that multiple players have had the technology, in some cases for a year or two, and held it back. Open-source versions such as Tortoise, and earlier work from the Suno team, already existed. He is surprised how little competition there has been.

Zooming out to capital, the foundation-model market is not a venture story dressed as one. VCs have put hundreds of millions into individual companies. Cloud providers, big tech, and Nvidia have put billions. The same is true in China, where pre-existing internet giants fund the stack. VCs are almost a bootstrap — tens of millions, sometimes hundreds — and real scale comes from elsewhere. Clouds have a reason to keep writing checks as long as the models drive cloud revenue. Microsoft’s last quarter, Gil notes, Azure was about $25 billion and grew roughly five percent from AI-related products, another billion to a billion and a half a quarter, five to six billion annualized and still growing. That utilization is the investment case.

He still expects more money in the market, just not as many new language-model companies from angels and VCs. Music, text-to-speech, video, image generation, physics models, biology, materials science, and robotics are where foundation-model funding should accelerate. The script will likely rhyme: venture does the first rounds, then large strategics who benefit from the domain take over. Applied Intuition, quietly doing simulation software and behavior modeling for years, is his example that interesting work is already happening off the LLM main stage.

Guo’s filter for what is both societally important and a business is blunt: look at capabilities we are still missing. She has long been interested in time series plus general knowledge and reasoning. Anomaly detection, infrastructure monitoring, security, healthcare, consumer behavior — all of them would be more valuable if models understood time series, and the problem is unsolved. Context windows and how you handle this kind of data are still open. She sees huge commercial applicability and room for architectural approaches that actually break through.

Language-model success also pulled the best machine-learning people from other fields into the top labs; some of them are now looking homeward. In robotics and biotech, where Guo has been spending time, several smart teams are posting leading results on traditional benchmarks. “Something in the water,” she says: multiple groups decide a domain can be solved with a more general foundation model plus some cleverness. Robotics is still, to most ML people, a data-collection problem. Internet video is not enough; you need embodied action and controls data. Routes diverge on simulation versus efficient real-world collection, and on software for many hardware types versus a vertical company. She is inclined to believe the domains get solved. The hard part is picking the product path.

Gil finds the market odd enough that he has started incubating again, which he would rather not do — he would rather back someone. A dozen verified, talented teams pile into one area; another obvious gap sits empty, and even clearly working products lack the fast-follows you would expect. His explanation is mimetic. Every tech wave spreads memes of what to build, and the memes are often right. There were a dozen search engines before Google, a dozen social networks before Facebook, and in mobile he personally knew a dozen photo apps that went viral and died. Camera Plus had filters before Instagram and charged $20 a download because it was an indie shop that did not want to get big. Instagram put compact format, filters, a feed, and social together and it stuck. The meme was correct; the product substantiation was not automatic. Robotics, biotech, and LLMs are harder still because the markets are huge.

Guo’s other explanation is missing cross-pollination. Putting these models on financial or accounting data looks like a straighter path to commercial value than robotics or last-generation biotech platforms, which experienced executives will tell you were never easy industries. The world is full of accounting software and professional accounting services. Accountants and controllers barely share a room with systems engineers and research scientists. Gil adds a why-now: a few breakthroughs made old problems look newly tractable, so a market of technologists all got nerd-sniped into the same interesting questions.

Treating video as just another modality to solve is, Guo thinks, a mistake. Controlling video for a commercial application is a different business from video generation and understanding inside a general multimodal model, and the persistence of that split might be two years or ten. Demand for commercially viable, shareable, high-quality video is unbounded, and it is not one form: people cut A-roll and B-roll. The jump from one-second artifact-ridden clips to Pika and Sora is extraordinary, and interfaces, controllability, length, and quality still have a long way to go. Making those assets commercially viable is a deep product problem with specific research inside it, and it deserves more investment.

HeyGen, one of her investments, grew quickly over the past year with prosumer and commercial traction on video avatars — a clone of you, or a spokesperson, speaking to camera. End users do not know how the stack works; they are creative and want full control. A recent release lets someone take existing footage of an avatar walking and gesturing and replace the words. The last generation of mobile and video apps taught a useful lesson: take some dimensions of creativity away, or lower the required quality on others, and ordinary people actually finish something usable. That is the path companies like HeyGen are on.

Gil has been writing about agent interfaces. Most earlier work copied ChatGPT chat or Copilot-style autocomplete. Cognition’s Devin demoed a different layout: four panes for the plan, the shell, the code being written, and chat. You can see what is coming, watch the code, redirect if it is going down the wrong path, and hand over API keys when asked. Teams who demoed other products to him a week or two earlier have already shipped in this direction. Nobody wants to sit and wonder whether the agent is doing the job. They want to see it, interrogate it, and put it back on the path so the work finishes faster.

His working model: today’s agent is a junior intern, eager, trying hard to please, still in need of direction. The four-pane UI is the intern’s update email. You reply that they should try something else, or ask whether they considered three other things. He thinks this is an intermediary, human-in-the-loop step. As base models and reasoning improve, some of that UI should go away, the way you stop reading status mail after an intern becomes reliable. The early forms are already exciting because they are new paradigms, not reskins of chat.

A lot of the big AI use cases, Gil notes, are surprisingly prosumer: HeyGen, ChatGPT, Perplexity, Suno. Guo says it almost has to be that way, given the pace of enterprise adoption. Prosumer apps can create direct user value without immediately fighting consumer social networks that already have locked-in network effects, and the products generate their own distribution. You still have to be smart about reaching people, but these companies are growing on the back of a product someone wants. Getting to millions of enterprise users in a year is very hard: security review, procurement, internal roadmaps, risk tolerance. Ten dollars a month that makes you more productive is a decision you can make in minutes.

The data, she says, deserves respect. Canva’s numbers are public: $1.7 billion in ARR, a billion of it prosumer. The argument for these companies is that they can grow into more professional use cases, and that rising capability is creating huge markets that did not exist before. She believes they will be sneaky-big. Enterprise is not empty — Harvey in legal AI is on the board — but the first wave’s main characters are individuals and small teams. Gil reaches for history. The 1990s internet started consumer, then B2B. Email entered personal life before the office. Mobile’s defining new companies were consumer: Uber, Instacart, WhatsApp. Maybe that is just what fundamental technology shifts do. The first dividend goes to the people who start using the thing.