API prices fell 200x. The LLM table still is not set.

In September 2024, Sarah Guo and Elad Gil skipped the usual guest and used a No Priors episode to talk through what everyone else was already arguing about in private. Guo, founder of Conviction, and Gil, the well-known operator-investor, teed up LLM consolidation, the chip layer, how far an AI company should push legal and content risk, and a few large transactions. The first question was the sharp one: are we done here? Is it too late? Has the language-model market already consolidated?

The tension they spent the next twenty-five minutes unpacking is easy to state and hard to resolve. API prices have crashed on the order of 200x in a year and a half, which is what a maturing, commoditizing market is supposed to look like. At the same time, the next training runs are so expensive that only hyperscalers and sovereign funds can write the check, and traditional venture is simply too small. Falling prices and rising capital intensity are happening together. Whether that means the table is set depends on which host you ask.

Gil opened with the shuffle already underway. Parts of Inflection, parts of Character.AI, and parts of Adept — the last of those into AWS — had seen core teams join larger enterprises. Many of those companies still exist, and some products are still running, but the set of independent frontier-model players had thinned. Alongside the talent moves, he saw enormous capital moats forming around the biggest foundation-model runs. These companies are raising billions, often tens of billions, and the only writers of those checks are hyperscalers such as Amazon and Microsoft, or sovereigns.

That scale, he argued, is why everyone is partnering up. The question for every other player is where the next round comes from, and who they sit next to. He offered a homework assignment: list the large firms that still lack a model partner — Apple, Samsung, and the rest — then draw lines to the remaining model companies and watch the industry’s outline appear. Hyperscalers have their own reason to fund this, because more model usage often becomes more cloud utilization. Incentives among VCs, clouds, other strategics, and sovereigns keep flipping. Without a breakthrough in architecture, training cost, inference, or post-training, he found it increasingly hard to see most companies staying competitive. It felt, to him, like a consolidation stage.

Guo did not dispute the mergers. She disputed the mood. Over the last year and a half, she argued, the market had become *more* competitive, not less — at least among the players who already have a capital moat and some breakaway scale. From the consumption side the pattern is continual, aggressive performance gains, benchmark races, price cuts, and real open-source opponents. Consolidation and “nobody is making money yet” can be true at the same time. That is the uncomfortable equilibrium the rest of the episode tries to name.

The 200x number came from David, a colleague on Guo’s team, who charted API pricing across models. Dollars per million tokens had fallen dramatically over 18 to 24 months. Part of that is distillation — large models teaching small ones — and part is newer GPU generations knocking down inference cost. Usage, and therefore the pool of revenue, is still rising. What is getting harder is competing as a model API business. When price falls faster than volume grows, you are pushed toward specialization: bespoke models, targeted post-training, vertical applications.

Gil’s blunt read of the mega-rounds was that the mostly unsaid business plan is still AGI itself. Models will show emergent behaviors, figure out how to make money, or become so obviously valuable that no one has to explain the bill. He did not call that logic wrong. He did say that on a two-to-three-year horizon the visible businesses are consumer — apps, subscriptions, or advertising, and nobody has really gone the advertising route in anger yet — and enterprise. Both are real today. The fight, Guo added, has already moved from research into product. Anthropic hired Instagram co-founder Mike Krieger as chief product officer; OpenAI hired Kevin Weil into the same seat. Labs are stacking a product moat on top of the capital and research moat, trying to make more than chat as an interface.

Lock-in is the other half of that product fight. Guo reached for the old AWS playbook: do not let customers think this is just a storage bucket. Prompt caching, JSON output interfaces, and fine-tuning get layered on until switching feels expensive. Anthropic’s prompt cache was, in her view, a clever move on both cost and latency. Once users live in those features, the model market is much less of a commodity. She is personally excited, as a consumer, for smarter chat with automatic context on more surfaces. Chat does not go away; the interface around it gets denser.

Is there still a path for a latecomer? Guo listed a few. Differentiated reasoning approaches, including groups betting on math and code for self-play, can still matter, and the capital required to reach competitive scale falls when you are repeating work others have already done, with hardware progress thrown in for free. Distillation and small-model fine-tuning remain an open question. Gil narrowed the frame: this whole debate is about language models. Physics, biology, materials science, and image generation each have their own markets, sometimes multimodal, often not. Around the core LLM sit two other axes that evolve almost orthogonally — reasoning models and agentic flows, and the infrastructure around the model: caching, long context, RAG. Evolution will happen on all three. The live question is only the core LLM layer.

Will the core LLM market really shrink to a handful of names? Gil reached for social networking. Everyone thought Friendster would win, then MySpace. When Facebook emerged, the consensus flipped to “this is a commodity, there will never be long-term differentiation.” Facebook took the core social graph, and then came Instagram, Twitter, Snapchat, and eventually ByteDance and TikTok. LinkedIn took professional identity; Twitter took news and real-time information. People called the end of social, and another giant arrived anyway. He expects the same pattern here: certain players grab parts of the market, and the map keeps getting redrawn.

Guo’s guess for video and audio is a split: general-purpose models for some jobs, specialized ones for others. Two variables decide it. How far do the capabilities themselves generalize, and does the tooling around the model need to be vertically integrated with it? A strong image-gen product might serve art, graphic design, and UI design — is that one model plus fine-tunes, or several specialists? She used the brain as a reference. It is a set of reasonably specialized modules for vision and emotion. The psychological literature has the construction-site case of a steel beam through the skull: one narrow kind of emotional or reasoning function disappears, and everything else is fine. How specialized versus how general these models become is still an open question, for AGI as much as for image gen.

The image and video market, she said, had heated up in the last couple of months. The default independent winner had looked like Midjourney, with OpenAI’s DALL-E and Gemini-side work as the multimodal backdrop. Now a growing set of companies is shipping high-fidelity models. She roasted her own prior: she keeps expecting video, image, and audio models to get rapidly better and then commoditize, and she is still surprised by the pace. When Sora landed, the question was who could possibly catch it. Runway and Pika are now close; mid-stage names such as Ideogram and Hotshot, plus what Luma Labs has been shipping, all look competitive. Five-person teams with not much capital are producing models that, five or six months earlier, the market narrative had ruled out. That, she said, was a real update.

Gil pointed to a second trend almost nobody had underwritten a year or two ago: models are shrinking relative to performance. Distillation is part of it, other tricks are part of it, but across the board more capability is coming off smaller and smaller weights. People knew there was compression headroom; they did not know it went this far. The brain again: a 20-to-30-watt device in your head, general-purpose, decent at image identification, still with room to go, in a compact space. Guo connected size to latency. Smaller, faster models plus new architectures such as Cartesia’s point toward real-time image, video, and audio. Almost none of that is in production at scale today. Generating images as you speak is a different product from waiting for a still to finish rendering, and she expected those experiences within a few months.

The second innovation axis, after chips, was how far a company should push on training data and on what the model is allowed to say. Gil went back to early Google. The company indexed the web, which from some contemporaries’ point of view was scraping other people’s writing and then making money on it. The fight often centered on the snippet — the short blur of text above the link. Opponents claimed those snippets were not fair use. News organizations and copyright holders sued. The arc took maybe a decade or fifteen years and settled into four outcomes: robots.txt, so a site could opt out of crawling; a finding that snippets fell under fair use; licensing deals for specialized feeds that filled Google’s OneBoxes; and, most ironically, news sites that pulled out of the index, watched traffic collapse, and asked to be crawled again. The penalty of being absent was larger than the penalty of being in. Google had a sharp copyright team, pushed the envelope thoughtfully, and came through reasonably unscathed.

Guo added Airbnb and Uber. Both ran straight at rules many people treated as regulatory capture — restrictions on short-term leases, taxi medallions. If those companies had waited for the rules to change, the services consumers now love would not exist. The method was to bet that demand was real, get to scale, learn the risk profile on the way, and only then negotiate policy from a position of usage. A lot of AI companies, she said, will have to wrangle the same sequence. The version of the question she hears constantly is whether Google will come after you for scraping YouTube to train video models. She would be shocked if there were a way to get to video-data scale that does not involve some YouTube. She also asked, given Google’s own stance on scraping and on SERP businesses, whether a hard line on YouTube data is legally hypocritical relative to Search.

Gil flipped the coin to Napster, which the music industry sued basically into the grave. He split risk into three kinds that are easy to mash together. First is legal and lawsuit risk, which are not always the same thing. Second is regulatory risk, where the rules are unclear — crypto is the classic, and AI has its own gray zones. Third is reputation: what outputs you are willing to allow. Grok was his live case. xAI has said, explosively, that it will not police outputs the way other labs do, for images and for text. That is closer to ordinary human speech than the tighter filters many companies have shipped, some of them trying to do right by users, some of them feeling politically driven. How much society actually cares that model output matches existing speech and creative norms is an experiment in progress. Many labs have curtailed more than the median of society would.

Guo added a philosophical cut: restrict generation, or restrict distribution? If you own a platform, controlling certain kinds of distribution is a responsibility. Generation sits closer to free speech. There is no off-the-shelf answer, and they did not pretend to have one.

Guo turned to semiconductors because training and inference both sit on chip performance. A new wave of chip and systems startups has appeared. The last one, six or seven years ago, was Groq, Cerebras, and a few others. This one includes MatX and Etched, some of whom would be on the podcast soon. Gil admired the earlier cohort’s foresight: they saw that the AI workload would need a different kind of computation. The hard part was placing a chip-and-systems bet that far into the future. Seven years ago it was not obvious that scaled Transformers would eat so much of the workload. The market evolved in unpredictable ways. The new cluster of companies is tightly focused on Transformer architectures and on area allocated to matrix math.

Whether they can beat AMD and Nvidia on performance-per-dollar is the unavoidable question. Both incumbents have been high-speed innovators, and AMD is making progress. Chip investing, as Gil boiled it down, is three things: the architectural bet you are willing to live with for years, the pace of delivery, and price-performance. Both hosts thought the bet was still worth making. Demand has also changed shape. Sovereign clouds — countries building their own AI infrastructure — are a new order book that a startup can actually sell into.

AMD’s acquisition of ZT Systems was the live puzzle. The market is divided on whether AMD can become competitive with Nvidia, so Gil laid out the missing pieces. On software, against CUDA, AMD had already bought Silo AI — a roughly $600 million acquihire of hundreds of AI engineers and researchers who had done a lot of work on AMD. On networking, AMD is in UALink, the open-source camp’s answer to NVLink. ZT, in his arithmetic, is a $1-to-$2 billion acquihire of about a thousand systems-design people, moving the business from individual chips and components to rack- and data-center-scale AI. That is what Nvidia is really selling now: full systems, data centers for research labs, via a multi-year strategy. The components of a reply are on the table. Whether AMD can assemble them is a different question.

They ran out of time with the big questions still open: whether a latecomer can still reach frontier language models, how far small teams can push image and video, and whether AMD can turn a pile of acquihires into a data center. Guo was already looking forward to talking with Etched, MatX, Cerebras, and the rest of the next chip wave on upcoming episodes. The market call they started with — is it too late? — had not been answered with a yes.

Gil closed with a public-service announcement that was only half a joke. He wants to buy a humanoid robot, or a Spot from Boston Dynamics, or another machine in that class. Suggestions, advice, or a unit for sale: ping him. The LLM table is not set. In the meantime, he is shopping for a robot.