Companies Like Jane Street, Meta Have Captured Most Of The Value In AI, Not App Layer: Semi Analysis’ Dylan Patel

SemiAnalysis founder Dylan Patel said on the Dwarkesh Podcast that most of the value generated by frontier AI models is captured by sophisticated users like Jane Street and Meta, not by model providers OpenAI and Anthropic. Patel cited Jane Street's exclusive contract with OpenAI for GPT-5.6 ultra fast mode and its status as one of Anthropic's biggest customers, noting the trading firm generates far more value from tokens than Anthropic's profit. He also mentioned Meta, rumored to be up to 10% of Anthropic's business, which gains efficiencies like 5% longer engagement time, making more money than Anthropic. Patel added that the app layer has generated very little value, while the model layer has shifted from negative gross margins to massive positive ones, on the path to $100 million per megawatt.

AI is creating a lot of value, but not necessarily in the places most people expect. That’s the picture SemiAnalysis founder Dylan Patel painted in a recent appearance on the Dwarkesh Podcast, where he laid out who is actually pocketing the money being generated by frontier AI models. His answer wasn’t OpenAI or Anthropic, and it wasn’t the wave of AI startups either. It was the customers using the models, particularly large, sophisticated ones who know exactly how to turn a token into a dollar. “Most of the value that these models generate does not get given to OpenAI and Anthropic,” Patel said. “Thankfully, so far, it is mostly just being given to the users.” Patel’s clearest example was Jane Street, the quantitative trading firm known for its outsized presence in options markets. “Jane Street, with their exclusive contract with OpenAI for GPT-5.6 ultra fast mode, they’re also one of Anthropic’s biggest customers, is generating way, way, way, way more value out of the tokens they’re paying for than Anthropic is generating in terms of profit,” he said. “Because they get to make money off the market.” The logic is straightforward once you sit with it. A trading firm paying for low-latency inference isn’t buying a chatbot subscription, it’s buying speed and pattern recognition that can be converted almost instantly into trading profit. Whatever premium OpenAI charges for “ultra fast mode” is trivial next to what Jane Street can extract from being marginally faster than everyone else in a market. Jane Street has, in fact, been named as a strategic investor in Anthropic’s own funding rounds, giving it a foot in the door with both labs at once. Meta was Patel’s second example, and arguably the more interesting one because it doesn’t involve trading floors or exotic contracts. “Meta, who at one point was rumored to be as much as 10% of Anthropic’s business, they’re generating way more efficiencies by optimizing their ad algorithms, getting engagement time 5% longer, all these things,” Patel said. “They’re making way more money off of using these models than Anthropic.” A 5% bump in engagement across a platform with billions of users translates into an enormous amount of ad revenue, dwarfing whatever Meta pays in API fees to access the underlying models. It’s the same dynamic as Jane Street, just running through an ad auction instead of a trading desk. Patel used this to sketch out a broader map of where value sits in the AI stack right now. “Where does the value go in AI? You’ve got the end user, which I think we all agree is generating more value than anyone else,” he said. “But then you’ve got the app layer. So far the app layer’s generated very little value.” That’s a notable claim given how much capital has poured into AI-native applications over the past two years, from coding assistants to customer service tools to note-taking apps. According to Patel, most of that layer still hasn’t figured out how to capture a meaningful cut of the value it helps create, largely because the underlying model capability is a commodity that can be swapped out, and because the users themselves, or companies like Jane Street and Meta with the infrastructure and distribution to exploit the models directly, end up scooping up the returns instead. The model layer itself, Patel argued, has gone through a dramatic reversal. “You’ve got the model layer, which up until a year ago was generating negative gross margins and is now generating massive positive gross margins, on the path to $100 million per megawatt,” he said. That’s a striking turn for an industry that was, until recently, burning cash on every query just to keep up with demand and stay competitive on price. It also lines up with what’s been happening on the ground at labs like Anthropic, which has reportedly been struggling to meet its own demand https://officechai.com/ai/anthropic-has-passed-openai-in-us-business-adoption-for-the-first-time-says-ramp-data/ even as it works through capacity constraints and rate limits across its consumer and API products. Positive gross margins at inference don’t necessarily mean profitability at the company level once training costs and compute build-outs are factored in, but it marks a meaningful shift from a period when every additional user was a drag on the balance sheet. Zoom out a year, though, and the picture looks entirely different. “If we go back a year ago, the hardware supply chain was generating all this gross margin while literally everyone else was losing money,” Patel said. “So the value capture has shifted around a lot.” That’s consistent with how the AI money has flowed over the past two years, first concentrated almost entirely around NVIDIA and the broader chip and cloud supply chain, then gradually spreading out as labs found pricing power and enterprise adoption caught up with the hype. The web of circular deals between chipmakers, cloud providers and labs, the kind Patel himself has previously described as an infinite money glitch https://officechai.com/ai/big-question-if-openai-oracle-nvidia-infinite-money-glitch-can-sustain-says-elon-musk/ , has only made tracing where the actual profit lands more complicated. What Patel’s framing ultimately suggests is that AI’s value chain hasn’t settled into anything close to a stable structure yet. Compute was the chokepoint a year ago, model providers are capturing real margin now, and the app layer is still searching for a way to hold onto its share instead of watching it flow through to whoever is best positioned to act on model output. For now, that means quant funds, ad platforms, and end users doing their own work with a chatbot are coming out ahead of the labs building the technology and the startups packaging it.