The following piece is a collaboration between CO and Tomasz Tunguz of Theory Ventures. TT is a well-known investor and blogger, and someone I’ve interviewed several times. Enjoy! The newsletter will return to its regular form on Monday!
— Alex
What’s the value of an AI model? Whatever the figure, it’s temporary. Newly crowned AI models can expect to spend 41 days atop the global leaderboards before being dethroned, offering a limited window for efficient monetization. Frontier labs have leaned on token pricing to make the math pencil out.
Lower-cost, highly-performant open-weight models are challenging frontier lab economics. If cheap open-weight models can quickly follow frontier labs’ performance gains, the latter group’s ability to charge prices sufficient to cover their training, staff, and inference costs comes into doubt.
This dynamic is partly why market sentiment regarding the leading AI labs – OpenAI and Anthropic – has shifted from ebullient earlier in the year to something more wary as we foray deeper into the third quarter.
Have we seen this saga before? Yes. DeepSeek’s R1 model challenged OpenAI’s then-market-leading o-series reasoning models, purportedly at a far smaller training cost, if you can recall all the way back to early 2025.
Today, Z.ai’s GLM-5.2 model and Moonshot’s Kimi K3 model are once again fomenting concern that frontier lab pricing methods – Fable 5 is comically expensive to run, while OpenAI’s GPT-5.6 model family did launch with a focus on cost efficiency – are unsustainable. (Moonshot’s decision to enact a 30% tax on third-party inference providers that want to sling K3 to their customers does somewhat close the gap between open- and closed-source AI pricing.)
The chance of having their midday meal eaten by Chinese AI labs is not the only risk facing American frontier labs. Not only must they contend with a rebound in American open-weight models (Nvidia’s Nemotron 3 family, the Gemma collection from Google), they face competition across two other vectors: Local inference, and first-party hyperscaler models with ready-built distribution channels.
The local angle
Theory Ventures recently led Ollama’s $65 million Series B, bringing the local model developer tool to $88 million. Why? The MacBook you are reading this article on, it turns out, is much more powerful than you would expect. In Intelligence Per Watt: A Study of Local Intelligence Efficiency, Stanford researchers found local language models can accurately handle approximately 88.7% of real-world, single-turn chat and reasoning queries.
Cloud open-source models trail closed state-of-the-art models by only a few weeks. Smaller, local-friendly models by only a few months. This means that you can run GPT 5.1 intelligence on your laptop, at speeds similar to what you would see in the cloud. Or run real-time dictation on your computer in 100s of milliseconds.
These local models will form part of the constellation of AI models that we all use. Some tasks are processed on our own machines; some are moved to the cloud as a function of the routing our employers implement.
Products like Ollama, which is used by more than 9 million people, make it simple to run AI locally, both for simple use cases, but also complex coding tasks with custom harnesses at home or at work.
The simple reality is no business in the world can afford state-of-the-art AI for every use case for every employee, and the inference market will segment based on latency, cost, and performance.
Hyperscalers go rogue
Microsoft agrees. Its recent security products pair homegrown AI with external models for a blended offering that it claims is as performant as competing services for half the cost. Microsoft is building models for coding (MAI-Code-1-Flash), alongside general-purpose models (MAI-Thinking-1) and more specialized models (image, voice, etc). The result? Microsoft writes that its MAI models are “powering many of [its] most widely used products to maintain or improve quality while using significantly fewer tokens, in many cases saving 50-90% of GPU costs.”
Microsoft is also building its own AI accelerators that pair well with its in-house models. Google is rocking a similar story with its TPU fleet and in-house models. Amazon? Well, it’s building AI accelerators and CPUs, even if its model efforts have been lackluster in recent quarters.
Every token that Microsoft serves with its own AI models is a token it doesn’t federate to OpenAI, Anthropic – or Mistral, SpaceXAI, etc. Today, MAI models are helpful in reducing its inference spend (blunting frontier lab revenue growth). Tomorrow, if they become true first-class AI citizens, those same models could become Microsoft customer favorites, given their potential for cost savings at comparative intelligence levels; Microsoft could steal enterprise momentum from its own portfolio companies (Redmond has backed both OpenAI and Anthropic, though with far fewer dollars in the latter case).
Thus Microsoft’s massive compute footprint, proximity to customer data and workflows, ability to swap out third-party models for its own in production settings, and now up-and-running LLM release cadence represent a real threat to business revenue for labs chasing the frontier.
Cross that headwind with the oft-debated open-weight threat to frontier lab margins and growth, and mix in improving local AI performance, and you have a cocktail of headwinds for Sam and Dario precisely at the moment they are looking to go public. Reports that OpenAI is still growing quickly underscore the point that the world’s leading labs are far, far from drying up and blowing away. But even with continued growth, imagine how much faster they would expand top line if they were not fighting zero-to-low-cost models, local compute, and increasingly aggressive hyperscalers that want to protect their own margins?
This level of sustained competition, segmentation within the broader market, and tremendous amounts of capital push innovation forward at an aggressive pace, enabling consumers of AI products, including businesses and individuals, to expect dramatically more performance per dollar.
Good luck, everyone!