Mira Murati's Design.
By Stephen Klein.
By Stephen Klein, contributor. This is his first “column” for News Items.
We have been told that the defining battle in generative AI is the United States versus China. I am no longer convinced that is the most useful way to understand it. The defining battle will not be intelligence versus intelligence. It will be intelligence divided by cost. And if we wanted to have a little fun with it, we might say the contest is not really Washington versus Beijing at all. It is Mira Murati versus the DeepSeek playbook.
Ms. Murati matters because she represents a smarter possible American response: not building the largest and most expensive intelligence possible, but designing intelligence around customization, efficiency and economic usefulness. To see why, we need to acknowledge something Silicon Valley would rather not admit: large language models are becoming commodities.
Commodities do not have to be identical, only substitutable enough that buyers can compare them and switch. That is already happening. Developers swap one model for another. Companies route different tasks to different models. Open-weight systems are customized and deployed privately. DeepSeek has even made its API compatible with the formats used by OpenAI and Anthropic, deliberately reducing the friction of switching. Once several products can produce a sufficiently good result, price becomes decisive. The frontier models may be the most powerful commodities ever created. They may also be the most expensive commodities in human history.
For years, American frontier laboratories have competed through scale: larger models, more compute, more expensive talent, extraordinary capital. The strategy worked brilliantly, and has created an economic trap. Each breakthrough compels financing the next before competitors catch up. OpenAI reportedly expects to spend roughly $600 billion on computing through 2030; its inference costs reportedly quadrupled during 2025, pushing adjusted gross margin from 40 percent down to 33. Revenue growth remains remarkable. But revenue does not answer the underlying question: can the cost of producing and delivering frontier intelligence fall fast enough to generate sustainable returns while cheaper competitors close the performance gap?
China, partly through necessity has been asking a different question: how much intelligence can we produce, and at what cost? Chip restrictions forced Chinese developers toward architecture and efficiency. DeepSeek’s mixture-of-experts design activates only part of the model per request. Moonshot’s new Kimi K3 contains a reported 2.8 trillion parameters but activates just 16 of its 896 experts per operation, converting compute into capability roughly 2.5 times more efficiently than its predecessor. The pricing makes it concrete: Moonshot lists K3 at $0.30 per million input tokens and $3 per million output; Anthropic lists Claude Sonnet 5 at $2 and $10, rising to $3 and $15 after August 31. The comparison is imperfect, but the direction is unmistakable. China is not merely competing to produce intelligence. It is competing to change its price.
Which brings us back to Mira Murati. On July 15, Thinking Machines Lab released its first model, Inkling, and said something no American frontier lab has been willing to say: “Inkling is not the strongest overall model available today.” That admission is the strategy. Inkling is an open-weight mixture-of-experts model, 975 billion parameters with only 41 billion active, positioned not as a champion but as an efficient foundation to be adapted with a 12-billion-active-parameter Inkling-Small coming to cut cost and latency further. Analysts rated it the most capable American open-weight model, while noting it still trails the best Chinese open models. That is not a footnote. That is the point. The efficiency race is real, it is being run at the frontier, and the United States is behind in it.
Some early evidence: In a collaboration with Bridgewater Associates, a model fine-tuned on specialized financial data scored 84.7 percent on financial reasoning benchmarks, beating the proprietary frontier models it was tested against on both accuracy and cost. The figure is self-reported, not independent. But it reveals the strategy: customize intelligence for a particular problem rather than repeatedly buying the most expensive general intelligence available. China arrived at efficiency through necessity. Murati is pursuing it by design.
Here is the uncomfortable part: the incumbents cannot easily follow her. Hundreds of billions in data-center commitments, chip contracts and energy agreements have been signed on the premise that ever-larger frontier models will command premium prices. Valuations assume it. Investors have priced it. A company that has promised the frontier cannot suddenly announce that the frontier is a commodity. A new entrant can choose efficiency by design; the giants must keep running the race they created, and keep paying for it.
The most revealing evidence may be what OpenAI and Anthropic are doing with the money: moving into implementation and consulting. OpenAI has launched the OpenAI Deployment Company with more than $4 billion in initial investment and is acquiring Tomoro, an applied-AI consulting firm with roughly 150 engineers. Anthropic has helped establish an AI services company with Blackstone, Hellman & Friedman, Goldman Sachs and others. If access to frontier intelligence were sufficient to produce valuable outcomes, the model makers would not need armies of consultants to help customers use it. Producing intelligence and creating value from intelligence are two different businesses.
The frontier laboratories have accomplished something historic. But the next race will not be won by the company that produces the most intelligence. It will be won by the company that delivers the right intelligence at the lowest sustainable cost.
Stephen Klein is the founder and CEO of curiouser.ai, a serial entrepreneur, and a part -time instructor at the University of California, Berkeley, teaching ‘Marketing in the Age of AI and AI Ethics’.