Watch the video with chapters & transcript →
AI TOOLS
In short: Stanford's AI Index puts the performance gap between the leading US and Chinese models at just 2.7% as of March 2026, even though US private AI investment ($286 billion in 2025) was more than 23 times China's ($12 billion).
Ani Björkström, the Stockholm-based technology and finance consultant behind the YouTube channel "Ani Björkström | AI for Finance," went back to primary sources — Stanford HAI's 2026 AI Index, DeepSeek's own technical report, the US Commerce Department's January 2026 export rule, and IEA data-centre energy figures — to answer who is actually ahead in the US-China AI race. Rather than repeating headlines about a cheap Chinese model crashing Nvidia's share price, the video traces what the underlying numbers do and don't say, and what that means for the tools finance professionals pay for and the AI vendors they might invest in.
The US leads on investment and output, but the performance gap between the two countries' best models has nearly closed. Stanford counted 50 notable US models launched in 2025 versus 30 from China, and its March 2026 comparison put the leading US model only slightly ahead — a 2.7% gap.
That lead is not stable. Stanford notes that the leading US and Chinese models have swapped places several times since the start of 2025, and a model that excels at one task, such as a programming problem, can fail at another, such as formatting a simple report. Because of this, Ani Björkström argues the public benchmark leaderboard tells only part of the story that matters to a paying business.
No — the $5.6 million figure describes an estimate for DeepSeek's V3 model based on an assumed chip-rental price, not the company's total spending. DeepSeek's own technical report explicitly excludes prior research and experimentation costs, and the figure applies to V3, not to R1, the reasoning model the company released around January 2025.
The video compares using that number to judge the whole industry to comparing an airline's fuel bill with the cost of building the airline. Even so, DeepSeek's R1 weights were released for download under an MIT license, giving other companies a real, usable alternative to paying for a hosted AI service — which is what triggered the roughly $600 billion single-day drop in Nvidia's market value described in the video's framing. Downloadable weights are not free to run, though: hardware, engineering, security and maintenance costs remain, so a hosted service can still be the cheaper option at scale.
Because both the US and China depend on the same physical supply chain to turn any model into a usable product. Stanford notes that Taiwan Semiconductor Manufacturing Company makes almost all of the world's leading AI chips, and many US-designed chips rely heavily on manufacturing there.
In January 2026, the US Department of Commerce announced case-by-case review of some powerful chips, including Nvidia's H200, under security conditions — tightening China's access to advanced computing power. But chip access is only half the physical constraint: the IEA estimates global data-centre electricity consumption was about 485 terawatt-hours in 2025 and projects roughly 950 terawatt-hours by 2030, a reminder that cooling, networking and power capacity take years to build, regardless of how fast a model ships.
Ani Björkström recommends tracking three signals rather than weekly headlines: the total cost of getting useful work done, access to operational (not announced) computing power, and how easily a customer can switch providers. A cheap model that needs 20 minutes of human correction can end up more expensive than a costlier one that doesn't.
She also flags a lock-in risk that applies to vendors from either country: once a company connects its documents, builds workflows and trains staff around one model, switching stops being "just choosing another chatbot." That is why she frames the decision as a three-way trade-off — capability, total cost, or control over data — rather than a single winner.
| Metric | United States | China |
|---|---|---|
| Private AI investment, 2025 (Stanford HAI) | ~$286 billion | ~$12 billion (state funding likely underestimated) |
| Notable models launched, 2025 | 50 | 30 |
| Leading-model performance gap, March 2026 | Slightly ahead | 2.7% behind the US leader |
| Key constraint | Depends on TSMC (Taiwan) for leading-edge chip manufacturing | Restricted from some advanced chips, incl. Nvidia H200, under Jan. 2026 US export review |
| Notable 2025 release | Hosted and open models from major US labs | DeepSeek R1 weights released under MIT license (downloadable) |
That figure is DeepSeek's own estimate for V3's training compute, based on an assumed chip-rental price, and it explicitly excludes prior research costs — it is not the company's total spending and does not apply to its R1 reasoning model.
Investors reacted to DeepSeek's low reported training cost as evidence that powerful AI could be built far more cheaply than assumed, raising doubts about the scale of spending on Nvidia's chips that the market had priced in.
In January 2026 the US Department of Commerce announced case-by-case review of exports of some powerful chips, including Nvidia's H200, under security conditions, adding to existing limits on China's access to advanced computing and manufacturing technology.
Read the full video transcript, with chapters, on the video page →
Want this working inside your finance team?