Why US Startups Are Quietly Switching to Chinese AI — The AI Race
Every headline about "the AI race" picks a winner within the first sentence. The actual data tells a stranger story than either "the US dominates" or "China is catching up" fully captures. In 2025, the US spent roughly 23 times more than China on private AI investment — and the gap in model capability between the two countries' best systems narrowed to just 2.7 percentage points anyway. Here's what's actually happening across the layers that make up this race: model capability, cost, chips, and strategy — with the specific numbers most coverage skips.
The AI race isn't one competition — it's several running in parallel: raw model capability, cost efficiency, chip manufacturing, and national strategy, and they're not all moving in the same direction.
Stanford HAI's 2026 AI Index — one of the most cited annual benchmarks in the field — deliberately declines to declare a winner. Instead, it describes a top tier of model developers, including Anthropic, xAI, Google, OpenAI, Alibaba, and DeepSeek, now clustered tightly together in raw capability.
As of March 2026 LMArena data, Anthropic's Claude Opus 4.6 led the blind-test leaderboard, with ByteDance's model trailing by just 2.7 percentage points — down from a 17.5 to 31.6 point gap as recently as May 2023.
📊 The Number That Reframes the Whole Conversation
The US spent approximately $285.9 billion on private AI investment in 2025. China spent approximately $12.4 billion — roughly 23 times less. And yet the performance gap between the two countries' best AI models is now down to 2.7 percentage points. Whatever "winning the AI race" means, it clearly isn't a simple function of who spends the most money. The more useful question, per the AI Index's own framing, isn't who's ahead — it's which specific dimension you're measuring, because the answer changes depending on whether you're looking at capability, cost, deployment speed, or hardware.
Spending vs. Capability — The Numbers Side by Side
💰 23x the Spending, 2.7 Points of Difference
Source: Stanford HAI AI Index 2026, LMArena leaderboard dataThe Efficiency Shift — Why US Companies Are Quietly Switching Models
The most concrete, measurable shift in 2026 isn't happening in a lab — it's happening in company billing dashboards. As frontier model prices rose, many US businesses started routing everyday tasks to cheaper, open-weight Chinese alternatives whenever a task didn't require the single best available model.
⚡ The Story Behind One Real Switch — Not a Hypothetical
Flo Crivello, CEO of AI startup Lindy — a roughly 25-person company — switched 100% of his company's AI traffic from Anthropic's Claude models to China's DeepSeek earlier in 2026. "We did it, and you could see that cost curve go down, like, crash to the ground," he told CNBC, estimating the move would save millions of dollars within months. His own words on why: "It's a matter of survival for the business." This is the efficiency shift in practice, not theory — and it's a pattern Harpreet Arora, head of agentic infrastructure at Vercel, described directly: "Price is doing the work here. When a task doesn't need the best model, teams are beginning to route it to the cheapest one that's good enough."
The Gap That Isn't Closing — Chip Manufacturing
🔬 The Most Durable Advantage in the Entire Race
While the model-capability gap has narrowed dramatically, the hardware manufacturing gap has not moved nearly as much — and most "AI race" coverage focused on chatbot benchmarks skips this entirely. According to Council on Foreign Relations analysis, US and allied manufacturing capacity for advanced AI processor dies is estimated at 35 to 38 times China's capacity, adjusted for quality. Huawei — China's leading domestic AI chip developer — has reported chip yields in the 5% to 20% range, compared to Nvidia's estimated 60-80% yield on its current Blackwell architecture. Huawei's manufacturing partner SMIC remains limited to a 7-nanometer process, roughly two to three generations behind TSMC's leading-edge 3-nanometer capability.
Semiconductor analyst Stacy Rasgon of Bernstein Research offered the clearest framing: "On a chip-level basis, Huawei has essentially caught up to where Nvidia was three years ago." Real progress — made on a treadmill, since Huawei's roadmap toward Blackwell-level parity by 2027-2028 depends on SMIC clearing manufacturing hurdles that export controls are specifically designed to prevent.
Two Different Strategies — Not Just Two Different Speeds
A 2026 U.S.-China Economic and Security Review Commission analysis, titled "Two Loops," makes a point most coverage misses: the US and China aren't running the same race with different amounts of money. They're pursuing genuinely different strategies.
🎯 Frontier-Model Betting vs. Open-Deployment Strategy
| Approach | US Ecosystem | China's Approach |
|---|---|---|
| Core bet | Superior compute yields transformative frontier capability | Open development + rapid, economy-wide deployment |
| Qwen (Alibaba) downloads, March 2026 | — | 942 million — 2x+ the next 8 competitors combined |
| Baidu Hugging Face releases | — | 0 in 2024 → 100+ in 2025 |
| ByteDance / Tencent open-source volume | — | 8-9x increase, 2024 to 2025 |
| Pricing approach | Premium frontier pricing, rising in 2026 | State-subsidized API access, aggressive pricing |
The Funding Race — Numbers That Would Have Sounded Fictional in 2023
⚠️ Frontier Labs Have Become Infrastructure Companies
In 2026, OpenAI closed a $122 billion raise at an $852 billion post-money valuation, anchored by Amazon, Nvidia, SoftBank, and Microsoft. Anthropic layered an additional $40 billion from Google and $5 billion from Amazon (bundled with a $100 billion AWS spending commitment), plus chip-supply agreements with Google and Broadcom reportedly worth hundreds of billions, with reported talks for a further $50 billion round at a $900 billion valuation. Both companies filed confidentially for an IPO in early June 2026. Microsoft CEO Satya Nadella addressed the broader concentration-of-power dynamic directly in a June 2026 essay: "The last thing any of us want is a world where every company across every sector is ceding value to a few models that eat everything they see... If all the value is accrued by only a few models, the political economy will simply not tolerate it."
What Most "AI Race" Coverage Leaves Out
⚡ 1. Frontier Cyber-Offense Capability Is Doubling Every Four Months
In 2026, two frontier models — Anthropic's Claude Mythos Preview, followed three weeks later by OpenAI's GPT-5.5 — became the first AI systems to clear a UK AI Security Institute 32-step, end-to-end simulated cyber-attack range that typically requires 20 hours of human red-teaming. The Institute's own assessment: frontier cyber-offense capability is currently doubling roughly every four months. Important caveat directly from the researchers: these evaluations lack active defenders or defensive tooling, so they don't yet prove real-world efficacy against hardened, actively defended targets — but the capability trajectory itself is a significant, underreported dimension of what "winning the AI race" actually means in practice.
⚡ 2. Export Controls Cut Both Ways — And Reasonable People Disagree About the Tradeoff
In mid-2026, a government-directed export control action temporarily restricted the release of some Anthropic and OpenAI models. Some tech executives and administration officials have argued such restrictions are necessary to protect national security and prevent adversary access to the most capable systems. Other analysts have pointed out that any window where leading US models are restricted domestically creates a real opportunity for Chinese alternatives — many already competitive on specific benchmarks — to gain adoption during that gap. Both are legitimate, documented perspectives on the same underlying tradeoff, and the AI race genuinely includes this policy dimension alongside the technical one.
⚡ 3. The AI Labs Are Now Political Spenders, Not Just Technology Companies
A structural shift most "AI race" content skips entirely: groups aligned with OpenAI and Anthropic have both become significant direct spenders in US domestic political campaigns as of 2026, reflecting what one independent tech researcher described as a spending pattern that "really mirrors the corporate competition" between the two companies. Whatever one's view of any specific policy position, the underlying fact — that frontier AI labs are now major players in electoral spending, not just model development — is itself a genuinely new and underreported dimension of how "the race" is actually being contested.
The Honest Framing — What's Genuinely Uncertain
✅ What the Data Makes Clear
- Model capability gaps between leading US and Chinese systems are genuinely narrowing, not stable
- Cost efficiency has become a real competitive axis, independent of raw capability
- The US retains a substantial, harder-to-close advantage in chip manufacturing capacity and yield
- China's open-source deployment strategy is producing measurable global adoption (Qwen's download numbers)
- Capital available to frontier labs has reached a scale that reshapes what "winning" even means financially
⚠️ What Remains Genuinely Contested
- Whether export controls net help or hurt US competitive position is unresolved among experts
- Benchmark scores don't fully capture real-world deployment, safety, or reliability differences
- Cyber-offense capability evaluations lack active-defender testing, limiting real-world conclusions
- China's chip manufacturing trajectory depends on SMIC overcoming constraints specifically designed to prevent it
- Neither Stanford HAI's own AI Index nor most serious analysts declare a single "winner" — treat any source that does with skepticism
For Readers Who Want the Deeper Strategic Context
If this topic genuinely interests you beyond the headlines, a well-regarded book-length treatment of the broader US-China AI competition offers useful context that a single article can't fully cover.
Affiliate disclosure: the Amazon link above is an affiliate link. We may earn a small commission at no extra cost to you.
🛡️ What Happens When the AI Race Moves to the Battlefield?
Commercial investments and public benchmark scores are only half the story. The real global conflict is actively playing out across classified defense networks, sovereign hardware pipelines, and strict trade restrictions. Read our complete breakdown of the 2026 AI War to discover how military tech and tightening export controls are rewriting the geopolitical rules of engagement.
Read the AI War Deep Dive →Frequently Asked Questions
What is the "AI race" and who's actually involved?
Simultaneous competition across multiple layers: frontier labs racing on capability (OpenAI, Anthropic, Google DeepMind, xAI in the US; DeepSeek, Alibaba's Qwen, Zhipu, Baidu, ByteDance, Tencent in China), a broader US-China national competition, and a chip manufacturing race (Nvidia, TSMC vs. Huawei, SMIC). Stanford HAI's 2026 AI Index declines to name a winner, describing the top model-developer tier as tightly clustered in capability. As of March 2026 LMArena data, Anthropic's Claude Opus 4.6 led, with ByteDance trailing by 2.7 percentage points.
Is China catching up to the US in AI, or is the US still ahead?
Both, depending on the dimension. Capability gap: narrowed to 2.7 points (from 17.5-31.6 in 2023) despite the US spending ~23x more ($285.9B vs $12.4B in 2025). Cost efficiency: several Chinese models match frontier performance at a fraction of the price. Hardware: the US retains a 35-38x manufacturing capacity advantage per CFR analysis, with Huawei's yields (5-20%) still well behind Nvidia's (60-80%). No single source declares a clean winner.
Why are US companies switching to Chinese AI models?
Primarily cost. Chinese open-weight models have been reported 60-90% cheaper than leading US models. Chinese models' share of US company tokens on OpenRouter exceeded 30% weekly since Feb 2026 (peaking at 46%), up from 11% the prior year. Lindy CEO Flo Crivello publicly switched his company's entire traffic to DeepSeek, calling it "a matter of survival for the business." This has unfolded alongside separate, contested government export control actions on some US frontier models.
What is China's actual AI strategy compared to the US?
Per USCC's 2026 "Two Loops" analysis: the US concentrates investment in compute-intensive frontier models, betting hardware superiority yields transformative capability. China, compute-constrained by export controls but state-backed, pursues open development and rapid economy-wide deployment. Evidence: Alibaba's Qwen hit 942M downloads by March 2026 (2x+ the next 8 competitors combined); Baidu went from 0 to 100+ Hugging Face releases in a year; ByteDance and Tencent increased open-source releases 8-9x.
What does the AI race mean for chip manufacturing?
This is the most durable, slowest-closing gap. Per CFR analysis, US/allied manufacturing capacity for advanced AI dies is 35-38x China's, quality-adjusted. Huawei's chip yields sit at 5-20% vs. Nvidia's 60-80% on Blackwell. SMIC remains at 7nm, 2-3 generations behind TSMC's 3nm. Analyst Stacy Rasgon: "Huawei has essentially caught up to where Nvidia was three years ago" — real progress, but constrained by export controls specifically designed to slow further gains.
