The US-China AI Gap Just Hit a Record-Low 6%

Code 2FOR1 gets you two WeAreDevelopers World Congress passes for the price of one — so don't make the trip to Silicon Valley solo.

San José, CA · September 23–25, 2026. 10,000+ developers, 500+ speakers, 20+ stages, and the full software development lifecycle in one place.

On stage: Kelsey Hightower, Thomas Dohmke (former GitHub CEO), Christine Yen (Honeycomb), Olivier Pomel (Datadog) — the people building the tools you use every day. Three days of AI, agents, cloud, security, and architecture, plus workshops, live coding, and the official Congress party.

Bring the builder you'd want in the room with you.

Get Your Passes

3 things that happened while you were busy

1.  The performance gap between US and Chinese AI just hit a record low.

Bloomberg Intelligence measured the shortfall between the top US and Chinese models at 6% in June, down from 9% in May, against a 10% to 15% average over the prior year. The trigger: Zhipu AI's GLM-5.2 topped the global ranking in agentic coding, and Chinese models took two slots in LiveBench's global leaderboard. BI's blunt framing: this raises real questions about how long US supremacy holds.

2.  Washington's response: accusations of theft.

As Moonshot's Kimi K3 and Alibaba's Qwen 3.8 landed, Trump administration officials accused Moonshot of stealing US technology, with the White House science office claiming the firm distilled Anthropic's Claude Fable model and illegally accessed Nvidia chips. Moonshot has not publicly confirmed the claims. The accusation itself is a tell: you do not accuse a competitor of copying unless they have caught up enough to bother.

3.  Kimi K3 got so popular it had to stop taking customers.

Demand for Kimi K3 was strong enough that Moonshot suspended new subscriptions within 48 hours of launch, citing overloaded compute. A great marketing headline, but analysts read it the other way too: it exposes the very real hardware constraints Chinese labs still work under, the one gap that has not closed.

China caught up on models. The real contest is now about chips.

The shrinking gap.  The timeline tells the story. US export controls in 2022 were estimated to put American AI three years ahead. Then, as SCMP put it, years turned to months, and in July, months turned to weeks. A DeepSeek moment that looked like a one-off in early 2025 now looks like the start of a trend that never stopped.

The money paradox.  Here is what makes it remarkable. Stanford research found a 23-to-1 spending gap between US and Chinese firms produced only a 2.7% performance gap. China is generating frontier-class capability at a fraction of the cost, largely through algorithmic efficiency and a fast-follow strategy rather than brute-force spending. Efficiency, not budget, is doing the work.

The remaining moat.  But models are only half the race, and the closing half. The Carnegie Endowment's Matt Sheehan estimates compute constraints alone should leave Chinese labs about 40% behind, with distillation making up roughly half of that. Nvidia's chips, and America's access to them, remain the one advantage that has not eroded. The Kimi subscription freeze was that gap showing through in real time.

The takeaway.  For you, the model race closing is straightforwardly good news: it means better tools at lower prices, whichever flag they fly under. The open-weight Chinese models pressuring closed-model pricing is why your AI bill keeps falling. The practical move is to stop assuming the best tool ships from California and start testing on merit. The prompt below helps you pick by capability and cost, not by country.

What people are actually watching and sharing

What people are actually watching and sharing

The two-slot moment.  The stat everyone is quoting: for the first time, Chinese models took two spots in LiveBench's global ranking in a single month. Screenshots of the leaderboard are all over the timeline.

The efficiency flex.  The Stanford number, a 23-to-1 spending gap yielding a 2.7% performance gap, is being passed around as the single most quotable line in the whole debate. It reframes the race from who spends most to who wastes least.

Distillation on trial.  The White House's claim that Moonshot distilled a US model to train its own has reopened the industry's messiest question: when a model learns from another model's outputs, where is the line between studying and stealing? Nobody has a clean answer.

Open beats closed on price.  Analysts note Chinese open-weight models now compete closely enough to pressure closed-model pricing worldwide. The winners of that pressure are not the labs. They are everyone who pays for tokens.

Direct a 20-second AI video like a filmmaker

FLUX 3 Video, Gemini Omni, Seedance: the new video models all reward the same skill, which is thinking in shots instead of sentences. Paste this into Claude or ChatGPT with your idea, and it writes the director-grade generation prompt for you. Use the output in whichever video tool you have access to.

Concept examples: "a street food vendor at dawn preparing the first order", "a product reveal for a handmade leather wallet", "rain starting over a rooftop garden". Advanced tip: generate the cinematic and phone-camera versions of the same concept, post both, and let your audience tell you which style your brand should speak in.