Chinese AI models narrowed the performance gap with US counterparts to a record-low 6% in June, down from 9% in May, according to Bloomberg Intelligence. The finding raises what Bloomberg called “questions about the sustainability of US technological supremacy” in artificial intelligence.
What is the AI performance gap? It’s a composite metric that compares how well the best Chinese AI models score against the best American models on standard benchmarks — reasoning, coding, math, and language understanding. A 6% gap means China’s top models are performing within 6% of America’s best. A year ago, the gap was wide enough that Chinese models were dismissed as fast followers. Now they’re within striking distance.
🔍 THE BOTTOM LINE
The gap isn’t narrowing because US models are getting worse — it’s narrowing because Chinese models are getting better, fast. Moonshot’s Kimi K3 launch and Zhipu’s steady improvements prove the Chinese AI sector isn’t a one-off success story. Bloomberg Intelligence’s data shows a consistent monthly trend: 9% in May, 6% in June. If that trajectory holds, the gap closes before year-end.
Moonshot Proves It’s Not a One-Off
The Bloomberg report frames the gap narrowing through two data points: Moonshot AI and Zhipu. Moonshot launched Kimi K3 last Friday, advertising it as the first open-source model of its size — more freely modifiable than Anthropic’s Claude or OpenAI’s ChatGPT. Analysts judged K3 competitive with both leading US models, which are also less freely customisable.
Moonshot had to temporarily pause new subscriptions after demand for K3 strained capacity. That’s not the behaviour of a second-tier product. It’s the behaviour of a model that people actually want to use.
This follows the pattern we tracked in our coverage of Moonshot’s IPO push and China’s broader open-weights strategy. The Chinese approach — release capable models with open weights, let developers modify them freely, and compete on accessibility rather than raw capability — is working. Bloomberg’s 6% figure is the quantitative proof.
Zhipu Is the Second Data Point
Bloomberg’s headline names Zhipu alongside Moonshot, and the framing matters: if only Moonshot were closing the gap, it could be a one-off. Two companies doing it simultaneously means it’s a sector. Zhipu, backed by Tencent and Alibaba, has been releasing models steadily throughout 2026 and has been climbing the benchmark leaderboards without the same splash as Moonshot’s K3 launch.
The implication for US AI companies is uncomfortable. When the gap was 20% or 15%, the narrative was “China copies, America innovates.” At 6%, the narrative shifts to “China is about to catch up, and may already have in some domains.” Bloomberg Intelligence — not a publication known for alarmism — is explicitly questioning whether US technological supremacy is sustainable.
The IPO Valuation Impact
The timing is brutal for US AI companies preparing to go public. The Guardian reported that implied values for Anthropic’s sharemarket debut slumped $232 billion to $1.56 trillion on IG’s trading platform from Friday to Tuesday. OpenAI dropped $160 billion to $1.16 trillion. These are market bets, not true valuations — both companies are still privately held and valued under $1 trillion each. But the direction is clear: the Kimi K3 launch and the Bloomberg gap data are putting downward pressure on the IPO narrative.
The story US AI companies want to tell investors is: “We have a technological moat, China is years behind, buy our stock.” Bloomberg’s 6% number says the moat is shallow and the years are months. Anthropic, already dealing with a $1.5 billion copyright settlement and a supply chain blacklist from the Trump administration, now faces the harder question of whether its frontier models are meaningfully better than what Chinese competitors give away for free.
Why the Gap Matters for Everyone
The 6% gap isn’t just a US-China story. It affects every country buying AI services — including New Zealand. When Chinese models are nearly as good as American ones and significantly cheaper (or free, in the case of open-weights releases), the commercial landscape shifts. NZ companies that currently default to OpenAI or Google APIs for AI features now have a credible alternative in Chinese open-weights models they can run themselves.
The trade-off is security and sovereignty. The US government is already pushing to restrict Chinese AI model adoption on national security grounds, and Five Eyes intelligence agencies have echoed those concerns. NZ sits inside Five Eyes, which means the pressure to avoid Chinese models will be real — even as the performance argument for using them gets stronger by the month.
The tension between “it’s good enough and it’s cheap” and “we’re told not to use it” is going to define AI procurement decisions for years. Bloomberg just put a number on it: 6%.
❓ FAQ
What does a 6% AI performance gap mean in practice?
It means the best Chinese AI models score within 6% of the best American models on standard benchmarks. For many use cases — customer service, content generation, code assistance — a 6% difference is barely noticeable. For frontier applications like scientific research or complex reasoning, the gap matters more.
Is Bloomberg Intelligence a reliable source?
Bloomberg Intelligence is Bloomberg’s research arm, providing analysis used by institutional investors globally. It’s not a hype-driven blog or a partisan think tank. When BI says the gap is 6% and questions US supremacy, that’s mainstream financial analysis, not geopolitical speculation.
Can Chinese models really compete with GPT and Claude?
Moonshot’s Kimi K3 is judged by analysts as competitive with leading US models. Zhipu is climbing the same benchmarks. The open-weights approach means Chinese models are also more freely modifiable — users can fine-tune, inspect, and deploy them without vendor lock-in, which US models don’t offer.
What does this mean for New Zealand?
NZ companies buying AI services now have a choice between expensive US models and nearly-as-good Chinese ones. But Five Eyes security concerns may limit adoption of Chinese models for government and infrastructure use. The 6% gap makes that trade-off harder to resolve.
🔍 THE BOTTOM LINE
Bloomberg Intelligence says the US-China AI performance gap is 6% and shrinking. Moonshot’s K3 and Zhipu prove it’s a sector-wide trend, not a one-off. US companies preparing IPOs are watching their implied valuations drop as the moat narrative erodes. For everyone else, the practical question is already shifting from “can Chinese models compete?” to “can we afford not to use them?” — and that question is going to get louder every month the gap keeps closing.
