Demis Hassabis says Chinese AI companies are about six months behind leading western labs and the response to DeepSeek's R1 in 2025 was a “massive overreaction”
Google DeepMind Chief Executive Officer Demis Hassabis said Chinese artificial intelligence companies haven't been able …
Bloomberg
Context & Ripple Effects
The DeepSeek debate has already centered on whether headline efficiency claims captured the full cost of building competitive models: Hassabis previously called the company’s reported training-cost figure a misleading fraction of total development cost. This latest assessment extends that skepticism to the broader market reaction.
The claim also lands as DeepSeek’s models have moved beyond lab comparisons into deployment across Chinese consumer and industrial products, including the rush by Chinese device makers and automakers to build on DeepSeek. That makes the perceived gap consequential for both competitive positioning and AI-policy arguments.
First-order effects
Hassabis’s assessment gives DeepMind and other leading Western labs a public argument that DeepSeek’s R1 did not erase their frontier-model advantage, while tempering a narrative that Chinese labs have already caught up.
His proposed frontier-AI standards body would put pre-release model review on the immediate policy agenda, asking participating labs to submit models up to 30 days before launch.
Second-order effects
Chinese AI companies and their commercial partners face greater pressure to demonstrate sustained model capability and deployment value, rather than relying on the R1 moment as evidence of parity.
A US-led review framework would force frontier labs to weigh the safety and legitimacy benefits of coordination against the competitive sensitivity of sharing unreleased models; the earlier six-month-gap assessment supplies part of the strategic rationale for that debate.
Third-order effects
If frontier-model oversight is organized through a US-based industry body, access to release legitimacy may become another competitive advantage for incumbent labs with policy relationships and the capacity to support review processes.
The deeper contest is likely to shift from a single benchmark or cost claim toward competing AI ecosystems: frontier research, domestic deployment, and the institutions each side builds to validate and govern models.
The trend: AI competition is increasingly being framed as a contest over institutional credibility and deployment ecosystems, not just isolated model releases.
Would @demishassabis be in favor of an AI pause to give society a chance to catch up? “I think so. I sometimes talk about setting up an international CERN for AI to figure out what we want from this technology.” @emilychangtv #BloombergHouse #WEF26 ⏯️ https://www.youtube.com/... …
On peer based collaboration within the AI community, “I think I am on pretty good terms with all the leaders in the leading labs.” @demishassabis @emilychangtv #BloombergHouse #WEF26 ⏯️ https://www.youtube.com/... [video]
I don't have a strong view on the correct policy outcome, but I do think this analogy doesn't fit. China already has a domestic chip industry and is deploying capable AI models. Export controls don't stop that reality. The strategic rationale behind selling advanced chips
the only way this policy was remotely defensible was under a fast takeoff scenario, and we're not in a fast takeoff. this is just anticompetitive behavior as performative security theater. china is training on ascend now. you're just building nvidia's competitors.
Dario will be talking to the WSJ live in Davos in about an hour, but he's already caused a bit of a stir this morning. 'I think it would be a big mistake to ship these chips. ... If you think about the incredible national security implications of building models that are