Huang Told London to Ditch 'Opaque' AI — He Backs the Rival

The man who sells the world its AI chips walked into Mayfair this week and told Britain that the two companies leading the AI race cannot be trusted with anything that matters. Jensen Huang, Nvidia's chief executive, looked at a room containing the Mayor of London and the government's AI minister, named Anthropic and OpenAI, and said their “opaque” models have no business running anything mission critical.

Not everything, mind you. Huang was careful about that. For ordinary work, he said, relying on an opaque service is “fine, frankly.” But for the systems a country or a company cannot afford to get wrong — the trust, regulatory, sovereignty and control cases — his message was blunt: “How could you have an opaque agentic system, be contained in an opaque box, run by an opaque company? It's obviously not sustainable.”

That is a startling sentence from the chief executive of a $5tn company about two of his biggest customers. Nvidia's chips train Claude and ChatGPT. Huang has just told one of America's closest allies to treat those products as boxes it cannot see inside — and to stop handing them the keys.

Photo: Nvidia CEO Jensen Huang in 2024. In London this week he urged Britain to stop outsourcing “mission critical” work to opaque AI models from Anthropic and OpenAI. Credit: © European Union / Peter Dasilva, CC BY 4.0 via Wikimedia Commons.

A warning, delivered in Mayfair

The setting is the story. The Wednesday event in Mayfair was attended by Lord Khan of Tooting, the Mayor of London, and Kanishka Narayan, the UK's AI minister — and it doubled as a spending announcement. Nvidia said it was taking minority stakes in five British start-ups as part of a $1.16bn investment push into the UK. Dealroom's numbers, cited in the reporting, say Nvidia joined 53 venture rounds worth more than $100m in the first eight months of the year, outpacing Sequoia and Lightspeed. This was not a courtesy call. It was a chequebook tour.

That same evening Huang had whisky with the prime minister, and the flattery flowed both ways. He told Andy Burnham that AI-driven growth was sitting right in front of Britain, called the country's position a Goldilocks moment, and said the UK was phenomenally well-positioned to act as a mediator between the US and China on AI. Then he twisted the knife on British risk-aversion: give an entrepreneur a one-in-a-million shot, he said, and they'll ask “so you're telling me there's a chance?” — but in Britain, somehow, there are 999,000 reasons to fail.

The company also used the trip to announce $250m of computing power — $50m a year for five years — donated to British AI-safety research institutes, earmarked for studying the safety and security of AI agents. It is the same London trip on which Huang told Bloomberg he would cut off chip supplies to any buyer building unsafe AI. Read together, the week paints a coherent picture: Huang wants Britain's AI future built on systems he can see into — and, not incidentally, on Nvidia's open ecosystem.

He was equally dismissive of the industry's doom-talk. Concerns about existential AI risk, he said, were not credible in science and not based on facts — safety is an engineering problem, not a legislative one. Geoffrey Hinton, the Nobel laureate, disagrees loudly enough to be backing a private member's bill that would prohibit the development of superintelligent AI in Britain. That is the counter-argument Huang flew straight past on his way to Berlin, where Nvidia's annual showcase lands next week.

The open alternative he didn't mention

Here is the part Huang left out of his Mayfair remarks. The same week The Times published his warning, it profiled the company his own money is backing as the answer: Reflection AI, the open-model startup founded in 2024 by two Google DeepMind alumni — chief executive Misha Laskin, who led reward modelling for Gemini, and chief technology officer Ioannis Antonoglou, who co-created AlphaGo.

Reflection's first model, Beam, debuted on October 5: a sparse mixture-of-experts system with 501 billion parameters in total but only 23 billion active per token, built for coding and agentic work and designed to run cheaply. Reuters reported it as competitive with Z.ai's GLM-5.2 and closing in on Alibaba's Qwen3.8-Max, with the full weights landing later this month under an Apache 2.0 licence. It is, in other words, exactly the kind of “open” Huang was selling in London.

We re-checked the London remarks against the open Business Matters relay of the Times' paywalled piece, published this morning. One honest caveat: the Times' own headline calls Reflection a “£19bn threat to OpenAI” while its body cites a $25bn valuation, so every number in this article is attributed, not blended.

The freshest detail in the Times profile is how aggressively the open pitch is now being sold to business. Laskin told the paper that renting from a closed provider is like renting from a landlord who can “hike up your rates or evict you” — while owning an open model is “your asset.” The startup has signed more than 40 distribution partners, including AWS, Google Cloud, Dell and CoreWeave, and struck a compute deal with SpaceX's Colossus 2 data centre earlier this year. Nvidia's own involvement runs deep: the Journal has its investment at about $800m as of March, and the Times reports the chipmaker is in talks to acquire Reflection or increase its stake — Laskin declined to comment.

The deal mechanics behind all of this we covered when the talks first surfaced — see our earlier report on Nvidia's $25B Reflection AI talks. What is new is the timing: Huang denouncing closed models in London while his money sits behind the most credible open challenger to emerge in the West. Nor is this a one-off bet — in September Nvidia agreed to buy the open-model platform Hugging Face for $12.93bn, and Huang's first-ever post on X was the “Open Weights and American AI Leadership” open letter, signed by more than 20 companies from Meta and Microsoft to Mistral and Hugging Face. Notably absent from the signatures: Anthropic and OpenAI, the very two labs he named in London. Our earlier piece on the open-weights alliance that conspicuously excludes OpenAI and Anthropic has the full signatory list.

Photo: The City of London, seen from Tower Bridge. Huang told a Mayfair audience — including the Mayor of London and the UK's AI minister — that Britain cannot afford opacity in its critical systems. Credit: The wub, CC BY-SA 4.0 via Wikimedia Commons.

The labs are not having it

Anthropic and OpenAI have heard all of this before, and their answer has two halves. The first is that nobody is proposing to ban open models — a straw man both labs keep knocking down. “Anthropic has never advocated for a ban on open-weights models,” chief executive Dario Amodei wrote in a July blog post, while insisting that all models, open and closed, should face mandatory safety testing before release. The second half is the substantive one: the big labs argue that open models lack the safety guardrails their products carry, and that handing frontier capabilities to anyone with a server is reckless.

There is a real debate buried under the posturing. Open models do give users data control — weights run on your own hardware, your data stays on your premises, nobody's cloud sees your prompts. For a bank, a hospital or a government, that is not a philosophical preference; it is a compliance requirement. But the labs' warning is not empty either: an open model cannot be patched, recalled or rate-limited once it is downloaded. Both sides are describing genuine properties of the technology. The fight is about which risk you fear more — the black box you cannot inspect, or the open box you cannot control.

Our take: referee and striker

Now the honest part. Huang's conflict of interest here is not subtle — it is brazen. He sells the picks and shovels to every lab on earth, denounces the two closed labs by name on foreign soil, and holds an $800m ticket on the open challenger that benefits from exactly that denunciation. If you wanted to design a self-serving keynote, this is what it would look like.

But the self-serving version happens to be pointing at something real, and there is a harder commercial logic underneath it. Open models run on commodity compute — which is to say, on Nvidia's chips, wherever the customer wants. Closed models, by contrast, are drifting toward custom silicon: OpenAI and Google have both poured money into designing their own chips, which would cut their reliance on Nvidia entirely. Every enterprise that chooses an open model over a closed API is a customer that stays on Nvidia's hardware. Huang is not just the referee of the open-versus-closed fight. He is playing striker for one side — and his side's business model needs open to win.

Whether it can is the open question, and it will decide who owns enterprise AI. Beam's benchmarks are genuinely impressive for a first model, but “closing in on” Chinese open models is not the same as beating Claude or ChatGPT, and 40 distribution partners is not the same as 40 million users. Huang's bet is that sovereignty — the thing governments and regulated industries cannot compromise on — is the wedge that pries the market open. Watching whether that wedge holds is now one of the most interesting races in tech. We will be keeping score.

Sources: The Times (via Business Matters open relay); Reuters; The Wall Street Journal (via Gizmodo); CNBC (via WOWTALE); The Indian Express; Startup Fortune; Analytics India Magazine; OfficeChai; Anthropic blog; Nvidia investor announcement; Dealroom.

By Daniel Rao. Daniel Rao covers US tech and AI policy for the TechPulse USA newsroom.