Sovereign AI: Why Cosine wants to bring data home

“You can see us in the same way that you look at OpenAI, Anthropic and Mistral,” Yang Li tells me, as we perch at the fringes of Acceldata’s Autonomous tech event in London, where he is due to speak.

“We build the underlying large language model [LLM]. First we train it to a graduate level of competency, then we focus on post-training in a specific field, like writing software. We’re an AI lab that builds the models end-to-end.”

Cosine was founded in 2022 and was one of the first seven companies backed by the UK government’s £500m Sovereign AI venture fund set up this spring (and topped up in the latest Budget). In my interview with Yang Li you will learn:

• Why Cosine is training a UK-built AI model with 35 organisations, from HSBC to BAE Systems
• Why on-premises AI could beat the cloud on cost
• What the £500m Sovereign AI fund offers scale-ups beyond capital

Why AI models should be homegrown

Of course, we’ve all got used to using LLMs like ChatGPT and Claude, both in our personal and professional lives. Yet, like social media platforms or cloud providers, they are tech that is primarily developed and run from across the Atlantic or in China.

The idea of Sovereign AI is to protect us if a decision were suddenly made to restrict access to this vital tech. The issue became live this summer when the Trump administration temporarily blocked foreign nationals from accessing two of Anthropic’s latest models over national security concerns.

Sovereign AI has become a policy goal in both the European Union and the UK. It is designed to give countries direct control over AI infrastructure, as a matter of security.

Li’s team of 30 people is based in the UK. Its model, Lumen Sovereign, has been trained on Isambard-AI, the government supercomputer housed at the University of Bristol.

Cosine has always seen businesses as its main clients, rather than the general public. And its LLMs are designed to solve specific business tasks, which dictates their ‘post-training’ in designated sectors.

It has assembled a coalition of around 35 UK organisations, including BAE Systems, Babcock, HSBC, Lloyds, NatWest, PwC and BT Group, who either supply data or specify requirements that get trained into the model.

Anything identifiable is replaced with a synthetic equivalent, says Li. “It’s not even anonymised, it’s just completely different,” he explains. If a bank shares its own management accounts, Cosine generates alternative versions with different formulas and formatting, for example.

Cosine works with the AI Security Institute and has appointed the British firm Advai as third-party auditor.

Why Yang Li says AI is like Uber

In fact, Li ruled out the consumer route early, based on his analysis of AI competitors’ underlying strategy, which he believes relies on subsidising early losses with deep pockets to gain market share.

“It’s an open secret that every big lab is heavily subsidising usage to gain market share,” says Li. “It’s similar to Uber’s aggressive practices when they first entered the market. So we just can’t compete on the consumer side.”

That insight pushed the company not just towards B2B, but also to the idea of on-premises deployment. This goes against the logic of the AI revolution that's occurred over the past two decades, which has been all about storing data in the cloud, rather than a local server.

But Li argues that “on prem” is the more durable economic model in the AI era, because you no longer rent the hardware and you don’t pay per token.

A rack of GPUs costs roughly $500,000 and, Li points out, supports 50 to 100 employees. “That’s it, other than electricity.” Because the cost is sunk, the incentive inverts: the more you use it, the cheaper each token becomes.

He describes staff hammering the GPUs by day, then overnight doing the “background passive work” (cybersecurity scanning, KYC reports, processing meeting notes) at no significant cost. Owning your own servers in a warehouse may seem like “a 1970s or 80s play”, says Li, “but we think it’s the next frontier, we’re going full circle.”

“Right now, most CFOs are at odds with their CTO and CIO,” he adds. “They want to token max. They want to experiment, they want to accelerate. But every token you use is being charged to you.” And the cost is only going one way.

Li believes that the cloud has delivered on speed and convenience, but not on cost. “If you ask any large bank or industrial company, has the cost saving of cloud compute materialised? I would say by and large they would all say ‘no’. In fact, the cost has ballooned far beyond what they were expecting and actually they would have probably saved money if they just stuck to expanding on premises. And that’s the same with AI.”

How the Sovereign AI fund helps scale-ups

Cosine applied to the government’s Sovereign AI fund when the programme opened. It was picked after an interview process alongside Callosum, Prima Mente, Cursive, Doubleword, Twig Bio and Odyssey.

The benefit of being part of the government’s Sovereign AI programme boils down to three things, Li explains: “capital, compute and contracts”. The capital comes as the government’s right of first refusal to invest up to £25m, triggered only if Cosine raises again on the venture market (which it hasn’t, so far).

“We have no intention of raising funding right now,” he tells me. “We don’t need it.” The company has taken just over $8m in outside funding to date, from Lakestar, SOMA Capital, Northern Powerhouse and Gaingels, having come through Y Combinator’s W23 batch.

In terms of ‘compute’, the Sovereign AI Unit’s backing brings a certain amount of free access to the Isambard-AI supercomputer. The third ‘c’ of ‘contracts’ is not promised. What the government provides is navigation.

“With DSIT and Cabinet Office help, we navigate the annals of Whitehall much more effectively,” says Li. Thanks to them, Cosine has worked closely with the MoD, the Home Office and HMRC.

Three good exits and a piece of advice

Li is just the kind of person the chancellor John Healey might have had in mind when he talked about wanting to back “bloody-minded entrepreneurs”.

Li read statistics at university, then joined Job&Talent — the Spanish machine-learning recruitment business now backed by Atomico and SoftBank — when it was, he says, “two slightly brave and slightly crazy Spanish people”.

After a successful exit came the ride-share platform Mobike, an orange staple on UK city streets, before e-bike sharing took off. Born in London to Chinese parents, Li spoke Mandarin and became the Chinese bike-sharing firm’s first employee outside China.

He launched every city it operated in beyond its home market — from Berlin to Bogotá — rose to commercial director, and watched the company reach 220 million monthly active users on $1.3bn of venture funding. Meituan bought Mobike for $2.7bn in 2018.

After a stint at a family office founded by the father of Formula 1 driver Lando Norris, Li was introduced to Mark Hipperson, co-founder and former CTO of Starling Bank, and helped build Ziglu, an FCA-regulated electronic money institution.

Li co-founded Cosine in 2022 with Alistair Pullen, whom he met through his wife. Pullen started coding at age eight and made £10,000 from an Angry Birds clone at nine, Li points out with a smile. The pair were amazed at what GPT-3 could do when it appeared (just ahead of the first consumer model) and quickly realised the potential of the technology.

“There’s definitely an element of luck,” says Li of his career as a serial entrepreneur, before quoting advice he was given by one of his advisors, Michael Seibel, the Twitch co-founder. “He said: the good founders listen to what I say. The best founders don’t listen to anyone.”