The government’s high-risk science investment group is funding a test lab for start-ups to prove their AI chips can compete with those made by Nvidia and position the UK to grab a share of a $1 trillion industry.

The Advanced Research and Invention Agency (Aria) is committing £50 million to the initiative. An initial £16 million grant is going to the Cambridge-based not-for-profit CommonAI organisation, which will host the testbed for the next two and a half years.

The facility, called the Scaling Inference Lab, aims to provide a shortcut to commercialising new technologies in a similar way to the National Quantum Computing Centre, which lets quantum firms house and test their systems. The lab will operate in cycles, proving new AI compute systems every six months.

Suraj Bramhavar, Aria’s programme director, said it needed to move quickly because of the speed at which AI compute technologies are emerging. “We want to provide a place that is a neutral playing field where anyone can come and bring their prototype and we will be a willing first customer,” he said. 

“We will create the infrastructure that allows you to plug in your new technology and test it on real workloads and get some early data. They [start-ups] can then use that data to increase their leverage with investors or potential customers or their own internal purposes. This non-profit, neutral aspect is not something any venture capital firm will fund.”

Aria’s decision shifts its focus away from funding research, under its £50 million Scaling Compute programme. However it remains committed to its goal to lowering the cost of developing AI capabilities by as much as 1,000 times.

It has backed 12 projects, including one run by the semiconductor start-up Fractile, which said this month that it would invest £100 million in its UK operations over the next three years.

Business newsletter

The business editor’s exclusive analysis of all the latest financial and economic news.

Sign up with one click

Another, the AI chip start-up, Callosum, founded by the Cambridge and Oxford PhDs Danyal Akarca and Jascha Achterberg, will be using the new test lab. Bramhavar said Aria would continue to support the first wave of projects, but they would not receive follow-on funding because there was “ample risk-tolerant capital” available from private investors to develop the training models.

Nick McKeown, a professor at Stanford University, a board member at Aria and a former senior vice-president at US chipmaker Intel, said the UK had an opportunity to build expertise in AI hardware. “The global market for AI chips, which is just part of the infrastructure, is going to be about $1 trillion within ten years and it is growing at 30 per cent a year,” he said. 

“Almost all of that is in the United States. If the UK could step up into just 5 per cent of that, it would be $50 billion of very high-margin, profitable revenue within ten years. It would make a sizable difference to the UK economy, lots of wonderful high-tech well-paying jobs, and we would have a seat at the table, an important part of the global infrastructure.”

Aria is in discussions with corporate partners for the testbed, with some big-name chipmakers likely to provide funding as part of their corporate venturing initiatives.

Bramhavar said he was “bullish” that the UK could develop alternative AI chip technologies to those supplied by the giant chipmakers such as Nvidia, AMD and Broadcom. 

“The harder part is because there’s so much profit being made in this ecosystem, specifically on the hardware side, it becomes very easy for the industry to just incrementally move the needle from one generation to the next because they don’t want to risk the massive amounts of profit that they know are sitting there next quarter,” he said.

“So when these alternatives do arise, the question is, can we actually get them adopted into industry writ large. That question is harder than actually unlocking the alternatives to begin with.”