Start-ups seek to challenge Nvidia’s dominance over AI chip market

Start-ups are rushing to capitalise on surging demand for the specialist chips that power artificial intelligence, as a shortage of Nvidia’s latest products presents a once-in-a-generation opportunity for new challengers to the dominance of the world’s most valuable semiconductor company.

A huge upgrade to Nvidia’s sales forecast, driven by AI, pushed its market capitalisation above $1tn in May but demand is expected to outstrip supply for its latest chips well into next year.

Among the companies developing alternatives include SambaNova, Graphcore and Tenstorrent which have together raised more than $3bn over the last several years, according to figures collated by Dealroom.co, which tracks private tech deals.

Yet few have yet made significant inroads against Nvidia, whose A100 and H100 chips have become the go-to for companies such as OpenAI and Inflection AI that must process massive volumes of data to create their AI services.

Cerebras, a Silicon Valley-based AI chip start-up that has raised $730mn since it was founded in 2016, this week announced it would build and run a network of supercomputers for Abu Dhabi-based tech group G42.

The deal will be worth “in excess of $100mn” if certain milestones are met over the coming months, according to Cerebras chief executive Andrew Feldman.

“AI right now has an insatiable demand for compute,” he said. “When you’re David fighting Goliath you look for cracks . . . [Nvidia’s] inability to meet demand is just such a crack.”

The deal with G42, a private company that works across multiple sectors such as healthcare, energy and cloud computing, is one of the biggest contracts of its kind for a would-be rival to Nvidia.

G42 plans to use some of the new computing resources itself, while also selling on any “excess capacity” to other customers through its cloud-computing arm alongside Cerebras.

“People can’t get the hardware they want, or it’s too expensive,” said Talal Alkaissi, chief executive of G42 Cloud. “The market is hungry for alternatives.”

Over the years, start-ups have variously claimed to outperform certain Nvidia products for particular kinds of workloads, including training the large language models that power chatbots such as ChatGPT and other “generative AI” systems capable of producing humanlike text and realistic imagery.

But AI researchers, and the start-ups that are turning their research into commercial products, still overwhelmingly prefer Nvidia’s technology, according to entrepreneurs, investors and analysts in the sector.

“None of these start-ups are making any significant amounts of revenue,” said Jakub Zavrel, whose company Zeta Alpha tracks references to specific processors in AI research papers for tech investor Air Street Capital’s State of AI report.

While Cerebras has seen an uptick in research citations this year, overtaking Graphcore, they number in the dozens compared with thousands of researchers who mention Nvidia chips, Zavrel said. He predicted that the latest chips from AMD were more likely to take share from Nvidia than any of its private rivals. Intel is also readying its own attack on Nvidia after acquiring another AI accelerator start-up, Israel-based Habana Labs, for $2bn in 2019.

At the same time, many of the cloud computing providers that buy chips to provide services to the new wave of AI companies and their enterprise customers are also developing their own semiconductors.

Amazon Web Services launched Trainium, its custom chip for machine learning, in 2020, while Google Cloud has been offering its TPUs, or Tensor Processing Units, to customers for five years.

Microsoft, which ended a relationship with Graphcore in 2020 after just a year, is also developing their own custom silicon for AI, further squeezing the opportunity for the start-ups that would hope to go to market via cloud providers.

To win the contract with G42, Cerebras had to go far beyond creating some of the world’s most powerful processors — already an engineering feat few venture capital investors are willing to fund — by constructing and operating the entire infrastructure needed to host them too. Alkaissi called it a “white glove service” from Cerebras.

Some AI investors argue chip start-ups must go even further to match Nvidia’s offering.

“It’s not just a matter of designing the best chips, manufacturing those chips and bringing them to market in a way that people want them,” said David Katz, a partner at Radical Ventures, an AI-focused tech investor. “Nvidia has invested for a very long time in an ecosystem that lives around those chips . . . that has won the hearts and minds of the engineers that are working at the bare metal level.” That includes software and support, in particular its Cuda toolkit for programming its chips.

Faced with such a daunting set of tasks, some start-ups have pivoted away from a head-on competition with Nvidia.

Celestial AI, a Silicon Valley-based start-up that raised $100mn in June, refocused on “complementing” rather than competing with Nvidia, according to its chief executive Dave Lazovsky, by developing optical technology for connecting AI processors with the high-performance memory needed to feed them data.

“Most of these AI start-ups that are trying to compete with Nvidia have no chance because they are fighting the wrong battle,” he said. “The bottom line is the memory requirements are going to keep growing about 100 times faster than the compute requirements.”

Fabrizio Del Maffeo, chief executive of Netherlands-based Axelera AI, is developing AI chips designed for cars, medical devices and security cameras, rather than the cloud and data centres where Nvidia’s most powerful chips are in such high demand.

“I always said since day one it’s crazy to go against a trillion-dollar company with unlimited resources,” said Del Maffeo.

Additional reporting by Richard Waters