AI Chip Unicorn Suddenly Exits Chinese Market, Massive Layoffs Draw Attention
-
On November 23, according to foreign media reports, UK AI chip unicorn Graphcore has laid off most of its employees in China and halted sales in the country. This marks another major setback for AI chip companies.
The company has confirmed this decision, citing the recent escalation of U.S. export controls on China. "Regrettably, this means we will significantly scale back our operations in China," a spokesperson said in an email. The company declined to disclose the number of affected employees.
Earlier, in an effort to cut costs, Graphcore announced layoffs in September last year. According to Dealroom data, Graphcore's workforce decreased from 620 employees in October 2022 to 418 in October 2023.
Three years ago, Graphcore was one of Europe's hottest startups, seen as a potential competitor to Nvidia, with a valuation nearing $3 billion. Now, as Graphcore faces critical survival pressures, the U.S. ban on China has undoubtedly worsened its situation. After laying off Chinese employees and abandoning the Chinese market, expanding its customer base and improving profitability will become even more challenging.
Graphcore was founded in 2016, designing AI chips for data centers and cloud computing. It quickly became one of the most promising startups in the UK's tech and semiconductor industry, even earning a reputation as one of the louder "Nvidia competitors" in the AI chip race in recent years.
Graphcore claims that its Intelligence Processing Unit (IPU) is better suited for AI-specific needs than GPUs. In November 2019, Microsoft, the world's largest software company, signed an agreement to purchase Graphcore's processors. By 2020, the startup had raised $222 million in funding, reaching a valuation of $2.8 billion.
However, investor attention and patience are not limitless. For AI chip startups, the challenge of migrating researchers and engineers from Nvidia's GPU-dominated ecosystem to their own is daunting. To make matters worse, Graphcore lost a major order from its key customer, Microsoft.
The core issue for Graphcore now is profitability. Financial reports show that its 2022 revenue was $2.7 million, a 46% year-on-year decline, while losses expanded by 11% to $204.6 million. In October of this year, Graphcore disclosed the need to raise funds to sustain operations, warning of "significant uncertainty" about its ability to continue as a going concern if it fails to secure financing by May next year amid mounting losses. Since then, the company has not announced any new funding.
According to previous reports by The Sunday Times, venture capital firm Sequoia Capital, one of Graphcore's most prominent backers, has written down the value of its stake in the startup to zero. Another supporter, Baillie Gifford, reduced its investment valuation from $11.9 million in July last year to $2.8 million in July this year.
Multiple venture capital institutions and industry insiders told Business Insider that Graphcore would either sell itself or raise new funding in a deal with sale-like characteristics, and it wouldn't come close to its peak valuation.
Graphcore CEO Nigel Toon previously mentioned that China is a potential growth market, especially after trade war restrictions hindered Nvidia's ability to sell products in China. Last October, he revealed in a speech that sales from China could account for "20% to 25%" of Graphcore's business.
"Elsewhere, the demand for AI computing continues to grow, and Graphcore is working with customers worldwide to meet their needs for powerful, cost-effective GPU alternatives," said a Graphcore spokesperson.
In the UK, Graphcore has recently faced a setback. Although the UK government currently places great emphasis on advanced AI and earlier this year pledged £1 billion to develop the domestic semiconductor industry and £100 million to establish a local chip reserve, Graphcore has been excluded from the £100 million fund. This is because the bid explicitly specifies GPUs, thus excluding systems built around IPUs.
Graphcore spokesperson Mackenzie stated: "This is the realization of the warning Graphcore issued in its open letter to the UK government—that the lack of technological diversity in national AI computing infrastructure could steer users toward GPU-compatible applications and limit exploration of new models and technologies designed for AI systems."
He added that the U.S. Department of Energy's national laboratories have already incorporated IPUs as part of their infrastructure. Ironically, UK researchers can apply to use Graphcore IPUs through the U.S. Argonne National Laboratory. "We also reiterate that if the UK government truly wants to nurture a domestic AI industry, it should consider procurement as a powerful way to demonstrate such support—we hope to see this in future initiatives," Mackenzie said.
Why is the UK's AI chip unicorn struggling so hard?
According to European tech media Sifted, four former employees working in machine learning engineering, software, marketing, and recruitment reported that the company's struggles resulted from a combination of bad luck and poor management. They believe that executives made poor calls in business and technical strategies, while the AI market moved in the wrong direction, leading to low morale and talent drain.
A former machine learning engineer stated that Graphcore's troubles began when it decided to pursue major clients like Microsoft: "Graphcore has found more success targeting startups now, but it took them years to realize this."
There was also confusion about Graphcore's primary target market and products. Another former employee from the marketing department said this affected the company's ability to sell inventory, adding that changes in business strategy were not transparently communicated to the team. "One week, we heard we were chasing this market, and a month later, we were chasing another. The goals kept shifting."
Former employees also cited software issues as a key problem, which contributed to the failure of the Microsoft deal.
A former software engineer at the company told Sifted that NVIDIA's software is considered the easiest to use in the industry: "For customers, using Graphcore's software was a challenge."
"Graphcore is a hardware-centric company with a somewhat homemade approach to software," said another former employee who worked in the marketing team. They believed that Graphcore's systems would be more accessible and better served if they utilized more external software providers.
"There were many issues with maintenance. The software wasn't ready—it had numerous bugs and problems," the former employee from the machine learning team said regarding the deal with Microsoft. "Of course, things sometimes fail, but as an engineer, you need to be able to overcome those failures. It was nearly impossible—things would fail, and you wouldn't know why. There was a big push at the end, but it was already too late for Microsoft."
They found it surprising how little urgency was shown when Toon announced the loss of such a major client during a Zoom call in the spring of 2021. "He mentioned we lost Microsoft without a proper review, which was unbelievable for such an important client."
Former employees say frustration over the lack of commercial appeal and efforts to better serve customers is now leading top talent to leave. This includes Graphcore's former hardware VP Jonas Olsson and senior VP of systems Ola Tørudbakken, who are now at Meta, as well as Graphcore's former North America GM Saurabh Kulkarni, now at Intel.
"(Graphcore employees) worry about how low the revenue is. If you're an engineer, other companies and recruiters contact you every day. So people are starting to leave," said a former employee who worked in the recruitment team. However, a Graphcore spokesperson stated that the company is also hiring new employees and that turnover is expected in a "dynamic" industry like AI.
After Graphcore lost the Microsoft deal, the ongoing surge in AI computing demand has fueled the market's appetite for AI chips. In the latest generative AI and large model boom, Nvidia GPUs remain the biggest winner, especially as the only "preferred" choice for training chips.
For most AI chip startups, more opportunities lie in the inference market. German generative AI startup Aleph Alpha has established a strategic partnership with Graphcore, hoping its next-generation processors will bring more energy efficiency savings in inference.
In May 2022, Graphcore announced that its next-generation chip will be launched sometime in 2024. For this AI chip startup facing survival pressure, this may be the last chance to prove its development potential to the market and investors.