|Will Nvidia’s huge bet on artificial-intelligence chips pay off?|
The unassuming chipmaking giant was early to the AI revolution—and remains ahead of rivals
Aug 1st 2021
“WE’RE ALWAYS 30 days away from going out of business,” is a mantra of Jen-Hsun Huang, co-founder of Nvidia, a semiconductor company. That may be a little hyperbolic coming from the boss of a company whose market value has increased from $31bn to $486bn in five years and which has eclipsed Intel, once the world’s mightiest chipmaker, by selling high-performance chips for gaming and artificial intelligence (AI). But only a little. As Mr Huang observes, Nvidia is surrounded by “giant companies pursuing the same giant opportunity”. To borrow a phrase from Intel’s co-founder, Andy Grove, in this fast-moving market “only the paranoid survive”.
Constant vigilance has served Nvidia well. Between 2016 and 2021 its revenues grew by 233%. In the three months to May sales expanded by a dizzying 84%, year on year, and gross margin reached 64%. Although Intel’s revenues are four times as large and the older firm fabricates chips as well as designing them, investors value Nvidia’s design-only business more highly (twice as much in terms of market capitalisation). Its hardware and accompanying software are used in all data centres that make up the computing clouds operated by Amazon, Google, Microsoft and China’s Alibaba. Nvidia’s systems have been adopted by every big information-technology (IT) firm, as well as by countless scientific research teams in fields from drug discovery to climate modelling. It has created a broad, deep “moat” that protects its competitive advantage.
Now Mr Huang wants to make it broader and deeper still. In September Nvidia confirmed rumours that it was buying Arm, a Britain-based firm that designs zippy and energy-efficient chips for most of the world’s smartphones, for $40bn. The idea is to use Arm’s design prowess to engineer central processing units (CPUs) for data centres and AI uses that would complement Nvidia’s existing strength in specialised chips known as graphics-processing units (GPUs). Given the global reach of Arm and Nvidia, regulators in America, Britain, China and the European Union must all approve the deal. If they do—a considerable “if”, given both firms’ market power in their respective domains—Nvidia’s position in one of computing’s hottest fields would look near-unassailable.
Mr Huang, whose family immigrated to America from Taiwan when he was a child, founded Nvidia in 1993. For its first 20 years or so the company made GPUs that made video games look lifelike. In the past decade, however, it turned out that GPUs also excel in another futuristic, but less frivolous, area of computing: they dramatically speed up how fast machine-learning algorithms can be trained to perform tasks by feeding them oodles of data. Four years ago Mr Huang, who goes by Jensen, startled Wall Street with a blunt assessment of his company’s prospects in what has become known as accelerated computing. It could “work out great”, he said, “or terribly”. Regardless, the company was “all in”.
Around half of Nvidia’s annual revenues of $17bn still comes from gaming chips. They have also proved excellent at solving the mathematical puzzles that underpin ethereum, a popular cryptocurrency. This has at times injected crypto-like volatility to GPU sales, which contributed to a near-50% fall in Nvidia’s share price in late 2018. Another slug of sales comes from selling chips that accelerate features other than graphics or AI to computer-makers and car companies.
But the AI business is growing fast. It includes specialised chips as well as advanced software that lets programmers fine-tune them—itself enabled by an earlier bet by Mr Huang, which some investors criticised at the time as an expensive distraction. In 2004 Mr Huang started investing in “Cuda”, a base software layer that enables just such fine-tuning, and implanting it in all of Nvidia’s chips.
A lot of these systems end up in servers, the powerful computers that undergird data centres’ processing oomph. Sales to data centres have increased from 25% of total revenues in early 2019 to 36%, contributing nearly as much as to the total as gaming GPUs. As companies across various industries adopt AI, the share of Nvidia’s data-centre sales going to big cloud providers such as Amazon and Google has declined from 100% to half that.
Today its AI hardware-software combo is designed to work seamlessly with the machine-learning algorithms collected in libraries such as TensorFlow (which is maintained by Google) and PyTorch (run by Facebook), boosting the algorithms’ number-crunching power. Nvidia has created programs to hook its hardware and software up to the IT systems of big business customers with AI projects of their own. All this makes AI developers’ job immeasurably easier, says a former Nvidia executive. Nvidia is also expanding into AI “inference”: running AI models, hitherto the preserve of CPUs, rather than merely training them. Real-time, huge AI models like those used for speech recognition or content-recommendation systems increasingly need the specialised GPUs to perform well, says Ian Buck, head of Nvidia’s accelerated-computing business.
This is also where Arm comes in. Owning it would give Nvidia the CPU chops to complement its historic strength in GPUs and more recently acquired abilities in network-interface cards needed to run server farms (in 2019 Nvidia acquired Mellanox, a specialist in such interconnecting technology). In April the company unveiled plans for its first data-centre CPU, Grace, a high-performance chip based on an Arm design. Arm’s energy-efficient chips would help Nvidia supply AI products for “edge computing”—in self-driving cars, factory robots and other places away from data centres, where power-hungry GPUs may not be ideal.
Transistors in microprocessors are already the size of a few atoms, so have little room to shrink and tricks such as outsourcing computing to the cloud, or using software to split a physical computer into several virtual machines, may run their course. So businesses are expected to turn to accelerated computing as a way to gain processing power without spending through the roof on ever more CPUs. Over the next five to ten years, as AI becomes more common, up to half of the $80bn-90bn that is spent annually on servers could shift to Nvidia’s accelerated-computing model, estimates Stacy Rasgon of Bernstein, a broker. Of that, half could go on accelerated chips, a market which Nvidia’s GPUs dominate, he says. Nvidia thinks the global market for accelerated computing, including data centres and the edge, will be more than $100bn a year.
Nvidia is not the only one to have spotted the opportunity. Competitors are proliferating, from startups to other chipmakers and the tech giants. Venture capitalists have backed companies such as Tenstorrent, Untether AI, Cerebras and Groq, all of which are trying to make semiconductors even better suited to AI than Nvidia’s GPUs, which for all their virtues can be power-hungry and fiddly to program. Graphcore, a British firm, is touting its “intelligence-processing unit”.
In 2019 Intel bought an Israeli AI-chip startup called Habana Labs and ceased work on the neural-network processors it had acquired as part of an earlier purchase of Nervana Systems, another startup. Amazon Web Services (AWS), the e-commerce giant’s cloud division, will soon start offering Habana’s Gaudi accelerators to its cloud customers, claiming that the Gaudi chips, which are slower than Nvidia’s GPUs, are nevertheless 40% cheaper relative to performance. Advanced Micro Devices (AMD), a veteran chipmaker that is Nvidia’s main rival in the gaming market and Intel’s in the CPU business, is in the process of finalising a $35bn deal to acquire Xilinx, which makes another kind of accelerator chip called field programmable gate arrays (FPGAs).
A bigger threat comes from Nvidia’s biggest customers. The cloud giants are all designing their own custom silicon. Google was the first to come up with its “tensor-processing unit”. Microsoft’s Azure cloud division opted for FPGAs. Baidu, China’s search giant, has its “Kunlun” chips for AI and Alibaba, its e-commerce titan, has Hanguang 800. AWS already has a chip designed for inference, called Inferentia, and has one coming for training. “The risk is that in ten years’ time AWS will offer a cheap AI box with all AWS-made components,” says the former Nvidia executive. Mark Lipacis at Jefferies, an investment bank, notes that since mid-2020 AWS has put Inferentia into an ever-greater share of its offering to customers, potentially at the expense of Nvidia.
As for the Arm acquisition, it is far from a done deal. Arm’s customers include all of the world’s chipmakers as well as AWS and Apple, which uses Arm chips in its iPhones. Some have complained that Nvidia could restrict access to the chip designer’s blueprints. The Graviton2, AWS’s tailor-made server chip, is based on an Arm design. Nvidia says it has no plans to change Arm’s business model. Western regulators are due to decide on whether to approve the deal with Britain’s competition authority, which had until July 30th to scrutinise the transaction and is expected to be among the first to do so. China, for its part, is unlikely to welcome an American takeover of an important supplier to its own tech firms, which is currently owned by SoftBank, a Japanese technology conglomerate.
Even if one of the antitrust watchdogs puts paid to the acquisition, however, Nvidia’s prospects look bright. Venture capitalists have become markedly less enthusiastic over time about backing startups taking on Nvidia and the tech giants investing in accelerated computing, says Paul Teich of Equinix, an American data-centre operator. Intel has overpromised many things, including accelerated computing, for years, and mostly undelivered. AWS and the rest of big tech have plenty of other things on their plates and lack Nvidia’s clear focus on accelerated computing. Nvidia says that, measured by actual utilisation by businesses, it has not ceded market share to AWS’s Inferentia.
Mr Huang says that it is the expense of training and running AI applications that matters, not the cost of hardware components. On that measure, he says, “we are unrivalled on price-for-performance.” None of Nvidia’s rivals possess its software ecosystem. And it has a proven ability to switch gears and capitalise on good luck. “They’re always looking around at what’s out there,” enthuses another former executive. And with an entrenched position, Mr Lipacis says, it also benefits from inertia.
Investors have not forgotten the near-halving of Nvidia’s share price in 2018. It may still be partly tied to the fortunes of the crypto market. Holding Nvidia stock requires a strong stomach, says Mr Rasgon of Bernstein. Nvidia may present itself as a pillar of the industry, but it remains an aggressive, founder-led firm that behaves like a startup. Sprinkle in some paranoia, and it will be hard to disrupt.