THE ECONOMIST: Nvidia is in danger of losing its monopoly-like margins. But don’t count it out yet

It is common lore in Silicon Valley that no company has a bigger moat than Nvidia, the world’s dominant supplier of chips for artificial intelligence.

That was true until January 27, when a Chinese firm called DeepSeek, bearing an AI model that it said cost less than $US6 million ($10m) to train, blew a nearly $US600 billion hole in the value of the semiconductor giant, marking the biggest one-day loss in the history of America’s stock market.

Think of it as the 21st-century equivalent of a shot from a trebuchet, the medieval contraption capable of reducing castle walls to rubble. Have Nvidia’s defences finally been breached?

The company has been no stranger to stock market gyrations, either upwards or downwards. Over more than two years of AI-related frenzy its market value not only soared above $US3 trillion, but also suffered eight of the 10 biggest single-day falls in stock market history.

That reflects a tug-of-war between two types of investors; those who believe Jensen Huang, Nvidia’s leather-jacketed rock star boss, is a seer and can sustain his firm’s monopoly-like profits indefinitely; and those who think the high prices of his firm’s products and shares are emblematic of AI hype, and are set to fall.

The sceptics currently have the advantage. DeepSeek’s ability to produce AI models almost as capable as America’s best ones, using what look to be fewer and less-advanced graphics processing units, challenges the belief that winning the AI race requires pouring gazillions of dollars into chips.

Microsoft, Meta, Elon Musk’s xAI, and a joint venture known as Stargate that includes OpenAI, maker of ChatGPT, have all announced blockbuster plans in recent weeks to build data centres stuffed with GPUs. But, according to an insider at one big Nvidia customer, the lesson being taken from DeepSeek is that, when it comes to training and running models, it may be possible to spend much less on processing power.

That would be bad news for Nvidia. Its most advanced chips generate gross margins of 90 per cent-plus, according to Jeffrey Emanuel, an investor who published a prescient bear case for Nvidia two days before the market rout.

To make matters worse, competition from rival chipmakers has been stiffening. Nvidia’s GPUs are almost indispensable for the training of large language models. In inference — the querying of those models — it still has a lead, but plenty more rivals. Startups such as Groq are designing chips specialised for ultra-fast inference. Nvidia’s biggest customers, the cloud giants Amazon, Google and Microsoft, are also making custom chips to reduce reliance on its GPUs.

“The markets hate the single-source supplier constraint,” says an executive at one cloud provider. Moreover, the latest “reasoning” models from DeepSeek and OpenAI derive their superior performance not from training but from more time spent on inference (known as test-time compute), which could tilt the market in favour of rival chipmakers.

Then there is the hype question. Since ChatGPT set off the AI gold rush in late 2022, Nvidia has been the ultimate “picks and shovel” play. But like investment in the early days of the internet, the AI boom has so far been based more on the belief that it will change everything than hard evidence that it can generate returns.

Plenty of investors remain bullish about AI’s chip champion, though. Some wonder whether DeepSeek has exaggerated its thriftiness in order to give its native country bragging rights over America, though others say its claims could well be plausible. Even if DeepSeek did rely less on powerful GPUs, there is speculation that it simply trained its models by studying the results of American ones built with the whizziest chips.

DeepSeek’s success may also goad American rivals to double down on their AI investments, at least for a while. OpenAI, which gobbles up GPUs to push the frontier of AI development, will respond to DeepSeek with “much better models”, its CEO, Sam Altman, swiftly promised.

As for competition, Nvidia bulls acknowledge that the threat from rival chipmakers is real, especially in inference. But its leadership is reinforced by its CUDA software, which is widely used in AI, and the networking gear that lets it string its chips together for maximum efficiency in data centres. “Its moat is very wide. It’s a system company, not a chip company,” says Umesh Padval of Thomvest, an investment firm.

All told, Nvidia’s market share and its margins are bound to fall, and perhaps sooner than investors had thought before DeepSeek came on the scene.

But there is still plenty of scope for it to remain a dominant player, which might explain why its share price bounced back by more than 8 per cent on January 28. And a wildcard remains: whether Mr Huang is truly the visionary his admirers believe him to be.

According to Daniel Newman of Futurum, a research firm, Nvidia’s boss is a leader who always tries to be one step ahead of the competition. Already he is looking beyond lLMs, which are in danger of being commoditised, recently laying out plans to develop what he called “physical AI”.

By that he meant providing the infrastructure for foundational models that understand the real world, in order to make humanlike robots, self-driving cars and the like.

Nvidia may be under siege, but it is too early to count it out.