Last week, I wrote about the flood of appeals coming to the Oversight Board about the bloodshed in Israel and Gaza. Now, the board is starting to take action on those appeals, launching expedited reviews of two cases: one involving footage of kidnapped Israelis and another showing the aftermath of the strike on the al-Shifa Hospital. The reviews were announced this morning, with a final ruling expected within 30 days.
It feels odd to say it, but the most successful tech company of the year is a hardware components business. Nvidia has had an incredible year, tripling its market cap over the course of 2023. The company’s processing systems are at the heart of ChatGPT and the new wave of large language models (LLMs). As companies like Meta scramble to catch up, it means spending billions on more hardware from the same supplier. In the last reported quarter, Nvidia’s data center business made over $14 billion in sales — more than four times what it did the year before.
Understandably, CEO and co-founder Jensen Huang has taken a victory lap in the press. A few hours before Elon Musk’s disastrous appearance at The New York Times’ DealBook summit, Huang was interviewed on the same stage about the company’s success and his tempered expectations for self-reasoning artificial intelligence. A profile of Huang in The New Yorker’s AI issue last month aptly compared Nvidia to companies that sold prospecting supplies during the gold rush.
As far as the U.S. is concerned, the problem is AI development in China.
Nvidia has so much good news that some really serious bad news can pass by almost unnoticed. On the most recent earnings call, the company announced it was anticipating a problem with its data center sales to China, on the heels of a new round of export restrictions from the U.S. Department of Commerce. According to Nvidia’s internal figures, sales to China represent as much as a quarter of its data center sales, or roughly $3.6 billion. The new restrictions, placed on the Nvidia GPUs that are hardest to replicate, could reduce that business to almost nothing.
Huang took an optimistic view of the restrictions at the DealBook summit: “We have to come up with new chips that comply with the regulation,” he said. “Then we’ll go back to market.”
But there is a bit of a cat-and-mouse game here, as Nvidia responds to regulations with new chip designs and the U.S. responds to those designs with new restrictions. As far as the U.S. is concerned, the problem is AI development in China in general, rather than some particular quirk of a GPU.
So far, the U.S. officials in charge have been talking tough. “If you redesign a chip around a particular cut line that enables them to do AI, I’m going to control it the very next day,” U.S. Commerce Secretary Gina Raimondo said at a conference last week. It’s hard to imagine a chip that’s weak enough to be exportable but still powerful enough to be useful. And if Nvidia can’t thread that needle, it would mean giving up the entire $3.6 billion.
China, meanwhile, is desperately trying to prop up Huawei as an alternate source — although the company seems unlikely to be able to compete on either scale or quality. But as long as they’re operating in completely segregated markets, “compete” isn’t even quite the right word. China’s LLMs are still far behind those of the U.S. — a point that was tested and confirmed this week by the folks at ChinaTalk. Huawei will do good business supplying Chinese companies while China’s overall AI development falls further and further behind.
If you see AI as a fight between the U.S. and China (as most of the U.S. government does), then this is a win — but it leaves Nvidia in a surprisingly vulnerable place. Without continual improvements, it will be hard to sustain the current AI hype, which will make it hard to sustain the intense demand for Nvidia’s products. The history of AI research is full of these lulls — known among researchers as “AI winters.” It’s easy to imagine another winter coming soon if OpenAI and others don’t live up to their initial promise.
That winter is the biggest threat to Nvidia’s fortunes over the next few years — and it’s a threat that AI researchers all over the world are fighting together.
