Unlock the Editor’s Digest for free
Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.
Dear reader,
Before Jensen Huang stepped on stage at Nvidia’s huge developer conference, commentators had taken to calling the event Woodstock for artificial intelligence. Huang clearly had something else in mind. With flashing lights, references to Taylor Swift and friendship bracelets at the after-party, this was an Eras Tour for the semiconductor crowd.
The US’s third most valuable company drew an enormous audience to San Jose this week, all eager to hear what’s next for generative artificial intelligence chips. Of course, even with more than 10,000 people cheering, it’s not easy to turn a tech conference into a rousing cultural event. Harder still when you are a chip design company. Tesla can throw metal at car windows and Apple puts on slick videos. But Nvidia’s two-hour keynote presentation mostly consisted of Huang, in his trademark leather jacket, talking in front of slides. There was, as he warned at the start, a lot of mathematics.
The focus of Nvidia’s GTC (as in GPU Technology Conference) was the unveiling of the next generation of AI chips. Blackwell chips, led by the B200, have been made on TSMC’s advanced 4-nanometre process and are designed to be more powerful and efficient than the Hopper chips that made Nvidia into a trillion-dollar company. The expectation is that this should enable developers to make even more powerful AI tools.
Nvidia claims that every major cloud provider is making plans to offer the chips when they become available later this year. They are not wildly more expensive. One analyst suggested $50,000 per chip, but Huang said in an interview that they would be in a similar range to Hopper chips. This will make it more difficult for rivals to compete.
The new Blackwell architecture encompasses a range of products that target both small corporate customers and large data centres. One of the most anticipated is the GB200 “superchip” — made from two GPUs and a CPU.
Other announcements included Nvidia Inference Microservices, aka Nim (Nvidia product names are varied and hard to keep track of). These are pre-packaged generative AI models on Nvidia’s Cuda software.
There was also Quantum Cloud, a service that lets researchers try out quantum computing software and expansion of the omniverse — a virtual second Earth on which companies can model possible business changes and test out how they might work in the real world. Plus, there were stronger partnerships with Microsoft, Google and Amazon Web Services. These are notable given all three companies are also hoping to compete in AI chips themselves.
Nvidia knows that billions of dollars are being directed towards the market it dominates. But those who say it cannot maintain its lead should remember that after the company was created in 1993 its early graphics chips were quickly overtaken by rivals. The company managed to reorganise in a way that allowed it to launch new products faster than its peers and build software alongside its hardware. That created an ecosystem that gave customers a reason to stick with the company. It has experience keeping ahead of hungry competitors.
Unfortunately for many of us who made the trip out to sunny San Jose, what Nvidia did not do very well was explain where AI would go from here. It is a question that trillions of dollars of market value rests on. That probably explains why Nvidia’s share price has barely moved this week — despite the plethora of new products on show.
Further reading
Lex does not believe that competitors have any hope of catching up to Nvidia any time soon. Earnings in February illustrated how far the company has extended its lead.
Back in January we looked at Dutch chipmaking equipment manufacturer ASML for clues about a new chip cycle. The capacity its customers think they will need in the next two years exceeds some analysts’ forecasts. That indicates a cyclical upswing in the chip sector.
This week in Lex
US credit card delinquencies are rising as Americans run through their savings. Regulators want to cap late fees. Small, standalone credit card companies would be most affected.
Sadly for Unilever, there does not appear to be a hidden gem in its jumble of businesses. Scooping out ice cream will extract a division that has few synergies with the rest but it is unlikely to unlock value for shareholders.
Thames Water may be struggling beneath the weight of an $18.3bn debt load but if it can find new investors and focus on cash preservation it can still avoid nationalisation.
Things I’ve enjoyed this week
The elevation of “algorithms” as a mystical force against which we are powerless is something tech companies do little to debunk. But it’s not true. Henry Farrell, professor at Johns Hopkins, explains why we are not all programmable zombies.
The New Yorker has found that many US chains, including McDonald’s, are being supplied with seafood from Chinese companies that use forced labour from North Korea. Something to think about when ordering your next Filet-O-Fish meal.
Enjoy your weekend.
Elaine Moore
Deputy head of Lex
Read the full article here