Nvidia CEO Jensen Huang came out on stage Wednesday with a measured dose of confidence and prudence. He told investors that the world’s most valuable chipmaker isn’t finished yet. Indeed, Nvidia is now investing a lot in a new class of data center chips that will continue its sensational streak.
Over the years, I’ve observed that semiconductor cycles can be incredibly rapid when supply chains constrict or new players enter the mix. This is something Nvidia is obviously aware of, as evidenced by its latest forecast. Huang hyped a new chip, Vera, as the company’s bid into a $200 billion market opportunity. But he openly warned of supply constraints, too. It was refreshing to be so honest, in an age of slick earnings statements. It also highlighted the fine balance Nvidia is having to walk – a race against the whirl of demand, and a battle against overpromising.
In terms of strategy, Nvidia’s focus is on diversification of customer base, technically speaking. The firm’s flagship AI chips used to be synonymous with the training of the biggest and most advanced AI models. They are used by nearly all the major data centres worldwide. The story, though, is gradually moving towards a new concept, known as inference workloads, with the trained models being actually run to provide answers, images or predictions. This transition creates opportunities for other silicon giants such as Google, Amazon, AMD, and Intel to compete with it. The big question is if it can persuade investors that the AI construction is sustainable to 2027 and 2028, as the story moves toward inference workloads and competitors like Google, Amazon, AMD and Intel silicon, wrote eMarketer analyst Jacob Bourne.
It’s a very interesting point for me. Inference ≠ Training. It requires lower latency, more energy efficient and frequently, on specialized hardware that is not always an Nvidia GPU. Players take a shot. However, Nvidia is not resting on its laurels. Huang touted the Vera chip as a big step forward, and it’s tailored to perform the next generation of data center AI tasks. It’s too early to tell if Vera can match the jaw-dropping results of its legacy, but that’s not stopping the company from putting its money where its mouth is.
Nvidia has also increased its investment in supply in order to not cause any disruptions. The semiconductor industry has been suffering from a global memory chip shortage for months, and Nvidia is buying up inventory ahead of the anticipated shortage. It’s a sensible and competent course of action. I have seen companies being surprised by a shortage that they didn’t expect to experience in past cycles. Nvidia has seemingly taken a lesson from those errors. The company also upped its cash dividend to 25 cents from one cent per share and announced an $80 billion buyback program. Those are all textbook indicators of confidence that are meant to provide comfort to long-term investors that management can find value at current levels.
But the market’s initial response was subdued. Extended trading saw shares decline 1.6 per cent. To me that’s more about the weight of the expectation than Nvidia’s fundamentals. As Jacob Bourne pointed out, the higher a player beats, the more it costs. Investors now don’t need to wonder if Nvidia will post good numbers this year. They are inquiring if the AI craze will continue in 2027 and 2028. That’s a much tougher question! But it’s not just about the engineering capabilities of Nvidia, it’s about whether enterprises and cloud providers will keep spending billions on new models after the easy-to-harvest rewards of AI have been had.
Then there’s the competition. Google’s TPUs, Amazon’s Trainium and Inferentia, AMD’s Instinct family and Intel’s Gaudi all vie for a share of the data center market. They are all capable on their own, none is a match for Nvidia’s ecosystem in the present day, but the combination is a threat. I have witnessed such a situation in the chip business before: a kingmaker becomes complacent and the market splits. Nvidia’s management obviously wants to avoid that, and that’s why Huang continues to insist on new architectures and a wider range of customers.
I have another issue that I have been mulling on whether Nvidia’s growth is becoming too reliant on a few hyperscale customers. The company doesn’t get specific with the individual sources, but it’s not hard to tell that Microsoft, Amazon, and Google are a significant share of the data center revenue. Nvidia may feel the pinch if any of these giants choose to do more of their own silicon design. But even the most enticing chips for AI could face muted demand if the overall economy slows, and corporate IT budgets contract.
