Comment by stego-tech

Add in the fact companies seriously invested in AI (and like workloads typically reliant on GPUs) are also investing more into bespoke accelerators, and the math for nVidia looks particularly grim. Google’s TPUs set them apart from the competition, as does Apple’s NPU; it’s reasonable to assume firms like Anthropic or OpenAI are also investigating or investing into similar hardware accelerators. After all, it’s easier to lock-in customers if your models cannot run on “standard” kit like GPUs and servers, even if it’s also incredibly wasteful.

The math looks bad regardless of which way the industry goes, too. A successful AI industry has a vested interest in bespoke hardware to build better models, faster. A stalled AI industry would want custom hardware to bring down costs and reduce external reliance on competitors. A failed AI industry needs no GPUs at all, and an inference-focused industry definitely wants custom hardware, not general-purpose GPUs.

So nVidia is capitalizing on a bubble, which you could argue is the right move under such market conditions. The problem is that they’re also alienating their core customer base (smaller datacenters, HPC, gaming market) in the present, which will impact future growth. Their GPUs are scarce and overpriced relative to performance, which itself has remained a near-direct function of increased power input rather than efficiency or meaningful improvements. Their software solutions - DLSS frame-generation, ray reconstruction, etc - are locked to their cards, but competitors can and have made equivalent-performing solutions of their own with varying degrees of success. This means it’s no longer necessary to have an nVidia GPU to, say, crunch scientific workloads or render UHD game experiences, which in turn means we can utilize cheaper hardware for similar results. Rubbing salt in the wound, they’re making cards even more expensive by unbundling memory and clamping down on AIB designs. Their competition - Intel and AMD primarily - are happily enjoying the scarcity of nVidia cards and reaping the fiscal rewards, however meager they are compared to AI at present. AMD in particular is sitting pretty, powering four of the five present-gen consoles, the Steam Deck (and copycats), and the Steam Machine, not to mention outfits like Framework; if you need a smol but capable boxen on the (relative) cheap, what used to be nVidia + ARM is now just AMD (and soon, Intel, if they can stick the landing with their new iGPUs).

The business fundamentals paint a picture of cannibalizing one’s evergreen customers in favor of repeated fads (crypto and AI), and years of doing so has left those customer markets devastated and bitter at nVidia’s antics. Short of a new series of GPUs with immense performance gains at lower price and power points with availability to meet demand, my personal read is that this is merely Jenson Huang’s explosive send-off before handing the bag over to some new sap (and shareholders) once the party inevitably ends, one way or another.