A start-up called Inflection AI has decided to build what may be, if completed, the world’s largest supercomputer. For this they plan to use no less than 22,000 NVIDIA H100. This is NVIDIA’s most powerful graphics card to date. This makes it clear that companies are taking full advantage of the artificial intelligence boom.
The construction of supercomputers is not something that was born overnight, but rather they have existed since the dawn of computing. We can say the same for those who use graphics cards as part of their processing power, since the GPGPU concept has already existed since the second half of the 2000s. The difference compared to yesteryear? The optimization of graphics chips for AI in recent years.
Inflection AI: the supercomputer with 22,000 NVIDIA H100
Inflection AI is a startup company, creator of the Inflection-1 artificial intelligence modelwhich is used in Pi Chatbot. A ChatGPT-style application with less capacity than this, but which is more oriented to be used as a personal assistant. Well, they have been the ones who have announced the creation of one of the largest supercomputers focused on AI. and not for the 700 nodes based on Intel Xeon CPUs, but for the 22,000 NVIDIA H100 GPUsfor which you will need 31 Megawatts of power to run. Such a number of graphics cards makes us quite skeptical, mostly due to the fact that there is a shortage due to the high demand for this NVIDIA hardware.
The cost of this Inflection AI supercomputer will be 1.5 billion dollars in totalHowever, this investment is necessary, given that in order to execute increasingly advanced AI models, the creation of very high-power systems is necessary. Without forgetting that the business model of these companies is to offer their services through cloud computing.
Will this affect the supply of the RTX 40 for PC?
No, we have not misplaced another piece of news. We have to keep in mind that the benefits of an H100 for NVIDIA are much higher than that of an AD102, GPU of the RTX 4090. So Jensen Huang and his team are more interested in selling their graphics card for high-performance computing for the simple fact of having upper margins. In other words, the same wafer can easily generate 20 times more income, allocating it to create chips for AI than for gaming.
This wouldn’t be a problem if it weren’t for both the H100 and all RTX 40s are built under the same TSMC 4nm node. This means that the chips leave the same factory and even more important, the same wafer. And of course, that suddenly NVIDIA is asked for 22,000 H100s to build the Inflection AI supercomputer, well, it is worrying, and even more so, if we take into account that the demand does not come from a single company, but from several at the same time.
We all know what happens when a more lucrative market appears than the main one for a product. What’s more, this was experienced recently with mining fashion. The less lucrative gaming market is taking a backseat thanks to the boom in AI supercomputers. While there could be talk of future shortages, the poor sales of the RTX 40 only make it clear that NVIDIA will only find big revenue in AI.