Today, Nvidia’s revenues are dominated by hardware sales. But when the AI bubble inevitably pops, the GPU giant will become the single most important software company in the world.

Since ChatGPT kicked off the AI arms race in late 2022, Nvidia has shipped millions of GPUs predominantly for use in AI training and inference.

That’s a lot of chips that are going to be left idle when the music stops and the finance bros come to the sickening realization that using a fast-depreciating asset as collateral for multi-billion dollar loans wasn’t such a great idea after all.

However, anyone suggesting those GPUs will be rendered worthless when the dust settles is naive.

GPUs may be synonymous with AI by this point, but they’re much more versatile than that. As a reminder, GPU stands for graphics processing unit. These chips were originally designed to speed up video game rendering, which, by the late ‘90s, was quickly becoming too computationally intensive for the single-threaded CPUs of the time.

As it turns out, the same thing that made GPUs great at pushing pixels also made them particularly well suited for other parallel workloads — you know, like simulating the physics of a hydrogen bomb going critical. Many of Nvidia’s most powerful accelerators — chips like the H200 or GB300 — have long since ditched the graphics pipeline to make room for more vector and matrix math accelerators required in HPC and AI.

If an app can be parallelized, there’s a good chance it’ll benefit from GPU acceleration — if you have the software to do it. This is why there are so few GPU companies. A GPU needs to be broadly programmable; an AI ASIC only needs to do inference or training well.

CUDA-X many reasons to buy a GPU

Since introducing CUDA, its low level GPU programming environment and API interface, in 2007, Nvidia has built hundreds of software libraries, frameworks, and micro-services to accelerate any and every workload it can think to.

The libraries, collectively marketed under the CUDA-X banner, cover everything from computational fluid dynamics and electronic design automation to drug discovery, computational lithography, material design, and even quantum computing. The company also has frameworks for visualizing digital twins and robotics.

For now, AI has turned out to be the most lucrative of these, but when the hype train runs out of steam, there’s still plenty that can be done with the hardware.

For example, Nvidia built cuDF and integrated it into the popular RAPIDS data science and analytics framework to accelerate SQL databases or Pandas, attaining a 150x speed up in the process. It’s no wonder database giant Oracle is so keen on Nvidia’s hardware. Any compute it can’t rent out to OpenAI for a profit, it can use to accelerate its database and analytics platforms.

Nvidia doesn’t offer a complete solution, and that’s by design. Some of its libraries are open sourced, while others are made available as more comprehensive frameworks and micro-services. These form the building blocks by which software developers can use to accelerate their workloads, with a growing number of them being tied back to revenue-generating licensing schemes.

The only problem: up to this point, these benefits required buying or leasing a pricy GPU and then integrating these frameworks into your code base or waiting for an independent software vendor (ISV) to do it for you.

But when the bubble bursts and pricing on GPUs drops through the floor, anyone that can find a use for these stranded assets stands to make a fortune. Nvidia has already built the software necessary to do it — the ISVs just need to integrate and sell it.

In this context, Nvidia’s steady transition from building low-level software libraries aimed at developers to selling enterprise-focused micro-services starts to make a lot of sense. The lower the barrier to adoption, the easier it is to sell hardware and the subscriptions that go with it.

It appears that Nvidia may even open this software stack to a broader ecosystem of hardware vendors. GPUzilla has begun transitioning to a disaggregated architecture that breaks up workloads and offloads them to third-party silicon.

This week, Nvidia completed a $5 billion investment in Intel. The x86 giant is currently developing a prefill accelerator to speed up prompt processing for large language model inference. Meanwhile, Nvidia signed a deal last week to aqui-hire rival chip vendor Groq that it — though it remains to be seen how the GPU slinger intends to integrate the company’s tech long term.

In addition to its home-grown software platforms, Nvidia has made several strategic software acquisitions over the past few years, acquiring Run:AI’s Kubernetes-based GPU orchestration and Deci AI’s model optimization platforms in 2024. Earlier this month, Nvidia added SchedMD’s Slurm workload management platform, which is widely deployed across AMD, Nvidia, and Intel-based clusters for HPC and AI workloads, ensuring a profit even if you don’t buy its hardware.

• Nvidia spends $5B on Intel bailout, instantly gets $2.5B richer
• AI faces closing time at the cash buffet
• Nvidia wasting no time to flog H200s in China
• Nvidia pledges more openness as it slurps up Slurm

GenAI is here to stay

To be clear, generative AI as we know it today isn’t going away. The cash that’s fueled AI development over the past three years may evaporate, but the underlying technology, imperfect as it is, is still valuable enough that enterprises will keep using it.

Rather than chasing the mirage that is artificial general intelligence, applications of the tech will be far more mundane.

In fact, many of Nvidia’s more comprehensive micro-services make extensive use of domain-specific AI models for things like weather forecasting or physics simulation.

When the dot-com bubble burst, people didn’t stop building web services or buying switches and routers. This time around, people aren’t going to stop consuming AI services either. They’ll just be one of several reasons to buy GPUs. ®