Who is buying & selling AI Compute?

From tech heads to investors to politicians, everyone is talking about Data Centers. What these conversations are really about is calibrating the AI buildout: are we investing too much, or too little? And what’s at stake if we get it wrong? To answer those questions, we need to take a closer look at the commodity at the center of the AI supply chain: Compute.

Compute is the resource that sits between the physical data centers and the AI applications they power. Supplying too little of it could constrain growth, and supplying too much risks leaving expensive infrastructure unused.

So who is actually buying compute, and how much of it do they need? We analyzed Ramp transaction data for an initial read on these questions, and our findings reveal a highly complex, concentrated and growing market:

• Small buyer base: Only 9.8% of Ramp’s SaaS-adopting customers paid a GPU compute vendor in the past year, compared with 78.4% of customers who paid to use an actual AI application, like OpenAI or Anthropic.

• Outsized spending: Despite its much smaller customer base, spend on GPU vendors reached nearly 80% of spend on AI applications over the same period, emphasizing just how costly compute infrastructure is.

• Crowded field of sellers: That spending spanned dozens of providers competing for the same small set of buyers. If demand keeps climbing, a field that crowded can sustain itself. If it doesn't, a lot of those providers may not survive.

• Complex buying structures: Compute is sold in many different forms, pricing structures and term commitments, even within the same provider. That makes the market challenging for buyers to navigate, and difficult for labs like us to get a clear picture of demand.

Together, these findings point to a market we need to watch closely and understand better as the AI buildout accelerates. Below, we explain how this market works, what our data shows so far, and what we’ll track next to assess whether investment is keeping pace with demand.

What is Compute?

Compute is the processing power that software runs on. CPU, memory and storage are all typical examples of compute. For AI applications, the most important compute resource comes from graphics processing units or GPUs. GPUs are the physical chips that silicon vendors like Nvidia sell and data centers house. The compute generated by this hardware can perform many calculations at once and is used by AI companies to train models and run inference.

Investment in the land, power and other physical infrastructure to run these GPUs represents the supply side of the AI economic equation. Tracking GPU buyer trends will tell us the demand.

Who is buying compute?

Any company that trains and operates its own AI model buys compute. The spend of these businesses is the direct indicator of GPU demand. That includes:

• LLM giants like Anthropic and OpenAI

• Horizontal AI applications like Voice and Image generators

• Vertical AI Applications building industry-specific models for legal, finance and other business domains

Among all Ramp’s SaaS-adopting customers, that’s a highly concentrated population: Only 9.8% of businesses paid a GPU compute provider in the past year, compared with 78.4% paying to use an AI application, like OpenAI and Anthropic. Both rates continue to rise as increased AI usage drives demand for the computing power to support it.

Despite its much smaller customer base, GPU vendor spend reached nearly 80% of all AI Application user spend in the year through August. Compute is just one of the costs of building and operating an AI application, so this underscores the scale of investment going into delivering AI products. The economic question is whether those products can generate enough revenue to cover that investment, their other costs, and a profit.

Who is selling compute?

The GPU compute market is highly fragmented, with providers packaging the same underlying resource into various pricing structures and bundling it with different levels of infrastructure. We’ve used transactions, vendor catalogs and purchasing patterns to break the GPU vendors out into 4 key groups:

• Hyperscalers, like AWS, Microsoft Azure, and Google Cloud, sell GPU capacity alongside their broader cloud services.

• Neoclouds specialize in GPU infrastructure. Our definition also includes marketplaces that connect buyers with GPU capacity from multiple providers.

• Wholesale GPU capacity providers sell larger blocks of capacity, often through dedicated contracts.

• Model Serving & Inference providers handle the infrastructure and run models for customers.

Among our customers, GPU compute is most commonly purchased via Model Serving & Inference with 8.7% adoption, followed by Neoclouds at 3.3% and Wholesale GPU Capacity at less than 1%. Hyperscalers were excluded from the adoption breakdown, since their GPU compute is often bundled with their larger offerings and not reliably extractable.

How are they selling it?

Even within each provider category, customers buy compute in several different ways:

• By size: Rent a single GPU, an instance, a node, or a larger cluster.

• By model: Choose among GPU models with different performance and memory capabilities.

• By commitment: Pay for capacity on demand or reserve it for a specified period.

These choices make the market difficult for buyers to navigate. Customers need to match hardware, capacity, and contract length to workloads that may be hard to predict.

That same complexity makes it difficult for labs like us to track actual demand. Cost per GPU-hour provides a common basis for comparison, but getting there requires normalizing across dozens of different purchasing structures.

How can we tell whether supply is meeting demand?

This initial analysis maps the market, but it doesn’t establish whether GPU compute is scarce or whether future demand will justify today’s infrastructure investment. To get closer to those answers, we’re developing metrics and indices that track:

• Cost per GPU-hour: How much customers pay for the same GPU model over time, accounting for differences in purchasing terms.

• GPU model mix: Which GPUs customers buy access to, and how that mix changes over time.