Over the past decade, compute was mostly considered as a technical resource: something developers bought from cloud providers like AWS or Azure to run software.

But with AI, compute is becoming scarcer and more expensive as it is needed to train models and run them in production (inference). Suddenly, compute is no longer abundant enough to meet demand. And when a resource becomes scarce and strategic, it often starts to be treated like an asset that needs to be managed financially.

And this is what is currently happening with compute which is becoming a priced, tradable, financeable, and optimizable asset.

But what does it mean concretely?

What does financialization look like in practice?

Concretely, here are the different ways compute is moving from a cloud resource to a financial asset.

Price discovery:

• What it means: Compute gets a transparent market price. Instead of opaque bilateral deals, people can see what an H100 hour, B200 hour, or inference token really costs.

• Concrete example: Compute Exchange shows buyers comparable GPU prices across providers and contract types.

Standardization:

• What it means: Compute is transformed from a messy technical resource into a more comparable unit: chip type, location, memory, latency, performance, availability. Without standardization, it is hard to trade or hedge.

• Concrete example: Shadeform helps users compare GPU supply across many clouds, regions, and hardware configurations

Procurement optimization:

• What it means: Buyers do not just buy from one cloud anymore. They compare providers, switch between them, reserve capacity, or buy spot capacity depending on price and availability. Compute becomes something to actively manage.

• Concrete example: Compute Exchange helps buyers source GPU capacity through spot, reserved, and forward contracts.

Capacity arbitrage:

• What it means: A company secures scarce GPU capacity, power, or data center access, then resells compute at a higher price or with better packaging. This is the neocloud logic. They are monetizing scarcity.

• Concrete example: CoreWeave turns secured GPU infrastructure into cloud compute sold to AI companies

Hedging and derivatives:

• What it means: Compute users and compute owners use financial contracts to protect themselves against price movements (hedging). Buyers hedge against rising GPU costs. GPU owners hedge against falling rental prices. This is the “pure finance” layer.

• Concrete example: CME + Silicon Data create GPU futures linked to compute price indices.

Asset backed financing:

• What it means: GPUs and data centers become financeable assets because their future cash flows can be modeled. If you can estimate future revenue from GPU rental contracts, you can borrow money against that infrastructure.

• Concrete example: Compute Labs turns GPU infrastructure into investable products backed by future compute revenue.

Secondary markets:

• What it means: Reserved or unused compute capacity can be resold, subleased, or traded. This turns compute commitments into transferable economic assets rather than fixed private contracts.

• Concrete example: SF Compute lets companies buy, sell, or sublease GPU capacity.

As you can see from these examples this is not just a theoretical idea. A whole ecosystem of companies is starting to emerge around it. This is what we will explore in the next section and in the dedicated landscape.

What does the compute financialization ecosystem look like?

I split this emerging ecosystem into four different categories:

• Category #1: Compute capacity & supply. These are the companies building or controlling the actual compute capacity. In practice, this is mostly the neocloud layer (CoreWeave, Lambda, Nebius…). They secure GPUs, data center access, power, networking, and customer demand, then package all of this into cloud compute capacity. The important nuance is that neocloud does not always mean they own the data centers themselves. Some lease long term capacity from datacenter partners. Some build their own sites. And many do both.

• Category #2: Compute procurement, resale & routing. Here you have the companies that make compute easier to access across providers. Instead of treating compute capacity as something fixed, buyers can treat it as something they actively source. Here you find mostly marketplaces and routing layers.

• Category #3: Compute market data & benchmarks. This category is about making compute easier to measure. If compute is going to be traded or financed, people first need to know what they are looking at. Kind of “metadata providers”.

• Category #4: Compute financial products & assetization. This is the most explicitly financial part of the landscape. Here, compute becomes something businesses can hedge, finance, or turn into an investment product. Pure financial products.

What are the major trends shaping the landscape?

Compute capacity & supply

• Neoclouds are becoming much more than GPU rental companies. At first, it is tempting to describe companies like CoreWeave, Lambda, Nebius (etc.) as “GPU clouds.” But the model is becoming heavier than that. These companies need to secure GPUs, power, data center capacity, networking, and large customer commitments. The value is not only in exposing an API to GPUs. It is in being able to bring scarce resources together.

• The offer is moving from raw capacity to full AI infrastructure platforms. Many neoclouds are not stopping at renting a GPU (they started with that). They are now adding storage, inference, bare metal, managed clusters, agent sandboxes, observability, and workflow tooling. Basically going up the stack (unsurprisingly).

Compute procurement, resale & routing

• GPU procurement is becoming an active workflow. In the past, many teams would simply buy from one cloud provider and accept the available price and capacity. But with compute becoming scarce, this is no longer enough. Companies like Compute Exchange, Shadeform, SF Compute, and Vast.ai are building around the idea that buyers need more flexibility. They want to compare supply, reserve capacity, negotiate better terms, and sometimes resell what they do not use. Compute procurement starts to look more like capacity management than cloud account management.

• Routing is turning compute optimization into a software decision. OpenRouter is a good example of this at the inference layer. The user does not necessarily care which provider serves the request, as long as the answer is good enough and not too expensive. This financialization is about making cost and performance decisions dynamically, at the moment the workload runs.

Compute market data & benchmarks

• The market needs reference prices before it can mature. One of the biggest problems in compute today is that prices are hard to compare. The same GPU can cost different amounts depending on the provider, region, contract structure, and performance profile. This is why companies like Silicon Data are helping turn compute into something more transparent.

• Benchmarks are moving from model quality to real-world economics. It is no longer enough to know which model scores best on a benchmark. Buyers also want to know what it costs to run, how fast it responds or how reliable it is. Companies like Artificial Analysis or Epoch AI translate technical performance into economic questions that not only provide answers to “how good is the model?” but also “how expensive is it to run?”.

Compute financial products & assetization

• The pure finance layer is just starting to appear. The clearest example is the work between CME and Silicon Data on GPU futures. The idea is that compute buyers and compute sellers will need ways to protect themselves against price movements. If you are an AI company, rising GPU prices can hurt your margins. If you own GPU capacity, falling prices can hurt your revenue. Futures are an attempt to create a financial tool around that risk. This is still early, but it is an important signal that compute is starting to be treated like a financial market.

• GPUs are starting to be treated as cash-flowing assets. Companies like Compute Labs and GAIB are pushing this idea more directly. The logic is simple: if a GPU can generate future rental revenue, maybe that revenue can be modeled, financed, or packaged into an investment product. But we’re also still really early here.

How does the funding activity look like on this market?

There’s already a lot of investment activity in this market. Here are a couple of comments analyzing the 96 funding events listed in this landscape:

• Unsurprisingly, most of the money goes to the companies building actual compute capacity. CoreWeave, Lambda, Nebius, Nscale, Fluidstack, and TensorWave are raising very large rounds because their business is capital intensive. They need to buy GPUs, secure data center capacity, access power, and build the operating layer around it.

• Strategic investors matter a lot in this market. NVIDIA appears repeatedly across the funding events, alongside infrastructure, semiconductor, and financial investors. This makes sense because in this market, capital is not the only thing that matters. The right investors can also help companies secure supply, build trust, and reach large customers.

• Europe is present, but the compute buildout still feels more constrained than in the US. Companies like Nscale, Nebius, and Fluidstack show that Europe is not absent from the market. But the scale of financing still looks smaller and more concentrated than in the US. Building data centers in Europe is also harder because of regulation, permitting, power access, and local constraints.

• The lighter software and financial layers raise much less. The lighter software and financial layers raise much less. Companies in the three other categories do not need the same capital intensity as neoclouds. They are not building the compute capacity directly. They are building the layer that makes this capacity easier to use as a market.