Adam Sioud

What

In autumn 2025, the idea was that there was still room to enter the compute markets early. OUDAU could capture a first-mover advantage and become a neutral allocation layer where brokers create liquidity.

I have always been interested in markets. When I looked at AI, I was drawn to ideas that combined the game theory of markets with what was happening in AI. I wanted to build something grounded in the real economy and capture value somewhere along the path from the infrastructure being built to the models that we, as end users, use. I subscribe to the idea that the model is the product, so I began asking what kinds of companies would exist in that world and how I could position my own company for it.

Research and post-training need compute. If you were not going to be a model lab or do post-training yourself, one directional bet was to support that world through its need for more compute. Many different companies would be training models and building products, and beneath all of them is the journey from electricity all the way to the token.

Finding any useful position in that journey could be lucrative because, once you are part of that flow, you can add services or other products around it. CoreWeave acquiring OpenPipe is a much larger version of that playbook: start with compute, then add more of what customers want around it. Allocation was one way into that layer, and the broader bet was on being somewhere in the compute stack while it was still being built.

In my opinion, if you could be an early adopter, move into the compute vertical, and get somewhere on the route from data centers to labs, you should try. Back then, I saw an opportunity in Europe, which was behind and still is now. Even if Europe did not train its own models at all, inference and agents would still need compute served here. If you could enter that layer as a broker or in some other way, there was money to be made. I remember Mistral starting its compute offering back then, and I think today we can see how big that bet is going to become. Maybe even their only bet ;)

So I ended up with this, OUDAU. It took inspiration from my master's thesis, especially how traders and a market exchange bring liquidity to a market, which is needed to make it useful and help it function better, as adoption is hard to bootstrap in a new market.

As a larger idea, the compute marketplace / exchange was something that would be iterated on and built over time. The thinking was that you also needed tools or offerings that gave more immediate value and could slowly onboard users onto the platform, making it easier to transition into a full marketplace / exchange at some point.

Cost visibility was one example. You could have tools that brought AWS billing, GPU spending, token billing, and other data points into a dashboard, giving you a broader idea of where your expenditure went. At the time, much of the FinOps work at companies focused on AWS billing. I thought those FinOps teams would also begin managing more of the spending on LLMs over time. That was part of the longer-term vision.

The Thesis

The framing of the allocation desk came from Dan Shipper's allocation economy: if models can increasingly do knowledge work, the scarce skill shifts from doing the work to directing it. The advantage becomes how well you allocate resources, not what you know.

At the most abstract level, OUDAU was meant to enable that allocation economy. It would be an exchange where allocators could route compute across providers and decide what should run where, when, and under what constraints.

I guess the idea was that brokering could take on much of the burden in an exchange environment. An exchange, marketplace, or futures market needs allocators and traders who actually trade it, including desks that may not already be connected to the underlying asset. But they need a reason to participate and a way to make money. That could come from spreads, hedging, insurance-like protection, or simply doing a better allocation job. You can build the marketplace, but you still need participants, liquidity, and movement for it to become a market.

Commodities teach one lesson: the financial layer ends up bigger than the physical market. Compute is entering that wave, where liquidity, futures, and hedging matter as much as training AI models. OUDAU is the neutral allocation layer built for this wave.

Compute is often compared to oil, but it's not a storable stock. It's temporal, priced in GPU-hours, consumed over time. Unused capacity is gone. It looks more like electricity than oil, one might say.

One estimate puts AI infrastructure spending at more than $7 trillion over the next 10 years, with Compute as a Service growing from around $30 billion in 2025 to more than $250 billion by 2034.

A comparable projection argues that hyperscalers, cloud providers, and Compute as a Service platforms that make the best use of capacity, capital, and energy will gain an edge during the estimated $6.7 trillion buildout.

As with the oil industry a century ago, compute is becoming a tradable commodity. Analysts expect a massive opportunity for new exchanges, hedging products, and specialist trading desks to emerge and capture that edge.

And this is exactly where things are moving: exchanges and brokers are arriving.

These early movers already prove that transparent pricing and smart routing between builders and providers create immediate value.

These numbers were useful for showing that there was a large market here. But, as far as I know, they were not what first made me interested in compute and making a marketplace.

Looking back, some of that interest may date back to George Hotz and his ideas around the decentralization of compute. But in general, just the George Hotz ethos, his main point was that you should not have to rent a machine; you should be able to submit a job and let the system find the compute. That system in my mind could be OUDAU. The stream I watched was George Hotz | Just Chatting | how to actually win? | tinygrad.org.

I think my interest in DRW and Cumberland's work on Canton is what eventually led me to Simeon Bochev's Trade GPU compute like you trade oil. Canton interested me because it was built for private financial workflows and could settle the asset and payment sides of a trade together, something I had researched extensively in my master's thesis. I mapped that idea onto compute. One version of the OUDAU argument was that the exchange could be built on Canton, with compute as the thing traded on it. Canton could be the exchange infrastructure, and compute could be what was traded on it.

In my mind, I thought that was how Compute Exchange was working. After a while, my sense was that they were not using Canton for the exchange itself, perhaps only in the long, long future. But that was how I connected the dots myself. The talk made the idea of compute as a market and tradable commodity very exciting. I also remember watching “AI is killing innovation. That has to change”, also by Simeon Bochev, CEO of Compute Exchange. It was cool and motivating.

Tinkering

With OUDAU, I kept returning to a few ideas around allocation and where compute sits in that picture.

A useful analogy is a terminal. Its value is not that it shows you numbers, but that it brings the information, controls, and feedback needed to make a decision and act on it into one place.

OUDAU proposed an allocation layer, a terminal-like interface that joins execution, evaluation and funding. Compute was the starting point because everything routes through it. But the broader idea was allocation across AI resources over time: compute, data, models, capital, talent, and the environments that generate feedback.

The scarce skill may become routing compute, capital and talent toward the right problems, not producing more undifferentiated output. The allocation economy does not replace the knowledge economy; it amplifies it. It shifts which kinds of knowledge work get rewarded: fewer rewards for output for its own sake, more rewards for directing effort toward work that creates real progress. Value pulls toward new discoveries and verifiable results, not toward repeating the same output.

This is not only a software question. Compute has to run somewhere, and that begins with electricity, grids, land, and financing.

Europe accounts for only 4–5% of global AI computing capacity, and there is discussion about the need to increase that share. It already has national supercomputing centers and EuroHPC clusters, but fewer frontier labs and less specialist talent to run large training jobs itself. You could therefore argue that Europe does not need the same data-center expansion as the US. But as inference, agents, and industrial AI grow, Europe will still need data centers close to where that work happens. That is where local capacity comes into play.

But to remember, which is perhaps a little in contrast and hypocritical, thinking about my thinking in January–February 2026, Europe’s capacity problem might not be that big if you think about industrial and enterprise AI use cases. The capacity problem would then become how to route demand from Europe, or other regions with less capacity, to other places. That becomes an arbitrage opportunity, a hedging opportunity, and something customers or buyers of compute would want to do in order to meet demand.

With grid connections constrained in the established data-center hubs, more of Europe’s next buildout may move toward regions with available power, particularly the Nordics, France, Iberia, and parts of Southern and Central Europe, as Ember’s work on European grids shows.

One problem is power and connecting to the grid. Another is who can finance and build the capacity. Especially in Europe, a data center is not something a few people can bootstrap. It is an industrial project: large upfront capital, grid access, anchor customers, long contracts, and patient financing. That helps explain projects such as Nscale and Microsoft in Norway and SoftBank’s planned buildout in France. The financing is also beginning to look more like project finance and securitization than venture capital, as Hogan Lovells describes. Simon Grimm argues that Europe’s compute buildout will happen country by country, with governments helping private projects secure power, grid connections, permits, and sites rather than waiting for Brussels to fund everything. The projects in Norway and France suggest that this is already happening.

As noted, Europe’s older industrial companies are already positioning themselves. Aker’s sale of Cognite to Schneider Electric brings industrial AI software into a company already strong in power, automation, and data centers. What is strange is that Aker is concentrating so heavily on data centers with Nscale while selling something as valuable as Cognite, which could have become a European competitor to Palantir. That may suggest that, in Aker’s mind, the data-center segment is where more of the future value sits.

The agentic future may create another hardware opening. Agents use more CPUs, memory, networks, and runtime infrastructure, and Europe’s telecoms and older infrastructure companies already own useful assets. Some may strike gold there rather than arriving late to the GPU buildout.

I wrote two articles in late December 2025. The first, Allocation Economy, reflects how I was thinking about models, agents, and where compute sits in that picture. The second looks at earlier attempts to commodify parts of computing, such as memory and CPUs, and asks whether GPUs will be different this time.

• Allocation Economy Dec 2025
• The Exchange That Tried to Commodify Computing Dec 2025

Further