Compute is moving closer to power and demand: notes from eight weeks on the road

What the NYSE and YOTTA told us about where AI inference is going, and why we're building Starfield.

Between late August and October we spent a lot of time in rooms where the future of AI infrastructure gets argued about. A panel at the New York Stock Exchange. YOTTA in Las Vegas. A podcast studio with Money Moves.

The people in those rooms build chips, data centers, power plants and models. They don't agree on much. But three ideas came up everywhere we went, and together they describe where AI compute is heading.

1. Power is the constraint, not chips

For the past few years, the AI conversation was about chips: who has them, who can get them, and how long the queue is. That conversation hasn't gone away, but it has moved. The question we heard most often wasn't "can you get GPUs?" It was "where will you plug them in?"

YOTTA describes itself as the show for the place where AI, energy and infrastructure meet, and the hallways matched. People talked about grid queues, bringing their own power, and modular data centers that can be built in a factory and set down where the power already is.

The reason is simple. Servers ship in weeks. A new grid connection for a large campus can take years. So new hyperscale capacity often waits on power, while smaller sites that already have a connection, spare capacity and space can be put to work sooner. We wrote more about this in Why AI capacity follows power and Spare power is not usable power.

2. Inference wants to sit close to the people and machines using it

Training a model is a big job you do once, in one place. Answering questions with it, which is called inference, is a job that runs all day, every day, wherever people and machines need answers.

That changes what matters. On Money Moves, Paweł put it as three things that decide where AI runs: latency, data privacy and ownership. A robot on a factory floor, a camera network or a hospital system can't always wait on a round trip to a distant region. Some data isn't allowed to leave the place it was created. And more organizations want a say in whose machines run their work.

As he said on the show: "If we want to benefit from these improvements in our day-to-day lives, it means that compute will have to go decentralized."

For the longer version, see Where should AI inference run? Three things decide.

3. Compute is becoming a market

At the NYSE, Paweł joined Dan Wright of Armada, Øyvind L. Vesterdal of Fossefall and a speaker from Cato Digital on a panel about the supply side of AI infrastructure. The conversation kept coming back to one idea: compute as an asset you can price, buy and take delivery of, like electricity or bandwidth.

That same afternoon, Armada launched its 10 MW Orion, and Fossefall ordered five Armada Leviathans for Nordic AI factories. Modular capacity isn't a slide in a deck any more. It's being ordered.

A market needs a fair price, and today most buyers don't have one. On theCUBE, also at the NYSE, Paweł described the problem: "We're always buying the sun. We're always paying that maximum premium. And this is because nobody knows how much your capacity is actually worth." Or, more simply: "You don't have to power up the sun to cook something in the microwave." More in Stop buying the sun and What is AI compute worth?.

What we're building

These three ideas are why Starfield exists.

We're building standardized compute close to power, connectivity and demand, connected through one programmable network for AI inference. The network is designed to place workloads by policy, on latency, data residency, model and load, and to generate an auditable record of commitment, placement and delivery for every workload. Fabric, the control plane, is in development.

In practice, that means we sit in the layer above the module and below the model. We don't build models, and we don't only build data centers. We turn many smaller sites into one network that runs AI work where it makes the most sense.

Aggregation, not construction

The next wave of AI capacity won't only come from one 120 MW site. It will also come from thousands of smaller ones, each close to power and close to the people using it, run as one network.

The internet spread communication to everyone, then commerce, then finance. Intelligence is next. We think it spreads the same way: outwards.

If you have power, sites or inference demand, we'd like to talk. Find us at starfield.net.

Paweł Czech is CEO of Starfield.