Stop buying the sun: right-sized AI capacity

You wouldn't fire up the sun to heat leftovers. Yet a lot of AI gets bought that way.

In theCUBE's studio at the New York Stock Exchange, our CEO Paweł Czech put it in one line: "You don't have to power up the sun to cook something in the microwave." It's also the simplest way we know to explain why we started Starfield.

This page is that idea, unpacked, for anyone who doesn't have 25 minutes for the full interview.

The sun is the default

When a team decides to put AI into a product, the first stop is usually the biggest name with the biggest capacity. That feels safe. Nobody gets fired for buying the sun.

But most of what AI does every day isn't the work the sun was built for. Training a frontier model is a sun-sized job. Answering a customer's question, reading a camera feed, summarising a document or checking a part on a factory line is a microwave job. It's small, it's constant and it happens close to where people are.

That everyday work is inference: running a model that already exists. It runs every hour of every day. It's bound by latency, so it suffers when every request makes a round trip to a distant region. And it's power-hungry, which matters when power is the thing everyone is queuing for.

Inference wants something different from training. When it's served from the same few giant sites, the buyer pays for the sun every time they warm something up.

What the microwave stands for

Paweł's point on theCUBE had three parts. Buyers need the right capacity, close to them, at a fair price. Each part is a choice most buyers aren't offered today.

The right capacity. Size the compute to the job, not to the biggest thing on the menu. A model that answers support tickets doesn't need the same setup as a lab training its next release. Paying for the bigger setup doesn't make the answer better. It makes the bill bigger.

Close to them. Distance costs time, and for some work time is the whole product. A robot can't pause while its request crosses a continent. A video system can't buffer a safety check. Some data also can't leave the place it was created, by policy or by law. Running inference near the people and machines that use it solves both problems at once.

At a fair price. This is the part nobody talks about. If you can't compare what capacity is worth, you can't tell whether you're overpaying. So most buyers overpay, and they call it the cost of doing AI. (We go deeper on this in What is AI compute worth? Nobody knows.)

Why the sun keeps winning

If right-sized capacity makes so much sense, why is the sun still the default? Two reasons.

First, capacity is concentrated. Most of it sits in a handful of very large campuses owned by a handful of companies. If that's where the capacity is, that's where the work goes, whether or not it fits.

Second, building more suns is slow. New hyperscale capacity can wait years for a grid connection. Meanwhile there's power, land and connectivity spread across thousands of smaller sites that are ready much sooner, but nobody runs them as one thing. One small site on its own is hard to buy from. A thousand small sites run as one network is a different offer.

Aggregation, not construction

That's the bet behind Starfield. The next wave of AI capacity won't only come from bigger campuses. It will also come from many small sites, close to power and demand, run as one network.

We deploy standardized compute close to power, connectivity and demand, then connect it through one programmable network for AI inference. Workloads are placed by policy, on latency, data residency, model and load. Every workload can generate an auditable record of commitment, placement and delivery, so a buyer can see what they asked for and what they got.

Put simply, we're building the microwave network: lots of right-sized capacity, close to where it's used, that you can buy without paying for the sun.

What this means if you buy AI capacity

You don't need a new strategy to start thinking this way. You need three questions for your next AI workload:

1. How big is the job, really? Is this training or inference? If it's inference, size for the steady load, not the peak headline.
2. Where do the people and machines using it sit? If the answer is "in one city" or "on one factory floor", ask why the compute is somewhere else.
3. Where does the data have to stay? Residency is easier to design in than to retrofit.

If the answers point to small, constant and local, you're cooking with a microwave. Buy one.

Watch the full conversation

Paweł made the case in full on theCUBE + NYSE Wired, with host Gemma Allen. The microwave line comes at 10:17.

If you have inference demand, power or sites and want to talk about right-sized capacity, use Request capacity or Partner with Starfield below.

Quotes

“You don't have to power up the sun to cook something in the microwave.”