Here's an experiment you can run in ten minutes. Open the pricing page of a major cloud provider and find the H100 rate. Then open a decentralized GPU marketplace and find the same card. As of this writing, you'll find something like $2.80-$3.50 per hour on the big clouds versus $1.60-$2.10 on open markets. Same silicon. Same VRAM. Similar interconnect. Often the exact same data center region.

Nobody at the hyperscaler is lying to you. But the pricing structure is designed so that you never ask the obvious question: what exactly am I paying for?

The four things you're actually buying

When you rent an on-demand H100 from a big cloud, your hourly rate funds roughly four things:

  • The GPU. Maybe 30-40% of what you pay. Capex, depreciation, power, cooling.
  • The sales layer. Enterprise account managers, solution architects, the whole dance. You pay for this even if you self-serve through a webpage and never talk to a human.
  • The real estate and network. Pristine facilities, premium interconnect, regions on five continents. Beautiful. Do you need it? For training a frontier model, maybe. For running inference on a fine-tuned 7B, almost certainly not.
  • Optionality. This is the sneaky one. On-demand pricing is expensive precisely because it's flexible — you can leave anytime, so you pay hotel-bar rates forever.

That fourth item is the heart of the scam. The sticker price isn't a hardware cost. It's a flexibility premium, and most users paying it never use the flexibility. They run the same workload for months at the same rate, like someone paying nightly hotel prices for an apartment they never leave.

"But commit to a reserved contract!"

The industry's answer to this is the 1-year or 3-year commit, which drops the effective rate by 30-60%. Notice the shape of that offer: to get a fair price, you must lock yourself in, pass a credit check, and negotiate. The list price exists to anchor you; the discount exists to make you feel like you won.

This is not a technology business model. It's an airline pricing model. And it works because the alternative — a real-time open market for compute — didn't exist at scale until recently.

What decentralization actually changes

Decentralized compute networks invert the structure. Instead of one operator pricing its own inventory, you get thousands of independent suppliers — data centers with spare capacity, crypto mining farms that repurchased during the last bear market, even folks with racks in garages — competing in an open order book. The marketplace takes a small cut. The rest of the "sales layer" margin simply evaporates.

The honest objections, and there are real ones:

  • Reliability varies. A random supplier's node can vanish mid-job. Decentralized networks handle this with checkpointing, replication and slashing, but it's engineering overhead the hyperscaler hides from you.
  • Verification is hard. How do you know the node actually ran your job on the GPU it claimed? This is exactly the problem zk-verification and attestation systems (including what we're building at OMC) exist to solve — and the networks that solve it credibly will eat the middle of the market first.
  • Not for everything. A three-week frontier training run with 10,000 tightly-coupled GPUs still belongs in a hyperscaler's fat interconnect. Inference, fine-tuning, batch jobs, research — that's where the decentralized price gap bites.

The part everyone misses

The scandal isn't that hyperscalers are expensive. It's that the default option — the one every tutorial links to, the one your credit card already works with — is the most expensive way to buy the same thing. Pricing opacity is a feature for the seller and a tax on everyone else.

Markets fix this the boring way: by making prices visible and comparable. That's happening now. Once buyers can comparison-shop compute the way they comparison-shop flights, the flexibility premium stops being a premium and starts being a choice.

That's not a scam getting exposed by regulators. It's a scam getting exposed by a search box.