Public GPU pricing is not one thing.

The task sounds simple: find the public on-demand price of one NVIDIA H100 SXM 80GB GPU.

The fourth Sorso View Compute Index capture shows why that task still does not resolve into one clean market table. Sorso View reviewed twenty-seven headline-panel providers. Fifteen produced comparable fixed public rates. Twelve were captured and disclosed, but did not enter the medians.

That split is the finding.

A missing comparable rate is not automatically a hidden rate. Enterprise infrastructure is often sold through sales teams, negotiated contracts, reserved capacity, and custom deployments. A sales-led model is not misconduct.

The buyer problem remains. Without a comparable public rate, a buyer cannot tell whether a quote is rich, whether a reliability premium is justified, or whether competition is compressing prices across the market.

What Issue 004 measured

Issue 004 held at 103.8 on the SVCI Public GPU Rate Index, unchanged from Issue 003. Coverage remained 0.7150 and the Compute Opacity Index remained 28.5.

The important reading is not that public prices moved. They did not move at public index precision. The important reading is that public comparability stayed thin.

Fifteen providers published a comparable fixed public H100 SXM 80GB rate. Twelve did not.

Visible is not the same as comparable

The twelve non-inclusions were not one behavior. They included sales-gated pricing, missing public H100 SXM rate cards, starting-from prices, marketplace and auction observations, host-varying prices, aggregator pass-through offers, signed-in-only observations, and product-mode-incompatible observations.

Those categories are not interchangeable. A provider that says contact sales is different from a marketplace that publishes live supplier offers. A public floor is different from a fixed list rate. A signed-in dashboard observation is different from a logged-out public rate card.

Sorso View does not collapse those outcomes into one accusation. It records them separately and keeps the panel member visible.

The late-round lesson

The final Issue 004 captures made the distinction especially clear.

Jarvislabs published a clean pricing-table rate: H100 SXM 80GB at $2.69 per GPU-hour on demand. That row entered the Tier 2 median.

Hyperbolic showed public and signed-in surfaces that did not resolve to one logged-out, comparable, fixed H100 SXM on-demand rate. The signed-in app showed an operational H100 SXM5 offer, while public documentation showed a separate dedicated-instance price. The row was captured and excluded.

Paperspace showed the opposite kind of trap. The visible H100 card displayed $2.24 per hour, but the footnote tied that figure to a three-year commitment. The same public surface stated the on-demand H100 price was $5.95 per hour. That row entered Tier 3 at $5.95, not $2.24.

Vast.ai published live market rates, ranges, and distributions. TensorDock showed H100 prices from $2.25 per hour while its own table said the typical hourly price varies by host. Shadeform displayed provider-attributed H100 offers from third-party clouds inside a GPU marketplace.

Those are public prices in the ordinary sense. They are visible. They are useful. They are not the same thing as a fixed first-party public on-demand rate.

That is why they did not enter the SVCI medians.

The buyer problem

The market can have many public numbers and still lack comparability.

A buyer who sees one H100 price as a fixed on-demand list rate, another as a commitment rate, another as a marketplace floor, and another as a quote-on-request product cannot compare them without knowing the pricing basis. The number alone is not enough. The unit, term, product mode, source class, and availability model decide whether the number means the same thing.

This is why Sorso View reports coverage beside price. The benchmark does not only ask what the visible rate is. It asks whether the visible rate is comparable.

What to ask before accepting a quote

A buyer comparing GPU quotes should ask five questions before treating a number as market evidence:

1. Is the price a fixed on-demand rate, or a floor, estimate, spot price, reserved price, or sales quote?
2. Is the unit per GPU-hour, per instance-hour, per token, per worker, or per cluster?
3. Is the product the same product mode: GPU instance, cluster, dedicated inference, serverless inference, or managed application?
4. Is the offer first-party capacity, marketplace inventory, host-set inventory, or aggregator pass-through?
5. Is the rate public without sign-in, or visible only inside a console, calculator, private quote, or sales workflow?

A fair comparison starts only after those questions have the same answer across providers.

The finding

Issue 004 did not produce a new headline move. It produced a cleaner market reading.

Public AI compute pricing is becoming more observable, but not necessarily more comparable. The absence of a fixed public rate is now a measurable part of the market, not a footnote outside it.