A GPU-hour is a part number, not a unit

Correction, 8 September 2026. As first published, this note stated that no source in the panel discloses a fabric, a power envelope or a thermal limit. That was measured against what the index had stored, and it was the wrong place to look for it. One source does publish those attributes: vast.ai returns measured deep-learning throughput, total FLOPS, GPU memory bandwidth, PCIe bandwidth, NVLink bandwidth, a power ceiling and a temperature ceiling on every offer. The collector was reducing each offer to 21 pricing fields before storage and discarding the rest, so the audit below saw an absence the market had not created.

The measurements in this note are unchanged and reproduce as published: the stored payloads did contain none of those terms, interconnect was inferred on every row, and 805 rows still contradict themselves. What was wrong is the attribution. For eight of the nine sources the gap is the market's; for vast.ai it was this project's.

The collector now stores those fields, from 2026-09-08. The section headed What an assay would have to measure should be read against that: for one venue in the panel, three of its four axes are already being published and are now being kept.

Abstract

Every price index rests on a unit assumed to be homogeneous. Brent and WTI are separate grades because crude is not one substance, and the difference between them is quantified by assay: density, sulphur, a published differential. The TCI unit is one NVIDIA H100 SXM 80GB GPU-hour, which is a part number. It states what silicon is installed. It states nothing about what an hour of that silicon delivers.

This note audits what the index records about the quality of the goods it prices. Of 7,798 stored observations, none carries an interconnect value read from a field that any source disclosed. 57.6% carry a constant written into a collector. The remaining 42.4% carry a value parsed out of a SKU or instance-name string. Across all 7,798 raw payloads, the number mentioning InfiniBand, RoCE, NVSwitch, TDP, watts, cooling, fabric or bandwidth is zero.

The inference has a measurable cost. 805 rows, 13.4% of every SXM row in the database, assert gpu_model='H100_SXM' and interconnect='PCIe' at the same time, which is a contradiction in terms: SXM and PCIe are mutually exclusive form factors. They originate in a fallback branch returning "PCIe" whenever a SKU name fails two substring tests.

No published number is affected, because interconnect does not enter the calculation path. What is affected is the audit trail, which asserts facts no source supplied. The consequence for the index is a bound on what it can claim: the spread across European H100 sellers on 7 September was 7.40x, from $2.16 to $15.98, and the index cannot decompose that figure into price and grade.

---

The unit problem

Practitioners who benchmark GPU fleets make a claim that bears directly on the construction of any compute price index: published specifications generalise across silicon, and delivered performance does not. The reasons given are the silicon lottery, heterogeneous power delivery, network topology that differs between nominally identical nodes, and cooling. The method used is to run short reference workloads shaped like real jobs, covering synchronised training steps, collective communication, sustained matmul and checkpoint I/O under load, then to treat those results rather than the datasheet as the node's capability.

If that is correct, a benchmark quoting a single price for an H100-hour is averaging across goods of different quality and presenting the result as a price. That is a serious charge against the construction used here, and the appropriate first response to it is measurement rather than argument.

The performance claim itself cannot be tested from this dataset. TCI operates no benchmark harness, holds no fleet access and has no delivered-throughput data. What can be established is narrower and still decisive: if quality varies across the panel, does the index capture anything that would allow it to adjust?

All figures below derive from data/eucri.db in the public repository, covering every source and model from 2026-07-18 to 2026-09-07, a total of 7,798 observations across nine providers. The relevant columns are gpu_model, gpu_count, interconnect and raw_json, which stores each source payload as received. Every query appears at the end of this note.

---

A field that is fully populated and never observed

interconnect carries a value on every row. Tracing where those values originate:

57.6% of the database carries a constant typed into a collector. The remaining 42.4% is parsed from a product name, which is a naming convention rather than a specification. Nothing is read from a disclosed field. No source in the panel publishes an interconnect as a field; see the correction above for what one of them does publish.

A column that is fully populated and never observed is worse than an empty column. An empty column advertises the gap.

---

What the market actually discloses

Searching all 7,798 stored payloads for the terms that would matter:

The 264 hits are the string "NVLink" appearing inside a GPU display name on two marketplaces, not a fabric specification. No payload stored up to this date describes a fabric, a power envelope or a thermal limit. As the correction above records, that was a statement about what the collectors kept, not about what every source offers.

The most structured source in the panel returns this:

{
  "armSkuName": "Standard_ND96isr_H100_v5",
  "armRegionName": "westeurope",
  "retailPrice": 132.232,
  "unitOfMeasure": "1 Hour",
  "meterName": "ND96isrH100v5",
  "productName": "Virtual Machines NDsr H100 v5 Series Windows",
  "currencyCode": "USD",
  "type": "Consumption"
}

A price, a region, a SKU string and a billing unit. That is the whole of what the market tells a price index about the good being sold.

---

Two axes in one column

The values recorded in interconnect fall into two groups that do not belong together. NVLink, NVL and PCIe describe the intra-node bus, meaning how GPUs inside a single node communicate. InfiniBand and Ethernet describe the inter-node fabric, meaning how one node reaches another.

A row reading NVLink says nothing about whether the node can reach another node at all, and a row reading InfiniBand says nothing about the bus inside it. Only Azure ever produces an inter-node value, 678 rows out of 7,798, so for the other eight providers the fabric is not merely unknown but unrepresentable: the column has no way to express it.

For single-GPU inference that gap is survivable. For the multi-node synchronised training that dominates large-scale demand, inter-node fabric is close to the whole story, and it is the axis on which the index is blindest.

---

Rows that contradict themselves

The Azure collector derives the field as follows:

def _interconnect(sku: str) -> str:
    if "noIB" in sku:
        return "Ethernet"
    if "isr" in sku:
        return "InfiniBand"
    return "PCIe"

Applied to the three H100 SKUs Azure publishes:

The ND H100 v5 series is built on H100 SXM5 GPUs with NVLink 4.0 between them.¹ The isr and is suffixes distinguish RDMA support, not the bus. The third row is therefore not a near miss. It is the fallback branch firing on a SKU it was never taught, returning an answer that contradicts the same row's own gpu_model.

Across the database, 805 of 5,998 SXM rows, or 13.4%, assert an SXM part and a PCIe bus simultaneously; 123 are H100 and 682 are A100. Rows that disagree with themselves are the one error class in this dataset detectable without an external reference, and they went undetected because interconnect sits outside the calculation path, where no test asserts anything about it. A field that nothing checks is a field that drifts. That observation generalises past this column: an audit trail is only as reliable as its weakest field, and the weakest field is whichever one no test has an opinion about.

---

What this means for the published spread

On 7 September, EU/EEA H100 SXM, in a single collection run:

A spread of 7.40x. Research Note 2026-03 attributed spreads of this kind to the seller population, meaning who is quoting rather than where. That account stands and is not withdrawn, but it was incomplete, because it assumed the good was constant across the panel. This note is the evidence that the index has no basis for that assumption.

The claim here is bounded. The index cannot decompose that spread, and a benchmark unable to decompose its own spread should say so rather than allow a reader to assume that the cheapest row and the dearest row represent the same good at different prices. The likely direction of the bias is worth stating: if the expensive end delivers more useful work per hour, then true price dispersion is narrower than 7.40x, and a median taken across undifferentiated grades sits at a level that cannot be located.

---

How older commodities solved this

Crude is priced by grade, and grade is defined by assay: API gravity and sulphur content, measured to a standard, with a published differential between benchmarks. Dry bulk freight is priced per route and per vessel description, and the Baltic indices specify deadweight, age, draft and speed, because a Capesize hour is not a Handysize hour. Electricity is priced by delivery point and delivery hour, since a megawatt-hour in the wrong place is a different product.

The sequence is consistent across all three: name the attributes that make the good non-fungible, measure them to a published standard, then quote the differential. Compute has completed the first step informally, in that fabric and thermals are widely understood to matter, and has not begun the second. TCI is not in a position to begin it either, and the reference-unit definition in factors.yaml is a part number precisely because a part number is the only thing the sources supply.

---

What an assay would have to measure

The reference-workload approach maps closely onto the four properties this dataset cannot see.

Synchronised training steps capture straggler behaviour. In a data-parallel step every rank waits for the slowest, so fleet throughput is set by the worst node rather than the average one. That is the failure a datasheet cannot express, and it is what the silicon lottery and thermal variation produce.

Collective communication measures the inter-node fabric, the axis eight of nine sources in this panel never mention and the one that dominates multi-node training.

Sustained matmul throughput separates peak from sustained. Published TFLOPS figures are boost numbers, and what a node holds under a long job is a function of its power envelope and its cooling, neither of which appears in any of the 7,798 payloads examined here.

Checkpoint I/O under load covers the storage path, which no rate card in this panel describes and which converts directly into wall-clock time on a large training run.

Those four measurements would constitute a compute assay. An index does not need to perform them itself: crude benchmarks do not run their own assays, they cite an independent assay measured to a published standard. What compute lacks is the standard and the independent assayer, not the technique.

---

What would falsify this

If reference workloads run across this panel showed delivered performance clustering tightly for a given part number, within a few per cent between the cheapest and dearest European H100, then the part number is an adequate unit, the disclosure gap is a tidiness problem rather than a measurement one, and this note overstates the case.

If a source in the panel begins publishing fabric and power alongside price, the audit above expires and should be rerun rather than cited.

If price dispersion proves uncorrelated with every quality attribute once one becomes measurable, then grade is not present in the spread, and the population account in Note 2026-03 was complete on its own.

---

Limitations

No performance data underlies this note. It is an audit of disclosure, and nothing in it demonstrates that delivered performance varies across this panel. It establishes that the index could not detect such variation if it existed.

interconnect is not in the calculation path, so no published print is affected. The defect lies in the audit trail, which matters under IOSCO P16 and does not move a number.

The provenance table derives from reading the collectors rather than from a stored field, which is itself the gap this note identifies.

Nothing here is a forecast, and none of it is investment advice.

---

Reproducibility

-- Where interconnect values come from, by source
SELECT source, provider, interconnect, COUNT(*) FROM observations
GROUP BY 1, 2, 3 ORDER BY source, provider;

-- Payloads mentioning any fabric, power or thermal term
SELECT COUNT(*) FROM observations WHERE lower(raw_json) LIKE '%infiniband%';
--   repeat for roce, nvswitch, fabric, bandwidth, gbps, tdp, watt, power, cool

-- Rows asserting an SXM part and a PCIe bus at once
SELECT source, gpu_model, interconnect, COUNT(*) FROM observations
WHERE gpu_model LIKE '%SXM%' AND interconnect = 'PCIe'
GROUP BY 1, 2, 3;

-- The spread across European H100 sellers
SELECT provider, tier, MIN(price_usd_per_gpu_hr), MAX(price_usd_per_gpu_hr)
FROM observations
WHERE gpu_model = 'H100_SXM' AND substr(ts_utc, 1, 10) = '2026-09-07'
  AND country IS NOT NULL AND country <> 'US'
GROUP BY provider ORDER BY 3;

The remedies indicated by these findings are recorded in the project changelog against the date of this note. Any change they produce in the calculation path will be announced in advance under the notice procedure in GOVERNANCE.md §1 and published on the notices page.

---

Sources

The framing of the unit problem and the four workload classes restate what fleet-benchmarking practitioners describe as standard practice. TCI has performed none of that work and claims no part in it. It is cited as the technique a compute assay would use, not as a result obtained here.

Prior notes referenced: 2026-03, Compute prices barely move. Where you buy moves everything, for the spread decomposition this note qualifies; 2026-01, When an index measures its own sampling, for the bimodal panel.