Urdais Compute Price Index Family (UCPI)#

Status: proposed methodology, version 0.1.2-draft. Prepared 12 September 2026, amended 13 September 2026. No production value, child specification, provider list, parameter set, or price history is established by this document. Nothing here describes live infrastructure.

This proposal follows the principles of the Urdais methodology framework. It is a shared primitive rather than a published output: the framework's output template applies to the child instruments that consume it, not to this document. Rules expressed as requirements describe the proposed design, subject to approval. Parameters marked unresolved are launch requirements, not defaults. No UCPI child may be published until this family methodology has an approved version, a child specification exists, source licensing is secured, and the items in Open Questions / Empirical Validation Required are closed in a versioned release.

Purpose and Scope#

The Urdais Compute Price Index family measures what a precisely defined unit of market-accessible accelerated compute costs. It is a family, not a single index: each child instrument fixes one compute product and one commercial context, and this document defines the rules every child inherits.

This methodology owns the observation model, the identity framework for compute products and sellers, the source hierarchy, the normalization rules, the availability and eligibility framework, the aggregation architecture, the diagnostics, the publication statuses, corrections, lineage, and versioning. It does not define any particular piece of hardware. The first child, UCPI-H100-SXM, is specified separately and serves as the first empirical test of this architecture.

Product Family Architecture#

The family resolves into child instruments, each a separately specified and separately published series:

Urdais Compute Price Index family → UCPI-H100-SXM, UCPI-H200, UCPI-B200, UCPI-A100-SXM4, UCPI-RTX-5090, and future compute instruments.

There is deliberately no headline index combining accelerators. A single number averaging H100, H200, B200, A100, and consumer cards would have no stable economic meaning: those products differ in compute throughput, supported precisions, memory capacity and bandwidth, interconnect, workload suitability, power envelope, lifecycle stage, supply conditions, and cluster architecture. An average across them would move whenever the mix of observed products moved, independently of any price change, and would therefore fail the first requirement of a benchmark, that it represent the economic reality it claims to measure.

A quality-adjusted compute composite that expressed different accelerators in common performance units is a legitimate future research question. It would require an explicit, defended equivalence methodology, and it is not attempted here.

Primary Measurement Question#

What does a defined unit of market-accessible compute cost?

Every word of that question carries weight. Defined means the compute product is pinned by a child specification before any price is compared; UCPI never assumes that compute is fungible. Market-accessible means the price must be one at which a buyer could actually obtain the product, not merely one that appears on a page. Cost means the mandatory cost of obtaining the defined product, not a headline rate that excludes unavoidable charges.

What UCPI Measures#

The economic target#

Five distinct price statistics could be called "the price of compute", and they are not interchangeable:

1. The typical advertised price: what sellers publish, whether or not anything is available at it.
2. The price offered by a typical participating supplier: what a buyer approaching the market could actually transact at now, across those who offer the product.
3. The capacity-weighted offer price: the price attaching to the typical available unit of capacity.
4. The transaction-weighted clearing price: the price at which purchased compute actually cleared.
5. The marginal cheapest available price: the best price findable anywhere at that moment.

UCPI's target concept is the second: the market-accessible price offered by a typical participating capacity source for the specified product. That is the object the primary question describes, and it is the object a buyer can act on.

The target concept and the statistic currently computed are not the same thing, and the document names both rather than letting the ambition stand in for the arithmetic. The computed statistic is stated in The observable proxy below and in Regional Aggregation.

The other four are rejected as targets for specific reasons. The advertised price (1) measures publication, not the market: a price displayed beside no capacity is not formed by the competitive forces of supply and demand. The cheapest available price (5) is determined by whichever seller is currently most marginal, is the statistic most sensitive to a single erroneous observation, and answers a shopping question rather than a market question; market evidence in the research appendix shows a marketplace's own floor price and its own median differing by between seventeen and one hundred and three percent across accelerator models on a single day, so the choice between them is material rather than cosmetic.

The rejection of a market-wide minimum is not in tension with taking a minimum inside each participant, because the two operate on different populations and mean different things. Within a participant, the lowest qualifying price is that participant's best accessible supply price for a product the specification has already pinned. Across participants, the median identifies the typical participant. The market-wide minimum across all participants, which UCPI rejects as a headline, would identify only the marginal cheapest offer in the market. The capacity-weighted price (3) is the more economically complete object and is the natural future extension, but capacity is not reliably observable across providers and must not be fabricated. The transaction-weighted clearing price (4) is the ideal, and is not observable: compute transactions are private, contracts are negotiated, and no consolidated tape exists.

The observable proxy, and the gap#

Because the target is not directly observable in the way an exchange price is, UCPI states its proxy and its limits explicitly rather than blurring them.

The published statistic is stated exactly, because a benchmark's description is part of its methodology and a broader phrase would overstate what is computed:

The regional median of each eligible participant's lowest-priced qualifying accessible offer for the child-specified product.

In product language, UCPI measures the typical participating capacity source's best qualifying accessible price for the specified compute product within a region. A capacity source is the infrastructure operator where the operator is reliably determinable, and the seller otherwise, as defined in Capacity-Source Identity and Collapse; it is the participant unit over which the median is taken. The word "typical" applies to the distribution across participants, not to the distribution of raw offers.

It is therefore none of the following, and must not be described as any of them: the median of all offers; the price of the typical available accelerator-hour; the price paid by the typical buyer; a capacity-weighted price; or a transaction-weighted clearing price.

Three consequences follow and must be disclosed wherever UCPI is used.

• UCPI is a supply-side accessibility statistic, not a volume-weighted market price. It answers "what does the specified product cost across the participants offering it" and not "what did the typical purchased accelerator-hour clear at". A participant with one available accelerator and a participant with ten thousand contribute equally, because their relative capacity is not reliably observable. Where most capacity concentrates with few participants, the two questions can diverge materially, and this limitation is not removable by any diagnostic; the diagnostics published with every observation record what is known about composition, and the section on Coverage and Composition is explicit that economic capacity concentration remains unknown where capacity is not observable.
• UCPI is not an index level. It is a price in currency per accelerator-hour. See Published Values.
• UCPI is not a cost model. It measures the price of renting a defined product, not the cost of producing compute, the total cost of ownership of a deployment, or the cost of any workload.

What UCPI Does Not Measure#

UCPI does not measure model or token pricing. Prices charged for access to hosted models through an API are a different market layer, formed by different economics, and are outside this family entirely. A separate model-price methodology may later consume compute economics analytically; a token price is never a UCPI observation.

UCPI does not measure the price of compute in general. Each child measures one specified product, and a child's value must never be presented as the price of accelerated compute, of AI infrastructure, or of a vendor's product line.

UCPI does not measure performance, performance per currency unit, or value for money. It measures price for a specified product. Comparing two children to infer relative value requires a performance equivalence this family does not establish.

UCPI does not measure hardware acquisition prices, colocation or power costs, or the enterprise contract prices negotiated privately at scale, except to the extent such a contract is admitted as transaction evidence under the source hierarchy for a specification-matched product.

Parent and Child Responsibility#

The boundary is deliberate, and rules belong on one side of it only.

This family methodology owns the economic object and the exact statistic computed; the conceptual observation model; the observation-type taxonomy; the identity frameworks for compute instruments, sellers, operators, capacity sources, and regions, including the capacity-source collapse principle and the role of marketplaces; the source-quality framework and its separation from observation type; the availability and executability semantics; the reconfirmation semantics; the procurement taxonomy; the service-tier framework including tenancy; the principles for bundles, multi-accelerator offers, and fractional accelerators; the distinction between fixed and usage-proportional mandatory charges; currency normalization principles; temporal and freshness rules; the aggregation architecture, the aggregation population, and the median convention; robustness principles; the percentile population; coverage and composition diagnostics; missing-data behaviour; publication statuses; corrections; lineage; and versioning.

A child specification owns the exact hardware identity and its identity-defining attributes; the eligible service tiers including its tenancy and virtualization requirements; the single eligible procurement mode; the topology class and minimum quantity; the bundle envelope; the eligible regions and the region taxonomy it uses; the availability evidence threshold; the freshness ages, which may differ by dimension; the numerical publication gates; the currency and rate parameters; the percentile interpolation convention; any instrument-specific normalization; the empirical choice among the seller-reduction alternatives the family permits; the child's first publication date; and the child's own version history.

A family-level identity problem is never pushed into a child. Where this document leaves a choice open, it names the permitted alternatives and the evidence required to choose among them.

A rule that mentions a particular accelerator belongs in a child, not here. Nothing in this document is written to accommodate H100 specifically.

Compute Instrument Identity#

An accelerator model name does not define a comparable product. Two offers both labelled "H100" may differ in form factor, memory, interconnect, tenancy, minimum quantity, region, and service level, and the market evidence in the appendix shows a single provider publishing three distinct H100 variants at three distinct prices on one page.

A compute instrument is therefore defined by an explicit attribute set, and the child specification states which attributes are identity-defining for it. The family recognizes three categories of attribute, and the distinction matters because they are used differently.

Identity-defining attributes determine whether two offers are the same hardware product at all. Differences here make offers incomparable and cannot be normalized away. They include the vendor; the accelerator family and model; the architecture generation; the form factor, which separates SXM, PCIe, NVL and equivalents; the accelerator memory capacity and type; and whether the offer is a full physical device or a fractional partition.

Tenancy is deliberately not an identity-defining attribute. Whether a full accelerator is exclusively assigned, virtualized, or shared at the host level is a property of how it is sold and operated, not of the hardware, so it belongs to the service framework below. Keeping it in both places would have made the same characteristic decide eligibility twice under two different rules.

Service and commercial attributes determine the commercial context of an otherwise identical device. Differences here are not normalized away either, but they are handled by fixing them in the specification rather than by excluding the offer: the procurement mode, the service tier, the minimum quantity and node topology, the minimum commitment duration, and the region.

Observation metadata is recorded for lineage, diagnostics, and future research but does not by itself determine eligibility: the facility or availability zone, the bundled host resources, the billing granularity, the seller's catalogue identifiers, and the observed capacity where disclosed.

A child specification must state, for each identity-defining and commercial attribute, the required value or the permitted set. An attribute the child does not pin is an attribute along which its observations may vary, and every such attribute must be named as a known heterogeneity in the child's own limitations.

Geographic Framework#

Compute is not location-independent. Latency, power prices, data-sovereignty constraints, network access, local demand, and local capacity all affect price, so region is an economically meaningful dimension rather than metadata.

Regions are represented in a hierarchy from finest to coarsest: facility or availability zone; seller region label; metropolitan area; country; and canonical UCPI region. Seller region labels are not comparable across sellers and are never used directly for aggregation; each is mapped to a canonical UCPI region through a versioned mapping that records the evidence and effective interval for the mapping. The finest geography the source discloses is always retained even when the published series uses a coarser region, so that a future finer series can be constructed without re-collection.

The canonical region taxonomy is unresolved and must be approved with the first child. It must be coarse enough that eligible sellers exist within each region and fine enough that the offers within a region are plausibly substitutable for a buyer.

Regional series are primary; a global series is deferred#

A child publishes regional series. A global series is not published until a defensible regional weighting basis exists.

A global price is a weighted average across regions, and the weights are the whole of the economic content. Weighting regions equally would assert that a region with negligible capacity matters as much as one with most of it. Weighting by observation count would measure where sellers publish rather than where compute is. Weighting by capacity or by transactions would be defensible and requires data that is not currently available.

Publishing an unweighted global number would therefore embed an unexamined assumption in the most prominent figure Urdais produces for this market. The family declines to do so. This follows the same discipline applied elsewhere in Urdais: where a valid aggregation input does not exist, the aggregate is withheld rather than approximated. Whether region belongs inside a child's public identifier or is exposed as a dimension of one identifier is addressed in Child Specification Requirements.

Seller, Operator, and Marketplace Identity#

One physical accelerator can be offered through several commercial paths, and counting each as an independent market participant would overstate both the breadth of the market and the influence of whoever is being double-counted.

Three roles are distinguished. The seller is the counterparty with which a buyer would contract. The infrastructure operator owns or operates the physical hardware. The marketplace is a venue through which third-party sellers list capacity, and which may or may not itself be a seller.

A seller is identified by its legal identity. Two trading names, brands, tiers, or catalogues under one legal entity or under common control are one seller, and resolve to one seller-level observation before any participant is counted. This is a consequence of what a seller is rather than an additional rule, and it matters most where a region has few participants, since two brands of one company would otherwise present as a market.

Deduplication is performed at the level of the entity whose capacity is being offered, that is the operator where it is determinable, and the seller otherwise. Where a marketplace lists many third-party sellers, the family treats each distinct determinable operator as one participant rather than treating the marketplace as one participant or as many.

UCPI does not claim reliable deduplication where the data does not support it. Resale, white-labelling, and brokered capacity are frequently undisclosed. Every observation records the attribution basis, the share of eligible observations whose operator could not be determined is published as a diagnostic, and no inference of operator identity may be made from pricing similarity or from commercial speculation.

Source Types and Hierarchy#

The hierarchy adapts the benchmark-administration principle that inputs should be ranked, with concluded arm's-length transactions preferred and firm executable offers ranked above indicative information. It is adapted rather than copied because the compute market has almost no observable transactions, and a hierarchy that admitted only transactions would make the benchmark impossible rather than rigorous.

Ranked from strongest:

1. Concluded transaction evidence: an invoice or executed contract for a specification-matched product, where lawfully obtained and usable.
2. Executable programmatic offer: a seller API or machine-readable interface where the offer could be acted on at the observed price.
3. Seller pricing page with contemporaneous availability evidence for the specified configuration.
4. Marketplace listing that is executable at the observed price.
5. Written seller quote for a specified configuration, with its validity period.
6. Licensed specialist dataset whose own methodology is disclosed and assessable.
7. Third-party aggregator or comparison site, for discovery and corroboration only.
8. Search results, press coverage, marketing material, and commentary, for discovery only.

Grades 1 through 5 may produce eligible observations subject to the observation-type rule below. Grade 6 may produce eligible observations only where the dataset's methodology is disclosed, assessed, and recorded, and never where it is proprietary and unverifiable. Grades 7 and 8 never produce a production observation. Search-engine snippets are not evidence of price.

Where sources conflict for the same seller, region, instrument, and grade, the higher grade prevails; where they conflict within a grade, the observation is flagged Conflicted, withheld from the calculation, and escalated rather than averaged.

Source quality and observation type are different questions#

The ranking above answers only one question: how reliable is this evidence? It does not answer the separate and equally important question: what economic object does this evidence represent? Collapsing the two into one ordinal list would let a more authoritative document displace a better-matched measurement, and a more authoritative document is not necessarily a better measurement of UCPI's target.

Every observation therefore carries an observation type independently of its source grade:

• Current accessible offer: a price at which the specified product can be obtained now.
• Firm written quote: a price a seller has committed to for a stated validity period.
• Concluded transaction: a price at which the product was actually bought at some past moment.
• Advertised non-accessible price: a published price with no current availability.
• Indicative or list price: a published reference price that is not an offer.

UCPI's economic object is a current accessible offer, so observation type, not source grade, determines admission to the headline statistic.

Transaction evidence#

A concluded transaction is the strongest possible evidence that a price existed, and it is not automatically a measurement of a current accessible offer. A transaction may record a negotiated clearing price, a historic contract, a bulk discount, a long commitment, or a customer-specific rate, none of which is the price now available to an ordinary buyer.

Concluded transactions are therefore corroborative evidence by default, and are not automatically headline inputs. They may support a separate transaction-price series as future work, which would measure a different and legitimate object.

A transaction may enter the current accessible-price calculation only where all of the following are established and recorded: it is specification-matched; the procurement mode, service tier, region, and commercial quantity and topology are the same; it is contemporaneous under the approved recency rule; and the price remains available to an equivalent buyer under materially equivalent terms. Where any of those cannot be shown, the transaction is not mixed into the offer-price distribution.

The family rule on recency is fixed now even though its value is not: a historical execution does not establish a current executable offer indefinitely. The permitted age, transaction_relevance_age, is an unresolved empirical parameter.

On the use of benchmark-administration precedent#

The evidence-quality framework above is informed by benchmark-administration principles cited in the research appendix, which rank concluded transactions above firm executable offers above other information for benchmarks anchored in transactions. Those principles inform evidence-quality governance. They do not determine whether an observation matches UCPI's economic object, which is a separate determination Urdais makes here, and they do not imply that transactions must dominate current offers in this market. The same source expressly contemplates benchmarks built on executable quotes and notes that in a thin market a confirmed offer may carry more meaning than an outlier transaction. No external body has reviewed or endorsed UCPI.

Raw Compute Offer#

A raw offer is what a source says, recorded before any normalization and never overwritten. It captures the seller and, where determinable, the operator and marketplace; the source, its tier, and its retrieval evidence; the claimed compute instrument and every hardware attribute the source states; the region as the source labels it, with any finer facility identifier; the service tier and stated service-level attributes; the procurement mode; the price exactly as published, with its currency and billing unit; the minimum commitment duration; the minimum quantity and the node topology or bundle size in which the product is sold; all mandatory fees and whether tax is included; the bundled host resources; the availability state and the evidence for it; the observation timestamp and the effective timestamp of the price where the source discloses one; and the source provenance sufficient to re-examine it.

The raw offer is retained unchanged. Every later step produces a derived record that references it.

Availability and Executability#

A posted price is not necessarily an actionable price. Market evidence in the appendix includes an accelerator advertised at an hourly figure on a live marketplace beside the statement that there are no current offers, with the platform itself labelling the number a recent-market figure rather than a price. Treating such a number as the price of compute would misdescribe the market in exactly the direction that matters most: it would report cheap compute precisely when compute is unobtainable.

Availability is recorded as one of: Available, where the product can be obtained now; Limited, where it can be obtained subject to a stated constraint; Waitlisted, where access requires queueing; Sold out, where no capacity is offered; Quote required, where no price is transactable without negotiation; and Unknown, where the source discloses nothing.

Only Available, and Limited where the child admits it, may produce an eligible observation. Sold out, Waitlisted, and Quote required are ineligible for the price statistic. Unknown is ineligible for the headline but is retained.

The availability and pricing gap between advertised and accessible prices is retained and published rather than discarded. The count and price distribution of advertised-but-ineligible offers, with their availability states and their age and source status, is published as a diagnostic alongside every observation.

The methodology records the gap and does not interpret it causally. A low advertised price beside no availability may indicate scarcity, and may equally indicate a page that has simply not been updated. Both are worth observing and the data cannot distinguish them without further evidence, so the gap is published for analysis rather than labelled a scarcity signal. Excluding these observations from the price while retaining them as a diagnostic preserves both the integrity of the statistic and the information.

An availability evidence standard grades how availability was established, from a programmatic confirmation that an order would be accepted, through an explicit contemporaneous status signal from the seller, to an inference from the absence of a sold-out indication. The minimum acceptable grade is a child parameter and is unresolved; the distribution of eligible observations across evidence grades must be published, because a child resting mostly on the weakest grade is making a weaker claim than one resting on the strongest, and users must be able to see which.

Procurement Modes#

A procurement mode describes commercial obligation and interruption characteristics, and nothing else. The modes are: on-demand, available immediately without commitment and not subject to interruption; interruptible, covering spot and preemptible capacity that the seller may reclaim; reserved or committed, priced against a commitment term; and negotiated contract, contracted privately on bespoke terms.

A marketplace is not a procurement mode. It is a venue, and it belongs to seller and source identity, as set out in Seller, Operator, and Marketplace Identity. A marketplace can sell on-demand, interruptible, reserved, and auction-priced capacity, so treating "marketplace" as though it were equivalent to "spot" or "on-demand" would conflate where a product is bought with what is being bought.

Where price is set dynamically by platform supply and demand, that is recorded as a separate attribute, price formation, with values such as posted or auction and dynamic. Price formation is orthogonal to procurement mode: an auction-formed price may attach to interruptible or to on-demand capacity. Where a genuine procurement form exists whose commercial obligation is itself defined by an auction, it is documented as such by the child rather than assumed into an existing mode.

Promotional, trial, and subsidized prices are not a procurement mode either; they are a pricing condition, recorded as an attribute and excluded from every child as set out below.

Procurement mode is specification-defining. A child fixes exactly one mode, and modes are never blended into one series. An on-demand hour, an interruptible hour, and an hour under a multi-year commitment are different products bearing different risk and different obligations, and a number averaging them would describe none of them. Where several modes are worth publishing for one instrument, they are published as sibling series that a user may compare deliberately, not merged into a single figure.

Promotional, trial, and subsidized prices are excluded from every child. They are not ordinary commercial prices, their availability is conditional in ways sources rarely disclose, and admitting them would let a marketing decision move a benchmark. They are recorded and counted as a diagnostic.

Dividing a committed-term price by the hours in the term produces an arithmetic hourly figure and does not produce an on-demand price. That arithmetic is permitted only within a series whose specification fixes that same commitment, and the resulting figure must always be published with the commitment that produced it.

Service Tier and Service Level#

Two prices for identical hardware are not comparable if one is a dedicated instance with an uptime commitment and the other is a shared, interruptible, best-effort instance. Market evidence in the appendix shows a single provider offering the same accelerators under two differently operated tiers on one page.

A service tier is a named bundle of service characteristics that a child requires. The family defines the dimensions a tier may constrain, beginning with tenancy and exclusivity, whose values include full-device exclusive assignment through passthrough, full-device virtualized, host-level shared tenancy, bare metal, and fractional partition. A child may require a full accelerator with exclusive device tenancy without asserting that tenancy changes the accelerator's hardware identity, and bare metal is not required of every full-device product: which virtualization and tenancy values qualify is a child decision, made on evidence about whether they move price. The remaining dimensions are: interruption policy; uptime or availability commitment; provisioning and access model; support and replacement commitments; intra-node interconnect, such as NVLink or NVSwitch presence; inter-node fabric, such as InfiniBand or Ethernet and its bandwidth; attached storage and local disk; host CPU and memory allocation; virtualization and isolation model; and compliance or sovereignty attributes where they are economically material.

The family does not make every dimension a hard rule. It requires that a child declare, for each dimension, whether it is a comparability requirement that an offer must satisfy to be eligible, or recorded metadata that is preserved but not enforced. Dimensions treated as metadata become named heterogeneity in the child's limitations. Which dimensions must be requirements for a given instrument is an empirical question: it depends on whether they measurably move price, and that is tested before a child launches rather than assumed.

Bundled Resources#

Accelerator rentals are rarely sold as bare devices. The market evidence shows two sellers offering the same accelerator with materially different host bundles, one with roughly a hundred and twenty-five gigabytes of host memory and twenty virtual CPUs, another with roughly eighteen hundred gigabytes and over two hundred virtual CPUs across the node. The advertised price is therefore not a pure accelerator lease price in either case.

Four approaches were evaluated. Accepting the bundle as part of the product keeps prices auditable and observed, but compares products that differ. Subtracting an estimated value for non-accelerator components produces a purer accelerator price and requires inventing component prices that are not observed. Standardizing bundle requirements keeps comparability and reduces coverage. Classifying incomparable bundles into separate series is precise and fragments the measure.

The family selects a combination of the first and third: bundles are accepted as part of the specified product, within bounds the child declares. A child states the host-resource envelope it treats as ordinary for the instrument, offers outside that envelope are ineligible rather than adjusted, and the bundle composition of every eligible observation is retained and its dispersion published.

Synthetic decomposition is prohibited. UCPI does not subtract an imputed price for virtual CPUs, host memory, storage, or networking in order to derive an accelerator-only price. Doing so would replace an observed price with a modelled one whose value depends on component prices Urdais would have to invent, and would forfeit the property that every published figure traces to prices someone actually offered. Whether a defensible decomposition is possible is recorded as a research question, not adopted.

Multi-Accelerator Offers and Node Topology#

Sellers offer the same accelerator as a single device, as a multi-device instance, as a complete node, and as a cluster. The market evidence shows a seller quoting a cluster product per accelerator-hour at three different prices for sixteen, sixty-four, and two hundred and fifty-six or more accelerators, alongside single-device instance pricing for a different product.

Dividing a node price by its accelerator count is arithmetic. It is not economic equivalence, because it conceals the mandatory purchase quantity, the intra-node interconnect, the host resources, the network fabric, and any cluster premium or discount. An eight-accelerator node is not eight independent single-accelerator rentals, and the quantity tiers in the evidence show that sellers themselves price the difference.

Per-accelerator normalization is permitted only within a declared topology class. A child specification fixes the node topology and minimum quantity it measures, and division by accelerator count is then a unit conversion within a fixed product rather than a comparability claim across products. Minimum quantity and topology are always retained on the observation, never normalized away, and a child that admits more than one topology class must publish them as separate series.

Shared and Fractional Accelerators#

Fractional products, including hardware-partitioned instances, time-shared devices, and virtualized fractions, are offered alongside full devices; the market evidence shows a partitioned product listed in the same catalogue as full accelerators.

A fractional or shared offer never enters a full-device child. Multiplying a fraction's price to a notional full-device price assumes that partitions aggregate linearly into a device, which is not true of throughput, of memory bandwidth, or of the commercial product being sold. Full-device dedication is an identity-defining attribute, so a fractional offer is a different instrument, not a differently priced one.

Fractional compute is a legitimate future family member with its own specification and its own unit. It is out of scope here.

Price Components#

The price entering an observation is the mandatory cost of obtaining the specified product for the specified duration. It excludes charges that depend on how the buyer uses the product rather than on obtaining it, including data egress, additional storage beyond the specified bundle, and optional support upgrades. It excludes optional extras entirely.

Usage-proportional and fixed mandatory charges are treated differently#

Mandatory charges are not homogeneous, and converting them as though they were would require an invented assumption. A charge of five hundred currency units to activate a service plus two per hour is not an hourly price until someone decides how many hours to spread the five hundred over, and that decision, not the market, would then determine the published figure.

Usage-proportional mandatory charges scale with the metered quantity and enter the normalized hourly figure directly. A mandatory platform fee expressed per accelerator-hour is simply added.

Fixed or one-time mandatory charges, including activation fees, setup fees, and fixed reservation charges, are never silently amortized. Where a child fixes an explicit minimum or commitment duration, the child may define an amortization rule over that contractual minimum, because the contract itself then supplies the horizon rather than the methodology inventing one. Where no such contractual horizon exists, the fixed charge is retained and disclosed separately, and the offer is either excluded from a pure hourly-price series because comparability cannot be achieved, or published as a multi-component cost. A generic assumed utilization horizon is never selected.

A mandatory minimum spend is a commercial constraint, not an incremental hourly charge. It is recorded on the offer and preserved, and whether it makes an offer comparable with offers that carry no such floor is a child decision, not an arithmetic one.

Prices are recorded exclusive of transaction taxes, with the tax treatment of the source recorded. Market evidence shows at least one seller publishing prices explicitly exclusive of sales tax, value-added tax, and goods-and-services tax, while others publish without stating a basis. Comparing a tax-inclusive price with a tax-exclusive one would introduce a jurisdictional artifact into a price comparison; where a source's tax basis cannot be established, the observation is flagged and its share published.

Credits, coupons, and promotional discounts are not applied. The ordinary commercial price is the observation, consistent with the exclusion of promotional procurement.

Unit Normalization#

The normalized unit is currency per accelerator-hour, and it means the mandatory cost of one full, dedicated physical accelerator of the specified instrument, for one hour, under the specified service tier, procurement mode, topology, and region.

That definition is narrow on purpose, and the constraints that make it meaningful are carried by the specification rather than by the arithmetic. Conversions are permitted where they are unit changes within a fixed product: per-second or per-minute billing converts to an hourly basis; a node price divides by accelerator count within a declared topology class; a committed-term price divides by the contracted hours within a series that fixes that commitment.

Conversions are prohibited where they would cross a product boundary: from a fractional device to a full device; from a shared device to a dedicated one; from one procurement mode to another; from one topology class to another; or from a quoted price whose mandatory quantity differs from the specification.

Arithmetic convertibility is not economic comparability. A monthly commitment divided by its hours yields an hourly number that describes a fundamentally different procurement product from an hour of on-demand access, and the family's response is to keep them in different series rather than to trust the division.

Currency Normalization#

Children normalize eligible prices to a stated index currency, expected to be USD. The native price and native currency are always retained, the conversion is always recorded with the rate and its source and timestamp, and a converted value never overwrites an observed one.

The rate source and fixing convention, the relationship between the fixing and the observation timestamp, the treatment of stale rates, whether triangulation is permitted where no direct rate exists, and rounding conventions are unresolved launch parameters. Where no acceptable rate exists for an observation, the observation is excluded rather than converted at a stale or substituted rate.

Temporal Observation Rules#

Proposed initial frequency: daily. One observation per child per calculation date. Real-time and intraday values are not proposed, and nothing here should be built to support them.

Compute prices do not update on a market calendar. Seller pages change irregularly and often without a disclosed timestamp; marketplace prices can move continuously; quoted and contracted prices arrive episodically. The family therefore distinguishes four times on every observation: the observation timestamp, when Urdais retrieved the source; the effective timestamp, when the price became effective where the source discloses it; the calculation cutoff, after which no input may enter that date's value; and the publication timestamp.

A calculation date is a calendar date. Because the market does not close, weekends and holidays are ordinary calculation dates, and any systematic weekend or holiday effect in seller updating behaviour is a measurement characteristic to be quantified in validation rather than corrected by assumption. Look-ahead is prohibited: a value for a date uses no input whose retrieval time is after that date's cutoff, and a later-discovered price is applied prospectively or through the correction rules, never inserted into a published date.

The calculation calendar and cutoff#

The family owns the calendar. A child inherits it and may depart from it only by an explicit, versioned amendment stating why its instrument's market requires a different convention. The first child inherits it unchanged.

A calculation date is a UTC calendar date D, and its collection window is the half-open interval [D 00:00:00 UTC, D+1 00:00:00 UTC). The calculation cutoff for D is the instant D+1 00:00:00 UTC, exclusive: an observation at exactly that instant belongs to D+1. There is no separate end-of-day time to interpret and no local time zone anywhere in the rule.

The controlling timestamp is Urdais's own observation time: the moment the source's response was received and recorded, and for availability evidence the moment that evidence was re-observed. A source-effective time never controls eligibility; it is retained where disclosed and is frequently absent. A request that begins before the cutoff and completes after it produces an observation in the next window, because the evidence did not exist before the cutoff, and a response received after the cutoff carrying a source timestamp from before it likewise belongs to the next window. This is the look-ahead prohibition made mechanical: a value for D uses no input observed at or after D+1 00:00:00 UTC.

Collectors may retrieve and retry throughout the window. Where a source is observed more than once in a window, the observation that enters the calculation is the last complete reconfirmation before the cutoff: the latest retrieval in the window at which every eligibility-relevant field was re-established. Earlier retrievals are retained as evidence and are never averaged. A source not successfully observed at any point in the window is Source unavailable for D, and its participants leave that date's coverage under Missing Data. Retries are an operational matter and never extend the window.

The daily observation therefore represents the most recent eligible state of each participant successfully reconfirmed within the window for that date. It is neither a snapshot at one instant nor an average over the day.

Calculation begins at the cutoff and uses only inputs observed within the window. It is reproducible: re-running the calculation for D at any later time from the retained inputs yields the same value, and no wall-clock time enters the result except the cutoff itself. Validation follows under Data Quality Rules. Each observation carries its calculation date D, the cutoff as its as-of instant, the time it was calculated, and the time it was published, and these are the calculation and publication timestamps already required of every observation.

Publication follows validation as soon as it completes, and no later than the next cutoff. The publication deadline for D is D+2 00:00:00 UTC. A value released before the deadline has status Published. A value not released by the deadline has status Delayed until it is released, and when released it carries its calculation date D and its actual publication time, so a late print is never presented as timely and a user always sees the calculation date separately from the publication time. A value that cannot be produced for D is Unavailable for D, and nothing is filled in.

A percentage change is computed against the immediately preceding calculation date only. Where that date has no Published level, the change is withheld rather than computed against an earlier date; a child states the further conditions under which it is annotated or withheld.

Corrections are unchanged. An input later found wrong for D is handled under Corrections and Restatements: the recalculation uses the same window and the same retained inputs with the corrected one, the restated value carries the original and the reason, and the original is retained as Superseded. A percentage change is a derived quantity of two published levels and follows the correction status of its endpoints. The cutoff never moves.

Every calendar date is a calculation date. Saturdays, Sundays and holidays use the same window and cutoff. A price that has not changed is fresh if it was reconfirmed within the window; a price that was not reconfirmed within the window is not fresh however recently it last changed; and availability must itself be observed within the window, as Freshness and Staleness requires. No market calendar applies, because this market does not close.

The publication deadline is the one parameter here that may be tightened once validation runtimes are measured; it may be shortened by amendment and is not to be lengthened to accommodate a slow pipeline.

Freshness and Staleness#

A seller page that has not changed may mean the price is unchanged, or may mean nothing was published and nothing was checked. Many sources expose no update timestamp, so the family records what was actually established rather than inferring.

Each observation carries a freshness state: Reconfirmed, defined below; Newly observed, where an eligibility-relevant field changed at this retrieval; Carried, where the source could not be retrieved this cycle and a prior value is being relied upon, with its age; Source unavailable, where retrieval failed and no value is relied upon; and Withdrawn, where the seller no longer offers the instrument.

Reconfirmation is multidimensional#

An unchanged price does not establish an unchanged offer. A page can continue to display an hourly figure after the product has sold out, moved to a waitlist, changed configuration, raised its minimum quantity, or changed its terms. Because UCPI measures accessible offers, a freshness rule that checked only the price would certify as current precisely the observations most likely to have stopped being offers.

Reconfirmed therefore requires the current collection cycle to re-establish every eligibility-relevant field that can change, at minimum: the price; the availability state and its evidence; the instrument identity; the procurement mode; the quantity and topology requirements; the material commercial terms; and any mandatory fees.

Slowly changing reference data need not be re-retrieved every cycle where its version and effective interval remain valid. The family therefore distinguishes three freshness dimensions, which a child may govern with different limits: price freshness, availability freshness, and reference-data freshness.

Carry-forward#

A carried observation may enter a calculation only where the price is within its permitted carry age, the availability evidence is also within its permitted age, and no known withdrawal or configuration change exists. A recent price paired with stale availability evidence is not a current accessible offer, and the combination is ineligible rather than admitted on the strength of the fresher half.

Carry ages are unresolved and must be set from measured seller updating behaviour, separately per dimension where the evidence supports it. A carried observation is never presented as current: its age is recorded, the carried share is published, and a child whose carried share exceeds its gate is not published as a normal value. Beyond the maximum age the observation becomes ineligible and the participant drops out of coverage rather than persisting indefinitely at a stale price.

Seller-Level Reduction#

The unit of aggregation is never the offer. A seller publishing a hundred configurations must not carry a hundred times the influence of a seller publishing one; catalogue verbosity is an artifact of how a business structures its product page and carries no market information.

Eligible offers are therefore reduced to at most one seller-level observation per seller, per instrument, per region, per service tier, per procurement mode, per topology class, per calculation date. Because the child specification has already pinned the product, the offers competing within such a cell describe the same product from the same seller.

The proposed default rule is the lowest eligible offer price within the cell. For seller s with eligible offer prices P_s1 … P_sn in the cell:

p_s = min_j P_sj

The reasoning is that the specification has fixed the product, so a lower price for an identical specification is simply the price at which that seller will supply it, which is what an accessibility statistic is about.

This rule is provisional, not settled. Its known bias is recorded here rather than discovered later: a seller offering the specified product across several qualifying zones or catalogue variants has more opportunities to produce a low minimum than a seller offering it once, so breadth of catalogue can lower a participant's observation without any difference in the price it would actually charge for the same thing.

The first child must therefore empirically compare at least three rules before the family adopts one: the seller minimum defined above; the seller median across eligible offers in the cell, which is less sensitive to a single cheap zone; and a fixed or canonical selection on a variant dimension, for example a stable zone where one can be identified or the seller's smallest offered quantity, which removes the breadth asymmetry on that dimension entirely and may be combined with the minimum across the remaining variants. A capacity-weighted price within the cell is a fourth candidate wherever capacity is disclosed. The data may not vindicate the minimum, and the family does not assume it will.

Capacity-Source Identity and Collapse#

Seller-level reduction alone is insufficient, because several sellers can offer the same operator's capacity. Counting each as an independent participant would give one physical pool of capacity several equal weights, which is precisely the distortion the offer-level reduction exists to prevent, one level up.

The participant unit entering regional aggregation is therefore the capacity source:

The capacity source is the infrastructure operator where the underlying operator is reliably determinable, and the seller otherwise.

Collapse rule#

Where several eligible seller-level observations map to the same determinable infrastructure operator for the same child, region, service tier, procurement mode, topology class, and calculation date, they form one capacity-source observation.

The representative price for that capacity source is the lowest eligible seller-level price through which that operator's qualifying capacity can actually be obtained. The rule follows from what UCPI measures: the statistic is about the accessible price of the specified product, and if the same operator's capacity is obtainable more cheaply through one channel than another, the cheaper channel is the accessible price for that capacity.

Two alternatives were considered and are recorded, because the choice is not self-evident. Operator-direct precedence, using the operator's own price and ignoring reseller channels, has the merit that the operator's own terms are the primary commercial relationship; it was not adopted because a buyer who can obtain the same specified product more cheaply through a reseller faces the cheaper price, and ignoring that would misstate accessibility. Excluding reseller channels entirely was not adopted because resale is a genuine part of how this market is accessed.

The risk in the adopted rule is stated rather than hidden: a lower resale price may reflect subsidy, a different bundle, or a different commercial term rather than cheaper access to the same product. The family's protection is that the child specification must already have pinned the product, service tier, topology, and commercial terms before any offer is eligible, so a materially different commercial product should have been excluded earlier. Whether that protection holds in practice is an item for empirical validation, and the collapse rule may be replaced on that evidence.

Undetermined operators#

Where the infrastructure operator cannot be reliably determined, the seller remains the participant identity, the attribution status is recorded as Undetermined, and the undetermined share is published with every observation.

Operator identity is never inferred from matching or similar prices, from hostnames or network characteristics, from geography, from similar configurations, or from commercial speculation. An attribution requires evidence, and the evidence is recorded.

Attribution changes over time#

Discovering later that two sellers share an operator is new knowledge, not a historical error, and it must not silently rewrite published history. Every attribution records its evidence, the time the attribution became known, and its effective interval.

A newly learned attribution applies prospectively from the date it becomes known. A demonstrated historical error, where the attribution in force at the time was wrong on evidence then available, is handled under Corrections and Restatements. The two are distinguished on the record, because conflating them would let point-in-time discipline erode every time the market became better understood.

Regional Aggregation#

The regional value is the median across capacity-source observations within the canonical region, with each participating capacity source weighted equally.

For final aggregation participants k = 1 … N with representative accessible prices x_k, sorted ascending:

UCPI_r = x_((N+1)/2) for odd N

UCPI_r = [ x_(N/2) + x_(N/2+1) ] / 2 for even N

The even-N convention is the arithmetic mean of the two central ordered observations. This is a definitional choice rather than an empirical parameter, and the family fixes it now so that no child may resolve it differently. All inputs are unrounded normalized prices; a rounded value is never an input.

A median is chosen over a mean because the distribution of compute prices is not symmetric and contains legitimate extreme values at both ends: genuinely scarce capacity priced high, and marginal or distressed capacity priced low. A mean would let one such value move the published price substantially, while a median reports the middle of the market and is unaffected by how extreme the extremes are.

At exactly two participants that reasoning does not apply, and the family says so rather than leaving it to the arithmetic. The median of two observations is their mean; there is no middle distinct from the extremes, each participant is pivotal, and a change in either price moves the regional value by half that change. The family does not treat this as a reason to withhold the value, because two independent participants form a price where one does not, but it is a reason to label the value, and Publication Gates defines the market-breadth qualifier that does so. Three is the smallest count at which the median is an order statistic distinct from both extremes, and therefore the smallest count at which the reason for choosing a median holds.

Equal participant weighting is a deliberate answer to a question, not a convenience. It asks what the specified product costs across the participants who offer it. It does not ask what the typical available accelerator-hour costs, which would require capacity weights, nor what purchased compute cleared at, which would require transaction weights. The What UCPI Measures section states this limitation, and it must accompany the published value.

The full calculation sequence#

The sequence is deterministic and is applied in this order:

Raw compute offers → instrument, region, procurement-mode and service-tier eligibility → availability and freshness eligibility → seller-level candidate offers → seller-level reduction → operator attribution → capacity-source collapse where determinable → final aggregation-participant observations → regional median → UCPI child price observation.

Alternatives are retained for validation rather than dismissed: capacity weighting where capacity becomes reliably observable, transaction weighting where transaction evidence becomes sufficient, and trimmed or winsorized means. None may be adopted without evidence that the required inputs exist and are reliable.

Global Aggregation#

No global value is published in this version. The reasoning is in Geographic Framework: a global price is a regional weighting, no defensible weighting basis currently exists, and an arbitrary one would be embedded in the most visible number produced.

When a basis does exist, the family will define it by amendment. The candidates are not of equal standing.

Preferred where observable: observed regional available capacity, or observed regional transaction volume. Each measures the economic weight of a region in the market the benchmark describes.

A weak fallback requiring explicit justification: regional participant or seller counts. A count measures how fragmented a region's supply is at least as much as how important it is, so a region served by many small sellers would outweigh one served by a few large ones irrespective of the compute involved. It is recorded as a candidate only because it is observable, and adopting it would require a documented argument that fragmentation is an acceptable proxy in the specific case.

Equal region weighting remains rejected: it asserts a claim about regional importance that is certainly false. No global value is introduced by this version.

Until then, a user comparing regions does so explicitly, using regional series that each state their own coverage.

Weighting Philosophy#

Weighting is not a technical detail, and the family states its philosophy plainly: the economic question is chosen first, and the weighting follows from it.

Equal-participant weighting asks what the typical participating capacity source charges for the specified product. Capacity weighting asks what the typical available unit of capacity costs. Transaction weighting asks what purchased compute cleared at. Cheapest-available asks what the best obtainable deal is. These are four different questions with four different answers, and a benchmark that does not say which one it is answering is not a benchmark.

UCPI answers the first, because it is the question the available data can support honestly, and it publishes the composition diagnostics that record what is known about the participant set. Where data later supports the second, it will be added as a distinct series with its own name, not substituted silently into an existing one.

Robustness and Outliers#

Compute price distributions contain data errors, stale prices, reseller markups, scarcity premiums, and unusually cheap marginal capacity. Only the first two are defects. Price dispersion is market information, and the methodology must not destroy it in the name of cleanliness.

The family therefore prefers explicit validation rules over statistical clipping. An observation is excluded because a rule identifies it as invalid, for example a unit mismatch, a misidentified instrument, a region-mapping failure, a currency error, an expired quote, or an unresolvable source conflict, and the rule and its reason are recorded. An observation is not excluded merely for being far from the others.

Statistical dispersion measures are used as diagnostics that trigger review, not as automatic deletion. A robust dispersion measure such as the median absolute deviation may flag an observation for examination, and the outcome of that examination is recorded as either a validation failure with a reason or a confirmed genuine price that remains in the calculation.

The choice of a median as the central measure already provides the necessary resistance to a small number of bad values without deleting them. Trimming and winsorization are recorded as alternatives requiring evidence that they improve the measure, and no trimming fraction is adopted here.

Coverage and Composition#

A price derived from too few participants has stopped representing a market. Benchmark design principles require that sample adequacy, market size, and the distribution of activity among participants be considered, and the family makes that concrete by publishing diagnostics with every observation rather than assessing them only internally.

Equal-weight concentration measures are not published, because they carry no information. With N equally weighted participants every weight is 1/N, so the largest participant weight is 1/N and the effective participant count computed as the reciprocal of the sum of squared weights is exactly N. Reporting either as a concentration measure would restate the participant count while implying that something about market structure had been measured. Economic capacity concentration is unknown wherever capacity is not observable, and saying so is more honest than publishing a tautology in its place.

Where capacity shares are independently and reliably observable, a genuine capacity concentration measure may be published as a separate diagnostic, since those shares can actually vary. Capacity shares are never fabricated in order to produce one.

Each published observation carries: the final aggregation-participant count; the eligible seller count; the count of participants with a determined operator and the undetermined-operator share; the reseller and operator-direct mix; the eligible offer count before reduction; the price distribution as tenth, fiftieth, and ninetieth percentiles and the interquartile range; the source-quality distribution; the observation-type distribution; the availability evidence-grade distribution; the fresh and carried shares; the availability and pricing gap described in Availability and Executability; the composition change described below; and capacity coverage where it is actually measurable.

These exist so a user can tell the difference between a price supported by fifteen participants and one supported by two, without inferring it from the number itself.

Market breadth#

Every published regional value carries a market-breadth qualifier with exactly two values. Minimum means the value rests on exactly two independent capacity sources, the smallest number that constitutes a market under this family. Normal means it rests on three or more. The qualifier is derived from the final aggregation-participant count and from nothing else; it is not a status, it never replaces the participant count, and it accompanies the value wherever the value appears. The rule that produces it is in Publication Gates.

At Minimum breadth two further disclosures are mandatory. First, that every participant is pivotal: removing either would leave no market price, and a change in either price moves the value by half that change. Second, that the dispersion diagnostics are withheld: with two observations every percentile and the interquartile range are interpolations between the two participant prices, and publishing them would publish those prices individually under the description of a distribution. The level, the participant count, and the qualifier are published. Whether the two prices may additionally be shown is a data-rights question for each contributing source rather than a methodology question, and a user should understand that the level itself, combined with public knowledge of one participant's price, determines the other.

Source concentration#

Economic breadth is counted in capacity sources, because that is the unit whose prices are set independently. It is not counted in data sources. Twenty independent hosts observed through one lawful and reproducible venue are twenty price setters; two vertically integrated clouds observed through two interfaces are two. The number of interfaces measures how Urdais observes the market, not how the market is structured.

Observing a market through few interfaces is nonetheless a real dependency, on that venue's availability, sampling, ordering, and terms, and it is disclosed rather than gated. Each published observation carries the contributing-source count, the number of distinct source interfaces through which the final participants were observed, and the largest-source participant share, the share of final participants observed through the single most-used interface. Neither is a publication gate. A child may not adopt a minimum interface count as a gate without a documented argument that interface concentration, rather than participant concentration, is what would make its value unrepresentative.

Percentile population#

Percentiles and the interquartile range are computed over exactly the final aggregation-participant observations that enter the regional median, after seller-level reduction and any determinable capacity-source collapse. The headline and the published distribution therefore describe the same population.

Where any dispersion measure is computed over a different population, for example across raw offers rather than participants, it is labelled with that population explicitly and is never presented alongside the headline as though it described the same thing.

The percentile interpolation convention must be deterministic and stated in the child specification. Percentile reporting at small participant counts is governed by the child's publication and diagnostic requirements, except that at Minimum breadth the family withholds percentiles and the interquartile range for the reason given under Market breadth above; no other minimum count is invented here.

Composition change#

A UCPI value can move because the set of participants moved, not because any price moved: a participant may enter or leave, a source may become unavailable, availability may change, or an operator attribution may change and collapse two participants into one. A user reading a change in the level is entitled to know whether it reflects prices or composition.

Each observation therefore records the participants added since the previous observation, those removed, and any change arising from capacity-source collapse. This is a disclosure requirement, not a matched-participant index: constructing a matched-participant price change analytic is recorded as possible future work and is not attempted here.

Missing Data#

Nothing is fabricated. Where inputs are absent, the family's response is to narrow the claim, not to fill the gap.

A participant that disappears from coverage leaves the calculation, and the change in participant count is published. A source that cannot be retrieved produces a carried observation within the freshness limit and an excluded one beyond it. A region that loses coverage below its gate produces a value that is not published as normal. Where only one eligible participant remains, no regional value is published, because a single participant's price is not a market price; it may be published as a participant-level diagnostic. Where exactly two remain, the value is published at Minimum breadth under Publication Gates, never silently as a normal value. An observation whose availability cannot be established is ineligible. Sources that conflict irreconcilably produce a Conflicted observation that is withheld and escalated. A price shown as requiring a quote is not a price.

Carrying an observation forward is always visible. Its age, its share of coverage, and its effect are published; a silent carry is prohibited in every case.

Publication Gates#

A child publishes a normal value only when the measure still represents its market. The gates are defined here as concepts; their numerical values are unresolved and must be set from measured coverage, not chosen so that a series publishes.

The gates are: a minimum count of participants with a determined operator, or equivalently a maximum undetermined-operator share; a maximum carried or stale share; a maximum share of observations resting on the weakest availability evidence grade; a minimum source-quality standard; and a maximum share of observations with unresolved bundle or comparability status. Equal-weight concentration measures are not used as gates, for the reason given in Coverage and Composition. The participant count is not a numerical gate above the structural floor. It governs the market-breadth qualifier instead, as set out below; that is a deliberate change from the first draft, which listed a minimum participant count among the numerical gates, and it is recorded in Methodology Versioning.

Participant count and market breadth#

The participant-count rule is structural rather than parametric. It follows from the arithmetic of the median rather than from any coverage measurement, and it applies now.

• No eligible participant. The region has no value. Unavailable, with the condition named as no eligible participant.
• Exactly one. A single participant's price is not a market price. Unavailable, with the condition named as single participant. The price may be published as a participant-level diagnostic.
• Exactly two. The value is published with status Published and market breadth Minimum. Two independent capacity sources are the smallest set in which a price is formed by more than one party, so this is a market price, and it is the thinnest one the family will describe as such. The median of two is their mean and each participant is pivotal, so the value is published only with the disclosures required under Coverage and Composition.
• Three or more. The value is published with market breadth Normal, subject to every other gate. Three is the smallest count at which the median is an order statistic distinct from both extremes, and therefore the smallest count at which the reason for choosing a median holds.

The boundaries fall at two and three because those are the counts at which the statistic changes character, not because either was found adequate in data. A larger count is statistically preferable at every level, and the family does not invent a threshold on that ground. The participant count is published with every value so that a user may apply their own, and the qualifier states which side of the structural boundary the value sits on.

Independence at the floor. The two participants at Minimum breadth must be independent capacity sources under Capacity-Source Identity and Collapse: distinct determinable operators or, where operators are undetermined, distinct sellers by legal identity, neither under common control with the other and neither observed on evidence to be reselling the other's capacity. Two regions of one provider, two brands or tiers of one company, two records of one seller through two interfaces, and a reseller together with its disclosed underlying operator each count once. Where operators are undetermined, independence rests on seller identity and carries the standing limitation that undisclosed shared operation cannot be detected; at two participants that limitation is at its most consequential, and the undetermined-operator share is published with the value as always. Independence is never inferred from prices, in either direction.

Transitions are deterministic. The qualifier is computed from the participant count on each calculation date, with no hysteresis and no waiting period. A region moving from three participants to two publishes at Minimum breadth that day; moving from two to one produces Unavailable; returning to two or to three restores the corresponding state on the day it occurs. Every such move is a participant entry or exit and is disclosed under the composition-change rule, so a user can see that breadth changed and why. A waiting period was considered and rejected: it would either publish Normal breadth over a region that has two participants or withhold a value from one that has three, and each misdescribes the market on that date.

All other gates apply at every participant count. Minimum breadth discloses how many parties formed the price; it exempts nothing.

Failing a gate produces a Delayed or Unavailable observation with the failing gate named. It never produces a value computed from a residual set and presented as if complete.

Published Values#

A UCPI child publishes a price level, not a rebased index. The published quantity is currency per accelerator-hour for the specified product.

Rebasing to an arbitrary base of one hundred was evaluated and rejected. Official price indices rebase because they chain heterogeneous matched specifications over long periods, where no single unit price is meaningful across the whole series. UCPI's unit is homogeneous by construction, because the child specification pins the product, so the level carries direct economic meaning that a buyer can act on, and converting it into an index number would discard that meaning and add an arbitrary base date without adding information.

Consistent with Urdais convention, percentage change is the headline change signal. The currency difference between two levels is a legitimate analytical quantity and is not the headline; the two are labelled distinctly and never presented interchangeably.

Each observation publishes the price level; percentage changes over defined periods computed from unrounded values; the instrument, region, procurement mode, and service tier that define it; the market-breadth qualifier; the coverage and concentration diagnostics; the as-of date; the status; and the family methodology version, child specification version, and parameter set.

Dispersion is published alongside the level rather than as a separate product. Retaining the percentile distribution and cross-participant spread preserves the information that future scarcity, fragmentation, and regional basis analytics would require, without committing to those products now.

Historical Integrity and Lineage#

Every published value must be traceable through:

UCPI Child Observation → UCPI Family Methodology Version → UCPI Child Specification Version → Final Capacity-Source Observations → Seller-Level Observations → Normalized Offers → Raw Compute Offers → Seller and Marketplace Sources

For any historical date it must be possible to answer which instrument definition applied; which sellers and operators were eligible and which were excluded with which reason; how sellers resolved into capacity sources and under which attribution evidence; what offers existed and on what commercial terms; what region mapping and service-tier classification applied; what availability evidence supported each observation; what currency conversion was used; what aggregation rule applied; which methodology and specification versions governed the result; and what information was actually available at that time.

Retain raw offers, rejected and superseded observations, source retrieval evidence and vintages, and the mapping and classification decisions with their effective intervals. A historical value must never change because a mapping, a taxonomy, or a parameter changed later.

Any series computed for dates before a child's first live publication is labelled reconstructed research history, carries its limitations, and is never presented as live history. Reconstruction is feasible only where historical source evidence was lawfully retained; because seller pricing pages are not archival and rarely expose history, a reconstructed UCPI history may not be feasible at all, and that limitation is to be established rather than assumed.

Corrections and Restatements#

Distinguish, each with its own record: an input correction, where a price, currency, instrument identity, region mapping, availability state, or commercial term was wrong; a classification error, where procurement mode or service tier was misassigned; a duplication error, where one operator's capacity was counted more than once; a staleness error, where a withdrawn or stale offer was treated as current; a calculation error, where these rules were misapplied; and a publication error.

These are not: a genuine price change, a participant entering or leaving coverage, or a methodology amendment with a future effective date.

Handling follows the Urdais convention. Inside the correction window, restatement, with each restated value carrying the original value, the reason, the detection and republication timestamps, and the status Corrected, and the original retained as Superseded. Outside the window, history is ordinarily left as published and the correction applied prospectively with a notice. An exceptional historical restatement may be made only where an error materially compromises the integrity, interpretability, or reproducibility of the series; it must be publicly documented, identify the affected observations, preserve the originals, record the reason and approval, and be published as a distinct event. This is a high bar and not a licence to tidy small errors.

The correction window and materiality threshold are unresolved. History is never silently overwritten.

Status Model#

Observation status, one per published date: Published, Delayed, Unavailable, Corrected, and Superseded, with the same meanings used across Urdais outputs.

Input status is separate and never collapsed into the headline status: Valid, an observation meeting every eligibility rule; Stale, resting on a carried price within the freshness limit; Ineligible, excluded by a named rule with its reason; Unavailable, where the source could not be retrieved; and Conflicted, where sources disagree irreconcilably and the observation is withheld.

Both are published, because a value resting on many carried observations is making a weaker claim than one resting on freshly confirmed ones.

Market breadth is not a status. A value at Minimum breadth has status Published, and the qualifier accompanies the status without substituting for it. The word Limited is reserved for the availability state of an input, defined in Availability and Executability, and is not used for breadth, so that a value's breadth and the availability composition of its inputs can never be read as one thing.

Conceptual Data Requirements#

These describe information the family requires. They are not database tables, and no schema is defined here.

• ComputeInstrument: vendor, family, model, architecture generation, form factor, memory capacity and type, full-device or fractional designation, the identity-defining attribute set, and effective interval.
• Seller: legal identity, seller type, marketplace relationship where applicable, geographic scope, and effective interval.
• InfrastructureOperator: identity where determinable, the evidence establishing it, the relationship to sellers, and an explicit undetermined state.
• Marketplace: venue identity, whether it is itself a seller, and how third-party sellers are identified on it.
• Region: canonical region identity, the seller region labels mapped to it, the finer facility identifiers retained, the mapping evidence, and effective interval.
• ServiceTier: tenancy, interruption policy, uptime commitment, provisioning model, support and replacement commitments, intra-node interconnect, inter-node fabric, storage, host resources, virtualization model, compliance attributes, and, for each, whether the child treats it as a requirement or as metadata.
• ComputeOffer: the raw offer as defined in Raw Compute Offer, retained unmodified.
• NormalizedOffer: the derived price on the normalized unit and currency, the conversions applied with their inputs, the preserved commercial constraints, and the normalization status.
• PriceComponents: the usage-proportional mandatory charges entering the normalized figure, the fixed or one-time mandatory charges held separately with any contractual amortization applied, any mandatory minimum spend, the excluded usage-dependent charges, and the tax basis.
• AvailabilityEvidence: the availability state, the evidence grade, the evidence itself, and the time it was established, recorded separately from price evidence so the two can age independently.
• EligibilityAssessment: the eligibility outcome, the rule applied, the exclusion reason where excluded, the source quality grade, and the observation type.
• SellerPriceObservation: the seller, the seller-level representative price, the eligible offer set considered, the seller-reduction rule applied, and the input status.
• CapacitySourceObservation: the capacity-source identity, being the operator where determined and the seller otherwise; the operator attribution status and its evidence; the contributing seller-level observations; the collapse rule applied; the final representative accessible price; and the input status. This is the object over which the regional median and all published percentiles are computed.
• UCPIChildSpecification: the instrument definition, region definitions and published series, service-tier requirements, the single procurement mode, instrument-specific normalization, the parameter set, the publication gates, and the child's version.
• UCPIObservation: child identifier, region, price level, percentage changes, the market-breadth qualifier, the full diagnostic set including the composition change since the previous observation, status, family and child versions, parameter set, and the calculation and publication timestamps.
• SourceReference: source identity, tier, retrieval evidence and timestamp, licensing basis, and retention rights.
• FXObservation: currency pair, rate, source, fixing time, and status.
• MethodologyVersion and CorrectionRecord, with effective dates and change records.

Source Specification#

Production requires: seller pricing interfaces, preferring machine-readable and executable ones; marketplace listing data with availability; seller contract and quote documents where lawfully usable; transaction or invoice evidence where lawfully obtainable; hardware reference data sufficient to resolve instrument identity; region and facility reference data; currency rates; and licensed specialist datasets where their methodology is disclosed.

Scraping is not a production data model where it is not permitted. Whether a given source may be collected, stored, and retained is a licensing and terms question that must be resolved per source before that source contributes to a published value, and a source that cannot be lawfully retained cannot support a reproducible benchmark.

Data Quality Rules#

Before a value is published, at minimum: every observation resolves to exactly one instrument, seller, region, service tier, and procurement mode; instrument identity is established from stated hardware attributes rather than from a product name alone; units and billing granularity are validated and conversions are within their permitted class; currency and tax basis are established; mandatory fees are included and optional charges excluded; minimum quantity, topology, and commitment are recorded and consistent with the specification; availability state and evidence grade are recorded; freshness state and age are within limits; region mapping is current and evidenced; no operator is counted twice where attribution is determinable; seller-level reduction applied exactly one rule; the aggregation produced a finite positive price; every diagnostic is computed and every gate evaluated; and the lineage record is complete.

A failed critical check blocks the affected observation, or the published value where it concerns coverage or aggregation. It never authorizes imputation. Numerical tolerances require production data and are unresolved.

Licensing and Reproducibility#

UCPI's reproducibility claim is bounded by what may lawfully be retained. Seller pricing pages, marketplace listings, and specialist datasets carry terms that may restrict collection, storage, redistribution, and historical retention, and those terms differ by source and change over time.

An observation is independently reproducible only where the evidence behind it may be retained and re-examined. Where retention is not permitted for a source, the limitation is recorded and the affected observations are marked as not independently reproducible. Securing lawful collection and retention for the intended source set is a launch requirement and a genuine constraint on achievable coverage; coverage that cannot be licensed is a published limitation, not an assumption.

Methodology Versioning#

The family carries its own version, independent of any child. Each published observation identifies the family methodology version, the child specification version, and the child parameter set.

A change to family rules is announced with a prospective effective date, a documented rationale, and an impact assessment, and does not alter observations before that date. A change to a child specification affects that child only. Editing this public page is not a production change.

Version history: 0.1.2-draft, 13 September 2026, calculation calendar. The family now owns the daily timing convention that the first two drafts left unresolved: a calculation date is a UTC calendar date whose collection window is the half-open interval from midnight to the next midnight UTC, the cutoff is the next midnight exclusive, Urdais's own observation time controls eligibility and a source-effective time never does, a request completing after the cutoff belongs to the next date, collectors may retry within the window and the last complete reconfirmation before the cutoff is the observation, calculation begins at the cutoff over window inputs only, publication follows validation with a deadline at the following cutoff after which the status is Delayed, percentage change is computed against the immediately preceding calculation date only, corrections follow the existing rules unchanged, and every calendar date including weekends and holidays is a calculation date. A child inherits the calendar and may depart only by explicit versioned amendment. The family also confirms, in response to the first child's request, that its third seller-reduction alternative, a fixed or canonical selection, extends to a fixed selection on the quantity dimension; the wording is clarified in Seller-Level Reduction. No production effective date. 0.1.1-draft, 13 September 2026, market-breadth amendment. The participant-count publication gate above the structural floor is resolved structurally rather than left as a numerical parameter: a region with exactly two independent capacity sources publishes with a mandatory market-breadth qualifier of Minimum, and a region with three or more publishes at Normal breadth, subject to every other gate. The amendment records why the median of two is their mean and why three is the first count at which the median's rationale holds; defines independence at the floor in terms of the existing capacity-source identity, with seller identity clarified as legal identity; makes breadth transitions deterministic without hysteresis; withholds dispersion diagnostics at Minimum breadth because they would disclose the two participant prices individually; adds the contributing-source count and the largest-source participant share as disclosed rather than gated source-concentration diagnostics, deciding that breadth is counted in capacity sources and not in interfaces; and reserves the word Limited for availability. The rationale and the scenario analysis that prompted it are recorded in an internal architecture document. No production effective date. 0.1.0-draft, 12 September 2026, initial research-backed proposal, amended in review on 12 September 2026 before merge: the published statistic named exactly as the regional median of each participant's lowest qualifying accessible offer rather than as a broader typical-offer phrase; the capacity source introduced as the aggregation participant with a deterministic operator-collapse rule and point-in-time attribution discipline; source quality separated from observation type, with concluded transactions made corroborative by default rather than automatically headline inputs; equal-weight concentration measures withdrawn as tautological and the percentile population fixed to the final participants; the even-N median convention fixed; reconfirmation made multidimensional with separate price and availability freshness; marketplace removed from the procurement taxonomy and price formation separated from it; tenancy moved from hardware identity to the service framework; fixed mandatory charges separated from usage-proportional ones with amortization permitted only against a contractual horizon; the advertised-versus-accessible gap stated as an observation rather than a causal scarcity signal; seller-count global weighting qualified as a weak fallback; and a source description of bounded carry-forward corrected. No production effective date.

Child Specification Requirements#

A child specification is approved separately and must state: the instrument identity and every identity-defining attribute value; the service-tier requirements and which dimensions are requirements rather than metadata; the single procurement mode; the node topology class and minimum quantity; the host-resource envelope treated as ordinary; the canonical regions and which series are published; the minimum availability evidence grade; the freshness limit; the seller-level selection rule; the publication gates with numerical values; the currency and rate convention; the instrument-specific normalization, if any; the named heterogeneities the child accepts; and its own version and first publication date.

Naming#

Three identifiers are distinguished. The public display name is what a user sees, of the form UCPI followed by the instrument designation, for example UCPI-H100-SXM. The canonical instrument identifier carries every identity-defining and commercial attribute, including those not shown publicly, and is what a calculation and a lineage record reference. The specification version identifies the parameter set in force.

A public display name is therefore not sufficient to identify what is being measured, and a published value must always appear with its region, procurement mode, and service tier rather than relying on the name to convey them. Region, procurement mode, and service tier are dimensions of a child rather than components of its display name, so that a user filtering by region compares like with like within one specification; whether any of them should additionally appear in a display name is a presentation decision for the child, not a methodology decision. The concrete naming details are unresolved and settle with the first child.

Open Questions / Empirical Validation Required#

No production value may be published under this draft. The following require real observations, licensing decisions, or testing, and are recorded rather than resolved by assumption.

• Canonical region taxonomy: the region set, the mapping from seller labels, and whether offers within a proposed region are plausibly substitutable.
• Global weighting basis: whether regional capacity, transaction volume, or another defensible basis becomes available, and what a global series would measure if it did.
• Seller-reduction rule: compare the seller minimum, the seller median, and a fixed or canonical-zone selection, with a capacity-weighted rule where capacity exists, and measure the multi-zone breadth bias, meaning how much a seller's catalogue breadth lowers its observation independently of what it would charge.
• Operator attribution and capacity-source collapse: what share of offers permits reliable operator determination; the measured price difference between operator-direct and reseller channels for the same capacity; how often collapse changes the participant count; and whether the adopted lowest-accessible-channel collapse rule admits materially different commercial products in practice.
• Transaction relevance age: how quickly a concluded transaction stops evidencing a current accessible offer, and whether specification-matched contemporaneous transactions are ever available in sufficient number to enter the headline at all.
• Transaction-price sibling feasibility: whether a separate transaction-price series could be constructed, and what it would measure.
• Composition effects: how much of observed period-to-period movement is participant entry and exit rather than price change, and whether a matched-participant analytic is worth constructing.
• Regional central measure: whether the median is the right central measure against alternatives, tested on observed distributions rather than assumed.
• Publication gates: the numerical values for the determined-operator count or undetermined share, carried and stale shares, source-quality standard, and evidence-grade composition. The participant count is resolved structurally in Publication Gates and is no longer a numerical gate.
• Availability evidence standard: what grades are achievable across sellers, what share of the market can support the stronger grades, and what minimum a child can require without emptying itself.
• Freshness limits: measured seller updating behaviour, including any weekend and holiday pattern, and the maximum defensible carry age separately for price, for availability evidence, and for slowly changing reference data.
• Service-tier requirements: which dimensions measurably move price for a given instrument, and therefore which must be requirements rather than metadata.
• Bundle envelope: the observed dispersion of host resources for a given instrument, and whether a defensible ordinary envelope exists.
• Topology and quantity: the distribution of minimum quantities and topology classes, and whether separate series are needed.
• Operator attribution: what share of offers permits operator determination, and the measured extent of undetected duplication.
• Source tiers and licensing: which sources are lawfully collectable and retainable, what tier each supports, and the coverage achievable under those constraints.
• Currency conventions: rate source, fixing, staleness tolerance, and triangulation.
• Correction window and materiality threshold.
• Reconstructed history feasibility: whether any point-in-time history can be built at all under available retention rights, without survivorship or look-ahead bias.
• Capacity weighting feasibility: whether capacity is ever observable reliably enough to support a distinct capacity-weighted series.

Validation programme#

Before a child launches, test on real point-in-time observations: seller and operator coverage by region; instrument-identity consistency and how often a product name alone would mislead; the achievable availability verification rate and evidence-grade distribution; the procurement-mode and service-tier distribution; bundle heterogeneity; the distribution of minimum quantities and topologies; normalized price dispersion within and across sellers and regions; disagreement between sources for the same product; detectable seller and operator duplication; stale-price and revision frequency; outlier behaviour and how often statistical flags correspond to real defects; regional basis; participant entry and exit and how much of period-to-period movement it explains; whether capacity data exists in usable form and whether a genuine capacity concentration measure can be computed; the sensitivity of the published value to the seller-reduction rule, the capacity-source collapse rule, the central measure, and the eligibility boundaries; weekend and holiday updating behaviour; and historical source-retention feasibility.

Parameters must not be chosen to make a child publish. A child that cannot pass defensible gates is not launched.

Research Precedents#

Primary sources retrieved and read on 12 September 2026 are summarized below. Each states what problem it informs and what this family adopts, modifies, or rejects. Market observations are recorded as evidence of market structure, not as a provider list, and no price quoted here is a Urdais value.

IOSCO, Principles for Financial Benchmarks#

The Principles for Financial Benchmarks, Final Report, July 2013, is the closest institutional precedent for administering a benchmark over a market without a consolidated tape. Principle 6 requires that benchmark design "achieve, and result in an accurate and reliable representation of the economic realities of the Interest it seeks to measure", taking into account the adequacy of the sample, the size and liquidity of the market, and "the distribution of trading among Market Participants (market concentration)". Principle 7 requires that data be "formed by the competitive forces of supply and demand" and "anchored by observable transactions entered into at arm's length", while clarifying that the principle "does not preclude the use of non-transactional data for such types of indices that are not designed to represent transactions and where the nature of the index is such that non-transactional data is used to reflect what the index is designed to measure". Principle 8 sets a hierarchy running from concluded arm's-length transactions in the underlying interest, through transactions in related markets, to "Firm (executable) bids and offers", and then other market information, and it notes that "there might be circumstances (e.g., a low liquidity market) when a confirmed bid or offer might carry more meaning than an outlier transaction". Principle 9 requires that the exercise of judgment be explained, while noting that a benchmark "based exclusively on executable quotes as contemplated by Principle 7 would not need to explain in each determination why it has been constructed with executable bids or offers, provided there is disclosure in the Methodology". Principle 12 requires the published methodology to state input selection, "priority given to certain data types, minimum data needed to determine a Benchmark", and the procedures governing determination under market stress.

Retrieval note. The publisher's canonical URL, cited above, refuses automated clients, returning an access challenge to plain requests and to browser navigation alike on 12 September 2026. The document was read in full from an Internet Archive capture of that canonical URL, and every quotation and principle number above is from that document. A reader with a browser session may reach the canonical link directly.

Adopt: the ranked data hierarchy with executable offers admitted as a legitimate primary input; the requirement that a benchmark built on executable offers disclose that fact in the methodology, which is why What UCPI Measures states the target and the proxy explicitly; the treatment of market concentration as a design consideration, made concrete as published diagnostics; the requirement to state minimum data needed, made concrete as publication gates; and the point that a confirmed offer may outrank an outlier transaction, which is why the hierarchy governs admissibility rather than mechanically overriding aggregation. Modify: the hierarchy is rewritten in compute-market terms, since the compute equivalents of transactions and firm quotes are invoices, executable programmatic offers, and marketplace listings rather than dealer submissions. Reject: nothing material; this is a principles framework rather than a rule set, and it is used as such.

U.S. Bureau of Labor Statistics, Producer Price Index for Data Processing and Related Services#

The BLS factsheet for NAICS 518210, dated 7 November 2008, describes how an official statistical agency prices a heterogeneous computing service. Prices come from a probability-selected actual contract, companies provide "net transaction prices for a specified service" monthly, and "the price-determining characteristics are held constant while the service is repriced", with the same selected contract repriced each period. The page was retrieved and read on 12 September 2026 through a fetch tool; the canonical URL refuses other automated clients.

Adopt: the matched-specification principle, which is the foundation of the parent-child architecture. Holding price-determining characteristics constant and repricing the same specified thing is exactly what a child specification does, and it is the reason UCPI pins the product before comparing prices rather than averaging across whatever is labelled similarly. Modify: UCPI cannot reprice one selected contract per seller, because it observes public offers rather than surveying billed contracts; it holds the specification constant across sellers instead of across time within one contract. Reject: the rebasing to an index number. An official index chains matched specifications over decades where no single unit price is meaningful across the series; UCPI's unit is homogeneous by construction, so it publishes the price level, as set out in Published Values. The factsheet is also notably silent on quality change, which for accelerated compute is the dominant difficulty, and it offers no guidance there.

Silicon Data, GPU rental price indices#

The provider's H100 rental price index page and its index-construction article, dated 12 March 2026, describe the closest existing compute-price benchmark. The index is published daily in USD per GPU-hour under a neo-cloud ticker, drawn from neo-cloud providers, hyperscalers, colocation markets, and private rental platforms, with observations "standardized for machine specs, rental terms, platform performance, and geography, statistical outliers removed, and each day's value independently validated". The construction article describes roughly one hundred and fifty thousand daily verified pricing records across forty to fifty countries and regions and fifty to one hundred platforms, covering lease types from on-demand and spot to reserved terms of one to sixty months, and states that the provider "uses a proprietary framework to normalize the prices of heterogeneous physical units to a standard configuration". Neo-cloud and hyperscaler readings are kept separate.

Adopt: the unit of USD per GPU-hour; the recognition that normalization must address specification, term, and geography rather than price alone; and the segmentation of structurally different provider types into separate readings rather than one blended number, which supports the family's decision to separate procurement modes and to keep regions distinct. Modify: UCPI separates by procurement mode and region as specification dimensions rather than as post-hoc segments. Reject: the proprietary normalization framework, and the removal of statistical outliers as a construction step. The normalization framework is where the entire economic content of a compute benchmark sits, and a framework that is not published cannot be independently assessed or reproduced; that is precisely the gap this family is written to close. Automatic outlier removal is rejected for the reasons in Robustness and Outliers: in a scarcity-driven market, extreme prices are frequently the most informative observations, and they should be validated against rules rather than deleted for being extreme. Finding of fact worth recording: the closest existing compute-price benchmark does not publish a reproducible methodology, so compute-specific methodological precedent is weak, and this family rests mainly on general benchmark-administration principles and on directly observed market structure.

SemiAnalysis, H100 price index#

The H100 daily and contract pricing index collects prices "from hyperscale clouds, neoclouds, GPU marketplaces, and rental routers/aggregators" combined with "analyst-run contract-pricing surveys", across contract lengths from on-demand to five years. It is published hourly, constructed as "a composition-jump-resistant weighted mean of rental prices" with hyperscale clouds weighted more heavily than marketplaces, applies a "per-step move cap" circuit breaker in which "an input printing an implausible jump is rejected rather than passed into the level", and carries a source's "last accepted price ... forward for a limited window (forward-fill), so no index is computed on a constantly shifting subset of sources". The document states that the window is limited but does not disclose its length.

Adopt: the recognition that composition change, rather than price change, is a primary risk in a compute index, which is why this family fixes the unit of aggregation at the capacity source and publishes composition diagnostics; and the use of analyst survey evidence for contract pricing, which corresponds to the quote and transaction tiers of the source hierarchy. Modify: weighting provider types differently is a defensible answer to a different question, and the family instead fixes the question first and weights sellers equally within a specification, publishing the diagnostics that reveal the difference. Adopt in principle: bounding any carry-forward to a limited window, which corresponds to the age limits in Freshness and Staleness, and the stated reason for it, that an index should not be computed on a constantly shifting subset of sources. Reject: the per-step move cap. A move cap suppresses exactly the large genuine price movements a compute benchmark exists to report, and it cannot distinguish a data error from a real scarcity event; the family uses validation rules with recorded reasons instead. Modify: UCPI requires the carry window to be published rather than merely stated to exist, applies separate limits to price and availability freshness, and publishes the carried share with every observation, because a carried price whose age is not disclosed cannot be assessed by a user.

Observed market structure#

The following were retrieved directly on 12 September 2026 as evidence of how the market is actually structured. They are not a provider list, no seller named here is proposed for inclusion, and the figures are illustrative of structure only.

A major GPU cloud's pricing page presents the same accelerators under two differently operated tiers on one page, with per-hour and per-second billing toggles, and lists three distinct H100 variants at three distinct prices, an SXM variant with eighty gigabytes of accelerator memory, one hundred and twenty-five gigabytes of host memory and twenty virtual CPUs at $3.49 per hour; a PCIe variant with eighty gigabytes of accelerator memory but one hundred and eighty-eight gigabytes of host memory and sixteen virtual CPUs at $2.89; and an NVL variant with ninety-four gigabytes at $3.19. A hardware-partitioned product appears in the same catalogue alongside full devices. This establishes that an accelerator model name does not identify a product, that form factor and memory are identity-defining, that the host bundle varies across products from a single seller, that service tier varies within a seller, and that fractional products sit in the same catalogues as full devices.

Another provider's pricing page lists an H100 SXM instance at $3.99 per GPU-hour bundled with two hundred and eight virtual CPUs, approximately eighteen hundred gibibytes of host memory and twenty-two tebibytes of storage, alongside a cluster product quoted per GPU-hour at three different prices for sixteen, sixty-four, and two hundred and fifty-six or more accelerators on terms of two weeks to one year, with reserved capacity available only by contacting the seller, and prices stated as exclusive of applicable sales tax, value-added tax, and goods-and-services tax. This establishes that host bundles differ by an order of magnitude between sellers for the same accelerator, that sellers price quantity tiers explicitly so per-accelerator division across topology classes is not neutral, that reserved pricing is frequently quote-only, and that tax basis differs and is sometimes stated and sometimes not.

A marketplace's pricing page publishes, for each accelerator, both a lowest available price and a median across the platform: for an H100 SXM, $1.73 and $2.02; for an H100 PCIe, $1.87 and $2.61; for an H200, $1.97 and $4.00. It describes prices as "set by supply and demand across 40+ data centers" across on-demand, interruptible, and reserved modes, and shows at least one accelerator with an approximate hourly figure labelled as a recent-market value beside the statement that there are no current offers. This establishes, from the market's own publication, that the cheapest available price and the typical price are materially different statistics, differing here by seventeen to one hundred and three percent on a single day, which is the empirical basis for rejecting cheapest-available as the economic target; and that prices are displayed for products with no current availability, which is the empirical basis for the executability requirement.

Taken together with the figures above, a single accelerator designation spans roughly $1.73 to $3.99 per accelerator-hour across three retrieved sellers on one day, under different tenancies, minimum quantities, bundles, and service tiers. That spread is not a price signal. It is the measurement problem this family exists to solve.

Sources sought and not retrieved#

Commodity price-assessment methodology, in which an administrator assesses a physical commodity price from a fragmented over-the-counter market using a defined standard specification and firm, repeatable bids and offers within a defined assessment window, is an obvious adjacent precedent for the executability and standard-specification questions. The primary methodology guide could not be retrieved on 12 September 2026, and nothing is attributed to it here. It should be obtained and assessed before this family is approved for production.