The reference price, defined precisely.

TCI Methodology — v0.3.0-dev

TCI (The Compute Indices) is a daily reference price for renting AI compute delivered from data centres physically located in the EU/EEA. Headline series: TCI-CRI-H100, in USD per GPU-hour with a EUR companion at the ECB reference rate.

TCI is a price-transparency benchmark, not a settlement benchmark. Every print is reproducible by any third party from public sources using the published code. It is not transaction-based and must not be referenced in a financial contract. The conditions that would have to be met before settlement use is credible are published in §9.

1. Unit definition

One TCI-CRI-H100 unit = one NVIDIA H100_SXM 80GB GPU-hour, on_demand (no term commitment), datacenter-hosted, delivered from the EU/EEA, per-GPU, ex-VAT, excluding storage and metered egress.

A class prices its reference variant only. Cross-variant normalisation by assumed factors was removed in v0.3.0: H100 SXM and H100 PCIe are different products, and H100→H200 normalisers disagree in sign depending on whether you divide by BF16 PFLOPs (+31%) or HBM bandwidth (−9%). A variant may enter a class only under a factor measured from same-venue, same-day, same-SKU pairs.

Excluded outright: variants not listed above, community/consumer hosts, interruptible/spot tiers (interruptible, spot, community), term-committed prices, offers below 2 GPUs, prices outside the sanity band [$0.25, $25.00] per GPU-hour.

Why the node floor is 2, not 8. Measured within one venue on one day, the per-GPU discount saturates at 2 GPUs (1×=1.000, 2×=0.951, 4×=0.916, 8×=0.916), so 2/4/8-GPU offers are mutually comparable within ~4% and need no adjustment, while the 1-GPU offer carries a ~9% small-order premium and is excluded rather than normalised away. The former 8-GPU floor admitted marketplace inventory on 1 collection day in 10 and discarded essentially all of the index's price discovery.

2. Data hierarchy and market segments (IOSCO P8)

Executable marketplace asks are preferred over list prices and carry a 2× weight multiplier. Overlay data (power prices) never enters the calculation. Competitor indices are never ingested.

The constituent distribution is bimodal — measured separation of 5.4 standard deviations between the neocloud/marketplace cluster and the hyperscaler catalog cluster. Averaging across that gap yields a number no one quotes, so series are segregated by market segment and never drawn across it:

3. Aggregation (exact algorithm)

Per UTC day and series. The unit of aggregation is the offer, not the provider.

1. Collect all observations for the day passing the unit filters above, restricted to the series' market-segment population.
2. Static (manually verified) entries older than 90 days are excluded as stale (warning from 45 days).
3. Provider weight = 2 if the provider has any executable offer, else 1. Capacity does not enter here: it is unobservable for every list source, so a capacity term at provider level is fiction that made a rate card disclosing nothing outrank a marketplace disclosing a real 2-GPU offer.
4. Concentration cap: no provider may exceed 25% of total weight; excess is redistributed pro-rata, iterated to a fixed point (the cap relaxes to 100/n when n providers cannot satisfy it). Note this cap is mathematically inert at n=4 — it forces exactly equal shares — and binds only for n≥5. Every print publishes whether it bound.
5. Spread each provider's share across its own offers in proportion to offer capacity. Capacity is genuinely observable here, between offers from the same venue.
6. Trim: clamp the k highest and k lowest offer prices to the k-th order statistic from each end, k by panel size (n≥5 → k=1; n≥10 → k=2; n≥20 → k=3). Count-based, because nearest-rank percentile winsorising is inert at this panel size — at n=6 both p5/p95 and p10/p90 resolve to (min, max) and clamp nothing.
7. Weighted median over offers: sort by price ascending; the value is the first price at which cumulative weight reaches 50%. A median over ~6 providers has a delta of 1.0 to one constituent and 0.0 to every other; a median over many capacity-weighted offers is locally smooth and always lands on a price someone actually quoted.
8. Publish gate: fewer than 5 qualifying providers or fewer than 5 qualifying offers → no value is published, flagged insufficient_sources / insufficient_offers. There is no fallback waterfall. A gap is credible; a fabricated print is fatal.
9. EUR companion = USD value ÷ the most recent ECB EUR/USD reference rate dated on or before the print date. The ECB publishes ~14:00 UTC, after the 11:00 UTC cut-off, so the EUR leg is T-1 by construction and says so rather than pretending otherwise. Providers quoting natively in EUR are converted from their native amount at print time.
10. Headline companion: 7-day mean of daily prints (requires ≥4 non-null days).

Constituent prices moving more than 30% day-over-day are flagged (jump) for manual review but are not excluded. A print with no executable input is flagged no_executable_input — a list-price-only print says so on its face.

3.1 Continuity series

The headline is the raw daily cross-section, not a chain-linked level. A Laspeyres link with a same-day divisor reset contributes exactly zero return on panel entry and exit, which makes in-panel up-moves permanent while down-moves taken via exit and re-entry are laundered out — a ratchet worth roughly +7.8% per cycle for a constituent at the 25% cap. A chained level is published beside the headline as a labelled companion, and the divergence between them is a published health metric with a 5% review trigger.

3.2 TCI-CRI-COMPUTE (chain-linked class composite)

The composite aggregates the class series into one level (base 100 at its first print), chain-linked so that reweighting never jumps the published level:

• At each weight review, class basket shares = each class's share of total observed qualifying capacity over the window. With ≥2 eligible classes, shares are capped at 75% and floored at 5% (pro-rata redistribution, as in §3 step 6).
• Daily: composite(t) = composite(t−1) × Σ share × (class(t) / class(t−1)), summed over classes with a published value on both endpoints, shares renormalised over those classes. A class that gaps drops out of that day's link; if no class links, the composite gaps (no_linkable_series).
• As the observed market migrates across hardware generations (H100 → B200), the basket follows mechanically — no methodology change required.

4. Series

EU-CRI-H100-CLOUD was retired in v0.3.0 and is not published. Its historical values remain in index_history.csv under the methodology version that produced them.

Segment size and the publication gate. A sub-population series can only print when its segment holds at least min_providers (5) qualifying providers. EU-CRI-H100-MKT and EU-CRI-H100-HS draw on segments smaller than that and therefore gap by construction rather than by circumstance. This is stated here because a permanent gap and a temporary one look identical on the dashboard, and the difference matters to a reader.

5. Revisions and corrections

Raw observations and published prints are append-only (enforced by database triggers). Errors are corrected in the next print as a new revision flagged correction; prior revisions remain queryable forever. The full constituent set for any print is available via python -m tci.run constituents --date YYYY-MM-DD.

6. Changes to this methodology

Any change to the parameters below or the calculation code requires a version bump, a CHANGELOG entry, and one publication's notice before taking effect — enforced mechanically via METHODOLOGY.lock in CI. See GOVERNANCE.md.

7. Current parameters (config/factors.yaml, verbatim)

# EU-CRI methodology parameters. Every number that affects a published print lives here.
# Changing anything in this file (or sovereign.yaml, index.py, normalise.py, weights.py)
# requires a methodology_version bump + CHANGELOG entry - enforced by CI (see GOVERNANCE.md).
methodology_version: "0.3.0-dev"   # '-dev' relaxes the lock guard pre-launch; drop it at launch

reference_unit:
  gpu_model: H100_SXM        # headline reference variant (class H100 below)
  term: on_demand            # no term commitment
  location: EU_EEA           # datacenter physically in the EU/EEA
  # v0.3.0 removed `node_gpus: 8`. The node-size floor now lives in filters.min_gpu_count
  # and is 2, not 8 - see the measurement recorded there.
  # Per-GPU-hour, ex-VAT, excluding metered egress. USD primary, EUR companion at the
  # ECB reference rate (T-1: the ECB publishes ~14:00 UTC, after our 11:00 cut-off, so
  # the EUR leg structurally cannot use same-day FX - see METHODOLOGY.md).

# Compute classes the index observes. A class prices its reference variant ONLY.
# v0.3.0: cross-variant normalisation factors are removed. H100 SXM and H100 PCIe are
# different products, and H100->H200 normalisers disagree in SIGN depending on whether
# you divide by BF16 PFLOPs (+31%) or HBM bandwidth (-9%). A variant may re-enter a class
# only with a factor MEASURED from same-venue, same-day, same-SKU pairs (see
# `measured_factors` below). No assumed factors are permitted in the calculation path.
model_classes:
  H100:   {reference_variant: H100_SXM,  variants: {H100_SXM: 1.0}}
  H200:   {reference_variant: H200_SXM,  variants: {H200_SXM: 1.0}}
  B200:   {reference_variant: B200_SXM,  variants: {B200_SXM: 1.0}}
  B300:   {reference_variant: B300_SXM,  variants: {B300_SXM: 1.0}}
  A100:   {reference_variant: A100_SXM,  variants: {A100_SXM: 1.0}}
  H100P:  {reference_variant: H100_PCIE, variants: {H100_PCIE: 1.0}}   # PCIe priced separately

# Market segment per provider. This is the tier-segregation control and it is
# methodology-affecting: the constituent population is bimodal (measured 2026-08-15,
# cluster separation 5.4 sd between the neocloud/marketplace cluster [2.16..3.85] and the
# hyperscaler cluster [5.44..7.36]). Averaging across that gap produces a number nobody
# quotes. Providers not listed default to 'neocloud'.
segments:
  marketplace: [vast.ai, runpod]
  hyperscaler: [aws, gcp, azure]
  neocloud:    [nebius, datacrunch, seeweb, scaleway, ovhcloud, hetzner, genesis_cloud,
                leaseweb]

# Which segments each published series draws from. The headline deliberately excludes the
# hyperscaler catalog tier - those are the two highest and least-moving prices in the panel
# and they belong in the series where they are the correct object of measurement.
series_populations:
  headline:    [marketplace, neocloud]
  marketplace: [marketplace]
  hyperscaler: [hyperscaler]
  sovereign:   [marketplace, neocloud]   # further filtered by config/sovereign.yaml

filters:
  min_gpu_count: 2           # v0.3.0: was 8. Measured within-venue same-day (vast.ai
                             # 2026-07-18) the per-GPU discount SATURATES at 2 GPUs:
                             # 1x=1.000, 2x=0.951, 4x=0.916, 8x=0.916. So 2/4/8-GPU offers
                             # are mutually comparable within ~4% and need no factor, while
                             # the 1-GPU offer carries a ~9% small-order premium and is
                             # excluded rather than normalised. The old 8-GPU floor passed
                             # 1 of 10 vast.ai days and threw away ALL marketplace price
                             # discovery (9 distinct EU H100 prices, +12.3% range).
  price_floor_usd: 0.25      # junk-listing guard: asks below this are excluded
  price_ceiling_usd: 25.0    # sanity ceiling
  datacenter_verified_only: true
  exclude_tiers: [interruptible, spot, community]

aggregation:
  unit: offer                # v0.3.0: aggregate over OFFERS, not one price per provider.
                             # A median over ~6 providers is a step function with delta 1.0
                             # to one name and 0.0 to the rest. A median over many
                             # GPU-count-weighted offers is the SOFR construction and is
                             # locally smooth. Density is what makes the median sound.
  estimator: weighted_median # over offers, weighted by GPU count (capacity)
  trim_k:                    # count-based trim: clamp the k highest and k lowest.
    - {min_n: 0,  k: 0}      # Percentile winsorising is INERT at this panel size: at n=6
    - {min_n: 5,  k: 1}      # nearest-rank p5/p95 AND p10/p90 both resolve to (min, max)
    - {min_n: 10, k: 2}      # and clamp nothing. The shipped [5,95] was a no-op for the
    - {min_n: 20, k: 3}      # entire history of the index. Count-based trimming binds at
                             # every n.
  min_providers: 5           # below this: value=null, flag='insufficient_sources'.
                             # A gap stays a gap - no fallback waterfall, no fabrication.
  min_offers: 5              # and at least this many surviving offers
  smoothing_days: 7          # headline companion series: 7-day mean (>=4 non-null required)

weights:
  capacity_cap: 64           # daily capacity term = min(observed qualifying GPUs, cap)
  default_capacity: 8        # ... where capacity is unobservable (all list-price rows)
  executable_multiplier: 2.0 # executable-quote sources weighted double list-price sources
  max_weight_share_pct: 25   # per-PROVIDER aggregate cap, applied after offers are summed
                             # within a provider. NOTE: this cap is mathematically inert at
                             # n=4 (max(25, 100/n) forces exactly equal shares - verified
                             # 0.0 deviation over 50k random weight vectors). It binds again
                             # only once the panel is wider; `cap_bound` is published per
                             # print so a reader can see whether it did anything.
  review:                    # scheduled, data-driven weight reviews (fixed rule, no discretion)
    anchor_weekday: 1        # ISO weekday a review takes effect (1 = Monday, i.e. weekly)
    window_days: 28          # trailing observation window ending the day before the review
    min_history_days: 5      # fewer collection days than this in the window -> bootstrap

composite:                   # EU-CRI-COMPUTE: chain-linked basket of class sub-indices
  base_value: 100.0          # level on the composite's first print
  min_class_share_pct: 5     # floor for an eligible class at review (applies with >=2 classes)
  max_class_share_pct: 75    # cap for a class at review (applies with >=2 classes)

# Continuity series. v0.3.0 DEMOTES chain-linking from the headline to a labelled companion.
# A Laspeyres link with a same-day divisor reset contributes exactly zero return on panel
# entry/exit, which makes in-panel up-moves permanent while down-moves taken via exit and
# re-entry are laundered out - a ratchet worth ~+7.8% per cycle for a 25% constituent.
# The headline is therefore the raw cross-section (reproducible, no drift, no ratchet); the
# chained level is published beside it and the divergence between them is a health metric.
continuity:
  publish_chained: true
  base_value: 100.0
  divergence_review_pct: 5   # chained vs raw divergence above this triggers a documented review

fx:
  max_age_days: 7            # refuse an FX rate older than this (flag the print instead)
  # HARD RULE enforced in code: fx_date <= print date. Two 2026-07 prints were computed
  # with an FX rate published on 2026-07-21, i.e. AFTER the print date - a look-ahead
  # introduced by backfill. Covered by a regression test.

staleness:                   # static (manually verified) provider entries
  warn_days: 45              # loader warns
  exclude_days: 90           # excluded from the index with exclusion_reason='stale'

quality:
  jump_flag_pct: 30          # day-over-day provider move that flags (never excludes)

# Factors measured from same-venue, same-day, same-SKU observations. Published for
# transparency; NONE of these currently enter the calculation path. A factor may only be
# promoted into `model_classes` after it is supported by >=3 independent venue-days.
measured_factors:
  node_size_per_gpu:         # vast.ai, EU H100_SXM, same-day ratios vs the 1-GPU offer
    "2": 0.9514              # n=3 venue-days
    "4": 0.9159              # n=1  - provisional
    "8": 0.9158              # n=1  - provisional
  fabric_infiniband_premium: 1.107   # Azure westeurope ND96isr vs ND96is_noIB, n=1 pair.
                                     # PROVISIONAL - measures Azure's internal price ladder,
                                     # not the market's. Not in the calculation path.

eu_eea_countries: [AT, BE, BG, HR, CY, CZ, DK, EE, FI, FR, DE, GR, HU, IE, IT,
                   LV, LT, LU, MT, NL, PL, PT, RO, SK, SI, ES, SE, IS, LI, NO]
# NOTE: CH (Switzerland) is deliberately absent - Switzerland is in EFTA but NOT the EEA.

8. IOSCO Principles mapping

TCI is a research publication. It is not investment advice and is not administered as a benchmark under EU Regulation 2016/1011; it may not be used as a reference price in financial instruments.

Author and administrator: Mark Rusch (rusch.mh@gmail.com). Conflicts: the author may hold positions on venues whose prices the index observes; see GOVERNANCE.md.