Index methodology

The indexes at a glance

TAI · Tantalum AI Materials Index. Tracks the physical inputs whose supply must scale or AI infrastructure breaks. A composite of two sub-indices: TAI-M (materials) and TAI-P (power), weighted 75/25.

SOV50 · Sovereignty 50. A risk indicator, not a return index. Quantifies supply-concentration exposure: if a single dominant producing country closed the spigot tomorrow, how much AI-relevant supply would be at risk? Published as a concentration score (HHI) plus a price leg.

SDX · Southern Diversification Index. A thesis index pairing with SOV50. Tracks the alternative supply base for AI-relevant materials outside the China + DRC nexus: Brazil, the Lithium Triangle, Chile's copper belt, and Australian rare-earth and lithium producers.

TCC · Tantalum Context Capacity. The memory index. How many long AI conversations fit in the memory of one flagship AI chip, with a companion any-vendor reading (TCC-X).

The memory index: TCC

Memory is now the most expensive and most supply-constrained part of AI hardware. One index tracks it from the angle no price index covers: how many long AI conversations a flagship chip can actually hold. It does not track memory prices, which other providers already publish.

TCC · Tantalum Context Capacity. How many 128K-token conversations fit in the HBM of the highest-memory NVIDIA data-center GPU with confirmed customer deliveries on the reading date (B200 at 192 GB until July 2025; B300 and Rubin at 288 GB since), counting conversation memory only. TCC rises when chips gain memory or when models get leaner, so a higher reading is always the better one. Base 100 = 1 January 2026. It is a stylized gauge: real servers also hold weights, use FP8 caches and page memory to CPU or SSD.

TCC-X, the any-vendor companion (added 2026-10-11). Same construction, but the reference chip is the largest-HBM merchant data-center accelerator from any vendor with confirmed customer deliveries: AMD MI325X at 256 GB through mid-2025, 288 GB (B300 / MI355X) from Q3 2025, and AMD MI455X at 432 GB from Q3 2026, when the first Helios racks reached Oracle. The headline stays on NVIDIA because that is where most inference runs; the gap between TCC and TCC-X is itself a reading: on 1 October 2026 TCC is 218.5 and TCC-X is 327.5. "Confirmed customer deliveries" means a named customer has units on its floor, not an announcement, a benchmark or a launch date.

Input: memory per conversation. For each lab's flagship open-weight model, the bytes of KV cache needed to hold one 131,072-token (128K) conversation for one user, read from the model's published config.json. Keys, values, MLA latents and sparse-attention indexer keys are counted at 16-bit; linear-attention and recurrent layers (Gated DeltaNet, KDA, Mamba-2, lightning attention) keep a fixed-size state counted at fp32 plus their short-convolution state; sliding-window and chunked layers count only up to their window; model weights are excluded. This is the demand side of TCC and is published inside it as "GB per conversation" (6.25 GB on 1 October 2026), not as a separate index. TCC = 100 × (HBM ÷ GB per conversation) ÷ the same ratio on the base date.

• Basket: one flagship per lab: the largest general-purpose open-weight model of the lab's newest generation. Code-only, vision-only, distilled and Small/Medium/Air/Flash tiers are excluded; a multimodal model that is a lab's general flagship counts. A lab enters once it has released a general-purpose open-weight model of at least 20B parameters under a licence that allows commercial use, and exits after 15 months without a new flagship. Constituent dates are the date the weights became downloadable, not the API launch. Current labs (1 October 2026): DeepSeek, Alibaba Qwen, Meta, Google, Mistral, OpenAI, Moonshot, Zhipu, MiniMax, NVIDIA. Eligible under the rule and on the watchlist for the next reading, pending a config read: Baidu (ERNIE 4.5), xAI (Grok 2.5), ByteDance (Seed-OSS), IBM (Granite 4), Cohere (Command A), Ai2 (OLMo 3).
• Aggregation: chain-linked geometric mean of each lab's change versus the prior reading, counting only labs present in both readings, so entries and exits never cause a jump. The headline "GB per conversation" is the base-date basket geometric mean (13.64 GB) multiplied by the chained basket change since the base date.
• Cadence: quarterly, plus an update whenever a lab ships a new flagship. History is reconstructed quarterly back to 1 January 2025.
• Cross-checks: a Llama-70B-class model (80 layers, 8 KV heads × 128) gives 42.9 GB (40.0 GiB) at 128K, matching the widely cited figure. At 1M tokens the formula puts DeepSeek-V4-Pro at 11.5% of DeepSeek-V3.2's cache; DeepSeek's model card states 10%, a gap consistent with undisclosed cache-precision choices.
• Limitations: open-weight models only, since closed models do not publish configurations. MiniMax-M3's block-level sparse index and the short-convolution states of linear-attention layers are approximated, and MiniMax-M1 is assumed to share MiniMax-Text-01's architecture. Nemotron 3 Ultra's MoE layers carry no per-conversation state and are counted at zero.

TAI as a composite

TAI was originally a single basket. Combining materials and power inputs in one basket creates a signal-to-noise problem: a TAI move could come from a rare-earth supply tightening or from a hurricane spike in natural gas, and the headline number doesn't tell you which. In May 2026 we split TAI into two openly published sub-indices:

• TAI-M (AI Materials): pure mineral exposure. Copper, the rare-earth basket, lithium, the Brazilian mining cluster, and a listed producer standing in for tantalum, antimony and niobium. Power deliberately excluded. Gallium, germanium and helium are monitored and cited in the weekly readout but not priced in, because no public daily feed exists for them.
• TAI-P (AI Power): electricity inputs for AI infrastructure. Uranium and nuclear utility equity (the leading-edge story: hyperscaler PPAs, record long-term uranium contract prices, PJM capacity auctions clearing at the cap), broad utility sector equity, natural gas as a residual.

Top-line TAI = 75% TAI-M + 25% TAI-P. Both sub-indices are published and updated independently, and an automated gate refuses to publish if the identity does not hold.

Sub-index re-basing, 2026-08-22. From the May 2026 split until this date, the automated job advanced the composite each week but never wrote the sub-indices back to file. The published TAI-M and TAI-P therefore sat frozen at their May values, and 0.75 × TAI-M + 0.25 × TAI-P no longer equalled the published TAI. Both sub-indices were re-based onto the composite as of 2026-08-22 and are persisted and compounded independently since. Their displayed history is splice-adjusted by a single constant factor; the untouched pre-re-base series is retained in the data file.

Constituents and weights (basket v1 (2026-10-11))

These are the weights the weekly job uses, read from the same data file the site renders. A reader with the Yahoo Finance closes for these instruments and last week's published levels can reproduce this week's levels to the watermark.

Why these. Copper is the actual COMEX future, not a miners ETF, so the US tariff premium and the mine outages of 2026 show up as copper and not as equity beta. Uranium sits only in TAI-P: until October 2026 it was in both sub-indices and carried more weight in TAI than copper did. Lithium is out of SOV50 because, on USGS 2026 data, it is the least concentrated material in the panel. The COPX and LIT ETFs are out of SDX because both carry large Chinese weights, which is the opposite of what an ex-China diversification index should hold; Glencore is out because its exposure is inside the DRC. AMG Critical Materials is in three baskets because it is the only liquid listed producer of tantalum (the Mibra mine, Brazil) and a producer of antimony and lithium, the thin markets the target list names. Lynas is the only scaled ex-China rare-earth separator and a US Department of Defense price-floor counterparty, which makes it the cleanest listed expression of the concentration premium SOV50 exists to capture. Non-US listings (Antofagasta in London, AMG in Amsterdam, PLS and Lynas in Sydney) carry FX in their returns; the index does not strip it.

Monitored, unweighted. Niobium (CBMM is private), Pensana (development stage), Syrah (Balama production deferred to Q3 2026), gallium, germanium and helium (no public daily price). They stay on the constituent panels so readers can see what the thesis covers that the instrument cannot.

Splicing. When a constituent enters a basket or changes instrument, it has no prior mark, so it contributes zero for its first week and the index carries forward on the other constituents; it contributes normally from the second week. The spliced tickers and the date are recorded in the data file. Levels never jump on a basket change.

Data sources

All prices come from Yahoo Finance's public end-of-day chart endpoint: COMEX copper (HG=F) and NYMEX Henry Hub (NG=F) front-month futures; the US-listed ETFs REMX, LIT, URA, NLR and XLU; and the producer equities listed above on NYSE, Nasdaq, the London Stock Exchange, Euronext Amsterdam and the ASX. The concentration panel is derived from USGS Mineral Commodity Summaries 2026. The memory index reads model configurations from Hugging Face and vendor specification sheets for HBM capacity. No paid feeds, no API keys. Quotes are delayed end-of-day, which is all a weekly index needs. Each ticker is tried on two Yahoo hosts and then on Stooq's daily-history CSV; the source that answered is recorded per ticker in the price file, and a ticker no source can price is carried at its last close and listed as stale in the recompute basis rather than silently dropped.

What the feed was before 2026-10-11. Until this date the job read Polygon.io's free tier, which serves US stocks and ETFs only. "Copper" was therefore the COPX miners ETF and "natural gas" was the UNG fund, while this page said "COMEX and NYMEX futures via Stooq". The page described a feed the job did not use (and Stooq's CSV endpoint, tried first as the replacement, no longer answers). That is corrected here, and the ticker panel on the home page now names the instrument behind every number.

Three tickers on the home-page price panel (GLD, SLV, USO) and the LIT lithium ETF are context only and carry zero index weight except where the table above says otherwise.

Editorial overlay

Each published value is a weighted base computation plus, when the desk chooses to apply one, a disclosed editorial adjustment. No adjustment has been applied since the August 2026 correctness pass; every published weekly value since then is the pure base computation plus the provenance watermark. Earlier versions of this page said the published values included a proprietary overlay that would make naive replication diverge within weeks. They did not, and the sentence is withdrawn.

If an adjustment is applied in future, it is bounded at ±1.0 index points per index per week, written into the data file with its date, size and one-line reason, and listed in the changelog. The signal categories a desk adjustment may draw on are export-control and sanctions announcements, country-specific supply disruptions, and major project or contract events. Momentum smoothing is not a permitted reason.

Multi-model adversarial validation

Research articles pass through a structured adversarial review across independent frontier models before publication: one model drafts, a separate model re-verifies every numerical claim and cited source against the underlying public references, and only output that survives both passes ships. The automated audit agent that ran this on every push was decommissioned on 2026-09-28; the review now runs inside the publishing task itself, before commit. The index values themselves are not subject to this review because they are mechanical; the weekly readout that interprets them is.

Watermark and provenance

Each published index value contains a deterministic micro-adjustment of approximately ±0.05 derived from a secret seed and the publication date. It does not affect any directional or magnitude read of the index. It exists so that if a third party published an index suspiciously close to ours we could check whether the values originated here.

The watermark does not compound (corrected 2026-08-22). The desk keeps an internal unwatermarked level for each index, compounds that, and applies the watermark only to the published figure.

What we publish, what we don't

Public: constituents, instruments, weights, the thesis behind every cut, the update cadence, the splice and overlay logs, and the data providers. The computation scripts live in the site repository.

Not public: the watermark seed, and the desk's research notes behind any future editorial adjustment.

Institutional clients can license the full historical data, the backtest when it exists, and a support relationship under separate agreement. Talk to the desk →

Update cadence

Each sub-index and headline value is recalculated weekly against the Friday US close. The recompute runs in continuous integration at 01:15 UTC Saturday, refreshing prices and recomputing in one job so the index and its stored week-over-week marks come from a single consistent settle. Vercel deploys the refreshed numbers within about a minute of that commit. The home-page price panel refreshes once each weekday evening after the US close and states the session its prices belong to.

The memory index is recomputed on the first business day of each quarter and whenever a constituent lab ships a new flagship.

Limitations (read this)

The path to more rigor

1. One year of live data. Publish weekly from January 2026 forward; document every methodology change with its date.
2. Reconstruct and backtest at the one-year anniversary. Rebuild the pre-August 2026 weekly series and extend each index back to 2020 from historical Yahoo Finance and USGS data, on the published v1 weights. Publish the backtest, including where it disagrees with the live series.
3. An ex-China / China spread panel for SOV50 (NdPr, Dy, Tb, gallium), updated monthly from published price-reporting prints, so the concentration premium is measured directly.
4. Independent methodology review by a commodity index quant, particularly on the thin-market proxy and SOV50's bucket weighting. Publish the review and the desk's response.
5. Institutional licensing pilot once steps 1 to 4 are done.

Changelog