PT-BR BETA 1.0

Tantalum Strategy

Benchmarks, research, and advisory for the materials economy underneath AI.

Wed, 02 Sep 2026 22:40 UTC
[ INDEXES · METHODOLOGY ]

Index methodology

All Tantalum Strategy indexes are calculated from publicly sourced inputs and published on a weekly cadence (Friday close, posted Monday morning UTC). Constituent materials and the thesis behind each index are open. Exact weights, proxies for materials without spot markets, and the calculation kernel are the editorial judgment of the desk.

The three 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?

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, the African Copperbelt, plus selective Australian and Angolan producers.

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:

Top-line TAI = 75% TAI-M + 25% TAI-P. Both sub-indices are published and updated independently. Readers can decompose any TAI move into its materials and power contributions.

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: it compounded both of them off the previous composite value every run and discarded the result. 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 — a reader checking our arithmetic would have found it did not close. Both sub-indices have been re-based onto the composite as of 2026-08-22 and are now persisted and compounded independently, so the identity holds every week and is checked by an automated gate that refuses to publish if it does not. Their displayed history is splice-adjusted by a single constant factor so the shape of the real published series is preserved at the new level; the untouched pre-re-base series is retained in the data file.

Data sources

Input prices are sourced from publicly available endpoints: COMEX and NYMEX futures via Stooq.com (delayed market close), ETF proxies (REMX for rare earth, LIT for lithium, URA for uranium, NLR for nuclear utilities, XLU for the broad utility sector, SOXX for semiconductor demand as a Ga/Ge proxy, VALE and FCX for Brazilian and DRC mining equity exposure) for materials without transparent spot markets, USGS quarterly reference prices for gallium, germanium, and helium, and EIA / Henry Hub for the natural gas component.

We do not use paid commodity-price feeds. All inputs are reproducible from public data so any reader can re-derive the headline number from the open inputs and the methodology described here.

v0 base-computation basket

The full constituent list above is the target methodology. The automated weekly base computation (the scheduled job that recomputes the headline numbers every Friday) currently runs on a narrower v0 operational basket of liquid ETF proxies — the subset for which there is a clean, free, daily-refreshing public price. This is disclosed here so the published weekly move is reproducible from exactly the inputs the job actually reads, not from constituents it does not yet track.

What the v0 job weights today:

Three tickers that appear on the home-page price panel — GLD (gold), SLV (silver), USO (oil) — are context only and carry zero weight in every index. They are shown for market color, not used in any computation.

Uranium's effective weight in TAI. Because URA sits in both sub-indices, its weight in the top-line composite is the sum of its weighted appearances rather than either individual figure. At the v0 weights, uranium's effective weight in TAI exceeds copper's. That is a consequence of the 75/25 structure meeting a ticker that legitimately belongs to both the materials and the power thesis, not a deliberate ranking of uranium above copper. We disclose it rather than bury it; whether to de-duplicate is an open weighting question for the next methodology review.

Week-over-week, not one day compounded. Until 2026-08-22 the weekly recompute multiplied the previous index level by a single day's price move, so four of five trading days never entered the index at all. The job now stores the raw price of every basket ticker at each recompute and computes the true week-over-week ratio against it. Where a stored mark is missing — a newly added ticker, or the first run after this change — it falls back to the single-day figure and records that it did so, and the basis used is published in the data file alongside the values.

As cleaner data sources come online (a true 5-day basket fetch, NLR/XLU equity legs, USGS-blended thin-market marks), the v0 basket converges toward the full constituent list above, and each change is logged in the changelog with its date. The editorial overlay and watermark layers described below apply on top of whichever base basket is in force.

Editorial overlay

Each published index value is the output of two layers: a weighted base computation (constituent price ratios × weights) and an editorial overlay applied at publication time. The base is open and reproducible from the constituent list. The overlay is proprietary.

The overlay scope is documented openly. It incorporates a desk review of three signal categories: (1) geopolitical flow — export-control announcements, sanctions, country-specific supply disruptions, (2) supply-chain news triage — major contract signings, project starts/stops, refinery commissioning, (3) technical-momentum filtering — smoothing transient moves and weighting longer-horizon trend confirmation. The exact kernel that combines these into a per-index adjustment, and the magnitude bounds applied, are editorial judgment of the desk.

This is intentional, not theatrical. Naive replication of the indexes from public commodity feeds alone will diverge from the published values within weeks because the overlay corrects for noise, applies editorial judgment on signal weight, and embeds desk-level context that pure price math cannot capture. Institutional licensing includes the kernel under separate agreement.

Multi-model adversarial validation

Editorial output passes through a structured adversarial review across independent frontier models before publication. One model drafts. A separate model independently re-verifies every numerical claim and every cited source against the underlying public references. Only output that survives both passes ships. Disagreement between the two passes is treated as a signal in itself: it either surfaces a fact that needs to be revised, a citation that needs to be dropped, or an editorial framing that overreaches what the sources actually support.

This applies to research articles and to the per-index editorial overlay alike. The desk publishes the framework openly because it is part of what readers are buying when they read Tantalum. The exact models, the verification prompt structure, and the disagreement-resolution kernel are not published — that's where the operational moat lives, not in the constituents or the cadence.

The point is not that this prevents every error. It tightens the loop between drafting and publication so that fabricated specifics (made-up citations, drifted dates, plausible-but-uncited numbers) get caught at the gate rather than in the wild. The desk treats reader trust as the load-bearing asset of the entire product.

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. This is invisible to readers and does not affect any directional or magnitude read of the index. It exists so that if a competitor or third-party were to publish an index suspiciously close to ours, we can verify whether the values originated from our publication by checking the watermark pattern. Standard provenance practice for proprietary data products.

The watermark does not compound (corrected 2026-08-22). Previously the adjustment was written into the stored index value, which the next week's computation then multiplied — so the watermarks accumulated into a slow random walk sitting on top of real basket moves. The desk now keeps an internal unwatermarked level for each index, compounds that, and applies the watermark only to the published figure. A provenance mark that alters the series it marks is not provenance; it is error.

What we publish, what we don't

Public: the constituent materials of each index, the source we use to track each one, the broad thesis behind every cut, the update cadence, and the data providers we read.

Proprietary: the exact weight of each constituent, the precise price proxy used for materials without a clean spot market (e.g., what we use as a niobium proxy when CBMM trades privately, or how we blend USGS quarterly reference prices into a weekly mark), and the calculation kernel. These are the desk's editorial judgment and are revised over time as the AI-materials thesis evolves.

Institutional clients can license the full methodology, constituents, weights, and historical data under separate agreement. Talk to the desk →

Update cadence

Each sub-index and headline value is recalculated weekly against the Friday US close and published Monday morning UTC. 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 auto-deploys the refreshed numbers within ~60 seconds of that commit.

Why Saturday (corrected 2026-08-22). Our price endpoint returns the last completed session. Queried on Friday evening it returns Thursday's close, so a job that ran Friday after the bell was marking the index on Thursday data while labelling it Friday. Just after midnight UTC on Saturday is the first moment the Friday settle is actually available. The daily home-page ticker now also states which session its prices belong to, rather than the wall-clock time we happened to read them.

Daily ticker values on the home page reflect the most recent close. The ticker label shows the date of the underlying data so readers can see when it was last refreshed.

Limitations (read this)

Honesty is the editorial position of the desk, including about our own indexes. Four limitations a serious reader should know about:

  1. No backtest validation yet. The indexes started publishing in early 2026. Until at least a year of live data has been recorded and a backtest against synthetic constituents can be run, treat the YTD numbers as directional, not validated. We plan to publish a backtest review on the one-year anniversary.
  2. Thin-market proxies are imperfect. Gallium, germanium, helium, and niobium do not have transparent daily spot markets. We use a combination of USGS quarterly reference prices and equity proxies (SOXX for semiconductor demand, VALE for Brazilian mining cluster) to track them on a weekly cadence. These proxies move for reasons other than the underlying material's actual supply-demand balance. We disclose the proxy choices on this page; we revise them as cleaner data becomes available.
  3. SOV50's price basket is narrower than its concentration thesis. The HHI panel scores nine materials — rare earths, niobium, gallium, germanium, cobalt, graphite, tungsten, antimony, magnesium — but the automated price leg currently reads two liquid proxies, rare earths and lithium. Lithium supply is comparatively diversified, which makes it the weakest available stand-in for a concentration premium; gallium, tungsten and cobalt, which score far higher on concentration, contribute nothing to the price reading because no clean daily public feed exists for them. (Antimony is the instructive counter-case: on 2026 USGS data its mine production is now the least concentrated in the basket, which the price leg equally cannot see.) This is the widest gap between thesis and instrument anywhere in our methodology, and 2026 is a bad year for it to persist: export-control divergence has pushed ex-China gallium and heavy rare earth prices to multiples of Chinese domestic levels, which is precisely the signal SOV50 exists to capture and precisely what a REMX/LIT basket cannot see. Reworking this basket is the desk's next methodology priority.
  4. Editorial weights, not optimized weights. The constituent weights inside each index are the desk's view of what matters for the AI-materials thesis right now. They are not the output of a portfolio-optimization model or a covariance-minimization procedure. As the thesis evolves and as new data sources become available, weights are revised. Major weight changes are noted on this page with the date.

None of this makes the indexes useless. It does mean they are best read as thematic indexes (closer to MVIS, Solactive thematic) rather than as benchmark indexes (S&P GSCI, Bloomberg Commodity). The numbers are directional and editorially honest. They are not yet ready for use as a settlement reference for a derivative or a fund.

The path to more rigor

The roadmap to make these indexes institutionally license-ready:

  1. One year of live data. Publish weekly from Jan 2026 forward. Document any methodology changes with dates. Build a track record.
  2. Backtest review at the one-year anniversary. Reconstruct each index back to 2020 using historical Stooq + USGS data. Publish the backtest, including draws and disagreements with the constituents we have today. Adjust weights where the backtest reveals signal/noise problems.
  3. Engage a commodity index quant for an independent methodology review. Particularly on TAI-M's handling of thin-market materials and SOV50's risk-bucket weighting. Publish the review along with our response.
  4. Document inclusion / exclusion criteria. Every constituent needs a written rationale for why it's in. Inspired by MVIS and Solactive's published methodologies.
  5. Institutional licensing pilot. Once steps 1–4 are done, offer paid access to full weights and historicals to a small set of institutional readers under separate agreement, with the methodology binding the parties.

Changelog

2026-08-22 (b). SOV50's concentration panel rebuilt on Mineral Commodity Summaries 2026 (USGS, published February 2026, 2025 production estimates); it had been running on the 2025 edition. Each material's HHI is now derived from that edition's country production table — sum of squared country shares — and the per-material rows carry the chapter they come from and the production basis used, so a reader can re-derive any figure. The basket headline moves 0.68 to 0.66, and its weighting is now stated rather than implied: materials are weighted by the buckets already published on the page (highest 3, high 2, moderate 1, normalised), which makes the headline reproducible from this page plus the cited chapters. Two things worth naming rather than burying. First, antimony is no longer a China-concentration story on the mine side — China is 36% of 2025 mine production against Russia's 29% and Tajikistan's 20%, taking antimony's HHI from 0.31 to 0.26, the lowest in the basket. That is a real structural change and it cuts against the direction of every other material here. Second, germanium is now flagged as an editorial estimate rather than a derived figure: USGS states plainly that most producers do not publicly report germanium production and that global data are limited, so there is no country table to compute from. It stays in the basket, labelled, instead of borrowing the credibility of the numbers around it. Gallium (0.98), niobium (0.87) and magnesium (0.77) remain the extremes.

2026-08-22. Correctness pass on the automated computation, prompted by an audit of the published numbers. Six fixes, all of which had been silently wrong: (1) year-to-date figures were written once in May and never recomputed — SOV50 was published at +14.6% when the level implied +23.0%, and SDX showed a loss while the level implied a gain; YTD is now derived from the level on every run. (2) TAI-M and TAI-P were never persisted, so the composite stopped reconciling with its own sub-indices; both are re-based and now compound independently, with an automated gate that blocks publication if 0.75 × M + 0.25 × P does not equal the composite. (3) The weekly move compounded a single day's price change instead of the week's; true week-over-week marks are now stored and used. (4) The provenance watermark was being compounded week over week; it is now applied to the published value only. (5) The daily change on the price panel was computed from a session's own open and close — an intraday move labelled as a daily one — and is now a genuine close-over-close change, with the session date shown. (6) The recompute ran Friday evening against a feed that returns the previous completed session, marking Friday's index on Thursday's close; it now runs after midnight UTC Saturday. Also: published history is no longer truncated at 26 entries, which had been quietly destroying the base-period record the one-year backtest depends on, and a fourth entry was added to Limitations describing how far SOV50's price basket falls short of its concentration thesis.

2026-06-06. Added a "v0 base-computation basket" section disclosing the operational ETF-proxy subset the automated weekly job runs on (vs the full target constituent list), and clarifying that GLD/SLV/USO on the price panel carry zero index weight.

2026-05-27. Methodology page expanded with a Multi-model adversarial validation section, documenting the editorial review process that runs between drafting and publication.

2026-05-25. TAI restructured from single basket to composite (TAI-M + TAI-P). SOV50 reframed as a risk indicator (not a return index). SOUTH ticker renamed to SDX (Southern Diversification Index) and reframed as a thesis index pairing with SOV50. Methodology page expanded with a Limitations section and a path-to-rigor roadmap.

2026-01-01. All three indexes go live with base value 100.