Six indexes for the materials and memory economy underneath AI.
Each index cuts the strategic-resource market at a different angle. We publish the constituents, the instruments and the weights behind each cut, and the data file the job writes. Institutional licensing on request.
The physical inputs whose supply must scale or AI infrastructure breaks.
Open the indexAll three indexes, side by side
Base 100 on 1 January 2026. Hover to read any value; click the legend to hide a series.
View the data as a table
| # | TAI | SOV50 | SDX |
|---|---|---|---|
| Latest · October 17, 2026 | 92.1 | 97.9 | 93.5 |
| 33 | 96.8 | 102.4 | 94.3 |
| 32 | 97.3 | 104.3 | 94.7 |
| 31 | 99.7 | 107.5 | 94.4 |
| 30 | 100.8 | 110.7 | 97 |
| 29 | 103.2 | 112.8 | 101.1 |
| 28 | 107.5 | 119.3 | 102.6 |
| 27 | 108.4 | 122.2 | 103.6 |
| 26 | 108.8 | 125.7 | 102.2 |
| 25 | 105.9 | 121.8 | 98.3 |
| 24 | 106.5 | 122 | 99.3 |
| 23 | 106.7 | 121.9 | 100.1 |
| 22 | 104.9 | 119.7 | 99.5 |
| 21 | 105.1 | 120.9 | 99.5 |
| 20 | 107.3 | 123.6 | 100.8 |
| 19 | 107.1 | 123.6 | 99.8 |
| 18 | 107.7 | 125.2 | 100.2 |
| 17 | 109.1 | 126.4 | 100.3 |
| 16 | 109.5 | 127.5 | 100.9 |
| 15 | 105.8 | 122.1 | 97.9 |
| 14 | 105.7 | 122.4 | 98.2 |
| 13 | 102.9 | 119.9 | 96.6 |
| 12 | 102.4 | 118.9 | 96.1 |
| 11 | 102.3 | 118.6 | 96.4 |
| 10 | 102 | 118 | 96.2 |
| 9 | 102.1 | 117.5 | 96.5 |
| 8 | 101.5 | 116.9 | 96.9 |
| 7 | 101.8 | 116 | 96.4 |
| 6 | 101.2 | 115.1 | 96.8 |
| 5 | 101.4 | 114.2 | 97.2 |
| 4 | 100.9 | 113.4 | 97.8 |
| 3 | 101.1 | 112 | 97.4 |
| 2 | 100.4 | 110.6 | 98.2 |
| 1 | 100.6 | 109.4 | 97.8 |
Tantalum AI Materials Index
What does the AI buildout physically require? TAI tracks the inputs whose supply must scale or AI infrastructure breaks. Composed of two sub-indices: materials (the things you mine) and power (the electricity that runs them). Weighted 75% materials, 25% power.
TAI is a composite of two sub-indices, published separately: 75% TAI-M (materials) + 25% TAI-P (power).
Pure mineral exposure to the materials AI infrastructure requires. Power inputs deliberately excluded so the signal stays clean. Uranium moved out of this basket on 2026-10-11: it is a power input and was being counted in both sub-indices.
| Category | Material |
|---|---|
| Conductors & magnets | Copper |
| Conductors & magnets | Rare earth basket (NdPr, Dy, Tb) |
| Battery storage | Lithium |
| Brazil mining cluster | Iron ore, copper, nickel |
| Thin markets (tantalum, antimony, lithium hydroxide) | AMG Critical Materials |
| Semiconductors & lithography (monitored, unweighted) | Gallium, germanium, helium |
Electricity for AI infrastructure. Nuclear is the leading-edge story (hyperscaler PPAs, record long-term uranium contract prices, PJM capacity auctions clearing at the cap). Utility equity and gas are tracked as supporting signals.
| Category | Material |
|---|---|
| Nuclear | Uranium |
| Nuclear | Nuclear utility equity |
| Utility equity | US utility sector |
| Gas (residual) | Natural gas (Henry Hub) |
Sovereignty 50
SOV50 has two readings. The price-style index value (base 100) tracks the assets that reprice when concentrated supply tightens: the rare-earth ETF, the scaled ex-China rare-earth separator (Lynas) and the listed tantalum and antimony producer (AMG). The HHI concentration score (separately published) measures the actual supply concentration of the basket, derived from USGS country-level production data. Use the price index for market reads; use the HHI for structural-risk reads. Lithium left the price leg on 2026-10-11: it is the least concentrated material in the panel and was diluting the signal.
0.00 = perfectly distributed across producers. 1.00 = single-country monopoly. DOJ classifies HHI > 0.25 as highly concentrated for antitrust purposes; supply-chain literature treats > 0.50 as extreme.
USGS Mineral Commodity Summaries 2026 (published February 2026), 2025 production estimates. Tantalum added 2026-10-11.
Method
Per material, HHI = sum of squared country production shares, computed over the countries listed in that commodity chapter's production table ("Other countries" counted as one producer, which can only overstate concentration). Basket HHI weights each material by the published weight bucket shown below (highest 3, high 2, moderate 1, normalised), so the headline is reproducible from this page plus the cited USGS chapters. With tantalum (bucket: high) the weighted basket reads 0.63 (previously 0.66 over nine materials).
| Material | Concentration | Weight |
|---|---|---|
| Rare earths | China ~69% | Highest |
| Niobium | Brazil ~93% | Highest |
| Gallium | China ~99% | Highest |
| Germanium | China (leading producer) | High |
| Cobalt | Congo (Kinshasa) ~73% | High |
| Tantalum | Congo (Kinshasa) ~52% + Rwanda ~16% | High |
| Graphite (anode) | China ~80% | High |
| Tungsten | China ~79% | High |
| Antimony | China ~36% | Moderate |
| Magnesium | China ~88% | Moderate |
Southern Diversification Index
SDX is the solution side of the SOV50 question. Tracks the alternative supply base for AI-relevant materials outside the China + DRC nexus: Brazil, the Lithium Triangle, Chile's copper belt, plus Australian rare-earth and lithium producers. Direct equity exposure to the producers absorbing Western capital seeking to diversify processing dependence. From 2026-10-11 the basket holds the producers themselves; the COPX and LIT ETFs are out because both carry large Chinese weights, and Glencore is out because its exposure is inside the DRC.
| Country | Holding | Exposure |
|---|---|---|
| Brazil | Vale (VALE) | Iron ore, copper, nickel |
| Chile | SQM | Lithium brine |
| Chile | Antofagasta (ANTO.L) | Copper |
| Australia | PLS Group, formerly Pilbara Minerals (PLS.AX) | Lithium spodumene, tantalum by-product |
| Brazil | Sigma Lithium (SGML) | Hard-rock lithium |
| Brazil | AMG Critical Materials (AMG.AS) | Tantalum and lithium (Mibra mine), antimony |
| Australia | Lynas Rare Earths (LYC.AX) | Rare earths, ex-China separation |
| Argentina | Lithium Argentina (LAR) | Lithium brine (Cauchari-Olaroz) |
| Brazil | Niobium (CBMM, private) | Monitored, unweighted: no listed instrument |
| Angola | Pensana (Longonjo) | Monitored, unweighted: development stage |
| Mozambique | Syrah Resources | Monitored, unweighted: Balama production deferred to Q3 2026 |
One index for AI’s memory bottleneck.
Memory is now the expensive part of AI. TCC measures how many long conversations fit on one flagship chip. It rises when chips gain memory or when models get leaner, so up is always good.
Tantalum Context Capacity
The memory wall has two sides: how much memory each chip carries (the supply side, set by memory makers) and how much each conversation consumes (the demand side, set by model architects). TCC puts both in one number that rises when either side improves. HBM per flagship NVIDIA GPU has been flat at 288 GB since Blackwell Ultra, so every gain in the headline TCC since then has come from leaner models. The companion TCC-X reading uses the largest-HBM accelerator from any vendor; it stepped up to 432 GB with AMD's MI455X in Q3 2026.
Computed for Rubin / B300 (288 GB of HBM), counting conversation memory only. A stylized gauge: real servers also hold model weights.
Read the methodologyAverage across the leading open models for one 128K-token conversation. 67% less than in Jan 2025.
Rubin / B300. Flat since Jul 2025.
AMD MI455X (432 GB). TCC-X reads 327.5: same construction, largest-HBM accelerator from any vendor.
| Date | Reference GPU | HBM (GB) | GB per conversation | Conversations/GPU | TCC |
|---|---|---|---|---|---|
| 2025-01-01 | B200 | 192 | 18.72 | 10.3 | 48.8 |
| 2025-04-01 | B200 | 192 | 15.87 | 12.1 | 57.3 |
| 2025-07-01 | B200 | 192 | 10.10 | 19.0 | 90 |
| 2025-10-01 | B300 | 288 | 13.44 | 21.4 | 101.4 |
| 2026-01-01 | B300 | 288 | 13.64 | 21.1 | 100 |
| 2026-04-01 | B300 | 288 | 9.75 | 29.5 | 139.8 |
| 2026-07-01 | B300 | 288 | 6.99 | 41.2 | 195.3 |
| 2026-10-01 | Rubin / B300 | 288 | 6.25 | 46.1 | 218.5 |
| Lab | Model | GB per conversation | Architecture |
|---|---|---|---|
| DeepSeek | DeepSeek-V4-Pro (1.6T) | 1.30 | 30 CSA layers (4x compression + indexer) / 31 HCA layers (128x) + 128-token window, 1 KV head x 512 |
| NVIDIA | Nemotron 3 Ultra (550B) | 1.52 | 9 attention layers (GQA 2x128) + 36 Mamba-2 layers (fixed state) + 36 MoE layers |
| Meta | Muse Glimmer 30B | 1.83 | 13 global + 39 sliding (2,048), GQA 2x128 (dense, multimodal) |
| Moonshot | Kimi-K3 (2.8T) | 4.08 | 24 MLA layers + 69 KDA linear-attention layers (fixed state) |
| OpenAI | gpt-oss-120b | 4.84 | 18 full + 18 sliding (128), GQA 8x64 |
| Mistral | Mistral-Large-3 (675B) | 9.21 | MLA 61 layers, latent 512+64 |
| Gemma-4-31B | 11.58 | 10 global (4 KV heads x 512) + 50 sliding (1,024; 16x256) | |
| Zhipu | GLM-5.3 (744B) | 12.48 | MLA 78 layers + indexer on 21 layers (shared on 57) |
| Qwen | Qwen3.8-2.4T-A95B | 12.94 | 23 full-attention layers (GQA 4x256) + 69 Gated DeltaNet layers (fixed state) |
| MiniMax | MiniMax-M3 (428B) | 16.11 | GQA 4x128, 60 layers (block-sparse attention reads, full cache kept) |
What we publish, what we don’t.
Constituents, instruments, weights, the thesis behind each cut and the splice log are public, and the computation scripts live in the site repository. Only the watermark seed is not published.