From copper ore to a single inference

Break AI into minerals, industrial gases, chipmaking equipment, power, racks, networking, models and applications. Reading along the chain separates supply constraints, customer ties and valuation exposure. Downstream businesses still depend on upstream equipment and energy.

Scroll down and the figure highlights as you go. You can also click any link in the figure to jump there.

Scroll down and the colour bar at the top follows you. Each link opens with its own composition bar, and a full-chain table sits at the end.

• Market cap and revenue use different bases. In accelerator design, market-cap weight and the selected revenue proxy differ by about 3.1×; the gap is not evidence of mispricing.
• Multi-link companies are counted at full market cap in each link. This cannot allocate business value or bound industry concentration.
• Unlisted does not mean no indirect exposure. OpenAI and Anthropic have no standalone listed shares; listed investors and suppliers have different risks and returns.

The next three sections take these apart one by one. In a hurry for the chain itself? Click the figure.

Node size = company market cap ÷ sum of market caps of all nodes in its link (a market-cap proxy, not market share).

Market cap reflects company valuation; revenue reflects specific activities and periods. Accelerator design illustrates the difference:

Bars describe how sample market cap is distributed. The revenue proxy mixes business scopes and fiscal periods; it is not pure accelerator shipment share and cannot establish overvaluation or undervaluation against market-cap weights.

Multi-link entities enter each link at full market cap. Link weights use the de-duplicated chain total, so the 18 weights can sum to more than 100%. These sample valuation weights are neither business shares nor bounds on industry concentration.

Microsoft has seats in custom accelerators, cloud and applications. Filings provide no separately traded market cap for these activities. This map makes no segment valuation; it retains full company value and marks repeated inclusion. The red hatching in the figure is the companies below.The table below lists them all.

Links with high multi-link market-cap weights: Models 96.5% · Advanced Packaging 96.4% · Advanced Foundry 96.3% · Accelerator Design 94.1% · Cloud & Compute 93.3% · Networking & Optics 83.5% · Applications & Data 72.4% · Grid & Electrical 69.1%. This indicates mixed business exposure; it does not quantify market-share bias or establish upper or lower bounds.

Ranked by each entity’s connections among 892 disclosed relationships, including signed agreements and existing business ties reported in filings. Listing status and link inclusion are separate dimensions. Counts depend on this map’s coverage, not total contract value, revenue or industry standing.

Anthropic ranks 2 and OpenAI 8. Their compute purchases, financing and infrastructure ties make them important buyers in this sample; relationship counts do not establish their share of industry demand.

• Without standalone listed shares, listed investors, suppliers and creditors can offer indirect exposure, with economics and risks distinct from owning the model companies.
• The whole map has 182 genuinely unlisted counterparties, 154 entities that are listed but outside this chain (WMT, AEP, Barclays), and 164 anonymous counterparties (appearing in filings as “a customer”). Do not read the three as one group.
• Anonymous counterparties are recorded separately for each discloser; identities may overlap or differ. Node counts cannot establish customer concentration without comparable transaction values and a complete denominator.
• CR3 here is the combined weight of the three largest market caps in the selected sample. Other ratios, such as customer revenue shares, retain their explicitly stated source basis.

By weight, base metals sit near the top of a data center bill of materials: copper busbars, transformer windings, rack structural steel, cooling piping. What this link sells is mining and smelting capacity. Pricing follows long-term contracts and spot, two or three processing steps removed from AI demand.

This link is an origin on this map; nothing feeds into it.

Copper, rare earths, aluminum and steel feed racks, busbars, transformers and magnets. This is a US-listed mining and refining sample; market-cap weight is neither material volume nor a measure of AI’s impact on commodity prices.

Entities directly connected to companies in this link but not on the 18-link list. Some are genuinely unlisted (OpenAI, Anthropic); others are listed but outside this chain. Dotted boxes are anonymous counterparties that appear in filings as "a customer". The number is how many relationships connect to this link.

• 10 listed entities, CR3 66.6%. BHP, Southern Copper and Rio Tinto take most of the weight; MP Materials, the rare earth name, holds only 1.1%. computed
• A rise in metal prices does not reach the data center directly. Copper and grain-oriented electrical steel enter transformer makers' costs first, then feed through quoting cycles into equipment lead times. That stretch is covered in link 10. C
• Mine capacity expands on its own cycle. A new mine typically takes 10 years or more from exploration to first production, and near-term orders do little to change supply. This chapter has not run filing-level verification; the readings rest on the market-cap snapshot. C

Generation needs fuel first. This link is the intake for the generation side: gas pipelines, LNG terminals, gas fields, plus three uranium mining and nuclear fuel names. It produces no electricity itself, yet it determines whether newly installed gas turbines can run at full load for years.

This link is an origin on this map; nothing feeds into it.

Gas production, pipelines, LNG terminals, uranium mining and nuclear-fuel processing sit upstream of generation. Fuel risk depends on location, transport and processing capacity, not commodity quotes alone. Generation equipment and operators are in link 11.

Entities directly connected to companies in this link but not on the 18-link list. Some are genuinely unlisted (OpenAI, Anthropic); others are listed but outside this chain. Dotted boxes are anonymous counterparties that appear in filings as "a customer". The number is how many relationships connect to this link.

• 12 entities worth $502.2B combined, No. 16 of the 18 links by market cap and 1.23% of the chain. computed
• Six pipeline and LNG terminal names, Williams, Kinder Morgan, Cheniere, ONEOK, Energy Transfer and Targa, hold 78.4% together; the three producers, EQT, Antero and Expand Energy, hold 12.5%. computed
• Three uranium and nuclear fuel names, Uranium Energy (mining), Cameco (mining and conversion) and Centrus (enrichment), hold 9.1% together; the entities that turn fuel into power sit in link 11. computed

Industrial gases support wafer fabrication, etching and cleaning, often through long-term on-site supply arrangements. Interruptions can cause downtime, yield loss or scrapped work in process; the outcome depends on the process, backup supply and duration.

This link is an origin on this map; nothing feeds into it.

This link includes industrial gases, specialty chemicals and materials. Linde and Air Products are the industrial-gas names in this sample. Qualification and supply reliability affect switching costs; interruption risk alone does not establish pricing power.

Entities directly connected to companies in this link but not on the 18-link list. Some are genuinely unlisted (OpenAI, Anthropic); others are listed but outside this chain. Dotted boxes are anonymous counterparties that appear in filings as "a customer". The number is how many relationships connect to this link.

• 9 entities, CR3 84.2%. Only two of them, Linde and Air Products, are true industrial gas suppliers. computed
• Downstream sit three silicon-side links: leading-edge logic, advanced packaging and semiconductor equipment. All three draw their consumables from here. C
• Reliable supply and process qualification raise switching costs. Not every interruption scraps an entire wafer lot, and this operational risk alone does not establish pricing power. C

EDA provides chip-design and verification tools; IP supplies reusable circuit blocks. These businesses depend less directly on new wafer fabs, but talent, verification, software and hardware delivery, licensing and export restrictions can still constrain growth.

This link is an origin on this map; nothing feeds into it.

EDA and IP growth depends on R&D, verification, licensing and export compliance. Direct dependence on new wafer fabs is lower, but supply is not unlimited. CR3 measures the sample’s market-cap mix, not EDA or IP business share.

No mid-cap watchlist names in this link. Large/mega and small/micro caps remain in their respective groups. The all-seats basis describes sample valuation weights.

Entities directly connected to companies in this link but not on the 18-link list. Some are genuinely unlisted (OpenAI, Anthropic); others are listed but outside this chain. Dotted boxes are anonymous counterparties that appear in filings as "a customer". The number is how many relationships connect to this link.

• 4 entities, CR3 97.8%: Arm at 64.4%, Cadence and Synopsys around 15% each, and Rambus, the HBM interface IP name, at 2.2%. computed
• This link files quarterly, so the figures can be checked quarter by quarter. Further down, within a single year, disclosure granularity fell in the same direction across semiconductor equipment (ASML stopped reporting quarterly net bookings; Onto stopped breaking out dollar amounts), optical interconnect (Lumentum collapsed into a single reportable segment) and the data-center physical layer (Vertiv stopped reporting quarterly orders). TSMC, asked to break out advanced-packaging capex as a separate line, explicitly declined: granularity held rather than fell. A
• The limits are written into export controls and license terms, not into production lines. A license can be revoked; a fab does not vanish. C

Accelerators are the chips built to do matrix multiplication and little else. Whoever's silicon gets written into the training script takes the largest gross-profit pool in the chain. This link also has a structure found nowhere else: several of its largest customers are themselves listed inside it.

NVIDIA is about 77.6% of the selected-vendor revenue proxy for the fiscal periods mapped to 2026Q2, versus 82.8% for 2025Q2, when Marvell did not disclose data-center revenue and only four names were included. Excluding Marvell gives about 79.1% currently. Business scopes and fiscal dates differ; this is not pure accelerator industry share or evidence of distorted market caps.

No mid-cap watchlist names in this link. Large/mega and small/micro caps remain in their respective groups. The all-seats basis describes sample valuation weights.

Entities directly connected to companies in this link but not on the 18-link list. Some are genuinely unlisted (OpenAI, Anthropic); others are listed but outside this chain. Dotted boxes are anonymous counterparties that appear in filings as "a customer". The number is how many relationships connect to this link.

• 11 entities worth $21.7T combined, 52.97% of the chain, No. 1 of the 18 links by market cap; CR3 61.1%. computed
• On accelerator-vendor filed revenue (NVIDIA Data Center + AMD Data Center segment + Broadcom AI semiconductor + Marvell data center + Cerebras): NVIDIA's share was 77.6% in the second quarter of 2026, against 82.8% in the second quarter of 2025. The Broadcom cell for the second quarter of 2026 now carries the FQ3 actual of $16.7 billion, replacing the earlier $16.0 billion guidance, 14.6% of the denominator, evidence grade A. The two denominators are not the same: in the second quarter of 2025 Marvell had not yet broken out data-center revenue, so the denominator had four names; excluding Marvell from the second quarter of 2026 as well gives 79.1%, a like-for-like drop of about 3.7 percentage points, with the remaining roughly 1.5 points coming from the wider denominator. A
• The loosening shows up in contracts first: as of 2026-05-03, Broadcom's remaining performance obligations went from $45.0 billion to $164.6 billion in a single quarter, including one long-dated custom accelerator contract the customer has no right to terminate; by 2026-08-02 the balance had risen again, to $179.2 billion. A
• That share excludes Google's and Amazon's in-house silicon, which never enters the denominator of the filed figures. On 2026-09-03, though, Qualcomm issued a warrant to an Amazon affiliate for up to 25 million shares at $161.26, vesting in tranches against signed commercial arrangements, binding purchase orders and actual purchases, with cumulative payments capped at $60 billion. Server-chip purchases that move to a listed vendor land in filed revenue once they happen, so this is one way the denominator of this chapter widens later; the 8-K says only “server chip products, technology, systems and manufacturing services” and does not say how much of that is accelerators. The $126.6 billion supply-side hardware net figure is a separate basis spanning drives, switching and power, and this chapter does not use it as a share denominator. B

A finished layout has to be printed onto silicon, and the capacity to print the most advanced layers is highly concentrated. TSMC raised its 2026 capex guidance in July. The money goes in on the company's own schedule, but the tightest step is not this one.

TSMC’s July 2026 company guidance puts annual capex at $60–64B. Spending, installation, yield and deliverable capacity are distinct stages. Being untiered in this map does not guarantee that capacity arrives on plan.

No mid-cap watchlist names in this link. Large/mega and small/micro caps remain in their respective groups. The all-seats basis describes sample valuation weights.

Entities directly connected to companies in this link but not on the 18-link list. Some are genuinely unlisted (OpenAI, Anthropic); others are listed but outside this chain. Dotted boxes are anonymous counterparties that appear in filings as "a customer". The number is how many relationships connect to this link.

• The Q2 2026 earnings call in July 2026 raised capex guidance to $60 billion to $64 billion, against $40.9 billion actual in 2025 on the 20-F narrative basis (the cash-flow statement shows $40.56 billion). Board-approved capital appropriations over the last four quarters grew 76.1% year over year. B
• Advanced packaging is not disclosed separately: wafer revenue was 86% of net revenue in 2025, while the residual other line grew 41.2% year over year, ahead of wafer at 30.1%. A
• CR3 for this link is 98.3%. Intel Foundry, UMC, GlobalFoundries and Tower hold 24.7% between them. computed

A logic die and several HBM stacks have to be assembled onto one substrate. That step is advanced packaging. Its capacity is limited by deliveries of thinning, dicing, bonding and molding equipment.

Advanced packaging coordinates assembly lines, substrates, HBM and equipment deliveries. CoWoS-related capacity is monitored as an expansion that takes time. Important Japanese and European suppliers do not imply zero US-listed exposure. TSMC’s overall process-capacity outlook is not a packaging-specific convergence date.

No mid-cap watchlist names in this link. Large/mega and small/micro caps remain in their respective groups. The all-seats basis describes sample valuation weights.

Entities directly connected to companies in this link but not on the 18-link list. Some are genuinely unlisted (OpenAI, Anthropic); others are listed but outside this chain. Dotted boxes are anonymous counterparties that appear in filings as "a customer". The number is how many relationships connect to this link.

• TSMC's chairman said in July 2026, on the second-quarter earnings call, that packaging capacity is tight enough to constrain customer growth, and that the back-end gap is wider than previously indicated. B
• Revenue is never broken out; the only figures are verbal, from earnings calls: the chairman put advanced packaging at slightly above 10% of net revenue in October 2025, and the CFO in January 2026 gave slightly above 10% for 2025 with the low teens expected for 2026. B
• Advanced-packaging equipment includes Japanese dicing, molding and bonding vendors and Dutch BESI in hybrid bonding. Major suppliers are not all in this US-listed sample; that does not eliminate indirect or diversified US-listed exposure. The link has 4 sample companies and market-cap CR3 of 99.6%. C

AI needs both HBM close to processors and nearline hard drives for high-capacity storage. Their bandwidth, latency, capacity expansion and capital-spending strategies differ, so their shortages need separate analysis.

HBM and nearline HDD serve different workloads; track bit output, shipments and capex separately. HDD supply depends on areal density as well as media/head capacity, drive configuration, yield and mix. A third-party density estimate is not an upper bound on industry supply growth. Announced HBM capacity plans are not commissioning guarantees.

No mid-cap watchlist names in this link. Large/mega and small/micro caps remain in their respective groups. The all-seats basis describes sample valuation weights.

Entities directly connected to companies in this link but not on the 18-link list. Some are genuinely unlisted (OpenAI, Anthropic); others are listed but outside this chain. Dotted boxes are anonymous counterparties that appear in filings as "a customer". The number is how many relationships connect to this link.

• SK hynix held a 56.4% HBM revenue share in the first quarter of 2026, an IDC figure cited in the 424B4 prospectus it filed on 2026-07-10; third-party estimates of that share are scattered from 50% to 62%. Counterpoint's reading for calendar Q2 2026, published on 2026-09-01, puts SK hynix at 50%, Samsung at 33% and Micron at 18% (a third-party estimate, not a filed figure; on that firm's own series SK hynix was at 64% a year earlier). A
• Micron's FQ3 (ended 2026-05-28) grew revenue +346% year over year while bit shipments rose only around 20%. The rest is price. Reading memory revenue straight through as a demand indicator gets it wrong. A
• Western Digital grew revenue 36% and moved capex only from $412 million to $418 million; Seagate went from $254 million to $569 million in two years. Both guide capex to 4% to 6% of revenue, and actuals in most years land below the range. At Citi's TMT conference on 2026-09-09, Western Digital's CFO said the long-term range itself is unchanged, still around 4% to 6% viewed over five-year periods, but that the past two years ran below it and the next two years may run slightly above it (spoken management guidance on an annual basis, not filed guidance). A

Equipment is the machine tool of chipmaking. This link holds more companies than any other on the map, yet conditions inside it are stratified: some names contracted for all of 2025, while in 2026 some next-quarter guidance implies revenue doubling.

Equipment submarkets differ. EUV lithography and ATE are constraints tracked separately here. ASML’s low-NA EUV capacity plan is forward-looking, not the whole lithography market or actual delivery count. Packaging equipment also involves Japanese and European suppliers.

Entities directly connected to companies in this link but not on the 18-link list. Some are genuinely unlisted (OpenAI, Anthropic); others are listed but outside this chain. Dotted boxes are anonymous counterparties that appear in filings as "a customer". The number is how many relationships connect to this link.

• ASML is the central EUV lithography supplier examined here; EUV’s competitive structure cannot be generalized to all lithography. The cited plan calls for about 65 low-NA EUV systems in 2026 and 30% more capacity in 2027; actual deliveries still need verification. A
• A verification metric is disappearing. ASML stopped publishing quarterly net bookings from the first quarter of 2026: the figure appeared four times in the last release that still carried it (Q4 2025) and zero times in the Q2 2026 release. A
• Dispersion is wide: revenue was −17.6% at Axcelis for calendar 2025 and −2% at Photronics for FY2025 (year ended October 2025), while Kulicke & Soffa guided its next quarter to an implied +111.2% year over year (a +99.9% to +122.5% range). In the latest reported quarter, traditional ball bonding was 68.4% of revenue and over 80% of the year-over-year increase; the advanced-packaging segment is only 8.9% of revenue and loss-making, so reading the number as AI exposure is a mistake. A
• 22 entities, CR3 74.6%. The weight concentrates in four front-end equipment makers. computed

Server assembly puts GPUs, memory, power supplies and chassis into a finished box and racks it. Shipments depend on how much silicon upstream is willing to allocate, which is why pricing power does not sit with the assembler.

This sample covers server brands and EMS/ODM companies. GPU and component supply affects delivery, while bargaining power also depends on customers, design capabilities and contracts. Taiwan-listed ODMs such as Hon Hai, Quanta and Wistron are outside this map.

No mid-cap watchlist names in this link. Large/mega and small/micro caps remain in their respective groups. The all-seats basis describes sample valuation weights.

Entities directly connected to companies in this link but not on the 18-link list. Some are genuinely unlisted (OpenAI, Anthropic); others are listed but outside this chain. Dotted boxes are anonymous counterparties that appear in filings as "a customer". The number is how many relationships connect to this link.

• 9 entities worth $605.3B combined, 1.48% of the chain. computed
• Super Micro grew revenue +93% year over year in fiscal Q4 2026 (quarter ended 2026-06-30), the same order of magnitude as NVIDIA. Gross margin for the quarter was 17.5%; NVIDIA's was 75.0%. A
• Celestica's CCS segment rose from 72% of company revenue in Q2 2025 to 81% in Q2 2026 (quarter ended 2026-06-30), with segment margin still at 8.66%. A

Hundreds of thousands of GPUs have to talk to each other, and light carries the traffic. Optical module assembly has many participants, and the global leader, Innolight, is listed on the China A-share market and does not appear on this map. The scarcity sits in one part inside the module: the indium phosphide laser chip. US-listed exposure to copper connectors, optical fiber and indium phosphide substrates (Amphenol, Corning, AXT) has been added to this layer.

This link separates switching/protocols, interconnect chips, optical modules/components, copper links and fiber. AXT offers InP-substrate exposure. Laser shortages and expansions require company- and product-level checks. CPO/LPO alter parts of the architecture; a third-party forecast date is not a guaranteed end to shortages.

Entities directly connected to companies in this link but not on the 18-link list. Some are genuinely unlisted (OpenAI, Anthropic); others are listed but outside this chain. Dotted boxes are anonymous counterparties that appear in filings as "a customer". The number is how many relationships connect to this link.

• In March 2026 NVIDIA invested $2 billion each into Lumentum, Coherent and Marvell inside 30 days. The first two bought purchase commitments and capacity access; the Marvell investment is a silicon photonics joint development. A
• The gap is around 30%: a third-party transcript of the company's earnings call and LightCounting put it at the same order of magnitude, and both are third-party readings. Capacity is already up 20% and the imbalance widened anyway; long-term contracts lock existing capacity through 2027. Broadcom's FQ3 2026 call on 2026-09-02, for the quarter ended 2026-08-02, added one more supply-side reading: management said demand for EML and CW lasers is far surpassing industry supply, and that Broadcom is more than tripling capacity year on year at its EML, CW, VCSEL and indium-phosphide plants in the US and Singapore. That is a qualitative statement about capacity under construction, not capacity already in place. C
• Pricing power lands with the name closest to the chokepoint: Lumentum expanded non-GAAP gross margin by 1,130 basis points in fiscal 2026 (year ended June 2026), reaching 50.4% in the fourth fiscal quarter; Coherent expanded 152 basis points in its own fiscal 2026, which also ended in June 2026. A
• 19 listed names, CR3 85.7%. Each of the four sub-segments has its own structure: the Ethernet-versus-InfiniBand protocol contest sits on the switching side, and CPO/LPO is an alternative path on the module side. Neither has changed the scarcity of laser chips. computed

This link covers switchgear, UPS, batteries, coolant-distribution units, cold plates and fluids. Boyd Thermal belonged to privately held Boyd before Eaton acquired it in March 2026; no standalone listed liquid-cooling company was delisted. US-listed exposure still depends on each company’s business mix.

Covers switchgear, UPS/batteries, air cooling/HVAC, liquid cooling/fluids and board-level power. Eaton acquired the privately held Boyd Thermal business, not a standalone listed liquid-cooling company. Backup generators are in link 11 and grid transformers/GIS in link 10. Untiered does not mean unconstrained delivery.

Entities directly connected to companies in this link but not on the 18-link list. Some are genuinely unlisted (OpenAI, Anthropic); others are listed but outside this chain. Dotted boxes are anonymous counterparties that appear in filings as "a customer". The number is how many relationships connect to this link.

• Eaton acquired Boyd Thermal for $9.549 billion in March 2026. Intangibles plus goodwill came to 114% of consideration; what was bought is capacity and customer relationships. A
• Vertiv stopped reporting quarterly orders from February 2026. The shares were +24.5% on the day of the announcement; the last disclosed quarter, published the same day, showed orders +252% year on year. Five months later revenue came in below expectations and the stock was −17.3% in a single session, against −2.0% for the Nasdaq 100 that day. Signed on 2026-09-01 and announced on 2026-09-02, Vertiv is acquiring UtilityInnovation Group, a microgrid and behind-the-meter power architecture company, for about $1.45 billion in cash at closing, subject to customary adjustments, plus contingent consideration of up to $1.15 billion payable in two instalments against 12-month and 24-month EBITDA targets. The company puts the $1.45 billion at roughly 13 times UIG's expected 2027 EBITDA (a non-GAAP measure, and a company projection for a future calendar year rather than an actual). The deal awaits antitrust clearance and is expected to close in the fourth quarter of 2026. B
• Modine is the only company in this link with a data center segment of its own. In the first quarter of fiscal 2027, the three months to 30 June 2026, the segment was 39.9% of revenue against 26.9% a year earlier, while its gross margin fell from 29.8% to 20.2%, a drop of 960 basis points. The denominator behind that revenue share is about to change: Modine expects to spin off its Performance Technologies segment and combine it with Gentherm on 1 October 2026 (record date 28 September; Gentherm shareholders approved on 10 September), and says it will classify the segment as a discontinued operation from the period the deal completes. The cash the spun-off entity pays Modine before the merger is expected to fall from $210 million to $159 million, with the final amount set at closing. After the spin-off Modine keeps only its Data Centers and Commercial HVAC segments and plans to rename itself Modexus Solutions, keeping the MOD ticker; the name change requires shareholder approval. A
• 19 entities, CR3 45.7%. Backup generator sets now sit in link 11; UPS batteries and immersion coolants are added in this edition. computed

Connecting a data center to the grid means completing everything from the high-voltage line to the switchgear. This link is not short of orders. It is short of an equipment delivery calendar.

Contractors and electrical-equipment vendors share this link; diversified names use full-company market cap, not transmission/distribution business value. Transformer/GIS lead times vary by specification, region and order. Third-party long-lead examples and 2027–2030 expansion plans are research inputs, not a uniform delivery schedule.

Entities directly connected to companies in this link but not on the 18-link list. Some are genuinely unlisted (OpenAI, Anthropic); others are listed but outside this chain. Dotted boxes are anonymous counterparties that appear in filings as "a customer". The number is how many relationships connect to this link.

• Substation-class transformer lead times run beyond 160 weeks, high-voltage breakers around 125 weeks. Of announced expansions, the earliest completes in mid-2027, with the bulk landing between 2028 and 2030 (lead times are third-party estimates from Wood Mackenzie and others, consistent across sources). C
• This link holds both contractors and T&D equipment makers: GE Vernova (Electrification segment), Eaton, Hubbell and Powell are counted across layers at full-company market cap, together 69.1%. There is no pure-play high-voltage transformer maker listed in the US, so the top weight's core business extends well beyond this link. computed
• Backlog is not orders: MasTec states that about 40% of its 18-month backlog comes from cancellable master service agreements, while at Quanta such agreements are 41% of total backlog and 33% of the 12-month figure. The two companies define backlog differently and the numbers are not comparable. A

This link is the chain's power outlet: gas turbines, nuclear and conventional generation, and on-site generation and backup power. Unusually, it writes its capacity path into public materials. Which year capacity is added, and how much can be built in a year, are both verifiable.

This link includes turbine OEMs, generators, on-site/backup equipment and SMR developers; fuel is in link −1. GE Vernova’s 53GW equipment orders and 116GW including cancellable reservations have different bases. Against 20GW annual output they imply static ratios of about 2.7 and 5.8 years, not delivery forecasts. Other suppliers, existing plants and new projects can change regional supply.

Entities directly connected to companies in this link but not on the 18-link list. Some are genuinely unlisted (OpenAI, Anthropic); others are listed but outside this chain. Dotted boxes are anonymous counterparties that appear in filings as "a customer". The number is how many relationships connect to this link.

• GE Vernova's equipment order backlog is 53 GW, or 2.7 years against 20 GW of annual output. Counting the 116 GW of reserved production slots, it is 5.8 years. Both readings belong on the page, because reservations can be cancelled. A
• Prices have already moved: GE Vernova states that equipment order pricing in the first half of 2026 ran more than 20% above the fourth quarter of 2025. B
• At the source cutoff, future commercial output and commissioning dates for SMRs remain uncertain. Track Oklo’s NRC licensing work separately from its DOE authorization path. NRC’s project page updated September 15, 2026 lists Aurora pre-application activities. Power delivery depends on each project’s authorization, construction and commissioning. B
• 13 entities, CR3 68.9%. GE Vernova alone holds 22.2%; the three backup-power names (Caterpillar, Cummins, Generac) together hold 40.3%. computed

This link leases already-energized megawatts on long-term contracts, mostly to cloud providers and model companies, on terms running from the teens into the twenties in years. Its market-cap weight is far smaller than the cloud link's (the two are not measured alike: more than 90% of the cloud link's market cap is the full-company value of giants that span several layers), yet it carries the longest-dated contracts in the chain.

REITs, hosts and mining-to-AI projects differ in contract term, capacity basis, energization and revenue recognition. Disclosed base-term values must not be added to revenue-derived estimates as one signed-contract total; queues and conditional classifications are not secured power. Full market caps of AMT, IRM and others cannot establish link business value or concentration bounds.

Entities directly connected to companies in this link but not on the 18-link list. Some are genuinely unlisted (OpenAI, Anthropic); others are listed but outside this chain. Dotted boxes are anonymous counterparties that appear in filings as "a customer". The number is how many relationships connect to this link.

• Read contract values deal by deal, separating base terms, renewals, options and revenue-derived estimates; estimates must not be added into one signed-contract total. Capacity must distinguish grid access, site power, IT load and billing status. Inconsistent bases cannot support a contracted-to-operating multiple. C
• Interest arrives 12 to 24 months ahead of rent: Cipher's quarterly interest was 269% of revenue in the same period, TeraWulf's 126%. A
• Google is not a lessee, yet it is the credit hub of this link: it backstopped $6.23 billion for the unlisted Fluidstack in exchange for warrants struck at one cent. A

Cloud is the link with the densest circular transactions in the chain: the party writing the check, the party selling the silicon and the party renting the rack are often the same set of names. In the previous link Google backstopped a host operator. Here, the backstop has become a standard term.

Oracle reported $664B RPO at August 31 in its September 10, 2026 release. Historical customer revenue shares in the FY2026 10-K do not reveal current RPO concentration. Cloud, compute purchases and investments require separate checks. Repeated full-company market caps describe sample valuation weights, not circular transactions or industry concentration.

No mid-cap watchlist names in this link. Large/mega and small/micro caps remain in their respective groups. The all-seats basis describes sample valuation weights.

Entities directly connected to companies in this link but not on the 18-link list. Some are genuinely unlisted (OpenAI, Anthropic); others are listed but outside this chain. Dotted boxes are anonymous counterparties that appear in filings as "a customer". The number is how many relationships connect to this link.

• Oracle’s September 10, 2026 release reports $664B of RPO for Q1 FY2027 ended August 31, up from $638B at May 31; the 10-Q filed September 11 expects about 13% of it to be recognized as revenue over the next twelve months, 37% in months 13 to 36 and 34% in months 37 to 60. Its FY2026 10-K said no single customer accounted for 10% or more of annual revenue. That historical recognized-revenue measure does not establish the customer mix of current RPO. A
• Backstopping unsold capacity went from one-off to standard term: NVIDIA to CoreWeave at $6.3 billion, Meta to Nebius at $15 billion. A
• CoreWeave's quarterly net interest expense was $640 million, equal to a quarter of revenue in the same period. A
• This link totals $13.44T, 32.81% of the chain, ranking No. 2 of the 18 links by market cap. computed

The model link places the orders that travel upstream: power, racks and silicon all originate here. What is buyable on US exchanges is only part of it, and the two entities that actually define this link are not on the list.

The listed sample includes SpaceX (xAI), Alibaba and Baidu at full-company market caps, not model-only valuations. OpenAI and Anthropic’s standalone listing status requires confirmed exchange trading. Confidential submission, a public S-1 and listing are distinct events; filing reports alone do not change classification.

No mid-cap watchlist names in this link. Large/mega and small/micro caps remain in their respective groups. The all-seats basis describes sample valuation weights.

Entities directly connected to companies in this link but not on the 18-link list. Some are genuinely unlisted (OpenAI, Anthropic); others are listed but outside this chain. Dotted boxes are anonymous counterparties that appear in filings as "a customer". The number is how many relationships connect to this link.

• OpenAI and Anthropic have no standalone listed shares. Among 892 disclosed company relationships recorded here, Anthropic has 25 ties and ranks 2; OpenAI has 18 and ranks 8. Counts include signed agreements and disclosed existing business ties, not individual legal contracts. computed
• Of Amazon's $62.6 billion in net income for the second quarter of 2026, $53.4 billion was pre-tax other income. That figure is on a different basis from after-tax net income and cannot simply be subtracted from it. The filing attributes it mainly to Amazon's investment in Anthropic, which is not listed. A
• SpaceX listed after combining with xAI and is the only one of the eight filing-level buyers that prints AI segment capex into its filings: $23.55 billion in the first half. A
• Only 5 listed entities sit in this link, CR3 96.5%. computed

This is one of the links closest to end payment: contracts and leases upstream are ultimately settled by subscription, licence and model API revenue. Its boundary is also the blurriest on the map: these 13 names are a sample, not the full set.

This link is a terminus on this map; nothing flows out of it.

Applications and data services have open boundaries; this is a representative sample. Companies mix AI with other activities, so revenue and profit require segment-level checks. A small market-cap weight does not mean immaterial revenue, and full-company valuation cannot establish AI application share.

No mid-cap watchlist names in this link. Large/mega and small/micro caps remain in their respective groups. The all-seats basis describes sample valuation weights.

Entities directly connected to companies in this link but not on the 18-link list. Some are genuinely unlisted (OpenAI, Anthropic); others are listed but outside this chain. Dotted boxes are anonymous counterparties that appear in filings as "a customer". The number is how many relationships connect to this link.

• Microsoft holds 72.4% of this link’s sample market cap and also appears in Accelerator Design, Cloud & Compute. Each link uses its full market cap; this cannot allocate application-business value or establish application market share. computed
• The sample is top-heavy: the bottom three names together are 0.1% of the link, and most real application revenue sits inside statements that never break it out. computed
• The model link is acquiring applications in the other direction: in June 2026 SpaceX signed a merger agreement with Cursor, which is not listed, and the deal closed on 2026-08-14 at an implied equity value of $60 billion. That figure is the deal's implied valuation and is not equivalent to a listed company's market cap. A

• Prices use Nasdaq historical daily closes; target close date 2026-09-25, with actual trading dates labelled per seat. Market caps are endpoint snapshots: retrieval date is not effective date, and the endpoint supplies no market-cap trading date. Basis: Nasdaq previous close × inferred quote-equivalent shares (MarketCap ÷ lastSalePrice); date derived from the source quote; retrieval status 175/180.
• Company-to-company network: compiled from filings and announcements up to the source-note cutoff 2026-09-22, built 2026-09-03, revised 2026-09-22; updated by hand, not daily.
• Narrative sources as of 2026-09-22, revised 2026-09-23. Facts retain their reporting periods. Market caps use the quote snapshot; company financial cards have a separately labelled filing-fetch date.
• Filed data: SEC XBRL companyfacts (primary filings, grade A). Revenue, profit and cash flow are trailing four-quarter sums; where a 10-K gives only the full year, the missing quarter is derived as full year minus the other three quarters of the same fiscal year, and year-to-date cash-flow figures are differenced by quarter; foreign filers (20-F / 40-F) are mostly on an annual basis and marked as such; amounts are in the reporting currency and no valuation multiples are computed for non-USD filers; filings not yet in companyfacts fall back to the latest available period, noted on the card.
• Node size = company market cap ÷ sum of market caps of all nodes in its link (a market-cap proxy, not market share).
• Multi-link entities enter each link at full market cap. Link weights use the de-duplicated chain total, so the 18 weights can sum to more than 100%. These sample valuation weights are neither business shares nor bounds on industry concentration.
• Above $10B: large/mega-cap short-side hedge research group. This is not a short signal and does not imply a risk-neutral combination with mid caps.
• Relationship amounts: the dataset’s v field is in billions of USD; capacity goes in gw / mw; pct is a share whose meaning follows the edge kind: share of revenue on customer edges, equity stake on capital edges, share of output or capacity on offtake / supply / PPA edges; an empty v means no amount was publicly disclosed. Valuations of unlisted entities (Anthropic’s Series H post-money of $965bn, for instance) are self-announced single-round post-money or deal-implied figures, not comparable to a listed market cap, and enter none of this map’s weights.
• Link tiers are research classifications based on disclosed orders, lead times and expansion plans: how long constraints may last and whether additional capacity can ease them. Tier 1: harder to ease in the near term; tier 2: expansion may help over time; tier 3: announced capacity milestones can be tracked. Plans are not guarantees, and an untiered link is not necessarily unconstrained.
• Counterparty classes: off-map entities are split by listing status into 182 genuinely unlisted, 154 listed but outside the 18 links, 36 subsidiaries of listed companies, 30 public bodies or sovereign funds, and 82 acquired; financial investors are kept separate and never enter the bars.
• Relationship counts use the dataset’s binding category for disclosed ties: signed contracts, completed transactions and existing customer or supplier relationships reported in filings, such as annual customer revenue shares. This is not a count of inspectable executed legal contracts. LOIs, frameworks, reservations, demonstrations, pending and unverified ties are excluded. Signed does not mean delivered or non-cancellable. Undisclosed identities remain anonymous; insufficiently sourced inferences are not treated as verified.

A/B/C identify source types, not a pass from independent verification of every relationship in this review. The 734 A-grade entries cite SEC or exchange filings; 85 carry accession numbers and 634 provide clickable URLs. An identifier or link does not establish that the document supports every paraphrased claim. The 262 B-grade entries cite company announcements, IR, websites or press; of 43 C-grade entries, 39 are structural inferences without independent primary sources and the rest third-party accounts. Missing links or weak support require further verification; grades do not upgrade relationship status. Narrative points have their own count and source grades.

• Seats: US-listed (ADRs included) with a filed business line or a filed contract in that link; each layer's specific test sits in its inclusion rule, and sampled layers claim representativeness only, never completeness.
• Cross-links: a company with filed lines in several links holds one seat in each, counted at full market cap every time (Rule 2).
• Edges: at least one end inside the 18 links, and a publicly checkable event (filing, company announcement or locatable press report); edges filled in from industry knowledge alone are graded C and say so.
• Known gaps: IC substrates and ABF, PCBs, passives, grid-scale storage, data-hall construction and EPC, T&D utilities, silicon wafers and photoresist (not covered in this version, mostly for lack of a US-listed pure play or because Japanese and European names dominate). Copper connectors and cable, fibre and glass, InP substrates, UPS batteries and immersion fluids were added on 2026-09-09. Entities the data pipeline did not cover: Qualcomm holds a seat in accelerator design but the source notes contain no filing-level record for it.

• US-listed only. Entities listed in Hong Kong, mainland China, Japan, Korea or Taiwan are absent except for a few ADRs, and those ADRs are counted at the issuer's whole-company market cap, not the US-listed portion alone, so bars for seats like SK hynix and TSMC run far longer than the float actually tradable in the US. Each layer's excluded non-US incumbents are listed under its inclusion rule.
• Sampled layers: Mining & Commodities, Energy & Fuel, Materials & Industrial Gases, Applications & Data. These links have open boundaries; their members are a representative sample rather than the full set, and CR3 and total cap describe the sample only.
• Prices and market caps may update at different times, and relationships and business structures also change. Seats whose historical close cannot be verified keep a previous value with a date-mismatch label. That value is not necessarily from the preceding trading day; check the seat’s actual date.
• Not personalised: this page is public research material, takes no particular investor's circumstances into account, and is not investment advice, an offer or a solicitation.
• No licence: we hold no investment-adviser registration in any jurisdiction and do not provide personalised advice.
• Data timing: closes have trading dates; the market-cap endpoint supplies a retrieval date but no effective date. Financial filings, relationships and narrative each have their own periods and update dates; they are not a synchronized snapshot.
• Third-party rights: company names and tickers belong to their owners and are used only to identify them; source links point to SEC EDGAR and the companies' own public pages.

Interest statement. 1) Positions: positions held by the people who compile and review this page, or their affiliates, in entities on the map are not disclosed one by one; readers should assume such positions exist and may run with or against the views here. 2) Business relationships: this site sells paid research reports on some entities on the map, identified by the paid-report links on this page; it has no employment, advisory or paid-promotion relationship with any entity on the map. 3) Other businesses: this site also operates EA sales and an AI API service; the instruments traded by the EAs may include entities on the map. 4) Revenue: this site earns from report sales, EA sales and AI API fees; this page is free and no entity has paid for it.

This page keeps no historical snapshots. When citing, include the version string "atlas 2026-09-25 · structure 2026-09-09 · graph 2026-09-22.1 · text 2026-09-23" and the section anchor (anchors are fixed by layer number, e.g. #sec-n10, and do not shift when layers are added or removed), and record the price close date and market-cap retrieval date separately. The source does not provide a market-cap effective date; the citation date is not a substitute for either data date.

“Multi-link share” sums the sample market-cap weights of multi-link companies; CR3 sums the top three in the selected sample. Neither measures business market share or bounds industry concentration. The mid-cap watchlist includes seats above $3B and at or below $10B. Large and mega caps form the short-side hedging research watchlist; small and micro caps are reviewed separately for higher risk. These site-defined groups are not trade recommendations or evidence that a hedge works.