From copper ore to a single inference

The figure on the leftThis chain is not decoration. It is this page's table of contents and its conclusion: take the word "AI" apart and you get mines, gases, lithography tools, transformers, gas turbines, racks, optical modules, models, and a chat box. Each link down is one step closer to the end product and one step further from physical constraints.

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. A full-chain table at the end lets you jump to any link.

• 1. Block size is market-cap share, not market share. In the accelerator-design link the two differ by more than 3×.
• 2. Companies spanning several links are counted more than once. Microsoft appears in accelerator design, cloud, and applications, at full market cap in each — so concentration in links containing giants is an upper bound.
• 3. Two of the busiest hubs in the network are not on any exchange. OpenAI and Anthropic rank in the top four by relationship count, and cannot be bought. It is this map's least convenient conclusion.

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).

That reads like a disclaimer. It is a convention that bites. Take the accelerator-design link:

So the bars here answer "how is tradable market-cap weight split within this link", not "who ships more". Read them as market share and you will be systematically wrong in every link that contains a giant.

Companies spanning several links are counted at full market cap in each, so shares in links that contain giants are upper bounds.

Microsoft designs its own accelerators, sells cloud, and sells Copilot. It holds a seat in all three links, each at full market cap. Splitting it is impossible: public filings do not disclose market cap by link. So this page keeps the double counting and marks it. The red hatching in the figure is the companies below.The table below lists them all.

Most affected links: Advanced Packaging 96.6% · Advanced Foundry 96.4% · Models 96.2% · Accelerator Design 95.2% · Cloud & Compute 93.2% · Networking & Optics 87.6% · Applications & Data 73.4%. Read concentration in these links as an upper bound.

Rank the 271 company-to-company relationships by how often each entity appears, and the top ten look like this. The two hatched bars are not on the 18-link list — they are unlisted companies.

Anthropic ranks 2, OpenAI ranks 4. They are the main source of demand at the bottom of this chain: follow any compute offtake, power agreement, or capital commitment above to its end and it often lands on them. On a map drawn on a US-listed basis, they can only be a footnote inside someone else's bar.

• What you can buy are these two companies' suppliers and creditors, not the companies themselves.
• The whole map has 62 off-map entities and 30 anonymous counterparties (appearing in filings as "a customer"). Off-map entities include the genuinely unlisted (OpenAI, Anthropic, xAI, Fluidstack) and the listed-but-outside-this-chain (WMT, AEP, Barclays); do not read them as one group.
• So every concentration figure and every share on this page is concentration within these 18 links, not concentration of the industry.

By weight, the largest item on a data center bill of materials is 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.

• 10 listed entities, CR3 66.7%. BHP, Southern Copper and Rio Tinto take most of the weight; MP Materials, the rare earth name, holds only 1.2%.
• A price increase has to clear a full quoting cycle before it lands: copper and grain-oriented electrical steel move transformer makers' costs first, and only then data center lead times. That stretch is link 10.
• Capacity expands on mine cycles. A new mine runs a decade from exploration to first production and does not flex with the order book.

Generation needs fuel first. This link is the intake for the generation side: gas pipelines, LNG terminals, gas fields, plus one uranium fuel name. 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.

• 7 entities worth $333.4B combined, the smallest of the 18 links at 0.85% of the chain.
• The weight stops midstream: Williams, Kinder Morgan, Cheniere and ONEOK, the pipeline and LNG terminal names, are 84% together; producers EQT and Antero are 14%.
• Uranium is a single name, Uranium Energy, at 1.7%. Nuclear fuel offers almost no tradable exposure in this link.

A fab consumes more than power and equipment. It runs on high-purity gases and specialty chemicals. Supply piping is laid straight into the site, follows the fab wherever it is built, and the contract often lasts as long as the fab does. If supply stops, work in process is scrapped in full. There is no intermediate setting called reduced output.

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

• 9 entities, CR3 83.9%. Only two of them, Linde and Air Products, are true industrial gas suppliers.
• Downstream sit the three tightest links in the chain: leading-edge logic, advanced packaging and semiconductor equipment. All three draw their consumables from here.
• Pricing power comes from supply that cannot be interrupted, not from volume. The top name here holds more than half the link; the top name in mining and bulk commodities holds under 30%.

A chip is compiled in software before it is built. EDA is that toolchain; IP is a catalog of ready-made circuit blocks, and buying an Arm CPU core saves years of in-house design. This link owns no plants, so no capacity ceiling exists. The only thing that can be restricted is the license.

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

• 4 entities, CR3 97.9%: Arm at 61.9%, Cadence and Synopsys around 20% each, and Rambus, the HBM interface IP name, at 2.1%.
• This link files quarterly, so the figures are verifiable quarter by quarter. Further down, disclosure granularity at leading-edge logic, semiconductor equipment and optical interconnect all moved lower in the same direction within a single year.
• The limits are written into export controls and license terms, not into production lines. A license can be revoked; a fab does not vanish.

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

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 $21.17T combined, 53.80% of the chain and the largest of the 18 links; CR3 63.3%.
• On filed revenue: Nvidia data center revenue was $89.0 billion in the second quarter of 2026, or 70.3% of the $126.6 billion de-duplicated supply-side net figure, against 69.9% a year earlier.
• The loosening shows up in contracts first: 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.
• That 70% reading excludes Google's and Amazon's in-house silicon, which never enters the denominator of the filed figures.

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.

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.

• Capex guidance was raised in July to $60 billion to $64 billion, against $40.9 billion actual in 2025. Board-approved capital appropriations over the last four quarters grew 76.1% year over year.
• 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%.
• CR3 for this link is 98.2%. Intel Foundry, UMC, GlobalFoundries and Tower hold roughly 20% between them.

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, not by floor space.

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 that packaging capacity is tight enough to constrain customer growth, and that the back-end gap is wider than previously indicated.
• Revenue is never broken out. The only figure comes verbally from the CFO: advanced packaging was slightly above 10% of net revenue in 2025, with the teens as a percentage expected in 2026.
• The physical chokepoint sits with equipment makers in Japan and the Netherlands, and US-listed exposure to that stretch is zero. Only 4 US-listed names sit in this link, CR3 99.6%.

AI needs two kinds of storage at once: HBM sitting against the compute, and nearline drives holding cold data. Both are short, but the shortages are opposite in character, and so is the direction of capex.

• SK Hynix held a 56.4% HBM revenue share in the first quarter of 2026, a figure printed in the body of its own US listing document.
• Micron's most recent quarter 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.
• Western Digital grew revenue 36% and moved capex only from $412 million to $418 million. Both drive makers hold capex at 4% to 6% of revenue.

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

• Lithography is the one absolute monopoly. ASML's capacity is fixed in units: roughly 65 low-numerical-aperture EUV systems in 2026, and 30% more in 2027.
• A verification metric is disappearing. ASML stopped publishing quarterly net bookings from the first quarter of 2026. The figure appeared four times in the prior release and zero times in this one.
• Dispersion is wide: 2025 revenue was −17.6% at Axcelis and −2% at Photronics, while Kulicke & Soffa guided its next quarter to an implied +111.2% year over year (a +99.9% to +122.5% range). Of that growth, 68.4% comes from traditional ball bonding; the advanced-packaging segment is 8.9% of revenue and loss-making, so reading the number as AI exposure is a mistake.
• 22 entities, CR3 74.9%. The weight concentrates in four front-end equipment makers.

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 stays upstream.

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 $559.4B combined, 1.42% of the chain.
• Super Micro grew revenue +93% year over year in its most recent quarter, the same order of magnitude as Nvidia. Quarterly gross margin was 17.5%; Nvidia's was 75.0%.
• Celestica's CCS segment went from 72% to 81% of company revenue in a year, with segment margin still at 8.66%.

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.

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.
• The gap is around 30%, with company management and industry researchers arriving at the same figure. Capacity is already up 20% and the imbalance widened anyway, with long-term contracts locking existing capacity through 2027.
• Pricing power lands with the name closest to the chokepoint: Lumentum expanded full-year non-GAAP gross margin by 1,130 basis points, reaching 50.4% in the most recent quarter; Coherent expanded 152 basis points over the same period.

This link builds the switchgear, UPS systems, coolant distribution units and cold plates inside the hall, and it is the part that actually sets the schedule. In 2026 its tightest sub-link left the public market.

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.
• Vertiv stopped reporting quarterly orders from February 2026. The shares were +24.5% on the day of the announcement, then −17.3% in a single session five months later when revenue came in short.
• Modine is the only company in this link with a data center segment of its own: the segment went from 26.9% to 39.9% of the business in a year while its gross margin fell 960 basis points.
• 18 entities, CR3 46.5%, the most fragmented link outside the silicon side.

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.

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.
• The largest weight in this link is Caterpillar at 57.4%, but what it sells into data centers is on-site generation and backup power, not grid assets.
• 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.

This link is the chain's power outlet: gas turbines, nuclear, and fuel. Unusually, it writes its capacity path into public materials. Which year capacity comes online, and how much can be built in a year, are both verifiable.

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.
• 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.
• Small modular reactors add zero incremental supply within 24 months, and Oklo has not yet filed a construction and operating license application with the NRC.
• 12 entities, CR3 70.2%. GE Vernova alone holds 33.2%.

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 a rounding error against the cloud link, yet it carries the longest-dated contracts in the chain.

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.

• Base-term contract value traceable deal by deal is roughly $160.9 billion, against 850 MW to 922 MW energized and billing, a contracted-to-live ratio of about 7 to 1.
• Interest arrives 12 to 24 months ahead of rent: Cipher's quarterly interest was 269% of revenue in the same period, TeraWulf's 126%.
• 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.

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.

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 remaining performance obligations grew 4.6 times in a year to $638 billion, while the same annual report discloses that no single customer accounts for more than 10% of revenue.
• Backstopping unsold capacity went from one-off to standard term: Nvidia to CoreWeave at $6.3 billion, Meta to Nebius at $15 billion.
• CoreWeave's quarterly net interest expense was $640 million, equal to a quarter of revenue in the same period.
• This link totals $13.56T, 34.46% of the chain, second only to accelerator design.

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.

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.

• Neither OpenAI nor Anthropic is listed. Across the 271 company-to-company relationships on this map, Anthropic holds 24 and ranks 2, while OpenAI holds 21 and ranks 4.
• Of Amazon's $62.6 billion in second-quarter net income, $53.4 billion was pre-tax other income, which the filing attributes mainly to its investment in Anthropic, which is not listed.
• SpaceX listed after combining with xAI and is the only company that prints AI segment capex into its filings: $23.55 billion in the first half.
• Only 5 listed entities sit in this link, CR3 96.2%.

This is the only link that collects money from outside the chain. Every contract and lease upstream is ultimately settled by the subscription and license fees paid here. 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.

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 is 73.4% of this link and is counted in accelerator design and cloud as well. Entities that span links are counted at full company market cap in each, so this weight is an upper bound.
• The sample is top-heavy: the bottom three names together are under 0.1% of the link, and most real application revenue sits inside statements that never break it out.
• The model link is acquiring applications in the other direction: SpaceX signed a merger agreement with Cursor, which is not listed, at an implied equity value of $60 billion.

• Market caps and prices: Nasdaq public quote endpoint, as of 2026-09-03, US market basis.
• Node size = company market cap ÷ sum of market caps of all nodes in its link (a market-cap proxy, not market share).
• Companies spanning several links are counted at full market cap in each, so shares in links that contain giants are upper bounds.
• Relationship amounts: the dataset's v field is in billions of USD; an empty v means no amount was publicly disclosed.
• Link classes (upstream inputs / chokepoint / open / infrastructure) come from the dataset's kind field, not from this page's judgement.
• Counterparty classes: the dataset's nodes.cls marks priv as "unlisted", but that set includes WMT, AEP, Barclays, MUFG — listed companies simply outside these 18 links. This page therefore calls them "off-map entities" and names only the genuinely unlisted in the text.

Every relationship carries a source. Open any company card to check them one by one: grade A comes from SEC filings, grade C is this map's inference — discount accordingly.

• 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.
• The "Applications & Data" link has an open boundary; the dataset's own note reads "此层边界无限,下列为代表性样本而非全集".
• Market caps change daily; the chain's structure does not. Read the structure, not the decimals.
• This page is public research material and does not constitute investment advice.

"Multi-link share" is the combined market-cap share of multi-link entities within the link; the higher it is, the softer that link's concentration reading.