From a copper mine to the chat box in your hand

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 287 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, with more relationships than every listed company except NVIDIA. 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.

The heaviest thing in an AI rack is not the chip. It is copper. Busbars, transformer windings, rack steel and cooling loops all start as ore dug up at this link. It is the one starting point no software can replace: if these 10 companies stopped at once, no power would reach the data center.

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

• The 10 companies here total $801.5B, 1.14% of the market-cap proxy across all 18 links; NVIDIA alone is 6.7 times that.
• Concentration is moderate: BHP, Southern Copper and Rio Tinto hold about 67% of the link between them, and rare-earth miner MP Materials is 1.2%.
• Its effects reach downstream by a long detour. Substation transformers now run a 160-week lead time, and swings in grain-oriented electrical steel and copper prices are among the causes; that story belongs to link 10, so just note it here.

Data centers need power, and power plants need something to burn. This link is the intake valve of generation: gas pipelines, LNG terminals, gas fields, plus a little uranium. It generates nothing itself, yet it decides whether a newly installed gas turbine can keep running at full load.

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

• Seven companies total $337.5B, the smallest of the 18 links at 0.48% of the market-cap proxy.
• The mix is lopsided: gas-linked names make up 98.3% of the link, and the only uranium name, Uranium Energy, is 1.7%.
• The money sits in the pipes: Williams, Kinder Morgan, Cheniere and ONEOK, the four pipeline and LNG names, are 84% of the link, while the two producers, EQT and Antero, are 14%.

A fab consumes more than power and tools. It also consumes high-purity gases and specialty chemicals that cannot stop for a single day. The business model is unusual too: supply pipelines run straight into the fab site and follow the plant wherever it is built. An interruption does not mean lower output. It means scrapped wafers.

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

• Nine companies total $443.8B, 0.63% of the market-cap proxy; only Linde and Air Products are pure industrial-gas players, and together they are 66.2% of the link.
• The three upstream links (mining, energy and fuel, base materials) come to $1.58T combined, just 2.2% of the 18-link market-cap proxy.
• Downstream sit the tightest links on this map: advanced foundry, advanced packaging and semiconductor equipment all draw their consumables from here.

Chips are not drawn. They are compiled. EDA is the compiler, and IP is the set of ready-made blocks: licence an Arm CPU core and skip years of in-house work. This link has no fabs and no capacity ceiling. Its scarcity comes from something else entirely: who is allowed to use it.

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

• Four companies total $423.3B, 0.60% of the market-cap proxy, smaller than the mining link.
• On the market-cap proxy, Arm alone is 59.0% of the link, Cadence about 20%, Synopsys about 19%, and Rambus, which sells HBM interface IP, 2.2%.
• This link reports quarterly and every number can be checked quarter by quarter. Later on the map, advanced foundry, semiconductor equipment and optical interconnect all cut back their disclosure in the same year.

This is the most expensive link on the map: on its own it accounts for nearly 30% of the market cap across all 18. An accelerator is the chip built to multiply matrices and do little else, and whoever's chip ends up in the training script takes the thickest slice of margin on the chain. It also has a feature no other link has: several of its largest customers sit 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.

• On the market-cap proxy, the 11 companies here total $20.85T, 29.5% of all 18 links and the largest on the map.
• On reported revenue, NVIDIA's data center line was $89.0B in the second quarter of 2026, 70.3% of the $126.6B de-duplicated supply-side total; a year earlier it was 69.9%.
• The slippage shows up in contracts first: Broadcom's remaining performance obligations jumped from $45.0B to $164.6B in a single quarter, including a long-term custom-accelerator contract the customer cannot terminate.

Once the layout is drawn, someone has to print it into silicon. Worldwide, the list of companies that can print the leading-edge layers is now essentially one name. TSMC raised its 2026 capex guidance in July, with the top of the range more than 50% above what it actually spent in 2025. The money is being added at the company's own pace, but its tightest spot is not the wafer.

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.

• On the market-cap proxy, TSMC is 79.0% of this link; Intel Foundry, UMC, GlobalFoundries and Tower together are only 21%.
• The 2026 capex guidance was raised in July to $60-64B, against $40.9B actually spent in 2025; board-approved capital appropriations over the last four quarters were up 76.1% year on year.
• Advanced packaging is not broken out: wafer was 86% of 2025 net revenue, while the remaining "others" line grew 41.2% year on year, ahead of wafer's 30.1%.

After a chip is cut from the wafer, the logic die and several stacks of HBM still have to be assembled onto one substrate. That step is advanced packaging. Think of a topped-out building: the structure is finished, but nobody moves in until the interior is done. TSMC's chairman said in July 2026 that packaging capacity is so tight it "limits my customers' growth", and that the back-end gap had grown wider.

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.

• Only four US-listed companies sit in this link: TSMC, Intel, ASE and Amkor. TSMC is 79% by market-cap proxy; the other three add up to just over 20%.
• TSMC never breaks out advanced-packaging revenue. The only figures come from its CFO on a call: slightly over 10% of revenue in 2025, low teens expected in 2026.
• The physical chokepoint is the thinning, dicing, bonding and molding tools. Their suppliers are Japanese and Dutch, and US-listed exposure to them is zero.

Storage is the chain's memory. HBM is the spice rack within arm's reach of the stove; nearline hard drives are the pantry in the basement. AI needs both, and both are short. But the two shortages are opposite in kind: one can be fixed with money, the other cannot.

• SK hynix held a 56.4% share of HBM revenue in the first quarter of 2026, a figure printed in the body of its own US listing prospectus.
• Micron's latest quarter grew 346% year on year, but shipped bits rose only about 20%. The rest was price. Read storage revenue as a demand gauge and you will get it wrong.
• Disks went the other way. Western Digital's revenue rose 36% while capex moved only from $412M to $418M; both drive makers keep capex pinned at 4-6% of revenue.

Equipment makers are the machine-tool shop of chipmaking. If fabs are bakeries, these are the people selling ovens, and a good year for ovens does not make every oven scarce. Twenty-two US-listed names sit here, and 2026 has treated them very differently: some guide the next quarter to more than double, others are still shrinking.

• Lithography is the one outright monopoly. ASML states capacity in units: about 65 low-NA EUV systems in 2026, with a further 30% planned for 2027.
• The most direct check is disappearing. ASML stopped publishing quarterly net bookings from the first quarter of 2026: the number appeared four times in the prior release and zero times in the current one.
• The upcycle was not shared evenly. In the same year, ion-implant maker Axcelis saw 2025 revenue fall 17.6% and photomask maker Photronics fell 2%, while bonder maker Kulicke & Soffa's next-quarter guidance implies growth of 111%.

Server assembly means bolting GPUs, memory, power supplies and sheet metal into a machine that boots, then racking it in a data center. It is the step closest to AI and the thinnest in margin: how much you can build depends on how many chips upstream is willing to release.

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 nine companies in this link add up to about $529B of market value, less than a tenth of NVIDIA alone (market-cap proxy).
• Super Micro's latest quarter grew 93% year on year, the same order as NVIDIA. Its gross margin for that quarter was 17.5%; NVIDIA's was 75.0%.
• Celestica's CCS segment went from 72% to 81% of company revenue in a year, and its segment margin is still just 8.66%.

Once the servers are built, hundreds of thousands of GPUs have to talk to each other, and they do it over light. Optical modules are the microphones, and plenty of firms assemble them; the global leader is Innolight, listed in mainland China rather than the US. What is actually scarce is the indium-phosphide laser chip inside.

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 put $2B each into Lumentum, Coherent and Marvell within 30 days: the first two bought purchase commitments and rights to future capacity, and the Marvell deal is joint development in silicon photonics.
• The gap runs near 30%, a figure company management and an industry research house reached separately. Capacity is already up 20% and the imbalance widened; long-term agreements lock the output through 2027.
• Pricing power lands on whoever sits closest to the chokepoint: Lumentum's full-year non-GAAP gross margin widened 1,130 basis points and reached 50.4% in its latest quarter, while Coherent's widened 152.

However good the chip, something has to feed it power and carry off its heat. This link builds the switchgear, UPS units, coolant distribution units and cold plates inside the hall, the things that actually hold up schedules. One thing changed here in 2026: the tightest sub-link stopped being available on its own in 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.

• In March 2026 Eaton closed its purchase of Boyd Thermal for $9.549B; intangibles plus goodwill came to 114% of the price. It bought capacity and customer relationships, not factories.
• Vertiv stopped reporting quarterly orders in February 2026. The stock rose 24.5% the day it said so; five months later a revenue miss took it down 17.3% in a single session.
• Modine is the only company here that created a data-center reporting segment: its share of revenue went from 26.9% to 39.9% in a year while the segment's gross margin fell 960 basis points.

Hooking a data center up to the grid is like getting a water meter installed in a new building, except this meter takes three years to arrive. This link covers everything between the high-voltage line and the switchgear: contractors, substation equipment, on-site backup power. Money is not the constraint here. The delivery calendar is.

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 transformers now carry lead times of over 160 weeks and high-voltage breakers about 125 weeks; of the announced capacity expansions, the earliest completes in mid-2027 and the bulk lands in 2028 to 2030.
• Caterpillar tops this link with a 57% market-cap-proxy share, but what it sells data centers is on-site and backup generation equipment, not grid assets.
• Backlog is not the same as orders: MasTec discloses that about 40% of its backlog sits in cancellable master service agreements, and Quanta puts the same figure at 41%.

Order a heavy-duty gas turbine today and you are joining a queue for 2030. This link is where the chain's electrons actually come from: turbines, nuclear plants, fuel. What makes it unusual is that the capacity path has been spelled out by the companies themselves: which year the expansions land, how many gigawatts a year they can ship, all of it public.

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 of 53 GW against 20 GW of annualized output is 2.7 years; including slot reservations, 116 GW works out to 5.8 years. Both readings belong on the page, because reservations can be cancelled.
• The price increase has already happened: GE Vernova says its equipment orders in the first half of 2026 were priced more than 20% above the fourth quarter of 2025.
• Small modular reactors will add zero incremental power to the grid within 24 months. Oklo has not even filed a combined licence application with the NRC.

This link rents out megawatts that already have power, the way a landlord rents retail space: traditional data-center REITs plus a wave of converted bitcoin miners, signing leases that run ten to twenty-five years. Its market-cap proxy is just 2.7% of the cloud link's, yet it carries the longest-dated contracts on 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 contracts that can be traced filing by filing come to about $160.9B, while capacity actually in service and billing is only 850 to 922 MW, a ratio of roughly 7 to 1.
• Interest arrives 12 to 24 months ahead of rent: Cipher's quarterly interest expense equalled 269% of that quarter's revenue, TeraWulf's 126%.
• Google is never the tenant here, yet it is the link's credit hub: $6.23B of quantified backstop for Fluidstack, an intermediary that is not listed, in exchange for warrants struck at one cent.

The cloud link looks like a chain store that has sold out its memberships but has not opened the doors. Its market-cap proxy is second only to accelerator design, the second-largest link on the map and the densest in circular deals: the party putting up the money, the party selling the chips and the party renting the racks are often the same handful of names. In the last link Google backstopped a host; here, backstops have become boilerplate.

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 rose 4.6-fold in a year to $638B, while the same annual report states that no single customer accounted for 10% or more of revenue.
• Backstopping unsold capacity has gone from one-off to boilerplate: $6.3B of it from NVIDIA to CoreWeave, $15B from Meta to Nebius.
• CoreWeave's net interest expense of $640M in a single quarter equalled a quarter of that quarter's revenue.

The model link is the chain's order counter: the power, the racks and the chips upstream all ultimately get ordered from here. But only part of it is visible in US-listed equities. Three companies account for 96% of this link's market-cap proxy, while the two firms that actually define the link appear on no list at all.

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 are both not listed, yet across the 287 inter-company relationships mapped here they sit on 25 and 28 respectively: Anthropic is second only to NVIDIA's 35, and OpenAI ranks fourth.
• Of Amazon's $62.6B of second-quarter net income, $53.4B was pre-tax other income that the filing attributes primarily to its investments in Anthropic, which is not listed.
• SpaceX, which absorbed xAI before listing, is the only company that prints AI-segment capital expenditure directly in its filings: $23.55B in the first half.

This is the only place on the chain where money comes in from outside. Every contract, lease and commitment upstream is ultimately settled against the subscription and licence fees collected here. It is also the fuzziest link: the thirteen names on the map 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 accounts for 73.8% of this link, but it is also counted in the cloud and model links; a company spanning links is booked at its full market cap in each, so the share is a ceiling.
• The bottom three names together make up less than 0.1% of the link. The listed sample is extremely top-heavy, and most real application revenue hides inside statements that never break it out.
• The model link is buying the application link back: SpaceX signed a merger agreement with Cursor, which is not listed, at an implied equity value of $60B.

• Market caps and prices: Nasdaq public quote endpoint, as of 2026-09-04, 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 (raw / chokepoint / open / added) 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.
• 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.