State of the Semiconductor Supply Chain 2026 — Executive Summary

This is the Executive Summary of State of the Semiconductor Supply Chain 2026, published free and unedited. Start with the object itself. A leading-edge processor is a piece of patterned silicon you can hold between two fingers, and to make one you need a lithography machine that exactly one company on earth knows how to build, a fab that in practice means one island, a substrate supplier, a photoresist supplier, a packaging house that most people have never heard of, and — this part is recent — a power and cooling supply chain sized for a building rather than for a circuit board. Every one of those is a company. Every one of those companies has suppliers of its own, and those suppliers have suppliers. That web is the subject of this report. Not the chip. The web behind the chip. wafergraph tracks 615 companies (459 public, 156 private) across 12 segments of the semiconductor and AI supply chain, with a combined tracked market capitalization of $44.84T across 446 priced companies (market data as of 2026-08-05). $44.84T. That is the value sitting on top of the dependencies counted in the pages that follow, and the dependencies are the part almost nobody has written down. Here is what makes this countable instead of arguable. A dependency is not an opinion. Either a company states that another company is one of its key suppliers or it does not. This report counts those statements and does the arithmetic on top of them: the chokepoint ranking, the geographic concentration, the market-structure concentration, the deal record, and what wafergraph's own readers researched last week. Every figure is reproducible from the committed dataset. Nothing here is a forecast. Nothing here is a recommendation. If a number in this report is wrong, you can prove it is wrong — and being able to prove it is the entire point of publishing the data underneath it. The chokepoint map. If you read one number in this report, read this one. TSMC is the single biggest chokepoint in the tracked supply chain: 98 of the 615 tracked companies list it among their key suppliers, more than any other firm, followed by ASE Technology (72), Broadcom (45), NVIDIA (45), and Vertiv (43). Chokepoint means something narrow here, and narrow is the point. It means supplier in-degree: how many tracked companies name that firm as a key supplier. Not revenue. Not market cap. Not reputation. A count of how many other businesses have stated, in public, that they depend on you. And because edge coverage is partial — no dataset captures every supplier relationship in this industry, and this one does not pretend to — every count in this report is a floor rather than a census. The true number is at least this high. It is usually higher. Now look at where the dependence piles up, because it does not pile up evenly. It piles up in four places: foundry (TSMC (98), SMIC (31), Hua Hong Semiconductor (29)), advanced packaging and OSAT (ASE Technology (72), JCET Group (29)), AI-silicon design (NVIDIA (45), Broadcom (45), AMD (23), Marvell Technology (15)), and AI-datacenter power and cooling (Vertiv (43), Schneider Electric (41), Eaton (40), Delta Electronics (40)). Stop on that last group for a second, because it is the one that surprises people. Each of those four firms is named as a key supplier by roughly forty tracked companies, and not one of them designs a chip, fabricates a chip, or packages a chip. They sell transformers, switchgear, power distribution and cooling. A map of this industry that stops at the fab door is missing an entire tier of it. Geographic concentration. Three places on this map can halt the rest of it, and they do not fail the same way, so they do not get mitigated the same way either. The first is Taiwan. Taiwan hosts the #1 and #2 chokepoints by supplier in-degree, TSMC (98) and ASE Technology (72), and accounts for 315 of the graph's total 1470 supplier in-degree (21.4%), second only to the United States (502, 34.1%). One island. The top two names on the entire list, and better than a fifth of every dependency recorded in this graph. The second is China, and China's position is underneath everything else — it is in the ground, not in the fab. Per USGS Mineral Commodity Summaries 2025, China is the dominant producer of several minerals critical to chip manufacturing: gallium (~99% of world low-purity primary), silicon metal (~85%), rare earths (~69%), indium (~70%), tungsten (~83%), and antimony (~60%), with export controls on gallium and germanium in force since 2023 and an antimony export ban on the US since December 2024. Those controls are not a scenario somebody modelled. They are already in force, on inputs that sit underneath all 615 companies in this report. The third is the Netherlands, and it is the narrowest of the three. It is home to ASML, which wafergraph classifies as a monopoly: the sole supplier of EUV lithography machines required for all leading-edge chips (supplier in-degree 11). One company. One machine. There is no second source. And look closely at that in-degree, because it is small and it should bother you: the chokepoint score counts documented dependence, and the handful of firms that buy EUV scanners mostly do not publish supplier lists. The score understates the company. That is not a one-off either — 59 of the 112 firms wafergraph classifies as a monopoly or a leader carry three or fewer recorded dependents. So read the ranking as a map of what the industry has disclosed about itself, which is what it honestly is, and not as a map of what actually matters. Those two things overlap. They are not the same thing, and this report would rather tell you that on page one than let you find it out yourself. Market structure. Concentration in this industry is not evenly spread, and at the top of the table it is not close. Named market-share estimates (wafergraph's market-share dataset, 2026-06-15 basis) show the tightest concentration in lithography, ASML at 89% of the segment by revenue (CR3 100%) — with its EUV position tighter still, as the sole supplier of the EUV scanners required for leading-edge nodes — and data-center AI accelerators, where NVIDIA holds an estimated 88% (CR3 96%). Two segments. Two companies. Close to nine of every ten dollars in each of them. HBM memory is the interesting middle: a three-way oligopoly (SK hynix 52%, Samsung 38%, Micron 10%; CR3 100%). Three companies, and no fourth: the three named shares account for the whole segment, with no long tail underneath them. That is who supplies the memory sitting beside the accelerators in the paragraph above. And then there is the other end of the same table, which nobody ever quotes. Data Center Operators (Colo) and Analog & Mixed Signal are the most fragmented segments tracked, with the largest named player holding 13% and 19% of their respective segments. Same industry. Same report. The word "semiconductors" is covering a segment where one firm holds 89% and a segment where the leader holds 13%, and if you are making a sourcing decision, a pricing decision or an allocation decision, which of those two rooms you are standing in changes the answer completely. Full CR/HHI tables for all 17 tracked sub-segments are in Section 4. Deal flow. Money is the other way to read a supply chain, and this record is blunt about what the industry has decided it is. The wafergraph deal corpus tracks 74 M&A, investment, capacity, subsidy, partnership, supply-agreement, and compute deals from 2015 through 2026, with $2.20T in disclosed deal value across 62 deals carrying a public value (12 deals in the corpus do not disclose a value). Deal count and disclosed value both peaked in 2024–2025 (15 and 15 deals respectively), driven by a wave of AI compute mega-deals — Stargate Project launched ($500.00B), OpenAI–Oracle $300B cloud deal ($300.00B), and OpenAI commits $250B to Azure ($250.00B) are the three largest single items in the corpus. Read that list one more time. The three biggest line items in a semiconductor supply-chain deal corpus are not acquisitions of semiconductor companies. Not one of them. They are commitments to build and buy compute. When the biggest cheques in an industry stop buying companies and start buying capacity, that industry has changed shape, and the deal record shows it before the org charts do. Attention. One last read, and it is a different kind of data than everything above it: not what the chain is, but what people are looking at. Over the 7-day window ending 2026-08-05, Avantor was wafergraph's most-researched company and Materials its most-researched segment, by relative reader-attention rank (raw counts are not published). Attention is a leading indicator of nothing in particular, and this report is not going to claim otherwise. It is simply worth knowing which part of the chain the people who study it for a living were reading about this week. Full ranked lists are in Section 6. Every finding here is drawn from pages that are free on wafergraph: the chokepoint rankings, the geographic-concentration analysis, per-segment CR1/CR3/CR4 and HHI market structure for all 17 tracked segments, the sourced deal corpus, reader-attention data, and the methodology behind every score.