The Chokepoint: How One Supply Chain Decides How Much AI the World Gets
The AI boom is not short of money. It is short of factories. Every frontier GPU passes through one Dutch lithography monopoly, one Taiwanese foundry, three memory makers, a packaging step that has been sold out for three years, and a glass-fibre cloth made mostly by one Japanese company. This guide follows a single AI chip from design to data centre, shows where it gets stuck, and explains who is racing to widen each bottleneck.
๐ญ Factories, not money
By late 2026 the companies buying AI chips have more money than the industry can turn into silicon. Nvidia describes its growth as supply-constrained. TSMC, which makes almost all of the world's AI accelerators, told analysts on 15 October that the gap between AI demand and supply is "still quite big", after a record quarter. The memory makers have sold their entire 2026 output of high-bandwidth memory, and the advanced packaging step that joins chips to memory has been sold out since 2023.
That makes the supply chain one of the real limits on AI progress. How many GPUs exist next year depends less on how many anyone orders than on how many wafers TSMC can pattern, how many packages it can assemble, and how many memory stacks SK hynix, Samsung and Micron can build.
๐บ๏ธ The journey of one AI chip
A modern AI accelerator such as an Nvidia Blackwell or Rubin GPU is not one chip but an assembly: two large logic dies, eight or more stacks of HBM memory, a silicon interposer that wires them together, and an organic substrate underneath. Each part comes from a different, highly concentrated industry.
๐ Follow the chip: click each stop
From design to data centre. The red stops are the tightest bottlenecks in 2026.
๐ฏ Single points of failure
Most industries have several suppliers at each step. The AI chip supply chain has steps where one company, or one island, does nearly all the work.
๐ How concentrated is each step?
Share held by the leading supplier. Hatched bars are estimates with wide ranges between sources.
EUV: ASML is the sole supplier. Leading-edge foundry share: widely cited estimate for the most advanced nodes. HBM: SK hynix share of HBM4 for Nvidia's Rubin, estimated 60โ70% (sources conflict). T-glass: Nittobo, estimated 70โ90%. CoWoS-class packaging: TSMC, with Amkor, ASE/SPIL and others taking overflow.
๐น๐ผ TSMC: the foundry everything goes through
Taiwan Semiconductor Manufacturing Company makes the logic dies for Nvidia's GPUs, AMD's Instinct chips, Google's TPUs, Amazon's Trainium, Apple's processors and nearly every custom AI chip discussed in Custom Silicon vs Nvidia. Even the hyperscalers' "independence" from Nvidia runs through the same fabs in Hsinchu, Tainan and Taichung.
Its third quarter of 2026, reported on 15 October, was another record: revenue of about $46.7 billion, up 41% year on year, net income up about 69%, and high-performance computing at 68% of revenue. TSMC guided to $52.3โ54.3 billion for the fourth quarter, about 44% growth for the full year, and kept 2026 capital spending at $60โ64 billion. Its 2-nanometre (N2) process jumped from 3% to 8% of wafer revenue in a single quarter, which one outlet called the fastest node ramp in TSMC's history, and N2 capacity is expected to more than double by 2027.
๐ TSMC in Q3 2026
Share of revenue. N2 shown for Q2 and Q3 2026.
TSMC Q3 2026 results as reported by CNBC, Quartz, TechTimes and BigGo Finance. One outlet put revenue at $46.1B.
๐ฆ The hidden bottleneck: advanced packaging
Since 2023 the tightest step has not been making chips but packaging them. AI accelerators need HBM memory right next to the logic die, joined by thousands of tiny wires. TSMC's CoWoS (chip-on-wafer-on-substrate) does this by placing the dies on a large silicon interposer, then mounting that on a substrate. Blackwell-class GPUs use CoWoS-L, a version with small silicon bridges that allows packages several times larger than a normal chip.
TSMC's CEO has said CoWoS is sold out through 2026. Capacity is growing fast, but estimates of how fast vary widely: one May 2026 report puts end-2026 capacity at 115,000โ140,000 wafers per month, while other (lower-quality) sources cite targets of 130,000 or 150,000. Estimates of Nvidia's share of that capacity range from about 45% to 60%. The next step is CoPoS, which uses large rectangular panels instead of round wafers so more big packages fit at once.
America has almost no advanced packaging today, so even chips made at TSMC's Arizona fabs are shipped back to Taiwan to be packaged. Two projects aim to change that. Amkor is building a packaging campus in Peoria, Arizona, with Apple and Nvidia as lead customers; its first phase is due in mid-2027, with production ramping in early 2028. TSMC said in April 2026 that it has begun building its own Arizona packaging plant, with CoWoS capability before 2029.
๐งฎ How many GPUs can the world make?
๐๏ธ From packaging wafers to GPUs
A back-of-envelope model. Every input is adjustable, and the defaults are illustrative, not reported figures.
Teaching model: GPUs per month = min(wafers ร share ร packages per wafer ร yield, HBM stacks รท 8). Real packages differ in size, many products use 12 stacks, and capacity is shared with AMD, Broadcom, Google and others.
๐ฆ ASML: the machine nobody else can build
Every leading-edge chip is patterned with extreme ultraviolet (EUV) lithography, and only one company in the world makes EUV machines: ASML of Veldhoven, the Netherlands. Each machine is the size of a bus, relies on mirrors from Zeiss and light from tin droplets hit by a laser tens of thousands of times a second, and costs hundreds of millions of dollars. The newest High-NA version prints finer features in a single exposure.
ASML reported third-quarter 2026 sales of about โฌ11.7 billion on 14 October, with system sales of โฌ8.4 billion: โฌ4.1 billion from EUV and โฌ4.3 billion from older DUV machines. It sold one High-NA system in the quarter. Intel has exposed more than a million wafers on High-NA machines; Samsung targets volume production with High-NA in 2028 and TSMC in 2030. ASML expects China, which can buy only DUV machines under export controls, to be about 20% of its 2026 sales, and it no longer publishes order bookings.
๐พ Memory: sold out before it is made
High-bandwidth memory (explained in The Memory Wall) is made by only three companies. All three have said their 2026 HBM is spoken for: Samsung's CFO said its 2026 HBM4 volume was sold out, Micron's entire 2026 HBM supply including HBM4 was sold before an October 2026 webinar, and SK hynix said in October 2025 that its 2026 output was committed. Estimates of HBM4 share for Nvidia's Rubin put SK hynix at 60โ70% and Samsung at 25โ30%, though other reports have Samsung qualifying first and targeting a larger share. Most forecasts expect tightness through 2027 and possibly 2028.
HBM is also the binding limit for China. Analysts argue that Chinese foundries could make more AI accelerator dies than they can pair with memory, after Chinese firms stockpiled HBM before export controls tightened.
๐งต The materials nobody talks about: substrates and glass cloth
Under every AI package is an organic substrate built from ABF film and reinforced with a special low-expansion glass-fibre cloth called T-glass, which stops large packages from warping as they heat up. Counterpoint estimates that one advanced AI chip uses roughly ten times more ABF material than a traditional PC chip. T-glass comes overwhelmingly from Japan's Nittobo, with estimates of its share ranging from about 70% to 90%.
Counterpoint's channel checks show a T-glass supply gap of more than 10% in 2026, worse than in 2025, and Taiwanese substrate makers put it at 10โ20%, rising above 40% for high-end cloth in the second half. Nittobo is spending ยฅ150 billion to triple capacity, but most of it arrives for 2027 shipments. Taiwan's Nanya, Taiwan Glass and Fulltech, and China's Grace Fabric, are trying to qualify as alternatives. It is a reminder that a trillion-dollar industry can be held up by a fabric.
๐ฅ The challengers: Samsung, Intel and Rapidus
Customers badly want a second source for leading-edge logic. Three companies are trying to provide one.
๐ Challenger scorecard (October 2026)
| Company | Leading node | Where | Named AI / HPC customers | Status and risks |
|---|---|---|---|---|
| TSMC (benchmark) | N2, ramping fast | Taiwan; Arizona (fab 2 volume production planned Q3 2027) | Nearly everyone | Sold out; N2 capacity to more than double by 2027 |
| Samsung Foundry | 2nm (SF2) | Korea; Taylor, Texas | Tesla (AI5 taped out; $16.5B AI6 deal through 2033); Taylor reportedly booked | 2nm yield reported near 60%, up from under 20% in late 2025; power efficiency still a concern |
| Intel Foundry | 18A, then 14A | Arizona, Oregon, Ohio (delayed) | Tesla, the first named external 14A customer; others reportedly evaluating | 18A yields improving; 14A depends on more external commitments in 2026โ27 |
| Rapidus | 2nm | Chitose, Hokkaido | None confirmed; courting US AI firms | Pilot line since April 2025; volume production targeted for 2H 2027; heavy Japanese government backing |
Sources: TrendForce, Tom's Hardware, TechTimes, Bits&Chips, Yahoo Finance, DigiTimes, The Register. Customer lists are as reported, not exhaustive.
โ ๏ธ What if a link breaks?
Concentration is efficient until something goes wrong: an earthquake, a fire, a contamination incident, a blockade or a sudden export ban. The explorer below shows, qualitatively, how disruptions at different points would ripple through to AI compute.
๐งฏ Disruption explorer
Pick a disruption. Impacts are qualitative and illustrative, not forecasts.
๐๏ธ Tariffs, subsidies and export controls
- US subsidies
- The 2022 CHIPS and Science Act funded new fabs, including TSMC's Arizona site. TSMC is accelerating Arizona's upgrade to 2nm, and its second Arizona fab is due to start volume production in Q3 2027 (per one report on the earnings call).
- US tariffs
- A Section 232 proclamation imposed a 25% tariff on certain advanced computing chips from 15 January 2026, with exemptions including chips for US data centres, US research and development, and startups. Broader second-phase tariffs were recommended but not confirmed. ITIF, a think tank, has argued semiconductor tariffs would cost the US economy heavily.
- Export controls
- Since October 2022 the US has restricted exports of advanced AI chips and chipmaking tools to China, and the Netherlands bars sales of EUV machines there. China's answer is SMIC making 7nm and 5nm-class chips with older DUV machines and multi-patterning, which 2025 reports said gives lower yields and higher costs than TSMC. Huawei's Ascend chips are built this way, and analysts see HBM as an even tighter limit. China's chip programme will get its own article.
๐ฐ๏ธ Timeline
๐ญ What to watch in 2027โ2028
- Packaging capacity. Whether CoWoS and the first CoPoS lines finally catch up with demand, and whether Amkor's and TSMC's Arizona packaging sites open on time.
- HBM4E and new memory fabs. Memory makers expect only limited relief before 2027.
- A real second source. Whether Samsung's 2nm yields, Intel's 14A customer list or Rapidus' first customers turn into high-volume AI chip production.
- T-glass and substrates. Nittobo's new capacity in 2027 and whether alternative suppliers qualify.
- Policy shocks. Phase-two US chip tariffs, new export rules, and Taiwan Strait risk.
๐ข Reading the numbers
- TSMC and ASML results come from news coverage of their 14โ15 October earnings, not the companies' own filings; outlets differ slightly (TSMC revenue $46.1B vs $46.7B).
- CoWoS capacity and Nvidia's share are analyst estimates. Some widely circulated figures (130kโ150k wafers/month, 60% share) come from low-quality aggregator sites.
- HBM and T-glass shares are estimates with wide ranges.
- Arizona timelines (TSMC fab 2 in Q3 2027; Amkor in 2027โ2028) come from single reports and company statements and may slip.
- China's output figures are 2025 plans and estimates; no confirmed 2026 production data was found.
- The calculator and disruption explorer are teaching tools with illustrative inputs.
๐ References
- Company results
- CNBC: TSMC posts record profit for a sixth straight quarter (15 Oct 2026)
- Quartz: TSMC Q3 2026 earnings
- TechTimes: TSMC 2nm jumps to 8% of revenue
- BigGo Finance: TSMC Q3 2026 earnings call
- ASML Q3 2026 earnings call transcript (Motley Fool)
- ASML Q3 2026 earnings call transcript (Benzinga)
- Preview: China's share of ASML sales and the missing bookings figure
- Packaging, memory and materials
- ClusterBid: GPU supply chain 2026, CoWoS allocation
- SupplyICs: CoWoS to CoPoS roadmap 2026โ2029
- Reuters (via WHBL): TSMC plans Arizona packaging plant by 2029
- The Next Web: Amkor's Arizona packaging campus
- KED Global: Samsung sells out 2026 HBM supply
- TechTimes: Micron maps HBM to 2030 (Oct 2026)
- Counterpoint: From T-glass crisis to capacity erosion
- Fusion Worldwide: Understanding the T-glass shortage
- Challengers and policy
- Tom's Hardware: Tesla AI5 tapes out at Samsung Foundry
- TrendForce: Samsung Taylor fab reportedly fully booked for 2nm
- TechTimes: Samsung 2nm yield near 60%
- Bits&Chips: Intel signals external customer wins for 14A
- DigiTimes: Rapidus courts US AI customers ahead of 2H27 production
- The Register: Rapidus lands $1.7B
- EY: Section 232 proclamation imposes 25% tariff on certain semiconductors
- ITIF: Economic consequences of Section 232 semiconductor tariffs
- Tom's Hardware: SMIC 5nm for Huawei (2025)
- Design & Reuse: China's AI chip ambitions limited by HBM supply
- Related on this site
- AI Chips & Compute Infrastructure ยท The Memory Wall ยท Custom Silicon vs Nvidia ยท The Interconnect Wars