35 tracked companies across 15 stages, from the software used to design a chip to the cloud racks that rent the finished GPU. Click any node to trace what it supplies and what supplies it.
A leading-edge fab costs on the order of twenty billion dollars and takes years to bring up. NVIDIA owns none — it designs the chip and pays a foundry to build it, which means its output ceiling is set by someone else's capacity decisions made years earlier.
A modern accelerator isn't one chip — it's a die bonded to stacks of memory on a shared substrate. That advanced-packaging step has repeatedly been the binding constraint on how many accelerators exist, rather than the wafers themselves.
High-bandwidth memory sits beside the processor and feeds it. Only a handful of firms can make it, so how much HBM exists — and who has contracted for it — effectively caps how many accelerators ship and how large a model each one can hold.
Stages run left to right in the order a chip passes through them, and an arrow means the company on the left supplies the one on the right. Definitions for each stage are in the glossary under the graph.
On 16 October Economy Minister Kung Ming-hsin said Taiwan plans to add AI servers and advanced semiconductors to its controlled goods list under the Foreign Trade Act, replacing a company-level blacklist with a market-wide restriction. This is a plan, not law: it needs a legal amendment, and no start date, penalty schedule or final product list has been set.
GlobalFoundries announced on 2026-10-08 a multi-year agreement, described as $2 billion, under which it will provide manufacturing services to TSMC for silicon interposers used in TSMC's CoWoS advanced packaging, including parts with embedded deep trench capacitors. GF will add fabrication capacity at its Malta, New York site, with volume production expected to begin ramping in the first half of 2028. The initial term is five years, with a framework to expand capacity as demand grows. GF describes it as the first U.S.-based source of these interposers. The interposer is the silicon layer that connects a GPU die to its HBM stacks inside a CoWoS package, so this adds a second interposer source for the packaging step behind NVIDIA's data-centre GPUs. It does not cover die placement, HBM integration, substrates or test, so it is not a near-term increase in total CoWoS output.
KED Global (Hankyung) reported on 2026-10-07, citing people familiar with the matter, that Samsung's 12-layer HBM4E passed customer qualification tests with NVIDIA and major hyperscalers. Neither Samsung nor NVIDIA has confirmed it, so it is recorded as rumoured. The report names neither the NVIDIA product nor a mass-production date, and gives no supply volume. If it holds, NVIDIA would have a qualified HBM4E source from Samsung alongside SK hynix, which matters most for next-generation parts where HBM supply, not logic wafers, limits how many accelerators can be built.
In its fiscal Q4 2026 prepared remarks, Micron said it is working with NVIDIA on the industry's first custom HBM4E implementation, NVHBM, for NVIDIA's next GPU generation and NVLink Fusion platforms. On the 30 September earnings call, president and CTPO Scott DeBoer said that both NVHBM and Micron's JEDEC-standard HBM4E family are being developed with base dies made on a foundry process, not in Micron's own DRAM fabs. Micron did not name the foundry. The Elec attributes the work to TSMC, which is unconfirmed by Micron. NVHBM places NVIDIA's memory controller in the HBM base die, so the base die becomes logic silicon from a foundry, a change in how the memory stack is built.
In its fiscal Q4 2026 prepared remarks (30 September 2026), Micron said it has signed 26 Strategic Customer Agreements, which it estimates at over 35% of revenue through 2030. Some now extend into 2031, and two existing agreements received one-year extensions to 2031. All SCAs carry take-or-pay contracted volumes. Customer financial commitments total $32 billion, mostly cash deposits, and remaining performance obligations are about $150 billion. Separately, Micron said it has completed agreements for the vast majority of its calendar 2027 HBM bit supply, with significant price increases year over year. It also said it has no line of sight to DRAM supply and demand returning to balance, and plans higher fiscal 2027 capex than previously planned to add cleanroom space.
Micron's product catalog now marks both of its 2GB (16Gb) GDDR7 parts — the 28Gbps MT68A512M32DF-28 and the 32Gbps MT68A512M32DF-32 — as end-of-life, first spotted on 22 September and confirmed by trade press over the following two days. 3GB modules become Micron's only GDDR7 offering. Micron only began shipping GDDR7 to NVIDIA in volume in early 2026 and was never the primary RTX 50 supplier, so the immediate effect is a narrower supplier base for the 2GB density used across most of the desktop RTX 50 range rather than a shortage. Micron has not issued a statement.
Nscale Ltd filed an S-1 with the SEC on 18 September 2026, intending to list on the NYSE under the symbol NSCL. The filing puts on the public record the shape of NVIDIA's position in the neocloud. NVIDIA became a holder of more than 5% of Nscale's outstanding share capital through its participation in the Series B financing. Separately, in connection with an October 2025 net lease for a facility in Ward County, Texas, NVIDIA agreed to guarantee up to $860,340,000 of Nscale's obligations under that lease; in consideration it received warrants for up to 9,476,700 Series B preferred shares with an aggregate value of $60,000,000 at an exercise price of $0.01 per warrant share. Contingent default warrants tied to the same guarantee remain outstanding after the offering and would issue non-voting shares on exercise. Under a follow-on subscription agreement of 9 March 2026, NVIDIA's $300,000,000 SAFE converted to equity at Series C completion. The structure is worth reading as infrastructure rather than as a portfolio holding: NVIDIA is underwriting the lease obligations of a customer that builds the data centres its GPUs are installed in, which is a different exposure from selling parts into an arms-length buyer. Nscale's filing describes hyperscale sites in Norway, Portugal, Texas and West Virginia, and a deployment of over 66,000 NVIDIA Vera Rubin NVL72 GPUs with Microsoft from late 2027.
Reports on 16 September 2026 described SK hynix in exploratory talks to manufacture memory at Intel's delayed Ohio campus — variously as a lease of part of the site, or as a venture with Intel and cloud buyers seeking to secure memory supply. SK hynix published a statement the same day saying that nothing has been finalised regarding cooperation with any specific companies mentioned or memory chip production in the United States, and that no concrete plans, arrangements or decisions have been made. Recorded here as rumoured for that reason. What would make it structural rather than a headline: it would be SK hynix's first US memory production, it would give a site whose completion has slipped to 2030-2031 an anchor tenant, and any deal could face a Seoul review on national core technology grounds. Nothing about available HBM or DRAM output changes on this week's information.
A supply-chain tracker published by Silicon Analysts, reported 15 September 2026, places SK hynix in shipment of the 16-high 48GB HBM4 part for NVIDIA's Vera Rubin platform. SK hynix has not confirmed this itself, so it is recorded as rumoured. The part is the one the company showed publicly for the first time at CES in January 2026 and said it would mass-produce in the third quarter, so the timing is consistent with its own stated plan rather than a surprise. The consequence, if it holds, is per-package capacity: 48GB against 36GB for the 12-high stack is a third more memory per placement, without a second placement's power, area or interposer cost. Sixteen layers is the hard step of the generation — thinner dies, tighter die-to-die spacing, warpage control and bonding alignment all get worse at once — which is why the position is not easily matched. On the same tracker, Micron has been in high-volume production of 12-high HBM4 since March and has shipped 16-high samples without announcing volume production of the taller part; Samsung has the technology but has called commercialisation of a 16-high product unnecessary on current customer demand.
ETNews reported on 2026-09-04, citing industry sources, that Micron Technology plans to raise HBM production capacity from roughly 40,000-50,000 wafers per month last year to about 100,000 by the end of 2026, adding up to 60,000 wafers per month, and to lift 12-high HBM4 to about half of HBM output by year-end. HBM packaging runs at its Tongluo (Taiwan) and Woodlands (Singapore) fabs, with investment being prepared at Hiroshima (Japan). Micron has not officially confirmed the figures.
Seoul Economic Daily reported, as carried by TrendForce on 2026-09-01, that Samsung Electronics has allocated about 70% of its memory production capacity through 2031 to long-term agreements, naming NVIDIA, Microsoft and Google as major customers. The report puts 36 GB HBM3E at about $2,100 on the spot market, roughly four to five times the LTA price. The figure covers memory capacity broadly, not HBM alone. Samsung has not confirmed it.
The Information reported, as carried by Investing.com on 2026-08-31, that ChangXin Memory Technologies (CXMT) has begun low-volume HBM3E production, with broader commercial availability targeted for 2027. Alibaba's T-Head and Cambricon are reportedly testing the memory for processors as early as next year. The report cites low yields and restricted access to advanced equipment, and places CXMT's HBM three to five years behind SK Hynix, Samsung and Micron. CXMT has not confirmed it.
The Information reported, as carried by AI Weekly on 2026-08-30, that the Commerce Department has drafted a rule to block Chinese AI firms from renting NVIDIA GPU compute through data centers in third countries including Thailand, Singapore, Malaysia and Japan, with a draft possibly shared with trade groups in September. Nothing has been proposed or published. Export-control lawyers say BIS cannot regulate remote access under existing law; the Remote Access Security Act that would grant that authority passed the House 369-22 in January and sits in the Senate Banking Committee. Bloomberg reported on 2026-08-07 that BIS was reviewing such offshore rental arrangements.
GRAPH LAST CONFIRMED 2026-07-07