TSMC TSM
Leading-edge foundry + CoWoS packaging
-- 1D
72.5%global foundry revenue, Q2 2026
SUPPLY-CHAIN MINDMAP CONNECTING...
Left of the centre supplies this company; right of it buys from it. Percentages are the curated allocation for that edge — the tape beside each name is live. Click a branch to collapse it, a leaf for the number, its basis and its sources.
drag to pan · ctrl/⌘ + scroll to zoom
MARKET SHARE BY LAYER CURATED · RESEARCHED 2026-09-12
NVIDIA DIRECT-CUSTOMER CONCENTRATION — Q3 FY2026 SHARE OF TOTAL REVENUE, BY DIRECT CUSTOMER
22%15%13%11%39%
Customer A · 22%Customer B · 15%Customer C · 13%Customer D · 11%All other customers (each <10%) · 39%
official Four direct customers = 61% of a roughly $57B quarter, all in the Compute & Networking segment. That is up from 36% a year earlier. The split is disclosed; the names are not.
speculation Who are they? Nvidia declined to say. Two structural facts narrow it: (1) 'direct customer' means whoever Nvidia invoices — typically an ODM/OEM such as Foxconn, Quanta, Dell or Super Micro that integrates boards and racks, not the hyperscaler that ends up running them; (2) end demand is concentrated in Microsoft, Meta, Amazon, Google and Oracle, whose combined 2026 capex is around $725–800B. Press coverage tends to assume the big four hyperscalers; the ODM reading is at least as consistent with the wording. Both remain unverified.
SUPPLIER CARDS LAYER SHARE · NAMED CUSTOMERS · LIVE TAPE
Broadcom AVGO
Custom AI ASIC co-design + switch silicon
-- 1D
60%AI compute ASIC design partnerships
SK hynix 000660.KS
HBM / DRAM
-- 1D
50%HBM revenue share, Q2 2026
Marvell MRVL
Custom AI ASIC co-design
-- 1D
35%implied residual of the ~95% Broadcom+Marvell duopoly
InnoLight 300308.SZ
800G/1.6T optical modules
-- 1D
35%800G module shipments, 2026 estimate
Samsung Memory 005930.KS
HBM / DRAM
-- 1D
33%HBM revenue share, Q2 2026 (21% in Q1)
Eoptolink 300502.SZ
800G/1.6T optical modules
-- 1D
25%implied — InnoLight+Eoptolink ≈60% of 800G SFP supply
Vertiv VRT
Power + precision/liquid cooling
-- 1D
23%precision cooling, global
Schneider Electric SU.PA
Data-center power + cooling
-- 1D
23%DC physical infrastructure — tied with Vertiv within 0.1pt (Dell'Oro)
Micron MU
HBM / DRAM
-- 1D
17%HBM revenue share, Q2 2026 — residual of reported SKH+Samsung
Coherent COHR
Optical modules + components (lasers)
-- 1D
16%all optical transceiver shipments, 2025 (LightCounting)
Samsung Foundry 005930.KS
Foundry
-- 1D
5.9%global foundry revenue, Q2 2026
SMIC 0981.HK
Foundry — mature + China leading edge
-- 1D
5.4%global foundry revenue, Q2 2026
BUYER LEDGER EVERY MAJOR BUYER BY NAME · NO "OTHER BUYERS" BUCKET
| BUYER | 1D | SUPPLIERS | WHAT | HOW MUCH / WHAT SHARE | CONF | SPECULATION / CAVEAT |
|---|---|---|---|---|---|---|
| MSFT | - | Nvidia, AMD, Marvell, Broadcom, Vertiv, GE Vernova, Caterpillar, Arista, Foxconn | GPUs; Helios racks; Maia custom accelerator; cooling; on-site gas generation | $110–120B capex (2026); two GE Vernova turbines + Caterpillar Solar Turbines on the Chevron gas project | research | Individual vendor spend is not broken out. Frequently named as a candidate for Nvidia's 'Customer A/B' — unconfirmed. |
| META | - | Nvidia, Broadcom, Schneider Electric, Quanta | GPUs; MTIA custom accelerator; power distribution; rack integration | $125–145B capex (2026), raised from $60–65B planned for 2025 | research | MTIA volumes undisclosed. Also a common guess for an unnamed Nvidia >10% customer. |
| AMZN | - | Nvidia, Marvell, Alchip, TSMC, Eaton | GPUs; Trainium accelerators (Marvell/Alchip co-design, TSMC wafers); electrical distribution | ~$200B capex (2026) | research | Trainium-3 unit volumes and the Marvell/Alchip split are not public. |
| GOOGL | - | Broadcom, MediaTek, Marvell, TSMC, Nvidia | TPU v7 (N3P dual-chiplet); new inference silicon; GPUs for GCP customers | $175–185B capex (2026); Marvell warrant lets Google buy up to $12.2B of MRVL shares | press | Marvell's exact TPU-stack scope is reported, not confirmed in detail; Broadcom remains lead partner. |
| ORCL | - | Nvidia, AMD, Crusoe, Vantage, Bloom Energy, Vertiv, Foxconn | ~$40B of Nvidia GPUs; 50,000 AMD MI450s from Q3 2026; leased campuses; on-site power | FY27 capex guided $80–100B; ~$638B RPO | official | The $40B Nvidia figure is reported rather than filed. Bloom/Vertiv site assignments are inferred from sector reporting. |
| PRIVATE | - | Nvidia, AMD, Broadcom, Oracle, Microsoft, CoreWeave | 10 GW Nvidia systems; 6 GW AMD Instinct; 10 GW Broadcom custom accelerators; Oracle-contracted capacity | >$300B/5yr Oracle contract; up to $100B Nvidia investment alongside the 10 GW LOI | official | Capacity commitments are letters of intent and multi-year frameworks, not delivered hardware. Delivery risk is real: the first AMD gigawatt only lands in H2 2026. |
| AAPL | - | TSMC | Leading-edge consumer SoC wafers | 17% of TSMC 2025 revenue (NT$645.1B ≈ US$20.5B), down from 22% in 2024 | research | Share fell because Nvidia's spend doubled, not because Apple cut orders. |
| NVDA | - | TSMC, SK hynix, Samsung, Micron, InnoLight, Eoptolink, Coherent, Amkor, Ibiden | Wafers + CoWoS; HBM4; 800G/1.6T optics; assembly/test; ABF substrate | 19% of TSMC 2025 revenue (US$23.2B); ~60% of TSMC 2026 CoWoS; ~70% of its HBM4 from SK hynix | press | The Samsung/Micron split of Nvidia's residual ~30% HBM4 is inference, not disclosure. |
| AMD | - | TSMC, Samsung, Micron, SK hynix | N2P wafers for MI450 XCD; 12-high HBM4 stacks | ~31 TB of HBM4 per Helios rack; shipments from late Q3 2026 | press | 'Majority from Samsung + Micron' is reported; no party has published an allocation table. |
| AVGO | - | TSMC | Wafers for every co-designed accelerator it ships | ~60% of AI compute ASIC design partnerships | research | Share is a research-firm estimate of design wins, not of revenue. |
| CRWV | - | Nvidia, Dell, Vertiv | GPU fleet; servers; cooling | Not disclosed | press | Sits inside the Nvidia–OpenAI–Oracle circular-financing web; exposure is hard to size from outside. |
| 2317.TW | - | Nvidia, Delta Electronics | GPUs and boards for GB300 NVL72 rack integration; rack power electronics | Not disclosed | press | As a named Blackwell manufacturing partner it is a plausible unnamed Nvidia >10% direct customer. Never confirmed. |
| 3231.TW | - | Nvidia | GPU board and rack integration | Not disclosed | press | Same 'Customer A–D' caveat as Foxconn. |
| 2382.TW | - | Nvidia | Rack integration | Not disclosed | background | Named in coverage of the 10-Q disclosure as part of the ODM class; no figures. |
| ANET | - | Broadcom | Tomahawk-class merchant switch silicon | Not disclosed | background | Merchant-silicon shares per switch vendor are not published. |
| TSM | - | ASML | EUV lithography scanners | 100% sole-sourced | background | Structural fact rather than a 2026 figure — the single hardest point of failure in the chain. |
LIVE HEADLINES NVIDIA AND ITS DIRECT COUNTERPARTIES
Loading headlines...
SOURCES 55 REFERENCES · TAGGED BY EVIDENCE TYPE
officialMotley Fool — Blackwell sales, and Nvidia's customer concentrationQ3 FY26 10-Q: Customers A/B/C/D at 22/15/13/11% = 61% of a $57B quarter
officialSEC — Nvidia Form 10-QPrimary filing containing the customer-concentration disclosure
officialCIO Dive — Oracle teams up with AMD and Nvidia50,000 AMD MI450 GPUs on OCI from Q3 2026
officialVantage Data Centers — Stargate site in Wisconsin (press release)Joint OpenAI / Oracle / Vantage announcement
officialSEC — Bloom Energy Q2 2026 results (8-K exhibit)Record $1.1B revenue, +166% YoY; FY guide raised to $3.9–4.2B
officialBloom Energy — 2026 Data Center Power ReportVendor-published view of data-center power constraints
researchTrendForce — Global foundry revenue 2Q26Q2 2026 foundry revenue $53.49B; TSMC 72.5%, Samsung 5.9%, SMIC 5.4%
researchAstute Group — SK hynix holds 62% of HBM; 2026 pivots to HBM4Alternative HBM share series — sources disagree on the exact split
researchDigiTimes — Nvidia and Apple orders exceed 40% of TSMC revenueNvidia 19% / NT$726.9B; Apple 17% / NT$645.1B of TSMC 2025 revenue
researchDaloopa — Nvidia customer concentration: a Big 4 previewConcentration rose from 36% to 61% in a year
researchTrendForce — SK hynix to supply about two-thirds of Nvidia HBM4Above the >50% previously expected
researchTrendForce — AMD's Helios rack challenges Nvidia with an HBM4 edgeMicrosoft joins as a Helios customer; ~31 TB HBM4 per rack
researchTech-Insider — AMD MI400 seriesMI450 XCD on TSMC N2P
researchSemiconductorX — AMD Instinct MI400, EPYC Venice, Helios supply chainSupply-chain mapping for the AMD rack stack
researchSemiconductorX — Nvidia AI GPU supply chain, HBM4, CoWoS, Vera RubinCoWoS and HBM as the binding constraints
researchIntuitionLabs — Nvidia GB200 supply chainComponent-level map of the Blackwell rack
researchOplexa — AI chip packaging bottleneck 2026Nvidia holds ~60% of TSMC's 2026 CoWoS allocation
researchTom's Hardware — The custom AI ASIC state of playBroadcom ~60% of design partnerships; Broadcom+Marvell ~95%; TPU v7 on N3P with Broadcom + MediaTek
researchFuturum — Marvell attaches across Google's TPU stackWarrant vesting toward a $120B purchase threshold
researchEnkiAI — Oracle AI data centers, $638B RPOBacklog and campus strategy
researchSemiAnalysis — How Oracle is winning the AI compute marketStructure of the Oracle lease-and-rent model
researchLightwave (Dell'Oro data) — Schneider and Vertiv dominate DC physical infrastructureTied within 0.1 percentage point
researchIntrol — Vertiv vs Schneider vs Eaton cooling comparisonVertiv ~23% of precision cooling; DLC revenue more than doubled
researchMarketsandMarkets — Top companies in the data-center power marketSchneider, ABB, Eaton, Vertiv, Delta ≈41–43% combined
researchTechRepublic (Omdia) — Data center power and cooling trendsNext-gen racks toward 370 kW; liquid cooling becomes mandatory
researchCignal AI / LightCounting — 800GbE optics shipmentsInnoLight 23.4% of all transceivers in 2025, Coherent ~16%
researchValueAdd VC — $725B AI capex 2026Amazon ~$200B, Google $175–185B, Meta $125–145B, Microsoft $110–120B
researchEnkiAI — Unpacking Meta's Nvidia dealMeta capex trajectory 2024 → 2026
pressThe Tribune — TSMC foundry share hits record 72.5%Confirms the Q2 2026 record share
pressEvertiq — SMIC narrows gap with SamsungSMIC $3.01B, +20% QoQ
pressSeoul Economic Daily — Samsung HBM share jumps to 33%Q2 2026: SK hynix 50%, Samsung 33% (from 21%)
pressCNBC — Nvidia set to supplant Apple as TSMC's top customerFirst change at the top of TSMC's customer list since about 2014
pressTechPowerUp — Nvidia beats Apple to become TSMC's largest customerUS$23.2B vs US$20.5B
pressCNBC — Nvidia's top two mystery customers were 39% of Q2 revenueNvidia declined to identify Customer A and B
pressDCD — Two unnamed customers were almost 40% of Nvidia's Q2 FY26 revenueExplains that direct customers are typically ODM/OEMs, not end users
pressSeeking Alpha — SK hynix secures 70% of Nvidia's HBM orders for Vera RubinHigher of the two reported allocations
pressTrendForce — Samsung and SK hynix tapped as Rubin HBM4 suppliersShipments could start March 2026
pressBigGo Finance — SK hynix Nvidia revenue tops ~$5.2BQ1 2026 HBM revenue from Nvidia, +62.6% YoY
pressKorea Herald — Samsung supplies HBM4 as AMD's Helios enters productionSamsung is the named HBM supplier inside Helios
pressCNBC — AMD launches Helios, adding Microsoft as newest buyerFirst AMD rack-scale system to target Nvidia NVL
pressNotebookcheck — Nvidia secures supply orders from Wistron and FoxconnNamed rack manufacturing partners
pressCNBC — Marvell pops on report it will help Google with custom AI chipsBroadcom shares fell on the same report
pressTNW — Google in talks with Marvell for AI inference chipsMulti-supplier TPU architecture
pressYahoo Finance — OpenAI is making its own AI chips with Broadcom's help10 GW of custom accelerators from H2 2026
pressFierce Network — Encyclopedia of AI dealsConsolidated ledger of the Nvidia 10 GW, AMD 6 GW, Broadcom 10 GW and Oracle $300B commitments
pressBuilt In — OpenAI's $1T infrastructure planDeal-by-deal breakdown of OpenAI compute contracts
pressData Center Frontier — OpenAI and Oracle's $300B Stargate dealFive-year contract tied to US Stargate campuses
pressBenzinga — Oracle capex could hit $100B as Stargate expandsFY27 capex guide $80–100B; ~$40B Nvidia GPU procurement
pressDCD — Oracle to lease Texas data center from CrusoeAbilene campus; Oracle installs GPUs and rents onward
pressAI Magazine — OpenAI, Oracle and Vantage's $15bn data centreFour hyperscale data centers, nearly 1 GW by 2028
pressTechCrunch — Microsoft and Chevron plan a large gas-powered data center projectTwo GE Vernova turbines plus Caterpillar Solar Turbines
pressIP-Fiber — InnoLight and Eoptolink dominate 60% of 800G SFP supplyRemaining ~40% split across Coherent, Lumentum, Broadcom
pressCNBC — Hyperscalers face capex scrutinyInvestor pushback on the 2026 spending step-up
speculationAnalyst thread — AMD/Samsung HBM analysisHBM capacity across all three vendors sold out through 2026; AMD allocation inference
speculationCurionic — Inside AI's circular financing webNvidia / OpenAI / Oracle / CoreWeave equity and revenue loops
How to read this. Two layers sit on this page. The supply-chain graph is curated and dated 2026-09-12: no market API publishes supply allocations, so each edge is tagged official (filing or company statement), research (market-research estimate — firms disagree, most visibly on HBM share), press (credible reporting neither side confirms), speculation (explicitly unverified — the identities behind Nvidia's Customers A–D, and the Samsung/Micron split of its residual HBM4) or background (structural, no 2026 document pulled here). Prices, percentage moves and headlines are live from Yahoo Finance on every page load. Gigawatt figures are contracts and letters of intent, not delivered capacity.