AI

The Enabler

Sells the machines that make chip-making possible

ASML — ASML Holding

Dutch company with a 100% global monopoly on EUV lithography — the school-bus-sized, $200M–$350M machines that use extreme ultraviolet light to print microscopic transistors onto silicon. Without ASML, advanced chip production stops worldwide.

Why it matters: The single deepest moat in this web. Nikon and Canon only compete in older DUV tech — for cutting-edge EUV there is no alternative on earth.

Outlook: Raised full-year 2026 guidance to a record €43–45B after a Q2 beat (€9.3B net sales). Boosting EUV production capacity ~30% per year; advanced machine backlog fully booked through 2027.

revenue $48B (research snapshot; the live map shows current data)

The Memory

HBM, DRAM, and NAND — the silicon beside every AI chip

MU — Micron

One of three companies on earth making High-Bandwidth Memory — silicon layers stacked like tiny skyscrapers that must sit physically touching the GPU so data moves fast enough to keep the AI math fed.

Why it matters: Memory is the current bottleneck of the AI buildout. FQ3 2026 revenue grew 346% YoY to $41.5B on HBM premium pricing.

Outlook: Standout S&P 500 performer of 2026: up 240%+ YTD into the trillion-dollar club. Over $100B in binding multi-year contracts — advanced HBM supply is effectively sold out through the end of 2026.

revenue $120B (research snapshot; the live map shows current data)

SKHY — SK hynix

South Korean memory titan that pioneered modern HBM and was Nvidia’s original partner for co-packing memory next to the H100/H200. Now trades on Nasdaq as SKHY, sitting in the trillion-dollar club alongside Micron.

Why it matters: Micron’s only real HBM rival at the leading edge (Samsung trails). Together they decide how fast the whole AI buildout can go.

Outlook: Record-high operating margins on HBM3e premium pricing. Running at absolute maximum capacity — production slots fully spoken for through 2026/2027 by Nvidia and tier-1 hyperscalers.

revenue $95B (research snapshot; the live map shows current data)

SSNLF — Samsung Electronics

The third HBM maker and the world’s biggest memory company overall — plus the only foundry even attempting to challenge TSMC at the leading edge. Trades in Korea; the US ticker is a thin OTC listing.

Why it matters: The sleeping giant: if Samsung ever catches up in HBM or foundry yield, the economics of this whole map shift. Off the watchlist because its OTC feed is unreliable — numbers here are static research.

Outlook: Racing to qualify HBM4 with Nvidia while Samsung Foundry chases TSMC at 2nm. The memory upcycle is lifting earnings, but it remains #3 in HBM behind SK hynix and Micron.

market cap $1.20T · revenue $220B (research snapshot; the live map shows current data)

SNDK — SanDisk

Pure-play NAND flash maker (spun out of Western Digital) building the high-capacity datacenter SSDs that store the oceans of data AI training and inference feed on.

Why it matters: The storage side of the memory squeeze: AI datacenters are hoovering up enterprise SSDs, and NAND pricing has gone vertical — revenue up ~250% YoY.

Outlook: Riding a historic NAND shortage: revenue up ~250% YoY as AI datacenters absorb enterprise SSD supply, with pricing power expected to hold through 2026.

SIMO — Silicon Motion

Designs the controller chips that run SSDs — the brains sitting between raw NAND flash and the system. Client SSD controllers are the cash cow; the MonTitan enterprise platform is the AI-datacenter bet, now ramping at five tier-one cloud providers.

Why it matters: The picks-and-shovels of the NAND boom, one level up from the flash makers: every SSD needs a controller, and AI datacenters are suddenly buying the enterprise kind. Small enough that the MonTitan ramp can re-rate the whole company.

Outlook: Record Q1 2026 revenue of $342M, up 105% YoY, with boot-drive solutions up ~750%; MonTitan entered volume production ahead of schedule with five tier-one CSPs ramping in H2 2026. Street targets have chased the AI storage cycle up to $350–450.

market cap $9B · revenue $1B (research snapshot; the live map shows current data)

The Foundry

Physically manufactures and packages the chips

TSM — TSMC

The world’s largest chip foundry. Designs nothing itself — it takes blueprints from Nvidia and Apple and physically etches them onto silicon. Crucially, it owns CoWoS, the exclusive packaging tech that fuses Nvidia’s GPUs and HBM memory into one module.

Why it matters: Every advanced chip in this web funnels through TSMC’s fabs and packaging lines. Even the memory makers ship to TSMC, not to Nvidia directly.

Outlook: Q2 2026 net profit surged 77% YoY to $40.2B; raised full-year sales growth guidance above 40%. Capex scaling to $60–64B, including a $265B multi-fab expansion in Arizona to hedge geopolitical risk.

revenue $145B (research snapshot; the live map shows current data)

The Designer

Designs the GPUs and owns the software moat

NVDA — Nvidia

Designs the world’s most powerful AI GPUs (H100, Blackwell, upcoming Vera Rubin) plus the InfiniBand networking and CUDA software that lock in 80–90% of AI development. Fabless: TSMC builds everything.

Why it matters: The center of the AI trade. Everything to the left exists to build Nvidia’s chips; everything to the right exists to buy them.

Outlook: Trailing revenue ~$253B; Street models nearly a double to ~$555B by early 2028 as hyperscaler capex stays at record highs through 2027.

revenue $253B (research snapshot; the live map shows current data)

AMD — AMD

The second source for AI GPUs (Instinct MI series) and the dominant server CPU vendor (EPYC). The main hope for hyperscalers who want a real alternative to Nvidia — and, like Nvidia, fabless: TSMC builds everything.

Why it matters: Every dollar the buyers column spends with AMD is leverage against Nvidia’s pricing. Watching AMD’s AI revenue ramp tells you how real the "second source" thesis is.

Outlook: Instinct MI-series ramps and fresh cloud wins have revenue growing ~28% YoY; the second-source thesis strengthens every time Nvidia supply tightens.

revenue $33B (research snapshot; the live map shows current data)

CBRS — Cerebras

Builds wafer-scale AI chips — instead of cutting a silicon wafer into hundreds of small GPUs, it ships the whole dinner-plate-sized wafer as one giant processor, wired for ultra-fast AI inference.

Why it matters: The radical architecture bet in the designer column: if wafer-scale inference wins even a niche, it chips at Nvidia’s margins from below.

Outlook: Public since its 2025 IPO with wafer-scale inference deals ramping (G42 the anchor customer); the stock trades on the architecture bet more than current earnings.

revenue $2B (research snapshot; the live map shows current data)

AVGO — Broadcom

Co-designs the hyperscalers’ in-house AI chips — Google’s TPU above all — and sells the Ethernet networking silicon that stitches AI clusters together. The quiet giant of the "custom ASIC" counter-bet against Nvidia.

Why it matters: Every hyperscaler that designs its own chip does it with Broadcom. If custom ASICs take share from merchant GPUs, AVGO is the main winner.

Outlook: AI revenue guided to roughly double again in FY2026 as TPU volumes and a third custom-ASIC customer ramp; ~48% total revenue growth with Ethernet networking attached.

revenue $65B (research snapshot; the live map shows current data)

MRVL — Marvell

Broadcom’s main rival in custom AI silicon: co-designs Amazon’s Trainium chips and sells the electro-optics that move data between AI datacenter racks at light speed.

Why it matters: The other way to bet on the custom-ASIC wave. Marvell winning or losing a hyperscaler design cycle moves the stock violently in both directions.

Outlook: Custom-silicon ramps (Trainium 3) and 1.6T optics drive ~28% growth; the stock swings on every hyperscaler design-win headline.

The Builders

Assembles the silicon into the switches and racks datacenters actually install

CLS — Celestica

The contract builder of AI infrastructure: designs and manufactures the 800G/1.6T network switches, liquid-cooled racks, and custom servers hyperscalers deploy — turning Broadcom’s switch silicon into installed hardware.

Why it matters: The invisible layer between chip designers and datacenters: when a hyperscaler orders a custom switch or an ASIC rack, Celestica is who actually builds it. AI networking share gains made it one of the market’s quiet monsters.

Outlook: 2026 revenue guidance raised toward $17–19B on AI datacenter demand; won a co-packaged-optics 1.6T switch program with a hyperscaler, and the cloud segment (~75% of revenue) keeps compounding on the 800G→1.6T transition.

market cap $45B · revenue $15B (research snapshot; the live map shows current data)

The Optics

The lasers, transceivers, and retimers that move data between chips at light speed

COHR — Coherent

Makes the lasers, indium-phosphide chips, and optical transceivers that carry data between GPUs — the picks-and-shovels of AI networking, from the laser diode up to the 800G/1.6T module.

Why it matters: Every AI cluster is bottlenecked by how fast light moves data between chips — Coherent supplies the lasers that make it possible, one layer below Broadcom and Marvell.

Outlook: Datacenter transceiver demand is the growth engine as clusters scale to hundreds of thousands of GPUs; the debate is margin recovery and vertical-integration payoff.

LITE — Lumentum

Photonics leader pivoting hard from telecom into AI datacenter interconnect — high-speed optical transceivers and the laser chips inside them.

Why it matters: A more focused, higher-torque bet on the same optical-interconnect wave — smaller than Coherent, so design wins move it harder.

Outlook: Cloud/AI transceiver orders are re-accelerating revenue after the telecom slump; execution on 800G/1.6T ramps is the whole story.

AAOI — Applied Optoelectronics

Small-cap optical component maker supplying transceivers to hyperscalers and cable operators — the high-beta, high-risk way to play the datacenter-optics ramp.

Why it matters: The lottery ticket of the group: a single large hyperscaler transceiver contract can re-rate the whole company.

ALAB — Astera Labs

Pure-play designer of the connectivity silicon that holds AI racks together — Aries PCIe retimers, Taurus smart-cable modules, Leo CXL memory controllers, and Scorpio fabric switches. When GPUs outrun the wires between them, Astera sells the fix.

Why it matters: The copper-and-PCIe side of the interconnect story: every faster GPU generation makes data movement the bottleneck, and Astera monetizes exactly that gap — one layer down from Broadcom, alongside the optics makers.

Outlook: Q1 2026 revenue grew 93% YoY to $308M with Q2 guided to ~$360M as Scorpio fabric switches ramp; the expanded NVLink Fusion collaboration with Nvidia opens rack-scale content ramping into 2027. Reports Q2 on Aug 4, 2026.

market cap $52B · revenue $1B (research snapshot; the live map shows current data)

The Neoclouds

Rent-a-GPU: they buy Nvidia’s chips and lease them to everyone else

CRWV — CoreWeave

The original "neocloud": rents massive Nvidia GPU clusters to AI labs and hyperscalers alike. Grew out of a crypto-mining operation — Nvidia is both its supplier and a shareholder.

Why it matters: The purest proof that GPU capacity itself is a business — and the bridge between Nvidia’s chips and the labs that can’t build datacenters fast enough.

Outlook: Contract backlog anchored by an ~$12B multi-year OpenAI deal on top of its Microsoft business; the debate is debt-fueled buildout pace vs contract durability.

market cap $40B (research snapshot; the live map shows current data)

NBIS — Nebius

European GPU cloud spun out of Yandex — Nvidia-backed, building AI datacenter capacity across Europe and the US, anchored by a multi-year Microsoft capacity deal.

market cap $45B (research snapshot; the live map shows current data)

IREN — IREN

Former Bitcoin miner that pivoted its cheap renewable power sites into GPU datacenters — the energy-to-compute arbitrage play, now selling capacity to Microsoft.

market cap $12B (research snapshot; the live map shows current data)

The Buyers

Hyperscalers spending record capex on AI compute

MSFT — Microsoft

Runs Azure, one of the three big clouds, and bankrolls OpenAI. One of the largest single buyers of Nvidia GPUs.

revenue $300B (research snapshot; the live map shows current data)

GOOGL — Alphabet (Google)

Search, YouTube, and Google Cloud. Builds its own TPU chips so it leans on Nvidia less than the other hyperscalers.

revenue $400B (research snapshot; the live map shows current data)

META — Meta

Facebook, Instagram, WhatsApp. Spends tens of billions a year on GPUs for recommendation engines and its open-source Llama models.

revenue $190B (research snapshot; the live map shows current data)

AMZN — Amazon

AWS is the largest cloud on the planet. Buys Nvidia GPUs at scale while pushing its own Trainium chips to cut costs.

revenue $680B (research snapshot; the live map shows current data)

The Model Labs

Where all that compute finally turns into intelligence

OPENAI — OpenAI (private)

Maker of ChatGPT and the GPT frontier models — the company whose compute appetite kicked off this entire supply chain. Private, but arguably the most important customer in the map.

Why it matters: Every dollar in this web ultimately chases the demand OpenAI proved. Its Azure bill and direct GPU orders ripple all the way back to ASML.

Outlook: Closed a $122B round at an $852B valuation (Mar 2026), then filed confidentially for a $1T+ IPO. Revenue run-rate ~$25B and targeting $30B for 2026 — with burn near $27B as the Stargate buildout accelerates.

valuation $852B · revenue $25B (research snapshot; the live map shows current data)

ANTHROPIC — Anthropic (private)

Maker of the Claude frontier models, backed by Amazon and Google. Trains on AWS Trainium clusters — making it the proof point for the custom-ASIC counter-bet against Nvidia.

Why it matters: The other frontier lab whose compute demand anchors this web — and whose Trainium bet ties Amazon, Marvell, and the anti-Nvidia thesis together.

Outlook: Raised a $65B Series H at $965B post-money (May 2026) — the most valuable AI company — with run-rate revenue crossing $47B, up from $9B at end-2025. S-1 filed for an October 2026 IPO.

valuation $965B · revenue $47B (research snapshot; the live map shows current data)

The End Market

Where the money ultimately comes from

Businesses

Enterprises buying AI through cloud APIs, copilots, and agents — automating work that used to take headcount. The budget that justifies everything to the left.

Consumers

Hundreds of millions of people paying ~$20/month for an AI assistant — the first mass-market subscription born from this supply chain.

Relationships on the map

• Coherent supplies Microsoft — Optical transceivers (~$2B/yr)
  Coherent’s 800G/1.6T optical modules wire together the GPU clusters inside Azure’s AI datacenters.
• Lumentum supplies Amazon — Optical transceivers (~$1B/yr)
  Lumentum’s datacenter transceivers connect AWS’s AI racks at light speed.
• Applied Optoelectronics supplies Alphabet (Google) — Transceivers (~$300M/yr)
  AAOI supplies optical transceivers to hyperscalers building out AI network fabric.
• Coherent partners with Broadcom — Optics for Ethernet
  Broadcom’s Ethernet switch silicon needs the optical modules Coherent and its peers build around it.
• Broadcom supplies Celestica — Switch silicon (~$2B/yr)
  Celestica builds its hyperscaler network switches around Broadcom’s Tomahawk-class Ethernet silicon.
• Celestica supplies Alphabet (Google) — AI racks & switches (~$4B/yr)
  Google is Celestica’s largest customer: custom TPU racks and datacenter network switches, built to order.
• Celestica supplies Meta — 800G switches (~$3B/yr)
  Celestica designs and manufactures 800G network switches and AI compute hardware for Meta’s datacenter buildout.
• Astera Labs supplies Amazon — Retimers & fabric switches (~$350M/yr)
  Amazon has been Astera’s anchor customer since the AI buildout began — AWS racks lean on Aries retimers and Scorpio fabric to keep custom-ASIC and GPU clusters fed.
• Astera Labs partners with Nvidia — NVLink Fusion
  Astera builds NVLink Fusion scale-up connectivity with Nvidia — letting hyperscaler custom chips plug into Nvidia’s rack architecture, with ramps slated for 2027.
• Astera Labs competes with Broadcom — PCIe switches & retimers
  Broadcom is the incumbent in PCIe switching and retimers; Scorpio is Astera’s direct shot at that franchise inside AI racks.
• Lumentum competes with Coherent — Datacenter optics
  Lumentum and Coherent fight for the same hyperscaler transceiver and laser design wins.
• ASML Holding supplies TSMC — EUV machines (~$15B/yr)
  ASML sells 100% of the advanced EUV printing machines TSMC needs to manufacture Nvidia’s chips.
• Micron supplies TSMC — HBM memory (~$35B/yr)
  Micron ships finished HBM memory blocks to TSMC’s packaging facilities — not to Nvidia directly.
• SK hynix supplies TSMC — HBM memory (~$45B/yr)
  SK hynix supplies cutting-edge memory stacks to TSMC’s foundry points for integration into the module.
• Samsung Electronics supplies TSMC — HBM memory (~$10B/yr)
  Samsung supplies HBM where it can win qualification — still chasing Micron and SK hynix at the leading edge.
• TSMC supplies Nvidia — Finished GPU modules (~$45B/yr)
  TSMC prints Nvidia’s silicon and fuses it with memory using proprietary CoWoS packaging, shipping completed $30K+ processing modules.
• TSMC supplies AMD — Chips & packaging (~$8B/yr)
  TSMC fabricates AMD’s Instinct GPUs and EPYC CPUs on the same advanced nodes Nvidia uses — the two rivals share one factory.
• TSMC supplies Cerebras — Wafer-scale chips (~$1B/yr)
  TSMC fabricates Cerebras’ dinner-plate-sized wafer-scale engines.
• TSMC supplies Broadcom — Chips & packaging (~$6B/yr)
  TSMC fabricates Broadcom’s custom AI ASICs and networking silicon on the same advanced nodes as Nvidia and AMD.
• TSMC supplies Marvell — Chips & packaging (~$2B/yr)
  TSMC fabricates Marvell’s custom ASICs and optical interconnect silicon.
• Nvidia supplies Microsoft — GPUs & networking (~$50B/yr)
  Microsoft buys Nvidia GPUs at massive scale to power Azure and OpenAI workloads.
• Nvidia supplies Alphabet (Google) — GPUs (~$20B/yr)
  Google Cloud offers Nvidia GPUs alongside its in-house TPUs.
• Nvidia supplies Meta — GPUs & networking (~$40B/yr)
  Meta is among the largest GPU buyers on earth for training and recommendation systems.
• Nvidia supplies Amazon — GPUs (~$30B/yr)
  AWS rents Nvidia GPU capacity to the world while developing its own Trainium alternative.
• AMD supplies Microsoft — Instinct GPUs (~$5B/yr)
  Microsoft deploys AMD’s Instinct accelerators in Azure as its hedge against Nvidia dependence.
• AMD supplies Meta — Instinct GPUs (~$4B/yr)
  Meta runs Llama inference on AMD Instinct GPUs alongside its Nvidia fleet.
• Broadcom supplies Alphabet (Google) — TPU co-design (~$15B/yr)
  Broadcom co-designs and supplies Google’s TPU accelerators — the main reason Google leans on Nvidia less than its peers.
• Broadcom supplies Meta — Custom ASICs (MTIA) (~$4B/yr)
  Broadcom helps build Meta’s in-house MTIA inference chips, Meta’s hedge against GPU pricing.
• Marvell supplies Amazon — Trainium co-design (~$3B/yr)
  Marvell co-designs and supplies Amazon’s Trainium AI chips — AWS’s in-house alternative to buying more Nvidia GPUs.
• Marvell supplies Microsoft — Optics & custom silicon (~$2B/yr)
  Marvell sells Microsoft the electro-optics that link AI clusters, plus custom silicon work.
• SanDisk supplies Amazon — Datacenter SSDs (~$2B/yr)
  SanDisk’s NAND flash and enterprise SSDs store the training data and model weights AI clouds run on.
• Microsoft supplies OpenAI — Azure compute (~$15B/yr)
  Microsoft has invested $13B+ in OpenAI and provides the Azure supercomputers its frontier models train and run on.
• Amazon supplies Anthropic — AWS compute (Trainium) (~$15B/yr)
  Amazon has invested $8B+ in Anthropic, which trains Claude on massive AWS Trainium clusters — the chips Marvell co-designs.
• Alphabet (Google) supplies Anthropic — GCP compute & TPUs (~$5B/yr)
  Google is also an Anthropic investor and serves Claude workloads on Google Cloud, including TPU capacity.
• Nvidia supplies OpenAI — GPUs (direct) (~$10B/yr)
  Beyond what it rents from Azure, OpenAI increasingly buys Nvidia systems directly for its own datacenter buildouts.
• OpenAI competes with Anthropic — Frontier models
  GPT vs Claude: the two independent frontier labs race on capability, safety, and enterprise adoption.
• OpenAI competes with Alphabet (Google) — ChatGPT vs Gemini
  ChatGPT is the first real threat to Google Search in decades; Google answers with Gemini across its whole stack.
• AMD competes with Nvidia — AI GPUs
  AMD’s Instinct line is the only merchant-silicon GPU alternative to Nvidia at scale; CUDA vs ROCm is the software front of the same war.
• Cerebras competes with Nvidia — Wafer-scale vs GPUs
  Cerebras attacks AI inference with one giant chip instead of clusters of GPUs — a bet that Nvidia’s networking advantage can be designed away.
• Broadcom competes with Nvidia — Custom ASICs vs GPUs
  Broadcom’s custom ASICs and Ethernet networking are the hyperscalers’ escape route from Nvidia’s GPU + InfiniBand lock-in.
• Marvell competes with Broadcom — Custom ASIC design wins
  Marvell and Broadcom fight for every hyperscaler custom-chip contract — Broadcom holds Google and Meta, Marvell holds Amazon.
• Micron competes with SK hynix — HBM leadership
  Micron and SK hynix (with Samsung behind them) fight for every HBM design win; both are sold out through 2026.
• Samsung Electronics competes with SK hynix — HBM & DRAM
  Samsung is racing to qualify HBM4 at Nvidia and reclaim the memory crown from its Korean rival.
• Samsung Electronics competes with TSMC — Leading-edge foundry
  Samsung Foundry is the only fab even attempting to challenge TSMC at 2nm — so far with far lower yields.
• SanDisk competes with Micron — NAND flash
  SanDisk and Micron fight for the NAND/SSD orders of the same AI datacenters.
• Silicon Motion supplies Micron — SSD controllers (~$150M/yr)
  Silicon Motion supplies the controller silicon inside client SSDs across the major NAND makers, Micron included.
• Silicon Motion competes with Marvell — Enterprise SSD controllers
  MonTitan puts Silicon Motion in direct competition with Marvell for the enterprise SSD controller sockets in AI datacenters.
• Silicon Motion partners with Nvidia — AI storage ecosystem
  Silicon Motion’s PCIe boot drives and near-GPU MonTitan controllers are built to slot into Nvidia’s AI server reference designs.
• OpenAI supplies Businesses — API & enterprise (~$12B/yr)
  Enterprise and API sales are 40%+ of OpenAI revenue and on track for parity with consumer by end of 2026.
• OpenAI supplies Consumers — ChatGPT subscriptions (~$13B/yr)
  Consumer subscriptions remain the larger half of OpenAI revenue — for now.
• Anthropic supplies Businesses — API & enterprise (~$40B/yr)
  Anthropic’s $47B run-rate is overwhelmingly API and enterprise — the purest B2B lab.
• Nvidia supplies CoreWeave — GPUs & systems (~$15B/yr)
  Nvidia supplies CoreWeave’s entire fleet and holds an equity stake — vendor, investor, and demand backstop at once.
• Nvidia supplies Nebius — GPUs (~$3B/yr)
  Nvidia supplies and invested in Nebius — seeding the neocloud ecosystem that buys its chips.
• Nvidia supplies IREN — GPUs (~$2B/yr)
  IREN converts its power sites into Nvidia GPU capacity.
• CoreWeave supplies Microsoft — GPU capacity (~$8B/yr)
  Microsoft rents CoreWeave clusters to absorb overflow AI demand it can’t build fast enough itself.
• CoreWeave supplies OpenAI — Compute contract (~$4B/yr)
  OpenAI signed a multi-year ~$12B compute contract with CoreWeave — capacity beyond what Azure provides.
• Nebius supplies Microsoft — GPU capacity (multi-year) (~$3B/yr)
  Microsoft signed a multi-year, ~$17B+ capacity deal with Nebius — validation that the neocloud model has hyperscale customers.
• IREN supplies Microsoft — GPU capacity (~$2B/yr)
  IREN’s ~$9.7B cloud deal with Microsoft turned the ex-miner into a contracted AI infrastructure provider.