Tokens flow downstream. Margins flow upstream. Here's the full supply chain, mapped with real financials, so you know where to put your money.
Every AI chip starts as a silicon wafer. Every GPU needs HBM memory. Every data center needs fiber optic cables. This is the foundation of the stack.
| Company | Product | AI Revenue | AI % | Op Margin | Growth |
|---|---|---|---|---|---|
| SK Hynix 000660 | HBM memory (#1, 60% share) | ~$30B | ~40%+ | 49% | +50% |
| Micron MU | HBM, DRAM, SSDs | $20.75B | 56% | 47% | +50% |
| Samsung 005930 | HBM (#2), DRAM, NAND | ~$60B semi div | ~18%* | 50%+ | +208% Q4 |
| Shin-Etsu 4063 | Silicon wafers (#1) | N/A | N/A | 29% | +6% |
| Corning GLW | Fiber optics for DCs | $6.3B | 38% | 19% | +35% |
| SUMCO 3436 | Silicon wafers (#2) | N/A | N/A | 0.3% | Declining |
HBM memory is the gold. SK Hynix prints money at 49% margins with 60% market share. Every H100 needs 80GB of HBM3e. Every Blackwell needs 192GB. Memory demand scales linearly with GPU shipments.
SUMCO (0.3% margin, declining revenue). Being the #2 wafer maker with no pricing power is a bad place to be.
Someone has to print the transistors. TSMC does it for almost everyone. Samsung and Intel are trying to compete, and both are losing money doing it.
| Company | Product | AI Revenue | AI % | Op Margin | Growth |
|---|---|---|---|---|---|
| TSMC TSM | Chip fab (#1, 65% share) | ~$21B | 17-19% | 51% | +32% |
| Samsung Foundry 005930 | Chip fab (#2) | N/A | N/A | Negative | Declining |
| Intel Foundry INTC | Chip fab (#3) | $307M ext. | 1.7% | -50% | Losses 4x'd |
TSMC is a monopoly in everything but name. 65% foundry share, 51% operating margins, every AI chip on the planet runs through their fabs. AI revenue CAGR of 54-56% through 2029. Capex jumping to $52-56B in 2026.
Samsung Foundry and Intel Foundry are both hemorrhaging cash trying to catch up. The gap is widening. TSMC's moat comes from yield rates: 80%+ at 3nm vs. Samsung at sub-60%.
This is where the highest margins live. NVIDIA designs the chips. Broadcom designs the custom ASICs. Arista builds the networking that connects GPU clusters. They don't make anything physical. They design it, license it to foundries, and collect rent.
| Company | Product | AI Revenue | AI % | Op Margin | Growth |
|---|---|---|---|---|---|
| NVIDIA NVDA | GPU design, networking, CUDA | $197.3B | 91% | ~60% | +65% |
| Broadcom AVGO | Custom ASICs, networking | $8.4B/q | 44% | 67% EBITDA | +106% Q1 |
| AMD AMD | MI300/400 GPUs, EPYC CPUs | $16.6B | 49% | 33% | +35% |
| Arista ANET | DC Ethernet switches | ~$6.1B | 68% | 43% | +29% |
| Marvell MRVL | Custom AI silicon, networking | ~$4.3B | 75% | 34% | +27% |
NVIDIA is in a class of its own: $216B revenue, $120B net income, 91% from AI. But Broadcom is the stealth winner. 77% gross margins, 70% custom ASIC share. They design the chips (Google TPU) that compete with NVIDIA. AI revenue doubled in one quarter.
Why NVIDIA's margins matter: 75% gross margin on $216B revenue = $162B in gross profit. That's more gross profit than most companies have total revenue. The CUDA ecosystem lock-in is the moat.
The landlords of compute. Cloud providers rent GPU time. Data center REITs sell physical rack space. Power utilities sell the electricity. Cooling companies keep it all from melting.
| Company | Product | Revenue | AI Segment | Op Margin | Growth |
|---|---|---|---|---|---|
| AWS AMZN | Cloud, Trainium chips | $128.7B | $10B+ custom chips | 35% | +20-24% |
| Azure MSFT | Cloud, OpenAI API | $75B+ | $13B AI run rate | ~43% | +34-40% |
| Google Cloud GOOGL | Cloud, TPUs, Vertex AI | $70B+ run rate | N/A | 23-30% | +28-48% |
| Company | Product | Revenue | AI Exposure | Op Margin | Growth |
|---|---|---|---|---|---|
| Vertiv VRT | DC cooling, power mgmt | $10.2B | Nearly all AI-driven | 22% | +28% |
| Equinix EQIX | Colocation (280+ DCs) | $9.2B | 60% of top deals | 20% | +5-11% |
| Digital Realty DLR | DC wholesale space | $6.1B | 20% of bookings | 26% | +10% |
| Constellation CEG | Nuclear power | $25.5B | 20-yr PPAs (Meta) | 12% | +8% |
| Vistra VST | Nuclear/gas power | $17.7B | PPAs (AWS, Meta) | 11% | +3% rev |
Vertiv has orders up 252% and a $15B backlog. Every GPU deployed needs cooling. Liquid cooling market growing at 31.5% CAGR through 2033. Power is the bottleneck below the bottleneck. Constellation and Vistra have signed 20-year PPAs with hyperscalers. Nuclear plants can't be built fast enough.
They train the models that create all the token demand. Their product is intelligence itself, sold per token. Every company in L7 depends on them. And every one of them is burning cash at industrial scale.
| Company | Product | ARR | Valuation | Margin | Growth |
|---|---|---|---|---|---|
| OpenAI Private | GPT models, ChatGPT | ~$25B | $852B | Negative (-$9B/yr) | 3x/yr |
| Anthropic Private | Claude, Claude Code | ~$19B (Mar 26) | $380B | Negative | ~10x/yr |
| DeepMind GOOGL | Gemini models | Low single-digit B | Part of Alphabet | N/A | N/A |
| Meta AI META | Llama (open-source) | $2-3B direct | Part of Meta | N/A | 22% (Meta) |
| Mistral Private | Mistral Large, Mixtral | ~$400M | $13.7B | Not disclosed | ~20x/yr |
All private or embedded in larger companies. Anthropic's IPO (possibly Oct 2026) would be the first pure-play frontier AI stock. Until then, you're investing through cloud providers (L4) or chip makers (L3). OpenAI burned $9B in 2025, projecting $17B in 2026. The bet is that models get cheaper to run faster than the competition catches up.
They buy GPU time wholesale and sell tokens retail. The spread between cost and price is the business. This is the fastest-growing layer by percentage.
| Company | Product | Revenue | Valuation | Margin | Growth |
|---|---|---|---|---|---|
| CoreWeave CRWV | GPU cloud for AI | $5.1B | ~$35B mkt cap | 61% EBITDA | +170% |
| Fireworks AI Private | Inference API | $315M ARR | $4B | ~50% gross | +416% |
| Together AI Private | Open-source inference | ~$300M-$1B | $3.3-7.5B | N/A | 3x+ |
| Groq Private* | LPU inference chips | ~$500M proj | $6.9B | N/A | N/A |
CoreWeave (CRWV) is the only public pure-play. 170% revenue growth, 61% EBITDA margins. But more than half of revenue comes from one customer (likely Microsoft). Customer concentration is the risk.
The addicts. Every company at this layer is 100% dependent on the layers below. Cursor can't function without inference tokens. If Anthropic raises prices or Together AI goes down, their product breaks. This is the Token Daddy thesis playing out.
| Company | Product | ARR | Valuation | Margin | Growth |
|---|---|---|---|---|---|
| Cursor Private | AI code editor | ~$2B | $29-50B | N/A | 10x+ |
| Databricks Private | Data lakehouse + AI | $5.4B | $134B | ~Breakeven FCF | +55-65% |
| Perplexity Private | AI search | ~$200M | $21.2B | N/A | +230% |
| Jasper AI Private | AI content marketing | ~$88M | ~$1.8B | N/A | -27% from peak |
Cursor at $2B ARR is 10x Perplexity's $200M. Jasper's revenue declined from its peak, proving that commodity AI features get eaten by platforms when ChatGPT shows up. All private. Watch for Databricks IPO.
Gross/operating margins by layer. The pattern: margins are highest where you're closest to the silicon and where you hold a monopoly position.
Money flows downstream. Margins flow upstream.
Applications (L7) are the most visible. Chips (L3) and memory (L1) capture the most value. The addiction creates the demand, but the dealer keeps the margin.
Public, investable companies ranked by conviction.
| Company | ARR | Valuation | Expected IPO |
|---|---|---|---|
| Anthropic | ~$19B+ | $380B | Possibly Oct 2026 |
| Databricks | $5.4B | $134B | Preparing |
| Cursor | ~$2B | $29-50B | TBD |