Session 1 asked what the profit is. Session 2 asks where in the chain it sits, and why.
Each link's revenue is the next link's cost. Where a company sits decides whether a demand boom shows up as its margin or as its volume.
| Link | Who | Sells | Biggest cost | Players / pricing power |
|---|---|---|---|---|
| Fab | TSMC | Wafer fabrication | Fabs (fixed) | Very few; high |
| Chip design (fabless) | NVIDIA, AMD | Chips plus software ecosystem | R&D (fixed), foundry fees (variable) | Two or three; NVIDIA very high |
| Cloud | AWS, Azure, Google, Oracle | GPU hours | Chips, power, data centers | Three or four; medium |
| Model lab | OpenAI, Anthropic | Subscriptions plus API per token | Training (fixed), inference compute (variable) | A handful; medium |
| Applications | SaaS of all kinds | Subscriptions | API fees paid to model labs | Countless; low |
NVIDIA's fiscal year ends in late January, so FY2026 is mostly calendar 2025. FY2019 = before anything; FY2023 = the year ChatGPT launched; FY2026 = latest.
| FY2019 | FY2023 | FY2026 | |
|---|---|---|---|
| Total revenue ($B) | 12 | 27 | 216 |
| Data Center | 25% | 55% | 90% |
| Gaming (consumer GPUs) | 55% | 35% | 7% |
| Professional Visualization | 10% | 5% | 1% |
| Automotive | 5% | 3% | 1% |
| OEM & Other | 5% | 2% | 0% |
| Line | Buyers | Unit price | Typical deal | Annual volume |
|---|---|---|---|---|
| Gaming | Tens of millions of gamers via 10 to 15 board partners and Dell / HP / Lenovo | $300 to $4,000 per card | One card | About 30 million cards (est.) |
| Data Center | About a dozen cloud providers and model labs; in FY2026 the top two direct customers were 22% and 14% of total revenue | H100 / B200 about $25K to $40K per GPU (est.) | One GB200 rack $3M to $4M (est.); OpenAI letter of intent for 10 GW, up to $100B | A few million GPUs (est.) |
| Professional Visualization | Design, media, engineering workstations | $800 to $16,000 per card | IT fleet refresh | Not disclosed |
| Automotive | 35+ automakers and Tier-1s | Not disclosed | Design-win pipeline $14B over six years | Not disclosed |
| FY2019 | FY2023 | FY2026 | |
|---|---|---|---|
| Cost of revenue (TSMC, memory, packaging) | 40% | 45% | 30% |
| R&D | 20% | 25% | 9% |
| SG&A (marketing not separated) | 9% | 9% | 2% |
| Operating income | 32% | 16% | 60% |
AMD's fiscal year equals the calendar year. Segments were redefined in 2022 after the Xilinx acquisition.
| 2019 | 2022 | 2023 | 2025 | |
|---|---|---|---|---|
| Total revenue ($B) | 7 | 24 | 23 | 35 |
| Data Center | ~30% (old segments) | 25% | 29% | 50% |
| Client | ~70% (old segments) | 25% | 21% | 30% |
| Gaming | (in Client above) | 30% | 27% | 11% |
| Embedded | none | 20% | 23% | 10% |
| Gross margin (GAAP) | 43% | 45% | 46% | 50% |
| R&D | 23% | 21% | 26% | 23% |
| MG&A | 11% | 10% | 10% | 12% |
| Operating income (GAAP / non-GAAP) | 9% / 13% | 5% / 27% | 2% / 21% | 11% / 22% |
| Largest single customer | Sony above 10% | One Gaming customer 16% | One Gaming customer 18% | None above 10% |
OpenAI is private. All figures are press reports and leaked statements; directional only.
| 2022 | 2023 | 2024 | 2025 | |
|---|---|---|---|---|
| Revenue ($B) | 0.03 | 1.6 | 3.7 | 13 |
| ChatGPT consumer subscriptions | not split | not split | ~75% | ~2/3 |
| Business and enterprise seats | not split | not split | small | ~40% and rising |
| API per token | not split | not split | ~25% | 15% to 20% |
| Anchor | Value |
|---|---|
| ChatGPT Plus / Pro | $20 / $200 per month |
| Enterprise | about $60 per seat per month, 150-seat minimum; a typical contract is 150 x 60 x 12 = about $110K per year |
| API flagship | $5 / $30 per million tokens (input / output) |
| API cheap model | $0.05 / $0.40 per million tokens |
| Users | 900M weekly active; 50M paying individuals; 1M paying businesses |
| 2024 | 2025 | |
|---|---|---|
| Revenue | 3.7 | 13 |
| R&D (includes training compute) | 7.8 | 19 |
| Sales and marketing | 1.1 | 5.7 |
| Inference compute | not split | 8.4 |
| Gross margin | ~40% | ~33% |
| Revenue share to Microsoft | 20% of revenue | 20% of revenue (capped at $38B from 2026) |
| Net loss | 5 | 38 (about 41 of it a one-time non-cash restructuring charge) |
Sources: NVIDIA and AMD figures from SEC 10-K and 8-K filings, rounded; unit prices and volumes are analyst estimates marked (est.); OpenAI figures from press reporting (CNBC, The Information, Bloomberg, FT-verified leaked statements) and are directional.
Seven blocks from the deck, walked as five whiteboard steps. The step number on each card is the order we draw it.
Mapping growth options: existing or new products, into existing or new markets.
Framework for strategic marketing: five steps from analysis to measurement.