First-Principles Question
A chip that costs $3,100 to fabricate sells for $25,000–$40,000. Gross margin: 73%. Market share of AI-training GPUs: north of 90%. Porter's Five Forces says this can't last — abnormal profit should attract entrants until it's competed away. So why hasn't it?
That is Nvidia in fiscal 2024 — TSMC's fabrication cost for an H100 GPU was roughly $900–$3,100; Nvidia sold it for $25,000–$40,000, and it captured an estimated 98% of the data-center GPU market (Yoffie & Von Bargen, 2024). Five Forces would flag "low threat of entry" as temporary. It has no vocabulary for why AMD, Intel, Google, and three well-funded startups all tried to attack this position in 2023–24 and still couldn't dislodge it. That vocabulary gap is what this session fills.
What Five Forces sees
- Five separate boxes: rivals, suppliers, buyers, entrants, substitutes.
- A static photograph of bargaining power at one point in time.
- An assumption that the industry's boundary is obvious and fixed.
What it misses
- Ecosystems: a "supplier" (TSMC) or a "buyer" (Microsoft, Amazon) can also be a co-investor, a complementor, or a future rival — all at once.
- Interdependence: a rival's best move depends on what you do, and vice versa. Five Forces has no room for a payoff matrix.
- Not knowing the game: sometimes the honest answer to "how will this play out" is a probability distribution, not a forecast.
Quick Recap: PESTEL vs. Porter's Five Forces
You met both lenses earlier in the module. Neither one, on its own, explains Nvidia. Keep them — but treat them as the floor, not the ceiling, of external analysis.
| Lens | Unit of Analysis | Can a Firm Change It? | Blind Spot This Session Fixes |
|---|---|---|---|
| PESTEL | The macro-environment (a nation, a society) | No — scan & adapt | Treats every factor as exogenous; can't model a rival's reaction to your move. |
| Five Forces | One bounded industry, one snapshot in time | Partially — position & defend | Assumes clean role separation (supplier / buyer / rival) and a knowable future. |
| Ecosystems + Game Theory + Uncertainty | A web of interdependent players, moves, and futures | Yes — shape, adapt, or hedge | This session. |
First-Principles Question
Why can't AMD just build a faster, cheaper chip and win?
By 2024 AMD's MI300X was priced at $10,000–$15,000 — a fraction of Nvidia's $25,000–$40,000 H100 — and was closing the raw performance gap. It still couldn't take meaningful share in AI training. The reason has almost nothing to do with the chip. It has to do with CUDA: a sixteen-year-old software platform that has quietly become Nvidia's real product.
A business ecosystem is not an "industry"
It is a network of firms — a platform owner plus its complementors — whose products or services only create full value together. Adner (2017) calls this the shift from thinking about "industry" to thinking about "ecosystem": who has to do what, in what order, for your value proposition to actually show up.
Platforms win on two-sided network effects
More developers writing CUDA code makes the platform more valuable to GPU buyers; more GPU buyers makes it worth more developer-hours to write CUDA code. Each side subsidizes the other's reason to join. This is precisely the dynamic Steve Jobs engineered when Apple reversed course in 2008 and opened the iPhone to third-party developers with a native SDK and the App Store — turning outside developers into complementors with every shipped app deepening the platform's lock-in.
The moat is switching cost, not performance
By 2024, Nvidia had shipped over 150 CUDA-based SDKs spanning gaming, robotics, life sciences, and quantum computing (Huang, quoted in the case). Every model a lab has already trained, tuned, and deployed on CUDA is a rewrite cost if it moves to AMD's ROCm or a custom chip. That rewrite cost — not the price tag on the chip — is the real barrier to entry.
The Ecosystem Bet Almost Didn't Survive
Jensen Huang, Chris Malachowski, and Curtis Priem founded Nvidia in 1993 — reportedly plotting the company over Grand Slam breakfasts at a Denny's. Huang introduced CUDA in 2006 as a way to let developers program GPUs in something close to C++. It was not an instant hit: downloads declined from 2010–2013, Nvidia's stock sagged, and the board began worrying about corporate raiders (Yoffie & Von Bargen, 2024).
What saved CUDA wasn't a plan — it was a bet made under genuine uncertainty. When an Nvidia researcher showed Huang why deep learning would need parallel computing, Huang committed the company overnight. VP Greg Estes recalled: "He sent out an e-mail on Friday evening saying everything is going to deep learning... By Monday morning, we were an A.I. company." Nvidia had almost no AI revenue at the time.
The bet sat idle for years. Crypto mining kept Nvidia's GPUs selling through 2017–18, but deep learning adoption was slow. Then, in November 2022, ChatGPT launched. OpenAI disclosed that training GPT-4 required roughly 25,000 Nvidia GPUs and $100 million in compute. Every serious AI lab needed CUDA-compatible hardware at once — and the ecosystem investment of the previous sixteen years paid off in a single fiscal year.
Nvidia's data-center revenue tells the story better than any paragraph can (right).
Nvidia data-center revenue, FY2020–FY2024 ($M). A 217% jump in FY2024 alone — the CUDA ecosystem bet, plus a decade of R&D no rival could compress into a single product cycle. Source: Yoffie & Von Bargen (2024), Exhibit 1a.
2026 Update: The Moat Under Live Fire
- Nvidia reportedly crossed a $5 trillion market capitalization in October 2025 — the first company ever to do so — on the strength of continued AI data-center demand.
- Nvidia announced a strategic partnership to invest up to $100 billion in OpenAI (September 2025) to deploy on the order of 10 gigawatts of Nvidia systems — a customer relationship that is now also, in part, an equity relationship.
- The U.S. government's export-control regime kept shifting: an "AI diffusion" rule, then a 2025 arrangement reportedly letting Nvidia resume some H20 chip sales to China in exchange for a revenue share to the U.S. Treasury on the order of 15%.
- OpenAI — Nvidia's own showcase customer — simultaneously signed a custom AI-accelerator deal with Broadcom in 2025, a hedge that echoes "reserving the right to play" (see the Uncertainty section below).
- Open compilers such as OpenAI's Triton and AMD's ROCm continued to mature, lowering the cost of writing code that runs on more than one vendor's chips — precisely the "multi-homing" threat modeled in the calculator below.
- The Department of Justice's 2024 antitrust inquiry into Nvidia's GPU allocation practices remained an open overhang into 2026, alongside continued EU scrutiny of AI-compute concentration.
Figures reported as of mid-2026 press coverage; treat exact dollar amounts as directional, not audited.
New Tool · Ecosystem Moat Calculator
Rate a platform on five dimensions. Defaults reflect CUDA circa 2023–24. Try nudging dimensions toward "Medium" to model 2026's contested reality.
Rivals must out-invest for a decade, not a product cycle.
First-Principles Question
If you were Jensen Huang, would you auction your scarce GPUs to the highest bidder?
In 2024 Nvidia could not make enough H100 GPUs to satisfy demand. Larry Ellison, Oracle's chairman, described a dinner where he and Elon Musk were "begging Jensen for GPUs... Please take our money; no, take more of it." An auction would maximize short-run price. But Huang also had to keep Microsoft, Meta, and Amazon — his four largest customers, worth nearly half his revenue — from accelerating their own in-house chip programs out of frustration, all while a Department of Justice antitrust probe watched how "fairly" he allocated supply. Every option changes what his rivals and customers do next. That interdependence is exactly what game theory is built to analyze, and Five Forces is not.
A move that is your best response no matter what your rival does.
A pair of moves where neither player can do better by unilaterally switching — even if a different pair would make both better off.
A move that is costly to reverse, so rivals believe you'll follow through. TSMC demanding customers pre-pay billions to reserve fab capacity is exactly this — it makes capacity commitments observable and binding.
Coopetition: when your rival is your customer
Microsoft and Meta each bought an estimated 150,000 Nvidia H100 GPUs in 2023 (Exhibit 7 of the case) — while simultaneously funding in-house silicon (Microsoft's Maia, Meta's MTIA) to reduce reliance on Nvidia. Amazon's VP put the logic bluntly: custom chips could be "half as expensive as running that same model with Nvidia." Every hyperscaler is playing both sides of the board at once.
The capacity game is not new
Courtney, Kirkland & Viguerie (1997) describe this exact structure in "oligopoly markets, such as those for pulp and paper, chemicals, and basic raw materials": whether to expand capacity depends entirely on whether your rival expands too. Swap "pulp mill" for "TSMC wafer allocation" and you have Nvidia, AMD, and every fabless chip designer in 2025–26.
You run chip design for a fabless AI-accelerator company. Each year you and one rival must independently decide whether to Expand next year's TSMC wafer reservation or Hold it flat. Reservations must be booked (and largely pre-paid) before you see the rival's move. Payoffs are an illustrative profit index — higher is better for you.
Round 1 of 3: The Opening Move
It's 2025. Both you and your rival must commit to next year's wafer reservation, sight unseen. What do you choose?
First-Principles Question
What if the honest answer to "what will happen" is: nobody knows — not even Nvidia?
Courtney, Kirkland & Viguerie (1997), writing in Harvard Business Review, argue that most strategists make a binary error: they treat every situation as either perfectly forecastable or completely random. In truth, uncertainty comes in four levels, and the right analytic tools — and the right strategic posture — change at every level.
A Clear-Enough Future
One forecast is precise enough to act on. Use the standard toolkit: market research, Five Forces, discounted cash flow.
Alternate Futures
A handful of discrete scenarios (a regulator rules one way or the other). Use decision analysis and game theory.
A Range of Futures
A continuum with no natural scenarios (demand could be 10%–30% penetration). Use latent-demand research and scenario planning.
True Ambiguity
Even the range of outcomes is unknowable. Use analogies, pattern recognition — and stay humble.
Three Postures, Mapped Onto the Nvidia Case
Huang's 2013 deep-learning bet. With almost no AI revenue on the books, Huang redirected the entire company toward deep learning on a hunch backed by one internal researcher's pitch — a level-4, "true ambiguity" decision that tried to bend the market rather than predict it. This is what Courtney et al. call shaping: playing a leadership role in establishing how the industry will operate.
AMD's pricing under Lisa Su. AMD did not try to out-shape Nvidia's ecosystem. It positioned the MI300X at roughly 40–60% of Nvidia's price within the structure Nvidia already built — a classic adapter strategy: win through speed and value inside an existing market, not by rewriting its rules.
Hyperscalers buying Nvidia GPUs while building Trainium, Maia, and TPUs. Microsoft, Amazon, and Google keep purchasing Nvidia's best chips (a no-regrets move — they need the performance today) while funding in-house silicon as a hedge (an option) against the day CUDA's moat narrows. Huang's own hedge is symmetric: "In the event that we have to shift from one fab to another, we have the ability to do it," he said of Nvidia's fallback relationship with Intel's foundry business.
Self-Check: Which Level of Uncertainty Is This?
Click a scenario to reveal the level Courtney, Kirkland & Viguerie's framework would assign — and why.