P1-C3 · Why NVDA Is Not Intel¶
One Core Sentence
The biggest historical miss wasn't a technology miss, it was a strategic miss. Jensen bet on CUDA in 2006, while Intel spent those 20 years protecting x86 profits.
AI Industry Knowledge — History → Technology → Supply Chain → Business → Application → Geopolitics
P1-C3 (Part 1, Chapter 3). After this chapter, you'll be able to explain that NVDA isn't "suddenly lucky" — it's a 20-year strategic play; Intel/AMD's miss wasn't a technology miss, it was a difference in strategic mindset.
1. The Question: Intel Was the Chip King in the 90s/2000s — Why Not in the AI Era?¶
Market cap in 2005: - Intel (INTC): $130B (CPU king, 80% data center share) - NVDA: $5B (gaming GPUs for consumers)
Market cap in 2026: - NVDA: $3,300B (~660x) - Intel (INTC): $80B (~0.6x, down, not up)
A 4,000x gap. This isn't luck — Intel had more money, engineers, and customers at the time. It lost on strategy, not technology.
💡 Tip: Click any ticker (e.g.,
NVDA/INTC) to enter its Multi-Source Profile — includes supply chain coordinates, upstream/downstream, and key data.
In C1, you already learned NVDA is the historical winner. C3 shows you exactly what Jensen did over 20 years, and the 4 steps Intel missed.
2. The Solution: Platform Strategy vs. Profit Protection — Two Corporate Mindsets¶
| Dimension | NVDA (Jensen's Platform Strategy) | Intel (Otellini / Krzanich's Profit Protection) |
|---|---|---|
| Long-term vs. Short-term | Sacrificed gaming GPU margins for data center long-term | Maintained x86 profits, didn't disrupt existing customers |
| Ecosystem vs. Product | CUDA given free to developers (locked in ecosystem for 20 years) | Sold CPUs, didn't invest in developer tools |
| Cross-Boundary vs. Defensive | 5 lines: data center / autonomous driving / robotics / AI / gaming | Defended CPU + server CPU, missed mobile / GPU / AI |
| Customer Education | Jensen personally evangelized GPU computing for 10 years | Intel waited for customers to buy CPUs |
| CEO Tenure | Huang founded in 1993, still CEO 33 years later — long-term view | Intel had 4 CEOs in 8 years, short-term quarterly focus |
Core: NVDA bet on "whatever works, we own the entire stack." Intel bet on "as long as x86 works, we make money."
3. How It Works: NVDA's 4 Key Decisions + Intel's 4 Key Misses¶
3.1 NVDA Decision 1 (2006): CUDA Launch — Free Programming Tools for Developers¶
In 2006, NVDA launched CUDA. At the time, GPUs were mainly used for game rendering. CUDA let developers use GPUs for general-purpose computing (scientific computing, image processing).
Short-term: No one used it. No revenue. Huge cost for the CUDA team. Long-term: Universities worldwide started teaching CUDA. Physicists, chemists, and biologists used GPUs for simulations. 6 years later, in 2012, the AlexNet team used NVDA GPUs + CUDA — because there was no other choice.
→ Platform lock-in isn't a technology gap; it's a 20-year time gap. AMD's ROCm (CUDA competitor) didn't launch until 2016, 10 years late, and still can't catch up.
3.2 NVDA Decision 2 (2014): All-In on Data Center¶
In 2014, NVDA's data center revenue was < $200M (vs. gaming at $4B). Jensen decided to go all-in on data center: - Cut gaming GPU margins (G80 → Maxwell made gaming cards cheaper) - Shifted most engineering teams to data center - Pushed Tesla / Volta / Hopper / Blackwell
Result: By 2026, data center revenue was $35B+/quarter, 8x the gaming business. Gaming went from core to side line.
→ Only a long-tenured CEO can make this kind of decision. A short-tenured CEO won't cut an existing cash cow.
3.3 NVDA Decision 3 (2019): Acquiring Mellanox — Buying InfiniBand¶
NVDA spent $7B to buy Mellanox. At the time, people thought it was "defensive (locking down the NVLink ecosystem)." In reality, it was entering the networking layer.
Result: AI training clusters have 1000+ GPUs; their interconnects rely on InfiniBand (Mellanox is the king). Now NVDA sells not just GPUs, but the entire data center fabric (GPU + DPU + networking). This raised the moat against hyperscaler custom ASICs — even if you build your own ASIC, you still need NVDA's network.
3.4 NVDA Decision 4 (2024+): Stargate Strategic Partnership¶
The $500B Stargate (OAI/Oracle/MSFT/SoftBank) uses NVDA chips. NVDA also invests in cloud providers (CRWV $36B stake, NBIS, CRWV customer guarantees).
→ NVDA doesn't just sell products; it shapes the market: investing in buyers so they have money to buy NVDA, creating circular financing.
3.5 Intel Miss 1 (2006-2010): No GPU Investment¶
Intel launched Larrabee (x86-based GPU) in 2009 — it failed and was canceled. For the next 12 years, Intel had no discrete GPU. It only released the ARC series in 2022, far too late.
3.6 Intel Miss 2 (2008): Mobile — Refused to Make Chips for iPhone¶
Otellini later said in an interview: "We calculated the ROI. iPhone volumes were too small, margins weren't high enough, so we declined." Apple turned to ARM (TSMC fab), and Intel lost the entire mobile era.
→ They misjudged ROI — not by looking at current volume, but by failing to see who would occupy the platform shift.
3.7 Intel Miss 3 (2014-2020): Data Center ML Acceleration¶
Intel acquired Nervana in 2016 ($350M, ML chip) and shut it down in 2019. It acquired Habana in 2019 ($2B, Gaudi), which still isn't competitive. Total of $10B+ spent on AI acquisitions, all with no follow-through.
→ In Intel's mindset, AI was an extension of the CPU. In NVDA's mindset, AI was a new platform. This determined who they bought and how they integrated it.
3.8 Intel Miss 4 (2020+): Foundry Lag¶
Intel brought back Pat Gelsinger as CEO in 2020, aiming to enter the foundry business (competing with TSMC). But Intel 18A (equivalent to 1.8nm) won't be in mass production until 2026, 1-2 years behind TSMC's 2nm. AMD / NVDA / AAPL all use TSMC, so Intel missed this wave.
4. vs. What You Already Know from C2¶
| Dimension | C2 Gave You | C3 Gives You More |
|---|---|---|
| Technology | Transformer + scaling laws made LLMs possible | Doesn't explain why a specific company occupies the position |
| Positioning | Not mentioned | NVDA's 4 decisions + Intel's 4 misses |
| Investment Significance | Knows LLMs might not face a winter this time | Knows the source of the moat — CUDA's 20-year ecosystem, not something that can be caught up in 1 year |
C2 = the technology reason. C3 = the company reason. Without C3, you won't know how deep the moat is, and you'll be easily scared off by noise like "AMD MI300 is catching up."
5. Try It: The Common Pattern in Intel's 4 Misses¶
Task: Summarize each of Intel's 4 misses (GPU / mobile / AI acquisitions / foundry) in one sentence, then think:
| Common Pattern | How to Use in Your Thesis |
|---|---|
| All were protecting existing profits + unwilling to cut the cash cow | Every incumbent behaves this way in a new paradigm |
| All were decisions by short-tenured CEOs | Use CEO tenure to predict long-term decision-making ability |
| All were looking at current ROI, not the platform shift | Look for companies willing to "lose now to win long-term" |
Self-check (if 3 items are true, proceed to P1-C4):
- You can use Intel's failure pattern to evaluate an incumbent (e.g., Cisco vs. new AI networking / IBM vs. cloud / Oracle vs. SaaS)
- You can explain "why AMD MI300 won't replace NVDA in the short term"
- You can name one company currently repeating Intel's miss pattern (your short / bearish candidate)
6. What's Next¶
You now know NVDA's position. But you still don't know how LLMs actually work internally — without that, your thesis is shaky (e.g., you can't answer "why NVDA's H200 sells for more than H100").
→ P1-C4 · Neural Networks / LLM Intuition uses 3 analogies to explain LLMs clearly, with no math.
7. Deep Dive (optional): AMD / Google TPU / Tesla Dojo / China ASCEND Competition¶
Click to see NVDA's 5 challengers' strength assessment
**AMD MI300X** — Launched 2023, good price-performance, but weak ROCm ecosystem. Got $5B+ orders from META / MSFT, but still a fraction of NVDA's $50B.
Google TPU — In-house developed for 10 years. v5e/v5p performance rivals H100, but only available on GCP — not sold externally, doesn't affect NVDA's external market.
Tesla Dojo — Musk pushed from 2024+. Custom chip + Dojo cluster. Primarily for internal use, not sold, limited impact.
AWS Trainium2 / Microsoft Maia / Google TPU — Hyperscaler custom ASICs. In the short term, they divert 20-30% of NVDA's internal workloads, but **the general market + agentic workloads still mainly use NVDA** (ecosystem decides).
**Huawei Ascend 910C / SMIC** — Domestic replacement in China. NVDA loses ~25% of China revenue, but other markets compensate. Long-term, the China market splits into two (domestic Huawei + external NVDA).
Key: NVDA's moat is CUDA ecosystem + Mellanox networking + hyperscaler strategic relationships, not a single chip. Competitors need 5-10 years for each of these 3 layers, and catching up on all simultaneously is nearly impossible.