In March 2000, Cisco Systems was the most valuable company on earth. It sat at a $275 billion market cap — number one in the top-20 US tech ranking — on the back of a thesis so clean it sounded irrefutable: "You can't build the internet without Cisco." The company sold the routers and switches that every enterprise, every carrier, every data centre had to buy. It was the picks-and-shovels play on the biggest infrastructure buildout in a generation.
Twenty-five years later, Cisco is worth $300 billion. Nominally above its 2000 peak. In real terms, after inflation, it has delivered a negative return across a quarter century. It drifted from #1 to #13. The infrastructure king became a middle-of-the-pack software vendor.
NVIDIA is now worth $4,500 billion. Number one on the same ranking. The thesis: "You can't train AI without NVIDIA." Infrastructure play. Picks and shovels. Biggest buildout in a generation.
The question is whether history rhymes.
The bump chart above shows every company's year-end rank in the top-20 US tech market cap list from 2000 to 2025. What it shows is unambiguous: the #1 infrastructure king has never stayed #1 across a full technology cycle. Not once in 25 years of data.
Cisco fell from #1 in 2000 to outside the top 5 by 2003, and has drifted between #8 and #16 since. Intel, #3 in 2000, fell out of the top-20 entirely by 2024. The only company in the top 5 in both 2000 and 2025 is Microsoft — and it survived by becoming a fundamentally different company.
| Company | 2000 Rank | 2000 Cap | 2025 Rank | 2025 Cap | Arc |
|---|---|---|---|---|---|
| Cisco | #1 | $275B | #13 | $300B | Drifted |
| Microsoft | #2 | $231B | #4 | $3,600B | Survived & grew |
| Intel | #3 | $202B | Out of top 20 | — | Collapsed |
| Oracle | #4 | $160B | #9 | $680B | Held (cloud pivot) |
| IBM | #5 | $150B | #14 | $260B | Drifted |
| NVIDIA | Not ranked | — | #1 | $4,500B | 12.5× since ChatGPT (3 yrs) |
The structural similarities between Cisco in 2000 and NVIDIA in 2025 are not superficial. They are five deep.
Both sat at #1 on a once-in-a-generation capex supercycle. The dot-com internet buildout drove enterprise and carrier spending on Cisco hardware at rates that were never going to last. The AI compute buildout is driving hyperscaler spending on NVIDIA GPUs at rates that analysts widely acknowledge will normalize.
Both had the "indispensable" thesis. Not "nice to have" — enterprises HAD to buy Cisco to connect their networks. They HAD to buy NVIDIA GPUs to train competitive AI models.
Both grew faster than underlying demand warranted. Cisco's revenues grew 50%+ annually through the late 1990s, fuelled by the expectation that internet traffic would require ever-more infrastructure. That expectation was correct — but it pulled forward years of future demand into two years of explosive buying.
Both faced the moment when the buildout normalizes. For Cisco, it was 2001. For NVIDIA, the question is whether it comes in 2026, 2027, or 2028.
The market applied the same multiple. At peak, Cisco traded at ~130× trailing earnings. NVIDIA has traded above 70× for the last two years. Both reflect the market pricing in an indefinite continuation of an extraordinary run.
The case that NVIDIA avoids the Cisco scenario rests on four genuine structural differences. Not spin. Not investor relations talking points. Four things that are actually different in the data.
Cisco's moat was hardware plus proprietary protocols. Both were eroded from below: Huawei commoditized the hardware, white-box switching commoditized the protocol layer, and open networking standards finished the job. NVIDIA's moat is CUDA — a 20-year software ecosystem where every major machine learning framework (PyTorch, TensorFlow, JAX) defaults to CUDA and has been optimized for it by thousands of contributor-years. You can build a cheaper GPU. You cannot easily rebuild the software ecosystem that surrounds CUDA. Software moats are harder to undercut than hardware moats.
Cisco's demand was largely one-time: enterprises bought routers when they built their networks. NVIDIA's demand has two components. Training — like Cisco's buildout — will normalize as frontier model development stabilizes. But inference is different: every company running AI in production generates ongoing, compounding GPU demand that grows with usage. The more AI enters production, the more GPUs are needed. That's a structural floor Cisco never had.
Cisco was threatened from below — commodity hardware that was good enough, sold for less. That kind of disruption is total: once commodity is good enough, the premium disappears fast. NVIDIA faces pressure from above — Google's TPU, AWS Trainium, Microsoft Maia, custom silicon from every hyperscaler. Custom silicon erodes at the margin rather than replacing NVIDIA wholesale. Hyperscalers will run their own silicon for workloads where it's optimized; they'll still buy NVIDIA for frontier training and for workloads where CUDA compatibility matters. It's market share erosion, not annihilation.
NVIDIA at $4,500B is 16× larger than Cisco was at peak. For NVIDIA to drift to Cisco's current position (#13, ~$300B), it would need to lose $4.2 trillion in market cap — a destruction of value with no historical precedent in absolute dollar terms. The percentage fall required (~93%) is roughly consistent with what Cisco experienced from its 2000 peak. But the absolute scale, the index weight, and the institutional ownership all create structural resistance that Cisco in 2000 did not have.
Intel's trajectory from the same 25-year dataset is the most relevant cautionary case. Intel was #3 in 2000 at $202 billion. By 2024 it had fallen out of the top 20 entirely.
The optimist NVIDIA case is not "stays #1 forever." It's "becomes the Microsoft of AI infrastructure" — the company that is genuinely indispensable to the underlying platform, not just for one cycle, but by continuously adapting. NVIDIA is moving in this direction: up the stack into networking (Mellanox), software (NeMo, Triton, cuDNN), and sovereign AI infrastructure partnerships. Whether the execution matches the ambition is the open question.
| Cisco (2000) | NVIDIA (2025) | |
|---|---|---|
| Rank | #1 | #1 |
| Market cap | $275B | $4,500B |
| Thesis | Picks and shovels for the internet | Picks and shovels for AI |
| Moat | Networking hardware, proprietary protocols | CUDA: 20-year software ecosystem |
| Peak narrative | "Can't build the internet without Cisco" | "Can't train AI without NVIDIA" |
| Supercycle | Dot-com internet buildout | AI compute buildout |
| Primary threat | Commodity hardware (Huawei, white-box) | Custom silicon from above (Google TPU, AWS Trainium, Microsoft Maia) |
| Software moat | None | CUDA ecosystem — 20 years deep |
| Persistent demand layer | No — one-time buildout | Yes — AI inference at scale |
| 25-year outcome | #13, $300B, near-zero real return | TBD |
The data points in one direction: the #1 infrastructure king at the peak of a capex supercycle has never retained its position across a full technology cycle. Zero exceptions in 25 years. The Cisco pattern is real.
But the magnitude and the timeline are different. The pessimist case is not "NVIDIA goes to zero" — it is "NVIDIA drifts from #1 to #4–8 over a decade as the AI capex supercycle normalizes, hyperscaler custom silicon captures 30–40% of the training market, and CUDA's moat proves insufficient to defend the entire business." That outcome would still be consistent with NVIDIA at $1–2 trillion in 2035. A halving from $4,500B as AI infrastructure spending normalizes is entirely consistent with what the data shows happening to every infrastructure king.
The Cisco scenario — drifting to #10–15 for 25 years — requires CUDA to fail as a moat. Possible, but it requires a specific sequence: a viable CUDA alternative gaining widespread adoption, then hardware that outperforms NVIDIA GPUs, then hyperscaler custom silicon expanding beyond training into general inference. None of that is impossible. None of it is imminent.
The most likely path: higher floor than Cisco, longer timeline, different threat vector. Not the Microsoft survival story. Not the Intel collapse story. Something in between — probably #3–6 in ten years, at a market cap that looks disappointing from $4,500B but would represent a defensible infrastructure franchise. The data says that's the best-case realistic scenario for any company that has ever held this position. Whether CUDA changes the rules is the question the next five years will answer.