Last price
$201.68
The beneficiary under threat — biggest Sequoia tailwind, sharpest structural headwind.
Live quote sourced from Yahoo Finance. Prices cited in narrative below reflect the original memo date and may be stale.
Every Sequoia-cited autopilot startup trains and serves on NVIDIA, directly or via a hyperscaler. The services-as-software wave multiplies inference demand — autopilots run constantly, not just when a human prompts them.
NVDA is simultaneously the single biggest financial beneficiary of the Sequoia thesis AND the single most structurally threatened name in the compute stack — by a disruption that is already underway, not theoretical. Your read on this stock turns on which of those two forces wins over the next 18-24 months.
Every Sequoia-cited autopilot — Crosby, Harvey, Harper, Rillet, Anterior — trains and serves on NVIDIA. Autopilot inference is structurally different from training: continuous, 24/7, per-user-interaction, not episodic. If autopilots capture services budgets at scale, compute demand compounds in perpetuity with adoption, not in bursts with training cycles. Training is lumpy; inference is a subscription.
| Workload | NVDA lock | Trend | Timeline |
|---|---|---|---|
| Foundation-model pretraining | Strong (CUDA + B200 + NVLink fabric) | Stable | Multi-year dominant |
| Large-scale fine-tuning | Strong | Eroding | 12–24 months |
| Frontier-model inference | Contested | Shifting to custom | Happening now |
| Mid-size model inference | Vulnerable | Custom wins | 6–12 months |
| Edge / on-device | Lost | Apple Silicon, Qualcomm | Structural |
| Robotics / physical AI | Strong | Stable | Multi-year |
Even losing 30% of inference share, NVIDIA's 70% of a market growing 5–10× dwarfs the current base. $500B of Blackwell + Rubin visibility is already booked.
4M+ developers, 500+ libraries, 15 years of compiled software. For a services-as-software startup shipping a fine-tuned model this quarter, the engineering risk of moving off CUDA is months — NVIDIA wins by default every time product velocity matters, which is always.
That is not a decelerating business. Skeptics have to model from that base.
Pre-optimized model microservices let an enterprise deploy production inference without an ML team. This is the picks-and-shovels vendor moving up-stack toward the outcome layer Sequoia cares about — selling the shovel plus the excavator plus the operator.
UAE, Saudi, France, UK, India, Japan, Korea are each building domestic training capacity for political-neutrality reasons. These buyers do not price-shop against Trainium. They buy NVIDIA at list.
Meta is no longer experimenting. It is pre-committed to a custom-silicon roadmap through 2030. Billions of inference dollars walking away from merchant GPU in a way that cannot be un-done.
Inference is less FLOP-bound, less interconnect-bound, more cost/watt-bound — exactly where custom ASICs win. The very wave NVIDIA most benefits from volume-wise is the one it most loses share-wise.
Top 4 hyperscalers ≈ 45% of data-center revenue. Those same 4 are the 4 biggest competitive threats. No analogous situation exists in semis history.
Forward P/E ~30× on FY27 estimates. Any deceleration in hyperscaler capex (watch Meta's Q2 commentary especially) triggers multiple compression on top of any fundamental miss. Meta's January 2026 ‘year of efficiency 2.0’ rhetoric already caused a 12% intra-month NVDA drawdown.
China export controls remain restrictive. EU AI Act is creating uncertainty on EU deployments. Novel (not yet actioned but worth tracking): White House review of whether AI-compute concentration is itself an antitrust concern.
NVDA is not an autopilot — it's the runtime. The thesis question: does the compute pie grow faster than NVIDIA's share shrinks? Answer: almost certainly yes in aggregate, with higher volatility than bulls model. Every major autopilot launch is bullish for NVDA; every hyperscaler custom-silicon deployment is bearish. Expect both kinds of news weekly. Unlike a SaaS incumbent, NVDA has no ‘platform moat being eroded’ risk — CUDA is stronger than ever. Its weak spot is more specific: its biggest customers are building alternatives, and will deploy them exactly where the Sequoia thesis says the volume will be.
Tailwind is structural and company-agnostic; the question is multiple, not direction.