Is NVIDIA Still Worth It Now That Its Biggest Customers Build Their Own Chips?
NVIDIA's data center revenue run-rate exceeds $140 billion annually. The custom silicon threat from Google, Amazon, and Microsoft is real. Here is how to think about both simultaneously.
The Numbers That Make This Story Real
NVIDIA's data center revenue grew from $3.6 billion in fiscal year 2022 to a quarterly run-rate above $35 billion by early 2026, an annualized pace exceeding $140 billion. For context, the entire semiconductor industry generated roughly $527 billion in revenue in 2023. NVIDIA reported fiscal Q4 2026 revenue of $39.3 billion, up 78% year-over-year, with data center comprising 88% of that. Gross margins came in at 73.5%, not software margins, but close. The hyperscalers, Microsoft, Google, Amazon, and Meta, collectively committed over $300 billion in capex for AI infrastructure in 2025 and 2026 combined. Most of that compute is NVIDIA.
The CUDA software ecosystem is NVIDIA's real moat, not the hardware. Hundreds of thousands of AI models, research papers, and production deployments have been built on CUDA over 15 years. Switching to a different compute platform requires rewriting that code, a switching cost measured in years and billions of dollars across the industry.
Why This Looks Different From the Fiber Optic Bubble
The most common bear argument is historical analogy: the AI infrastructure buildout mirrors the fiber optic overbuild of 1999-2000, which ended with massive overcapacity and a decade of depressed returns. The analogy is worth taking seriously, but there is a critical structural difference. The telecom overbuild created capacity that was available but not yet used. The AI capex cycle is building infrastructure to meet demand that already exists. Every GPU deployed by Google for Gemini inference, every Microsoft Azure GPU cluster running Copilot, every Meta recommendation system is generating revenue from day one. The capacity utilization dynamic is fundamentally different from the fiber overbuild.
The Critical Caveat: Custom Silicon
The bear case on NVIDIA is not about demand, demand for AI compute is not in question. It is about who captures the economics of satisfying that demand long-term. Google's TPU v5, Amazon's Trainium2, Microsoft's Maia 100, and Meta's MTIA chip are all custom AI accelerators designed to reduce NVIDIA dependence for specific workloads. At hyperscale volumes, custom silicon can offer dramatically better cost-per-inference for specific model architectures. As the AI compute mix shifts from training-heavy to inference-heavy, the hyperscalers' custom silicon investments become more valuable relative to NVIDIA. The market will likely bifurcate: NVIDIA dominates training, custom silicon captures an increasing share of commodity inference at hyperscale.
Valuation
NVIDIA trades at roughly 35x forward earnings, extraordinary in absolute terms but defensible against a company growing revenue at 70-80% annually with 73% gross margins and a software ecosystem moat that has no realistic near-term replacement. The relevant comparison is what multiples the market historically assigned to companies that genuinely dominated transformational technology cycles: Microsoft during the Windows/Office era, Cisco during the early internet buildout, Intel during the PC manufacturing ramp. Each sustained elevated multiples for years while delivering earnings growth that ultimately justified the premium. On a 5-year view that captures the CUDA moat and the training compute cycle, the answer is probably yes. On a 1-2 year view where custom silicon share gains are the marginal catalyst, there is real downside risk to current expectations.