NVDA — NVIDIA Corporation

Business summary

NVIDIA designs GPUs and the system architecture (NVLink, NVSwitch, Spectrum-X networking, Quantum InfiniBand) that surrounds them. Data Center is now ~85% of revenue, dwarfing Gaming and Pro Visualization. The company's competitive position rests on four reinforcing layers: (1) silicon — Blackwell (B200/B300) shipping and Rubin in development; (2) systems — racks and clusters (NVL72) sold complete; (3) networking — InfiniBand + Spectrum Ethernet; (4) software — CUDA, cuDNN, TensorRT-LLM, NIM microservices, Omniverse. The software moat is the durable one; the silicon lead resets every generation.

The key strategic question for the next 18-24 months is custom-silicon substitution: how aggressively will Google (TPU on Broadcom), AWS (Trainium), Meta (MTIA), and Microsoft (Maia) displace merchant NVIDIA accelerators? The consensus answer is 'meaningfully but not dramatically' — hyperscalers will pursue dual-sourcing for inference and select training workloads, while keeping NVIDIA as the primary for cutting-edge training. The risk is that custom silicon catches up faster than expected.

Connected theses

T002 — The semi capex cycle is being underwritten by hyperscalers, not consumers — and equipment is the cleanest expression · core
Participation in the AI leader, sized to allow upside without overcommitting given custom-silicon erosion risk.

Key metrics

Data Center revenue (FY26 est)	$160-180B Up from $115B FY25. Implied 40-55% YoY growth.	Consensus
Gross margin (non-GAAP)	~74-75% Industry-leading; sustainability depends on competitive intensity + product mix (Blackwell vs Hopper).	Recent quarters
Hyperscaler capex pacing	$600-700B aggregate 2026 Roughly 40-50% flows through NVDA in some form.	Goldman + Citi est

Valuation snapshot

Price	$222.32
Market cap	$5,400B
Forward P/E	32.0×
EV / EBITDA	24.0×
FCF yield	2.6%

Priced as the durable leader. Multiple has compressed from FY24-25 peaks but still well above hardware-vendor norms. Valuation supportable if Data Center growth holds >30% YoY through 2027; vulnerable if growth decelerates faster.

Evidence

2026-Q1 secondary Blackwell ramp + Rubin commentary
Blackwell shipping at scale into 2026; Rubin announced for 2027 with continued annual cadence.
2026-Q1 secondary TSMC capex raised to $52-56B
TSMC FY26 capex raised — implies strong NVDA + AVGO order book.
2026-Q1 secondary Custom-silicon substitution accelerating at hyperscalers
Google TPU + AWS Trainium + Meta MTIA + MSFT Maia all ramping. Custom share of AI compute growing.

Catalysts

2026-08 FY27 Q2 earnings (calendar Q2 2026 results) high
What to watch: Data center revenue YoY, Rubin commentary, China revenue, gross margin trend
2026-Q3 Annual GTC announcements / product cadence updates medium
What to watch: Rubin specs, networking strategy

Falsifiers

Hyperscaler aggregate capex guidance cut >15%
armed · Quarterly hyperscaler earnings
A hyperscaler announces second-source ASIC as primary AI accelerator
armed · Hyperscaler earnings + AI event announcements
Material gross-margin compression (sub-70% non-GAAP) for 2 quarters
armed · 10-Q margin disclosures
Export controls expand materially beyond current China scope
armed · BIS rule changes

Agent notes

Hold. NVDA is the leader but not the value. The position is sized to participate without overcommit; AVGO is the matched hedge. The bear case (custom-silicon eating share) is real but on a longer timeline than the bear narrative implies. Watch hyperscaler capex commentary as the leading indicator — capex direction is the read-through.

Educational notes

📚 merchant GPU vs custom ASIC

A merchant GPU (NVIDIA H200, B200) is a general-purpose chip you can buy off-the-shelf to run any AI workload. A custom ASIC (Google TPU, AWS Trainium, Meta MTIA) is a chip designed by a hyperscaler — often using Broadcom or Marvell to do the silicon engineering — for their specific workloads. ASICs trade flexibility for efficiency: a Trainium can run AWS's specific training stack 1.3-2× more efficiently than a comparable NVIDIA GPU, but it can't run anyone else's framework. The current equilibrium: hyperscalers use NVIDIA for new training runs (where flexibility matters) and custom ASICs for stable, well-understood inference workloads. The question is how fast that boundary moves.

Open questions

What's the NVDA share of inference workloads vs training? Inference is the larger long-term TAM.
How does Rubin pricing compare to Blackwell on $/TFLOP?
What's the realistic timeline for custom-silicon parity at the leading edge (training)?