Give me - from the perspecitve of a deep PE / hedge fund investor - what is the bullet point of analysis they would do to analyse Nvidia? No analysis needed. But priortity should be descending. Do web search to find relevance of some new things that come up.
Based on current 2026 market data and the perspective of a deep PE/hedge fund investor, here is the priority-ordered diligence bullet list for Nvidia:
1. Hyper-scaler AI capex cycle durability (the single biggest question) - CreditSights now estimates ~$750B of 2026 capex across the top five hyperscalers — up ~67% YoY, a third straight year above 60%. - AI infrastructure is ~$450B+ of that. NVDA's revenue is effectively a derivative of this. Key question: is this a one-time build-out or a sustained multi-year investment curve? - Watch capex intensity (Oracle 86%, Meta 54%, Microsoft 47%) — can these persist without AI monetization proving out?
2. The custom ASIC substitution threat (the structural bear case) - Hyperscaler in-house silicon (Google TPU, AWS Trainium, Microsoft Maia, Meta MTIA) is the fastest-growing segment. - ASIC server shipments projected at ~27.8% of the market in 2026, growing ~45% YoY vs. ~16% for merchant GPUs. - Economic trigger is real: Midjourney's switch from NVDA to Google TPU cut compute cost 65%. - Broadcom is the central enabler, guiding ~$56B AI revenue FY26 — the key long on the other side of this trade.
3. CUDA moat — how wide, how durable, and where it doesn't hold - 15+ year ecosystem, 4M+ developers, all frameworks (PyTorch/JAX/TensorFlow) optimized CUDA-first. - Training share still >90% — hard to substitute. Inference share is only 60–75%, where cost-per-token beats ecosystem fit. - Diligence must separate training (defensible) from inference (contested).
4. Market share trajectory vs. TAM growth (narrative vs. math) - NVDA ~81–87% of merchant AI data center chip share in 2026 (dependent on denominator). - Share is structurally declining (92% → low-80s) but absolute revenue is exploding ($47.5B FY24 → $75.2B in a single quarter FY27). - The correct frame: can TAM growth outpace share loss and margin compression simultaneously?
5. China exposure — a $5–13% revenue tailwind that has gone to zero - China was once ~13% of revenue; Bernstein pegs NVDA at ~40% share there in 2025, down from ~95%. - Despite US approvals and Beijing nods for H200 (March/May 2026), not a single H200 delivery has been made — zero revenue realized. - Huawei ascends; Chinese chipmakers recapitalized via IPOs (Moore Threads, MiniMax). This is a permanently impaired, possibly terminal, revenue line — and a rising long-term competitive threat.
6. Supply chain — gross margin and revenue are gated by HBM + CoWoS, not demand - HBM3E is now the binding constraint (sold out for 2026; SK Hynix ~50% of NVDA supply, ~2/3 of HBM4). - NVDA controls ~60% of TSMC CoWoS capacity — a moat for NVDA and a ceiling for competitors. - As an investor, understand: revenue prints only as fast as three Korean fabs can bond DRAM. Bull case must price the HBM ramp; bear case can short HBM providers directly (MU, Samsung).
7. Gross margin durability (~85%) in a share-shifting market - NVDA's 79–88% gross margin funds R&D and TSMC priority; AMD is 65–68%, Intel ~58%. - Question: as inference shifts to cheaper ASICs/AMD at 30–40% price discounts, does NVDA hold pricing power on training/sovereign/enterprise, or does blended margin compress?
8. Customer concentration risk - A handful of hyperscalers (Microsoft, Meta, Alphabet, Amazon, Oracle) account for the dominant share of data center revenue. - Each is simultaneously NVDA's largest customer and its largest future competitor (via custom silicon). This is an unusual and fragile dynamic.
9. Blackwell → Rubin product cycle execution - Q1 FY27 data center revenue of $75.2B confirms Blackwell ramp is working. - Rubin (R100) begins CoWoS-L allocation in 2026; the generational transition is the next execution test.
10. Valuation / multiple compression math - At these scale levels and a 75%+ market share ceiling, the key is whether earnings growth compounds fast enough to absorb multiple compression as the growth story matures from hyper-growth to secular staple.
Overarching thesis tension: NVDA is simultaneously the most dominant franchise in semiconductors (CUDA + supply-chain control + pricing power) and structurally over-earning on a customer base that is actively and rationally trying to defund it. The deep-value PE/hf framework reduces to one question — what is the normalized, steady-state earnings power once the build-out phase of AI ends and custom silicon reaches scale?
Want me to pull the actual financials (revenue, gross margin, FCF, customer concentration) and build the normalized-earnings math against these risks?
Are each of these 1 slide?
No — the way I structured it is a 10-point checklist on a single consolidated slide. Each bullet is one item on one slide, not ten separate slides.
If you want it split out, here's how it could map:
Option A — Single slide: one dense slide with all 10 priorities in descending order (what I gave you).
Option B — Ten slides: one slide per priority, each with its own title, data point, and the key investor question.
Option C — Three to four theme slides: 1. Demand/commercial (hyper-scaler capex, custom ASIC threat, market share vs. TAM, China) 2. Moat/competitive position (CUDA training vs. inference, gross margin durability, customer concentration) 3. Supply chain (HBM + CoWoS gating, Blackwell→Rubin execution) 4. Valuation (multiple compression, normalized earnings power)
Which format do you want — and if multiple slides, do you want the full analysis fleshed out per slide or just the prioritized headlines?