NVIDIA (NVDA) — The Compute Tax on the AI Era
1. Who They Are
NVIDIA (NASDAQ: NVDA, CIK 0001045810) designs accelerated computing hardware and the software stack that runs on it. Founded 1993, fiscal year ends late January (FY27 = year ending ~Jan 2027). SIC 3674 (semiconductors). Santa Clara HQ. ~42,000 employees. CEO Jensen Huang since founding.
What started as a gaming GPU company is now the dominant supplier of compute to the AI buildout. Stock closed $202.81 on 2026-07-17, market cap ~$4.92T, 52-week range $164.07–$236.54, all-time closing high $235.47 on 2026-05-14.
2. How They Make Money
Three buckets, one of which is the entire story:
| Segment | FY26 Revenue | % of Total | YoY |
|---|---|---|---|
| Data Center | $193.7B | 89.7% | +68% |
| Gaming | ~$11B | ~5% | — |
| ProViz / Auto / OEM | ~$11B | ~5% | — |
| Total | $215.9B | 100% | +65% |
Q1 FY27 (quarter ended 2026-04-26, filed 2026-05-20) — verified directly from SEC XBRL 10-Q:
| Metric | Q1 FY27 | Margin |
|---|---|---|
| Revenue | $81.615B | — |
| Gross profit | $61.157B | 74.93% |
| Operating income | $53.536B | 65.6% |
| Net income | $58.321B | 71.5% |
| Diluted EPS | $2.39 | — |
| Operating cash flow | $50.344B | — |
Balance sheet (2026-04-26): Total assets $259.5B, equity $195.5B, cash $13.2B, AR $40.7B, PP&E net $12.4B, total liabilities $64.0B. Net cash is modest relative to the capex cycle — this is a capital-light royalty-on-silicon business funding its own growth.
Unit economics: Gross margin compressed in FY26 to 71.1% GAAP (from 75.0% in FY25) due to (a) mix shift from Hopper HGX cards to full Blackwell datacenter racks and (b) a $4.5B H20 inventory/purchase-obligation charge from China export controls. Q1 FY27 GM recovered to ~75% as Blackwell scaled. Net margin crossed 70% in Q1 — software-like margins on hardware.
Valuation derived from verified Q1 + web price:
| Multiple | Value | Basis |
|---|---|---|
| P/E TTM | ~31.1x | Yahoo EPS $6.53 TTM |
| Forward P/E (Q1 annualized) | ~21.2x | $2.39 × 4 = $9.56 run-rate EPS |
| P/S FY26 | 22.8x | $4.92T / $215.9B |
| P/S forward (Q1 annualized) | 15.1x | $4.92T / $326B |
The gap between TTM and forward is the Blackwell ramp — the market is pricing the run-rate, not the FY26 print.
3. What's Their Moat
CUDA. 18 years of compounding software lock-in. Every AI framework — PyTorch, JAX, TensorFlow, vLLM, Megatron — is written CUDA-first. The switching cost isn't the chip; it's the millions of CUDA developers and the libraries (cuDNN, TensorRT, NCCL, Triton) that make NVDA hardware the default compilation target.
Networking flywheel. Mellanox (acquired 2020) gives them InfiniBand. Spectrum-X Ethernet, NVLink, NVSwitch, BlueField DPUs — NVDA sells the full rack, not just the GPU. A hyperscaler buying Blackwell is buying an NVDA-defined system architecture. That's why GM held above 70% even through the H20 charge.
Annual cadence. Blackwell → Blackwell Ultra (GB300, H2 2025) → Rubin R100 (sampling Q4 2026, volume Q1 2027) → Rubin Ultra (H2 2027, 600kW Kyber racks) → Feynman (2028). A one-year architecture cycle is a moat — competitors are on 2–3 year cycles.
4. Why Now
- Blackwell is sold out through mid-2026. Backlog stretches into 2027. Demand visibility is unusual.
- Margin inflection just printed. Q1 FY27 GM 74.93%, net margin 71.5%. The H20 charge is behind them; Blackwell at scale is accretive. FY27 is set up to be the year margins re-expand toward 75%.
- Stock is 14% off the highs. $202.81 vs $236.54 52-week high. Prediction markets show only 7% conviction for a weekly close above $210 in July — sentiment has cooled into a setup where the next catalyst (Rubin sampling) is 1–2 quarters out.
5. Tailwinds
- Hyperscaler capex still climbing. Big 4 committed to multi-year $300B+ AI infra spend. NVDA is the tax collector.
- Sovereign AI. MBZUAI-scale deals proliferating. National GPU stacks as strategic infrastructure.
- Enterprise AI cycle just starting. Blackwell rack economics push inference downstream from cloud to enterprise.
- Annual roadmap = compounding switching costs. Every Rubin buyer is re-committing to CUDA.
- FCF machine. $50.3B operating cash flow in Q1 FY27 alone. Self-funding the capex cycle.
6. Headwinds
- Customer concentration. ~61% of revenue from 4 hyperscalers. Any one of them sneezing on capex cadence moves the stock.
- Custom silicon competition. Trainium (AMZN), TPU (GOOG), MTIA (META) — hyperscalers actively building alternatives. "Everyone is desperate to stop paying NVIDIA prices." Neither thesis resolved — volumes rising in parallel.
- China export controls. H20 charge was $4.5B. NVDA guiding Q1 with zero China data center revenue assumed. Binary risk — tightening loses a multi-billion-dollar TAM; loosening is pure upside.
- Geopolitical concentration. TSMC dependency. Taiwan scenario = supply chain extinction event.
- Valuation sensitivity. 31x TTM sounds cheap for the growth, but a single quarter of capex-pause narrative compresses the multiple fast. Beta 2.21.
7. When Should I Sell
Three concrete triggers that flip the call:
- Hyperscaler capex deceleration printed in guidance. Watch MSFT/AMZN/GOOG/META capex commentary — if 2027 guide flattens, the NVDA run-rate thesis breaks.
- Custom-silicon share disclosure. If a hyperscaler discloses Trainium/TPU at >15% of their AI compute share, NVDA's "platform not supplier" framing erodes.
- Gross margin breaks below 70% for two consecutive quarters without a clear one-off (H20-type) explanation. That's the signal pricing power or mix is structurally rolling over.
Secondary: Rubin sampling slippage past Q1 2027, or a Taiwan supply-chain event.
8. Why Buy
- Forward P/E ~21x on Q1 run-rate — for a company growing revenue 85% YoY with 75% gross margins and a CUDA moat, that's growth-at-a-reasonable-price, not a momentum multiple.
- Rubin sampling in Q4 2026 is the next leg. Annual cadence means a new architecture catalyst every 12 months — this is a compounding event stream, not a one-shot.
- FCF self-funds the cycle. $50B/quarter operating cash flow vs $12.4B PP&E — they're not raising capital to grow.
- The H20 overhang is cleared. Worst case on China is now in the numbers (zero assumed). Asymmetric: downside is priced, upside from any loosening is free.
9. Why Avoid
- Customer concentration is the bear case. Four buyers, one product cycle. If hyperscalers coordinate capex discipline — and they have every incentive to — NVDA's run-rate is exposed.
- Custom silicon is the structural risk. CUDA moat is real but not infinite. Trainium and TPU are gaining at the inference layer specifically, which is where the unit economics live (decode is memory-bandwidth-bound). If hyperscaler inference workloads migrate in-house, NVDA's 71% net margin is the line of attack.
- Geopolitical tail risk. Taiwan concentration is un-hedgeable. A 20–30% position-level drawdown on a Taiwan headline is the cost of holding this name.
- Insider selling is consistent. $3.3B in executive sales over recent periods, all 10b5-1. Huang selling 225k at $175–183 while stock sits at $202 isn't bullish either. Neutral-to-cautious.
- Near-term sentiment soft. 80.5% prediction-market conviction NVDA touches $192 in July. Technicals not supportive short-term.
10. Sources
- NVDA Yahoo Finance quote — price, market cap, P/E, EPS
- Capital.com NVDA market cap — $4.92T as of 2026-07-19
- MacroTrends price history — ATH $235.47 on 2026-05-14
- SEC EDGAR 10-Q XBRL (Q1 FY27) — verified income/balance/cash flow, period 2026-04-26
- NVIDIA FY26 Q4 release
- NVIDIA Q3 FY26 release
- NVIDIA 10-K via MarketScreener — H20 $4.5B charge
- Axis Intelligence NVDA stats
- stockanalysis.com forecast — consensus $301.62, Strong Buy
- Benzinga analyst ratings — $309.13 consensus
- vcpscanner institutional holders — 53.8% institutional
- Tom's Hardware Rubin roadmap
- AI-Infrastructure KB roadmap
- Daloopa customer concentration
- 24/7 Wall St — biggest threat
- Intellectia NVDA July 2026
- stocktitan Form 4 — Huang 225k
- TECHi insider selling
Memo by axelrod for Sweat Equity Holdings. Not personalised advice. Numbers carry their period; verify before acting.