NVIDIA Corp.
Combined average of Moat (AI Resilience), Growth, and Valuation scores.
Moat Score
CUDA software ecosystem and 10-year hardware lead in AI compute.
Nvidia's moat isn't just "fast chips", it's the Full-Stack Software Advantage:
- CUDA Software Ecosystem: With over 4 million developers, CUDA is the industry standard. Moving to another hardware provider requires rewriting massive amounts of code.
- Innovation Velocity: Hopper → Blackwell → Rubin on a one-year cycle. The Aug 26 PR names Vera Rubin ramping into full production, with racks at CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure and Nebius — the product cycle the moat is supposed to keep winning, not a reason to mark the moat up.
- Infiniband Networking: Their integration of networking (Mellanox) allows them to sell high-margin full-racks, not just individual GPUs. Spectrum-6 switches are part of the Vera Rubin platform in the same print.
Moat Verdict
NVIDIA's moat remains predominantly AI-resilient — CUDA, proprietary compute-optimisation data, and infrastructure-layer embedding all deepen as AI spend grows. The Q2 FY2027 beat ($96.2B vs a $91B guide) does not change that, and it does not repair the soft spot: regulatory lock-in stays weakened because the Q3 outlook still assumes zero China Data Center compute, and bundling stays weakened until a filing reverses the French finding or the DOJ look. NVIDIA remains the infrastructure layer of the AI economy; the print confirmed demand, not a freer hand to bundle or to sell into China.
84.0 resilient · 65.0 vulnerable · 80/20 = 80.2 · + 2 strength · = 82
Open a moat to read its note.
CUDA is the canonical learned-interfaces moat in semiconductors — 15+ years of developer mindshare, every ML PhD candidate learns CUDA, every major ML framework (PyTorch, TensorFlow, JAX) is CUDA-first by default. Switching to ROCm or any alternative is a multi-year rewrite. AI demand strengthens this moat rather than commoditising it: more AI workloads to write means more CUDA-coupled code, wider switching costs. Routed to resilient via aiExposure override — the AI wave is protecting this interface, not threatening it.
Customer ML training and inference pipelines are deeply embedded against CUDA-specific business logic — Megatron, DeepSpeed, vLLM, NCCL, cuDNN, cuBLAS are not portable abstractions. Every AI lab's production training stack is CUDA-coupled workflow at the code level. AI demand strengthens this lock-in by adding more CUDA-specific framework code to every codebase; routed to resilient via aiExposure override.
NVIDIA does not derive moat from public data access.
GPU architects, CUDA kernel engineers and AI systems researchers are scarce, but AMD, Google's TPU team, Broadcom's custom-silicon group and the hyperscalers now field credible accelerator teams. Deep bench, not an irreplaceable scarcity. Re-rated from strong to intact in the proof-point pass.
GPU + NVLink + Spectrum/InfiniBand networking + CUDA sold as full racks (Vera Rubin, Spectrum-6) delivers full-stack value competitors cannot match today. The DOJ look at the CUDA + hardware + networking bundle and the French dominance-abuse finding are a live risk to that bundle, but no remedy has been imposed, so the moat is not eroded — it is real, contested by regulators, and held at intact rather than strong until that risk resolves. Previously weakened. Routed to resilient: AI demand deepens the rack-level bundle.
NVIDIA learns from its own compute-optimisation work and from co-design with the largest labs, but customer training runs happen on customer clusters, and NVIDIA does not own that data. Real engineering knowledge, not a dataset rivals cannot replicate. Re-rated from strong to intact in the proof-point pass.
NVIDIA never held a regulatory moat. The China Data Center exclusion is a headwind already charged in the growth keyRisk (no China compute in the outlook), and the antitrust inquiries (DOJ on the CUDA-plus-hardware bundle, the French dominance finding) are a risk to the bundle, noted under bundling. Previously weakened, which scored regulation NVIDIA never benefited from as a lost lock — the convention fixed for Apple, Cadence and Synopsys.
4M+ CUDA developers create the largest and most entrenched AI developer community — switching has a multi-year rewrite cost.
Every major AI training and inference build is embedded in NVIDIA infrastructure: Data Center revenue reached $89.0B in Q2 FY2027 (+117% YoY), roughly 80% or more of AI accelerator spend by most industry estimates, and switching a production stack off CUDA is a multi-year rewrite.
CUDA is the de facto standard platform for AI compute — the PyTorch/TF ecosystem is CUDA-first by default.
The largest buyer of leading-edge wafers and advanced packaging: industry estimates put NVIDIA at roughly 60% of TSMC's CoWoS capacity, several times any custom-silicon programme, so it secures allocation first and spreads the largest R&D budget in the category across the most accelerator units. Hyperscaler ASICs buy at a fraction of that volume, one customer each.
Buyers are enterprises choosing on switching cost, integration and performance, which the other pillars rate. The name carries reputation, not a price premium it could hold on brand alone.
Combined average of Moat (AI Resilience), Growth, and Valuation scores.
Moat Score
CUDA software ecosystem and 10-year hardware lead in AI compute.
Growth Score
Q2 FY2027 (quarter ended July 26, reported Aug 26) printed revenue $96.221B, +18% from Q1's $81.615B and +106% from $46.743B a year ago, beating NVIDIA's own $91B guide. Data Center was $89.0B, +18% QoQ and +117% YoY. GAAP and non-GAAP gross margins were both 75.0%. GAAP diluted EPS $2.46, non-GAAP $2.22; GAAP operating income $63.734B. The Q1 falsifier — miss $91B by more than 5% — did not fire. Q3 is guided at $108.0B ±2% with GM 74.0% ±50 bps, and the outlook still assumes no Data Center compute revenue from China. Vera Rubin is named as ramping into full production. Cash conversion is the new residual: Q2 free cash flow $21.341B vs Q1 $48.554B, accounts receivable $63.059B vs $38.466B at Jan 25, inventories $31.575B vs $21.403B, long-term debt $32.366B vs $7.469B after issuing about $24.9B of debt in the quarter. $26.0B was returned to shareholders; ~$99B of buyback authorization remains; next dividend is $0.25 on Oct 1 (record Sep 10).
Valuation Score
After-hours ~$218 on Aug 26 (regular-session close $209.66 was pre-print; market cap ~$5.3T at the after-hours tape on ~24.19B basic shares). Unchanged ladder $130 / $260 / $430. At ~$218 the stock is 16% below the $260 base and 68% of the way from bear to base — piecewise 73, same as the Aug 10 card. The Q2 beat and $108B Q3 guide do not move the ladder; China is still zero in the outlook, so the old base-case H200 $15–20B increment is retired rather than earned. Live valuation will recompute against the tape; this static 73 is the after-hours print against the held corridor.
The Ecosystem Moat (CUDA)
Nvidia's moat isn't just "fast chips", it's the Full-Stack Software Advantage:
- CUDA Software Ecosystem: With over 4 million developers, CUDA is the industry standard. Moving to another hardware provider requires rewriting massive amounts of code.
- Innovation Velocity: Hopper → Blackwell → Rubin on a one-year cycle. The Aug 26 PR names Vera Rubin ramping into full production, with racks at CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure and Nebius — the product cycle the moat is supposed to keep winning, not a reason to mark the moat up.
- Infiniband Networking: Their integration of networking (Mellanox) allows them to sell high-margin full-racks, not just individual GPUs. Spectrum-6 switches are part of the Vera Rubin platform in the same print.
Moat Verdict
NVIDIA's moat remains predominantly AI-resilient — CUDA, proprietary compute-optimisation data, and infrastructure-layer embedding all deepen as AI spend grows. The Q2 FY2027 beat ($96.2B vs a $91B guide) does not change that, and it does not repair the soft spot: regulatory lock-in stays weakened because the Q3 outlook still assumes zero China Data Center compute, and bundling stays weakened until a filing reverses the French finding or the DOJ look. NVIDIA remains the infrastructure layer of the AI economy; the print confirmed demand, not a freer hand to bundle or to sell into China.
84.0 resilient · 65.0 vulnerable · 80/20 = 80.2 · + 2 strength · = 82
Open a moat to read its note.
CUDA is the canonical learned-interfaces moat in semiconductors — 15+ years of developer mindshare, every ML PhD candidate learns CUDA, every major ML framework (PyTorch, TensorFlow, JAX) is CUDA-first by default. Switching to ROCm or any alternative is a multi-year rewrite. AI demand strengthens this moat rather than commoditising it: more AI workloads to write means more CUDA-coupled code, wider switching costs. Routed to resilient via aiExposure override — the AI wave is protecting this interface, not threatening it.
Customer ML training and inference pipelines are deeply embedded against CUDA-specific business logic — Megatron, DeepSpeed, vLLM, NCCL, cuDNN, cuBLAS are not portable abstractions. Every AI lab's production training stack is CUDA-coupled workflow at the code level. AI demand strengthens this lock-in by adding more CUDA-specific framework code to every codebase; routed to resilient via aiExposure override.
NVIDIA does not derive moat from public data access.
GPU architects, CUDA kernel engineers and AI systems researchers are scarce, but AMD, Google's TPU team, Broadcom's custom-silicon group and the hyperscalers now field credible accelerator teams. Deep bench, not an irreplaceable scarcity. Re-rated from strong to intact in the proof-point pass.
GPU + NVLink + Spectrum/InfiniBand networking + CUDA sold as full racks (Vera Rubin, Spectrum-6) delivers full-stack value competitors cannot match today. The DOJ look at the CUDA + hardware + networking bundle and the French dominance-abuse finding are a live risk to that bundle, but no remedy has been imposed, so the moat is not eroded — it is real, contested by regulators, and held at intact rather than strong until that risk resolves. Previously weakened. Routed to resilient: AI demand deepens the rack-level bundle.
NVIDIA learns from its own compute-optimisation work and from co-design with the largest labs, but customer training runs happen on customer clusters, and NVIDIA does not own that data. Real engineering knowledge, not a dataset rivals cannot replicate. Re-rated from strong to intact in the proof-point pass.
NVIDIA never held a regulatory moat. The China Data Center exclusion is a headwind already charged in the growth keyRisk (no China compute in the outlook), and the antitrust inquiries (DOJ on the CUDA-plus-hardware bundle, the French dominance finding) are a risk to the bundle, noted under bundling. Previously weakened, which scored regulation NVIDIA never benefited from as a lost lock — the convention fixed for Apple, Cadence and Synopsys.
4M+ CUDA developers create the largest and most entrenched AI developer community — switching has a multi-year rewrite cost.
Every major AI training and inference build is embedded in NVIDIA infrastructure: Data Center revenue reached $89.0B in Q2 FY2027 (+117% YoY), roughly 80% or more of AI accelerator spend by most industry estimates, and switching a production stack off CUDA is a multi-year rewrite.
CUDA is the de facto standard platform for AI compute — the PyTorch/TF ecosystem is CUDA-first by default.
The largest buyer of leading-edge wafers and advanced packaging: industry estimates put NVIDIA at roughly 60% of TSMC's CoWoS capacity, several times any custom-silicon programme, so it secures allocation first and spreads the largest R&D budget in the category across the most accelerator units. Hyperscaler ASICs buy at a fraction of that volume, one customer each.
Buyers are enterprises choosing on switching cost, integration and performance, which the other pillars rate. The name carries reputation, not a price premium it could hold on brand alone.
Growth Analysis
Growth Drivers
Key Risk
The $91B Q2 test passed. The next hard test is Q3 FY2027 at $108.0B ±2% with GM 74.0% ±50 bps and still no China Data Center compute in the outlook. Falsifiable: miss that $108B guide by more than 5% (below ~$102.6B), print GM through the 73.5% floor, or a top-3 hyperscaler cuts FY2027 AI capex guidance by more than 10% on a single print. A second residual: Q2 free cash flow fell to $21.341B from Q1 $48.554B as receivables hit $63.059B and inventories $31.575B, funded in part by ~$24.9B of new debt — if that working-capital draw repeats in Q3 while revenue is still guided up, the demand print is running ahead of cash.
Score Derivation
93.1 base + 2.7 trajectory − 5 risk = 91
Base 93 (35–50% CAGR, midpoint 42.5%, baseFromCagr) + 2.7 trajectory (2 of 3 drivers accelerating) + 0 stable margins (printed 75.0%; Q3 GM 74.0% ±50 bps is mix, not a compression charge) − 5 moderate keyRisk (a hyperscaler capex cut and the Q2 working-capital draw). China is zero in the Q3 guide already, so it sits in the estimate and is not charged again here; the capex-cycle risk is graded moderate for the name with the widest margin and the least-exposed balance sheet in the AI-hardware cohort, below SMCI, DELL and MU = 91. Do not bump because they beat. The old author string that added TAM expansion points is retired — primaryType does not score.
Price Scenarios (12–24 Months)
Valuation Multiples
| Trailing P/E (GAAP) | ~33× |
| Forward P/E (NTM) | ~24× |
| PEG Ratio | ~0.5× |
| Price / Sales (FY27) | ~13× |
| Price / H1 FCF | n/m as TTM |
At ~$218 the forward multiple is still ~24× on the last-sourced ~$9 NTM EPS, PEG ~0.5×, which is the same GARP frame as Aug 10 — the print confirmed the $91B guide rather than cheapening the stock. Trailing GAAP ~33× on ~$6.54 TTM EPS. Do not treat H1 FCF $69.9B as a clean TTM: Q2 conversion halved as receivables and inventory absorbed cash. The $108B Q3 guide with GM 74.0% is the multiple the tape has to underwrite next; China is still not in it.
Approximate figures as of August 26, 2026.
Where We Are vs Targets
Loading live price…
Export controls re-escalate targeting Blackwell/Rubin-class chips; hyperscaler in-house ASICs capture 20%+ of AI training workloads; working-capital draw repeats so the $108B guide prints without cash.
- U.S. imposes new export restrictions on Blackwell/Rubin-class chips to allied nations, removing $15B+ in annual revenue
- Google TPU v6 and Amazon Trainium3 capture 20%+ of hyperscaler AI training by end of 2026, pressuring NVIDIA market share below 75%
- Q3 misses $108B by more than 5% or GM prints through 73.5%, and the Q2 pattern — FCF $21.3B, AR $63.1B, inventories $31.6B, ~$24.9B of new debt — repeats
Q3 lands near $108B ±2% at GM ~74%; Vera Rubin stays in production at the named clouds; FY2027 tracks toward ~$390B from the guided run-rate with China still zero in the model. NVIDIA Enterprise software reaches $5B+ ARR.
- Q3 FY2027 revenue lands near the $108.0B ±2% guide, confirming the Blackwell-to-Rubin handoff after Q2 beat $91B at $96.2B
- China Data Center compute stays out of the model — the Q3 outlook assumes none, so base does not count H200 as $15–20B of FY2027 revenue
- Vera Rubin remains in full production at CoreWeave, Google Cloud, Azure, OCI and Nebius, extending the cycle into FY2028 without needing a China reopening
Vera Rubin cycle exceeds the current run-rate; sovereign AI buildout accelerates; China Data Center compute returns to the guide; software inflects above $10B ARR.
- A subsequent guide includes China Data Center compute after a sourced policy change — that is upside, not the base
- Vera Rubin yields exceed roadmap targets and sovereign/national AI factories (Korea, Japan, Europe HPC named in the PR) add a recurring government layer
- NIM / NVIDIA AI Enterprise scale to $10B+ ARR, re-rating toward software multiples on a higher-margin mix