Nvidia
Nvidia
One-line summary: Dominant AI-accelerator (GPU) company; "probably the greatest company in the first part of this century" per a competitor (andrew-feldman). Tracked here mainly for the durability of its moat at inference — Feldman argues CUDA is "not important now and has no role whatsoever in inference," and that two of the three leading frontier models already train without CUDA.
What it is
Designs the GPUs (H100/H200/B200/GB300) and the CUDA software stack that underpinned the first wave of AI training. The GPU architecture is, per Feldman, "an extremely good architecture and extremely efficient at building very slow tokens" — i.e. cheap at low speed, but its cost and power-per-token rise steeply as you push for fast tokens. Nvidia is the single largest consumer of TSMC CoWoS capacity (>50%) and the anchor demand behind the HBM shortage.
Why it matters to artificial-intelligence
For the AI thread (the "compute markets / GPU supply" subdomain), the durable claim from this source is the decay of CUDA as an ecosystem moat, framed as a fact about where frontier labs actually run, not about NVDA's price. Per andrew-feldman, CUDA "was really important in the creating of the AI landscape" but is now irrelevant at inference and shrinking at training: of the three leading frontier models, Gemini (Google TPUs) and Anthropic's Claude (Trainium) use no CUDA, while only OpenAI's GPT trains in the CUDA environment. The architectural why — GPUs being optimized for cheap slow tokens, with cost/power-per-token rising as you push for speed — is the hardware constraint that the thread's inference-economics sources (Turley's gpu-as-zero-sum-constraint, the chatgpt-super-assistant-vision pricing forecast) discuss from the demand side. See cuda-moat-erosion-at-inference and open-vs-closed-source-model-economics.
Why it matters to stock-market
Nvidia is the incumbent every alternative-silicon thesis is implicitly short. Two SCOPE-relevant claims from this source:
- CUDA-moat erosion at inference. Feldman: CUDA was decisive in creating the AI landscape but "has no role whatsoever in inference," and a model can be moved from GPUs to Cerebras "in 10 keystrokes." A year ago every frontier model was CUDA-built; today two of three are not (Gemini on TPUs, Anthropic on Trainium) — "a hemorrhaging of share." See cuda-moat-erosion-at-inference.
- Export-control posture. Feldman explicitly comes down against Nvidia's stated position (give China access to keep them on US-designed product); he favors limiting diffusion of "our most precious technologies" to an "industrial enemy," accepting that some markets get foreclosed.
- Valuation / catch-up-trade view (dan-loeb, Third Point, May 2026). Against the moat-erosion bear case, Loeb makes the long-side valuation argument: NVDA is "a catch up trade... at 15 times 27, 12 times 28 for the most dominant, very fast growing company at its size." He reviewed Third Point's whole semicap/hyperscaler book expecting to take profits and instead concluded "it's the most attractive sector right now — it's where the bulk of our capital is invested," conditional only on the AI cycle not "rolling over in 31 or 32." A useful counterweight to the Feldman CUDA-erosion framing: the moat-at-inference question and the 12-month valuation question are separable.
Key facts
- Moved from 400mm² to 800mm² die over ~5–6 years "for this exact reason" — bigger chips process more information in less time (Feldman's framing for why wafer-scale is the logical extreme). From andrew-feldman in 2026-05-21-odd-lots-why-cerebras-ceo-andrew-feldman-built-the-world-s.
- GPU memory is HBM-class (high capacity, slow) → slow inference; the cost/power-per-token rises as you push for speed ("like miles-per-gallon falling as you drive faster"). From andrew-feldman.
-
50% of TSMC CoWoS capacity consumed by Nvidia (~800–850K wafers/year, 2026). See cowos-packaging-capacity-crunch.
- CEO Jensen Huang named by Feldman alongside Hock Tan (Broadcom) and Lisa Su (AMD) as great CEOs of the era.
- Q1 FY2027 (Feb–Apr 2026, reported May 2026): $81.6B revenue (+85% YoY), beating consensus $78.9B. Net income $58.3B vs estimate $42.9B. Data Center: $75.2B (+92% YoY, +21% QoQ) — record. DC compute: $60.4B (+77% YoY). DC networking: $14.8B (+199% YoY) (InfiniBand, Spectrum-X Ethernet, NVLink). Q2 FY2027 guide: $91B vs Street estimate $86B — first time guidance has exceeded consensus this cycle. Three straight quarters of acceleration. Blackwell (including GB300 Ultra) accounts for large majority of DC compute; adopted by "every major hyperscaler, every cloud provider, and every model builder." Jensen Huang: "Demand has gone parabolic. Agentic AI has arrived." China DC: zero revenue (ongoing). From 2026-05-26-autoresearch-semis-ai-infra-macro-scan-may-23-26-2026.
- H20 charge: $4.5B charge for H20 inventory/purchase obligations (prior period event); guided zero China DC revenue going forward. From 2026-05-22-autoresearch-china-blackwell-compute-routing-may-2026.
- China export-control status (May 2026): US cleared ~10 Chinese firms (Alibaba, Tencent, ByteDance) for H200 purchases (BIS case-by-case review, Jan 15, 2026 policy change). But not a single H200 has shipped — Beijing blocked domestic firms from buying, steering them to Huawei Ascend. China's self-block is currently the binding constraint, not US export policy. From 2026-05-22-autoresearch-china-blackwell-compute-routing-may-2026.
- B30A (China Blackwell): Downgraded single-die Blackwell chip for China market; $6.5–8K price target; June 2026 production target. Commercial viability unclear given China's domestic-chip-first policy. From 2026-05-22-autoresearch-china-blackwell-compute-routing-may-2026.
Groq deal and DOJ risk
- Nvidia-Groq $20B deal (December 2025): Licensing + acqui-hire. Groq's LPU (Language Processing Unit) — primary inference chip architecture faster than CUDA for certain workloads — licensed to Nvidia, key personnel including CEO acqui-hired. Structured as licensing to potentially avoid HSR pre-merger review. Effect: eliminates primary inference competitor. Senators Warren/Blumenthal letter (March 20, 2026) to Jensen Huang questioning whether it violates antitrust law. From 2026-05-27-autoresearch-regulatory-antitrust-semis-ai-may-2026.
- DOJ Nvidia investigation (ongoing): Subpoenas issued 2024/early 2025 for market-dominance concerns (difficulty switching GPU suppliers, customer penalties for non-exclusive use, RunAI acquisition). Groq deal provides second DOJ vector. Trump DOJ posture toward tech monopoly enforcement is uncertain but has been less aggressive than Biden. Tail risk: forced Groq license divestiture, behavioral remedies, or stock uncertainty during investigation.
- China-Russia SAMR/FAS bilateral antitrust MOU (May 25, 2026): China's State Administration for Market Regulation (SAMR) and Russia's Federal Antimonopoly Service (FAS) signed a 2026–2027 bilateral antitrust cooperation memorandum, formalizing coordinated enforcement capacity against US tech companies. Named targets: Nvidia (SAMR probe ongoing since December 2024, Mellanox acquisition condition violations, preliminary breach September 2025 — NOT closed despite Trump-Xi May 2026 summit) and Qualcomm (SAMR Autotalks acquisition investigation opened October 2025, NOT closed). From 2026-05-30-autoresearch-regulatory-antitrust-tech-biotech-utilities.
Strengths (from a thesis-input perspective)
- Still the default for training; CUDA's training role is "shrinking" but not gone (GPT still trained in CUDA).
- Extremely cheap per token at low speed — owns the cost-optimized end of inference.
- Scale advantage in supply chain (CoWoS, HBM allocation).
Weaknesses (from a thesis-input perspective)
- CUDA moat is being routed around at inference and increasingly at training (TPU, Trainium, wafer-scale).
- Architecturally expensive/power-hungry at the fast-token end where engaged/agentic workloads are migrating.
- Export-control exposure to China; competitor CEOs publicly disagree with Nvidia's access-maximizing stance.
Open questions
- cuda-moat-erosion-at-inference — how fast does CUDA's training/inference lock-in decay, and does it re-rate NVDA's terminal multiple?
Sources
- 2026-05-21-odd-lots-why-cerebras-ceo-andrew-feldman-built-the-world-s
- 2026-05-28-podcast-invest-like-the-best-dan-loeb-lessons-from-30-years-of-investing — Loeb's NVDA catch-up-trade / "most attractive sector" valuation view.
- 2026-05-30-autoresearch-regulatory-antitrust-tech-biotech-utilities — China-Russia SAMR/FAS bilateral MOU; DOJ Groq probe status; no formal enforcement action.