Jensen Huang — Founder, President & CEO
Co-founded NVIDIA in 1993; 33-year CEO; BSEE Oregon State, MSEE Stanford. Ex-LSI Logic, AMD.
KAAMOS RESEARCH
CONTENTS
01 Summary
02 Thesis
03 Financials
04 Projections
05 Scenarios
06 Valuation
07 Company
08 Risks
VIEW
KEY FIGURES
FY26E Revenue
$199.9B
+53% YoY
FY26E EPS
$4.41
+51% YoY
FY26E GM
73%
−2pp YoY
AI accel. share
78%
DCF (base case)
$140
1 · Summary & Verdict
30% growth, 64% EBITDA margins, 51% FCF margins, 78% accelerator share — a superb business now trading at peer-median multiples. Our Scenario Fair Value of $217 sits ~5% above the $207 price; we initiate Neutral.
■We initiate Neutral: superb business, but at $207 the peer-median multiple has caught up to the operating outperformance. NVIDIA delivers 30% NTM revenue growth (vs. 20% peer median), 64% EBITDA margin (vs. 55%), 51% FCF margin (vs. 30%), and ~78% share of merchant AI accelerator units. At $207 the stock trades at ~26x EV/EBITDA NTM and ~35x P/E NTM — a modest premium to peer medians of 23.0x and 31x, capturing most of the operating outperformance. Our Scenario Fair Value of $217 (Bull $298 × 25% + Base $221 × 55% + Bear $98 × 20%) sits ~5% above the $207 price. View shifts to Positive below ~$170 (where 12-month risk/reward turns clearly favorable) and to Cautious above ~$260 (where the bull case is fully priced).
■Annual product cadence (Hopper → Blackwell → Rubin → Rubin Ultra) is itself a moat custom silicon cannot match. GB200 NVL72 rack-scale systems are in volume production at multi-million-dollar ASPs. Rubin platform unveiled at GTC March 2026 samples late CY2026 with Rubin Ultra following in 2027. Each new generation resets the price-performance bar before competitors can catch up. We model Data Center revenue from $172B (FY26A) to $410B (FY31E) — a 19% CAGR with operating margin sustained above 60%.
■Sovereign AI is a structurally less price-sensitive buyer class that extends the capex cycle. Multi-billion-dollar national AI commitments — Humain (Saudi Arabia), G42 (UAE), Mistral/Iliad (France), the SoftBank–KDDI–NTT consortium (Japan), Singapore, India, Korea — emerged in FY25–26 as a buyer class less price-sensitive than hyperscalers. We estimate $80–120B in cumulative sovereign bookings FY27–31E, materially diversifying NVDA away from the top-4 hyperscaler concentration (~55% of FY26 revenue).
■Software monetization (AI Enterprise + NIM) is the optionality the market has not yet underwritten. AI Enterprise (~$4,500/GPU/year) and NIM inference microservices began monetizing the CUDA stack at scale in FY25. Management has pointed to a $10B+ ARR milestone as the next disclosure threshold. At 8–12x recurring revenue, a $10B ARR software business is worth $80–120B (~$3–5/share) — and validates a separate-of-hardware multiple component, the same dynamic that re-rated MSFT Azure and AVGO software in prior cycles.
Historical valuation bands · vs 3 multiples that matter for this name
As of 2026-06-15 · 8-year weekly window
NVDA at $207 trades roughly 1σ BELOW its own 8-year mean on both P/E (31x vs mean 65x) and EV/EBITDA (30x vs mean 48x). The familiar framing of 'NVDA is expensive vs the S&P' is true, but vs its own history this is the most undemanding multiple in years — the bear case requires the multiple to compress AND earnings to disappoint, a harder set-up than the screen suggests.
P / E (TTM)
32.3x
vs 65.0x mean
z = −0.91σ
EV / EBITDA (TTM)
30.3x
vs 48.0x mean
z = −0.96σ
P / Sales (TTM)
20.1x
vs 21.4x mean
z = −0.15σ
See § 6 Valuation for the per-multiple analysis and historical band charts.
| Summary financials | FY23A | FY24A | FY25A | FY26E | FY27E | FY28E |
|---|---|---|---|---|---|---|
| Revenue ($B) | 27.0 | 60.9 | 130.5 | 199.9 | 259.8 | 322.2 |
| Gross margin % | 56.9% | 72.7% | 75.0% | 73.0% | 74.0% | 73.5% |
| EBITDA ($B) | 7.1 | 34.5 | 82.6 | 125.8 | 167.3 | 207.6 |
| Net income ($B) | 4.9 | 29.8 | 71.8 | 108.6 | 142.8 | 177.3 |
| Diluted EPS ($) | 0.20 | 1.19 | 2.93 | 4.41 | 5.83 | 7.30 |
| FCF ($B) | 2.9 | 27.6 | 65.2 | 97.9 | 131.8 | 166.8 |
THE THREE RISKS THAT MATTER
AI capex digestion (2027–28)
The principal bear case. Hyperscaler AI capex grew >70% in 2025 to $400B+; a 2027–28 digestion phase as deployed capacity catches demand is the primary downside trigger. Our bear case models flat to -5% revenue growth FY28–FY30 and gross-margin compression to 55% — implying ~$98 18-month value.
Hyperscaler custom silicon at the inference layer
Google TPU, AWS Trainium, Microsoft Maia and Meta MTIA together are a multi-billion-dollar in-house alternative, growing fastest in inference. A successful generation prioritized over NVDA buy commitments compresses incremental dollar growth at the top of the customer base.
Export controls + China
Successive U.S. export controls (Oct 2022, Oct 2023, Dec 2024, Jan 2026) have restricted advanced AI chip sales to China. H20 / B30 / CW30 compliant SKUs generate limited revenue at lower margins; Huawei Ascend is the structural domestic substitute. Further tightening effectively zeros the China data-center revenue line.
2 · Investment Thesis
1. CUDA software gravity is a 20-year compounding moat that custom silicon cannot replicate quickly
NVIDIA's defensible advantage is not its silicon — it is the integrated CUDA platform, now nineteen years old, with 5 million+ registered developers and libraries spanning training (cuDNN, Megatron, NeMo), inference (TensorRT, Triton, NIM), HPC (cuBLAS, cuFFT, RAPIDS), robotics (Isaac), and digital twins (Omniverse). AMD's ROCm has made meaningful progress but remains a generation behind in framework support and library maturity. Hyperscaler custom-silicon stacks (XLA for TPU, Neuron for Trainium) work for proprietary internal workloads but offer no portability. The cost of switching from CUDA at the application layer — measured in engineering-quarters per workload — is the principal reason we expect NVDA share of merchant accelerator units to remain above 70% over our forecast horizon.
- 5M+ registered CUDA developers; hundreds of millions of CUDA-capable GPUs in the installed base
- AI Enterprise + NIM monetize the stack via per-GPU subscription at ~$4,500/GPU/year
2. Annual product cadence keeps performance leadership a generation ahead
NVIDIA committed publicly to an annual rather than two-year cadence: Hopper (2022) → Blackwell (2024) → Rubin (sampling late CY2026, volume 2027) → Rubin Ultra (2027) → 'Feynman' (~2028). This tempo is one that hyperscaler custom-silicon programs (18–24 month internal cycles) and AMD's MI400 cannot match in lockstep. Each generation resets the price-performance bar, forcing competitors to chase a moving target. The shift to rack-scale GB200 NVL72 systems — sold at multi-million-dollar ASPs with NVDA capturing the entire networking and system value — extends per-customer dollar capture by 5–7x versus the loose-GPU model. NVL576 (Rubin generation, 2027) extends this further still.
- GB200 NVL72 ASP ~$3.0M per rack (CY26); NVL576 Rubin racks launch at ~$5.5M ASP
- Annual cadence vs. AMD MI400 (2026) and hyperscaler ASICs on 18–24 month cycles
3. Sovereign AI is a new buyer class that extends the AI capex cycle through hyperscaler digestion
Multi-billion-dollar national AI infrastructure programs — Humain (Saudi Arabia), G42 (UAE), Mistral/Iliad (France), the SoftBank–KDDI–NTT consortium (Japan), Stargate (UAE/OpenAI/Oracle/SoftBank), plus India, Korea, Singapore, Indonesia, Norway, Sweden — represent an entirely new buyer class that did not meaningfully exist 18 months ago. We estimate $80–120B in cumulative sovereign commitments across FY27–31E, materially diversifying NVDA away from the top-4 hyperscaler concentration of ~55% of FY26 revenue. These buyers are policy-driven, multi-year, and structurally less price-sensitive than the hyperscalers — smoothing the capex curve during any 2027 hyperscaler digestion period.
- Cumulative sovereign AI bookings $80–120B FY27–31E
- Reduces top-4 hyperscaler concentration from ~55% to ~42% of revenue by FY31E
4. At a modest premium to peer median, the structural moat justifies peer-plus multiples but no longer leaves an asymmetric discount
NVDA trades at ~26x EV/EBITDA NTM and ~35x P/E NTM versus peer medians of 23.0x and 31x — a modest premium that captures most of the operating outperformance NVDA delivers (30% growth vs. 20% peer median, 64% EBITDA margin vs. 55%, 51% FCF margin vs. 30%). Our base case implies the FY27E exit multiple compresses to the peer median (23.6x EV/EBITDA, 31x P/E) — no premium baked in at the 18-month horizon. The probability-weighted 18-month Scenario Fair Value of $217 (Bull $298 × 25%, Base $221 × 55%, Bear $98 × 20%) sits ~5% above the $207 spot — fair value with a modest upside tilt, the basis for our Neutral view.
- Implied P/E at our $217 18-month FV: 31.0x — exactly the peer median
- Bull case at 32x P/E NTM with $15B software ARR is the principal upside lever
3 · Financial Analysis
NVIDIA’s financial transformation between FY2023 and FY2026 is among the most consequential at this scale of business in the modern technology era. Revenue grew from $26,974M in FY2023A to $199,870M in FY2026A — a 7.4x increase in three years, or a 95% CAGR. Gross margin expanded from 56.9% to a peak of 75.0% in FY2025A before normalizing modestly to 73.0% in FY2026A as Blackwell ramp inventory and a lower-margin China SKU mix temporarily compressed margins. Operating margin moved from 20.7% to 61.3%, and net income from $4.9B to $108.6B.
Three things drove this transformation. First, the Data Center segment scaled from 56% of revenue in FY23A to 86% by FY26A as Hopper (H100) and then Blackwell (B100/B200/GB200) displaced gaming-cycle revenue dynamics with multi-year AI infrastructure demand. Data Center revenue alone grew from $15.1B to $171.9B, an 11.4x increase. Second, ASPs rose materially as NVIDIA shifted from selling discrete GPUs (~$25–35K H100 cards) to selling rack-scale GB200 NVL72 systems at multi-million-dollar prices, with full networking and software stack included. Third, operating leverage emerged as opex (R&D + SG&A) grew roughly with mid-teens efficiency: R&D rose from $7.3B to $17.5B (2.4x) while revenue grew 7.4x, so opex as a percent of revenue fell from 36.6% to 11.7%.
Free cash flow generation has been the cleanest metric of the franchise. FCF grew from $2.9B (FY23A) to $97.9B (FY26A), with FCF margin moving from 11% to 49%. NVIDIA is now among the three most cash-generative businesses in the public equity market alongside Apple and Saudi Aramco. Capital intensity remains low — capex was $8.5B in FY26A (4.3% of revenue) — because NVIDIA is fabless and outsources manufacturing to TSMC. The bulk of FCF has been returned via share repurchases ($55B in FY26A) and a token dividend, with the balance accruing to cash and marketable securities (~$78B at FY26A end).
Segment evolution. Data Center is now overwhelmingly the principal revenue line, accounting for 86% of FY26A revenue and projected to reach 88% by FY28E. Gaming ($15B, 7.5%) is mature and cycling on the RTX 50 series Blackwell launch. Professional Visualization ($2.4B, 1%) is steady. Automotive & Robotics ($2.0B, 1%) is the fastest-growing non-Data-Center segment, with DRIVE Thor ramping into Mercedes, Volvo, BYD, Li Auto, XPeng, Polestar and others. OEM & Other ($8.6B, 4.3%) includes legacy and miscellaneous.
Geographic mix. Reported geography reflects ship-to billing rather than ultimate end-demand: the United States (~50% of FY26A revenue) and Singapore (18% — an intermediary hub used by cloud procurement entities) together comprise ~68% of reported billings, while true end-demand is more globally distributed. Taiwan (~13%) reflects shipments to local ODM partners. China (~10%) reflects compliant H20/B30/CW30 SKU sales under U.S. export-control constraints and is sized to decline structurally given Huawei substitution.
Balance sheet. NVDA is essentially debt-free: $8.5B long-term debt against ~$78B cash and marketable securities at FY26A end, with net cash growing to ~$454B by FY31E in our base case. Working capital is a material absorption of operating cash flow each year as receivables and inventory grow with the business; ΔNWC averages ~$8B/year through the forecast period. The capital structure — equity-only, net cash, very low capex — is the principal reason the DCF WACC sits near pure cost of equity (10.5%) with minimal debt benefit.
Capital return. NVIDIA has been a meaningful capital returner since FY24: share repurchases totaled $9.5B (FY24A), $34.1B (FY25A), $55.0B (FY26A), and we model $75–108B annually FY27–31E. The dividend remains token ($0.04–$0.10/share through the forecast period) but supports passive index inclusion. Net diluted share count declines from 24.6B FY26A to 23.7B FY31E (~4% net reduction over five years) — modest given the share-repurchase magnitude, because stock-based compensation issuance offsets a portion of buybacks.
4 · Projection Assumptions
We model FY2027E–FY2031E as a five-year explicit period anchored on three macro drivers: (1) hyperscaler AI capex trajectory; (2) the Rubin product transition in CY2027; (3) sovereign and enterprise AI demand as a partial offset to any hyperscaler digestion. We assume gross margin normalizes from a Blackwell-ramp 73% (FY26A) toward 71% by FY31E as memory content per system rises, and operating margin remains anchored above 60% on continued opex leverage.
Revenue trajectory (base case). Total revenue grows from $199.9B (FY26A) to $468.4B (FY31E), an 18.6% CAGR with step-down growth: 30% FY27E, 24% FY28E, 18% FY29E, 12% FY30E, 10% FY31E. The growth profile reflects (i) Blackwell at full production and Rubin ramping in CY2027, (ii) maturation of the AI capex cycle through FY29E, (iii) modest digestion in FY30E, and (iv) recovery into Rubin Ultra / Feynman cycles in FY31E.
Data Center (86% → 88% of revenue). We model Data Center revenue growing from $171.9B FY26A to $409.8B FY31E, a 19% CAGR. Specific drivers: hyperscaler capex of ~$400B (CY25) → $520B (CY26) → $580B (CY27E) → $550B (CY28E digestion) → $620B (CY30E re-acceleration), with NVDA capturing 35–40% of the silicon + networking dollar; GB200 NVL72 ASPs of ~$3.0M per rack (CY26) declining to ~$2.4M by CY30 as competition emerges; Rubin NVL576 racks launching at ~$5.5M ASP in CY2027; networking (NVLink + InfiniBand + Spectrum-X) growing from $24B FY26A to $57B FY31E at ~14% of DC; software (AI Enterprise + NIM) reaching ~$37B FY31E (9% of DC); sovereign AI bookings of $20–30B incremental annual revenue across FY27–31E.
Gaming (7% → 5%). Revenue grows from $15.0B FY26A to $22.5B FY31E (8% CAGR). RTX 50 series Blackwell-derived GeForce GPUs launched early CY2026 with strong reception; DLSS 4 with multi-frame generation supports cycle ASP elevation. The segment is increasingly mature; we model low-to-mid single-digit unit growth and modest ASP improvement.
Pro Visualization (1%). Steady ~10% growth from $2.4B FY26A to $4.2B FY31E. Omniverse digital-twin platform and RTX workstation products serve architecture, manufacturing, and media/entertainment.
Automotive & Robotics (1% → 2.5%). Grows from $2.0B FY26A to $11.7B FY31E (42% CAGR) as DRIVE Thor ramps and Robotics (Isaac, Jetson) commercial traction emerges. Existing OEM wins anchor the base; we model new wins by FY29E. Robotics revenue from Jetson modules + Isaac platform grows from ~$200M FY26A to $2.5B FY31E.
Geographic assumptions. United States grows from $99.9B (50%) FY26A to $215.5B FY31E (hyperscaler demand the principal driver). Singapore held at 18% (intermediary). Taiwan held at 13% (ODM partner shipments). China declines from 13% to 9% as export controls bind. EMEA grows fastest, from 3% to 8%, driven by French sovereign AI, Nordic data-center expansion, German enterprise adoption, and UK financial-services AI factory deployments.
Margin trajectory. Gross margin: 74% FY27E → 73.5% FY28E → 71.0% FY31E. Compression reflects (i) rising HBM content per system as HBM4 scales, (ii) increasing share of lower-margin China-compliant SKUs, and (iii) general competitive pressure as AMD MI400 ramps. Operating margin: held above 60% throughout — 62.6% FY27E, 62.6% FY28E, 62.1% FY29E, 61.5% FY30E, 61.2% FY31E. EBITDA margin held above 60%.
Operating expense assumptions. R&D grows 25%, 20%, 13%, 8%, 8% across FY27–31E, outpacing revenue growth through FY28E (heavy Rubin investment) before re-leveraging. SG&A grows in line with revenue. Total opex falls from 11.7% of revenue FY26A to 9.8% FY31E.
Capital expenditure. Rises from $8.5B (4.3% of revenue) FY26A to $19.0B (4.1%) FY31E, reflecting in-house data-center investment for Rubin testing, expanded campus infrastructure, and capitalized software. NVIDIA remains fabless and outsources wafer manufacturing to TSMC.
Tax rate. 15% FY27–FY28E, rising to 16% FY29E onward as OECD Pillar 2 implementation and U.S. tax reform discussions normalize the rate from the unusually low ~12–14% range driven by FDII benefits and one-time items.
Capital return. $75B–$108B in annual share repurchases through FY27–FY31E, with a modest dividend ($0.05–$0.10/share). Diluted share count declines from 24.6B FY26A to 23.7B FY31E (~4% net reduction).
Output metrics (base case). FY27E EPS $5.83 → FY29E EPS $8.52 → FY31E EPS $10.51. FY27E FCF $132B → FY29E FCF $197B → FY31E FCF $243B. Ending net cash position $454B by FY31E.
5 · Scenario Analysis
We frame three scenarios with explicit probabilities. The Bull case (25%) extends AI capex through 2030 and rewards software monetization; the Base case (55%) assumes the live model with a normalization into FY30 digestion; the Bear case (20%) prices in hyperscaler digestion plus custom-silicon share take plus China zero. Probability-weighted 18-month value across the three scenarios is $217, which sits ~5% above current $207 spot — the basis for our Neutral view.
Bull case — $298 implied price (18 months). Probability 25%.
Thesis: AI capex extends durably through 2030, Rubin Ultra outperforms on price-performance, sovereign and enterprise demand offsets any hyperscaler softness, and software ARR scales to $15B+ exiting FY28E. Multiple expansion to 32x P/E NTM (vs. peer median 31x) on the validation of recurring software revenue.
Key parameters: Revenue CAGR FY26–FY31E 25% (vs. 19% base); FY29E revenue $446B; FY29E EPS $9.31; gross margin FY29E 75%; operating margin FY29E 64%; software ARR exit FY28E $15B+; sovereign AI bookings FY27–29E $120B+ cumulative; NVDA accelerator share FY29E 80%; exit P/E 32x.
Catalysts required: (i) Rubin Ultra sampling on schedule Q1 CY2027 with >2x price-performance vs. Rubin baseline; (ii) explicit software ARR disclosure ≥$10B run-rate within 4 quarters; (iii) sovereign AI bookings >$50B cumulative announced by end of FY27E; (iv) hyperscaler CY2027 capex guidance up ≥15% versus CY2026; (v) China export-control resolution allowing modest compliant SKU volume growth.
Base case — $221 implied price (18 months). Probability 55%.
Thesis: live model, FY27 growth ~30%, normalization through FY28–FY29, modest digestion in FY30. Gross margin compresses from 75% peak to 71% in line with management commentary. Exit multiple held at 26x P/E NTM (vs. ~35x current at $207 spot — i.e., the base case bakes in modest multiple compression to peer median by FY28E).
Key parameters: Revenue CAGR FY26–FY31E 19%; FY29E revenue $380B; FY29E EPS $8.52; gross margin FY29E 72.5%; operating margin FY29E 62.1%; software ARR exit FY28E ~$10B; NVDA accelerator share FY29E 75%; exit P/E 26x.
Rationale: This is our most-likely view, anchored on company guidance and the explicit-period build of segment, geography, margin, and capital-return assumptions documented in the projections section. The base case implies NVDA trades at peer-median multiples for a structurally above-median business — i.e., no premium for the superior growth, margin, and FCF profile. DCF base intrinsic value is $136; trading comps at peer median imply $176 (EV/EBITDA NTM) and $181 (P/E NTM); football-field weighted base is $155, ~25% below current $207 spot — i.e., the NTM trading-multiple lens alone does not currently underwrite the spot price, and the $62/share bridge to the 18-month Scenario FV is the 18-month cash-flow build plus assumed peer-median exit multiple.
Bear case — $98 implied price (18 months). Probability 20%.
Thesis: hyperscaler AI capex peaks in FY27 and rolls over in FY28–FY29 as deployed capacity absorbs demand; hyperscaler custom silicon (TPU v7, Trainium 4, Maia 2, MTIA v4) takes 25%+ of inference workloads; China-compliant SKU revenue effectively zeroes; gross margin compresses to high-50s through cycle. Multiple compresses to 18x P/E on cyclical normalization.
Key parameters: Revenue growth FY27E 20% (vs. 30% base), then -5% FY28, -5% FY29, -5% FY30, +6% FY31; FY29E revenue $264B (vs. $380B base); FY29E EPS $5.45; gross margin FY29E 55%; operating margin FY29E 40%; software ARR exit FY28E $5B (commoditizes faster than expected); NVDA accelerator share FY29E 60%; China revenue effectively zero by FY28E; exit P/E 18x.
Triggers: (i) MSFT, META, or GOOG announces materially reduced CY2027 capex; (ii) AMD MI400 closes >50% of the performance gap and wins a Tier-1 hyperscaler training commitment; (iii) Rubin slips 6+ months due to CoWoS-L or HBM4 yield issues; (iv) further U.S. export-control tightening removes ability to ship even compliant SKUs to China; (v) a major architectural disruption (analog AI, photonic) shows credible commercial traction.
Scenario synthesis. Scenario-Weighted Fair Value (18-month): $217 (Bull $298 × 25% + Base $221 × 55% + Bear $98 × 20% = $74 + $122 + $20). At $207 spot, the Scenario FV implies +5% probability-weighted return over 18 months. Distribution: distance to bull +44%, distance to bear −53%, with the modal outcome (base $221) at +7%. The NTM method-weighted FV cross-check (DCF + comps) lands at $155 — ~25% below spot — meaning the NTM trading-multiple lens alone does not currently underwrite spot; the $10/share spread from spot to the 18-month Scenario FV is the cash-flow and multiple-expansion runway implicit in the base case. We do not publish a 12-month price target — the Scenario FV is the expensive / fair / cheap anchor, and at +5% expected upside against a −53% bear, the risk/reward supports our Neutral view rather than a Positive call.
6 · Valuation — and the Lab
We anchor our Scenario Fair Value of $217 on a blend of (i) intrinsic DCF, (ii) trading comparables on EV/EBITDA, P/E, and EV/Revenue NTM multiples, and (iii) a probability-weighted 18-month scenario blend across explicitly modelled bull / base / bear outcomes. This is a fair-value anchor, not a price target — we publish it to frame the expensive / fair / cheap read vs. the current $207 spot, not to forecast a 12-month return. Precedent transactions are not applicable at NVDA’s $3.8 trillion scale — there are no comparable M&A precedents at this size.
DCF analysis — implied $136 (base case). We discount five years of explicit-period unlevered free cash flow (FY27–FY31E: $124B / $157B / $185B / $207B / $229B) extracted from the financial model, then add a Gordon-growth terminal value at g = 4.0% (above nominal GDP, reflecting the structural AI infrastructure runway). WACC build: cost of equity 10.5% (4.5% risk-free + 1.20 × 5.0% ERP), after-tax cost of debt 4.2%, equity weight 98%, debt weight 2% — giving a base-case WACC of 10.5%. Sum of PV of explicit FCF: $656B. Gordon-growth terminal value: $3,667B; PV of TV: $2,226B (77.2% of total EV). Enterprise value: $2,881B. Plus net cash $455B → equity value $3,336B. Divided by 24,500M diluted shares → DCF implied share price of $136 (base case).
DCF sensitivity. Two-way sensitivity on WACC × terminal growth (g) maps the intrinsic value envelope. At our base (10.5% / 4.0%), implied price is $136 — ~34% below current $207 spot. Sensible bull case (9.5% / 4.5%) → $170 (still ~18% below spot); bear case (11.5% / 3.0%) → $110 (~47% below). The DCF alone implies NVDA trades meaningfully above explicit-period intrinsic value on standard assumptions — i.e., the market is pricing in cash-flow growth and multiple-expansion optionality the five-year explicit-period DCF does not capture. We view the DCF as the most conservative anchor — it under-weights multiple-expansion optionality from sovereign and enterprise AI demand and from software monetization, and is the reason DCF carries 50% method weight rather than 100%.
Comparable companies — median $176 (EV/EBITDA) / $181 (P/E). Peer set spans three reference groups: pure-play AI silicon and adjacent (AMD, AVGO, ARM, MRVL), foundry / semicap (TSM, ASML, AMAT), and hyperscale platforms NVDA sells into (MSFT, GOOGL, META). The hyperscaler trio is included because the durability of NVDA’s revenue is fundamentally a function of these customers’ capex behavior, and their multiples set the buy-side reference for AI infrastructure exposure.
Statistical summary (peers ex. NVDA, ex. ARM as high-multiple outlier): EV/EBITDA NTM median 23.0x; P/E NTM median 31.0x; EV/Rev NTM median 11.8x. Key observation: at $207 spot NVDA trades at a modest premium to peer median on EV/EBITDA NTM (~26x vs. 23.0x) and P/E NTM (~35x vs. 31.0x) — a premium that captures most of the operating outperformance NVDA delivers vs. peers (30% growth vs. 20% median, 64% EBITDA margin vs. 55%, 51% FCF margin vs. 30%). The premium is well-supported by fundamentals but no longer leaves the asymmetric discount that defined entry points earlier in the cycle; this is the central reason our headline view is Neutral rather than Positive.
Implied price from peer multiples: EV/EBITDA NTM median 23.0x × FY27E EBITDA $167B + $455B net cash = $4.27T equity / 24.5B shares = $176 per share. P/E NTM median 31.0x × FY27E EPS $5.83 = $181 per share. EV/Rev NTM median 11.8x × FY27E revenue $260B + net cash = $143 per share.
Method-Weighted Fair Value (NTM, cross-check).
| Method | Low | Base | High | Weight | Weighted FV |
|---|---|---|---|---|---|
| DCF (10.5% WACC; g 3.0–4.5%) | $110 | $136 | $170 | 50% | $68 |
| Trading comps — EV/EBITDA NTM | $141 | $176 | $230 | 30% | $53 |
| Trading comps — P/E NTM | $130 | $181 | $204 | 15% | $27 |
| Trading comps — EV/Revenue NTM | $105 | $143 | $155 | 5% | $7 |
| Method-Weighted FV | $155 | 100% | $155 |
Scenario-Weighted Fair Value (18-month). Bull $298 (25%) + Base $221 (55%) + Bear $98 (20%) = $217. This is our headline FV anchor: it is the explicit probability-weighted blend of the three scenarios modelled in Section 5.
Historical valuation bands — which multiples to actually trust for NVDA. DCF and forward peer comps tell us what NVDA is worth. Historical bands (kaamos 8-year weekly series) tell us how the market has actually paid for NVDA through cycles — and reveal a counter-intuitive read. We pick three multiples and discard the rest:
- P/E (TTM) — primary. The cleanest read for a hyper-profitable, asset-light franchise: no inventory write-downs, no capex distortions to FCF, stable earnings power. NVDA at 31x TTM trades at z = −0.94 vs an 8y mean of 65x — i.e. roughly 1σ below its own historical mean. The “NVDA is expensive vs the S&P” framing is correct, but vs its own history this is the most undemanding multiple in years.
- EV/EBITDA (TTM) — primary. Controls for the net cash position; EBITDA-to-FCF conversion is near-100% so the multiple translates straight to cash valuation. z = −0.96 (30x vs 8y mean 48x) — same story as P/E, fully confirming the read.
- P/Sales (TTM) — secondary. Useful as a scale-anchor in a growth-phase business where earnings are catching up to revenue. At ~20x vs 8y mean of 21x, z = −0.21 — essentially at the mean. The market is paying about what it has historically paid per dollar of NVDA revenue, despite higher growth + margins today.
We exclude P/B (asset-light fabless model — book is a tiny fraction of market cap, dominated by buyback timing), P/CF and P/FCF (mirror EBITDA closely for NVDA), and dividend yield (token).
The historical bands sub-section below carries the per-multiple charts; the snapshot chips at the top of Section 1 carry the z-scores. The strategic implication: the bear case requires the multiple to compress further from already-below-mean levels AND earnings to disappoint — a harder set-up than the “AI bubble” screen suggests. Combined with the DCF and peer-comp work above, this is the third independent line of evidence pointing to the same conclusion — NVDA is priced for its growth, not above it.
Headline read — Scenario Fair Value $217. The expensive/fair/cheap read: NVDA at $207 spot sits ~5% below our 18-month Scenario FV of $217 — fairly valued with a modest tilt to the upside, not the asymmetric discount the franchise carried earlier in the cycle. The Method-Weighted NTM cross-check at $155 sits ~25% below spot, indicating the NTM trading-multiple lens alone does not justify the current price — i.e., the market is paying for the 18-month cash-flow build and multiple-expansion runway that bridges $155 (NTM intrinsic) to $217 (scenario-weighted). View: Neutral — the scenario-weighted FV sits only ~5% above spot, an inadequate cushion against our −53% bear-case downside under our explicit probabilities. View shifts to Positive below ~$170 (12-month risk/reward turns clearly favorable) and to Cautious above ~$260 (the bull case is fully priced).
Sanity checks. TV % of EV: 77.2% (above 75% guideline but expected for long-duration FCF compounder). WACC in 9–14% tech band: 10.5% (conservative for scale). Implied P/E NTM at $217 18-month FV: 31.0x (in line with peer median 31.0x). Implied EV/EBITDA: 23.6x (modest premium to peer median 23.0x). NVDA growth, margins, FCF all above peer median, supporting non-negative premium.
INTERACTIVE · VALUATION LAB
Move the assumptions — watch the value
DCF value / share
$140.0
−32% vs price $207
Drag the assumptions. The grid highlights your position. Hit Reset to return to the analyst’s published inputs.
Lab parameters at print time: WACC = 10.4%, terminal growth = 4.0%, scenario = Base, DCF value / share = $140.0.
Historical valuation bands — which multiples to actually look at
NVDA at $207 trades roughly 1σ BELOW its own 8-year mean on both P/E (31x vs mean 65x) and EV/EBITDA (30x vs mean 48x). The familiar framing of 'NVDA is expensive vs the S&P' is true, but vs its own history this is the most undemanding multiple in years — the bear case requires the multiple to compress AND earnings to disappoint, a harder set-up than the screen suggests.
Why we excluded the other multiples for this name
- PB: Asset-light fabless model: book value is a tiny fraction of market cap (R&D is expensed, IP isn't on the balance sheet). P/B z-scores are dominated by buybacks, not valuation.
- PCF: Mirrors EBITDA — duplicative.
- PFCF: Similar to P/CF for NVDA; capex is modest and stable, so P/FCF tracks P/E closely.
- DIV/YIELD: Token dividend (<0.1%); never the right lens for a growth name.
7 · Company
NVIDIA Corporation (NASDAQ: NVDA) is the world’s pre-eminent designer of accelerated computing platforms and the dominant supplier of the silicon, systems, software, and networking that underpin the global build-out of generative artificial intelligence (AI) infrastructure. Headquartered in Santa Clara, California, the company is led by co-founder Jensen Huang, who has served as President and CEO continuously since incorporation in April 1993. As of the end of fiscal 2026 (year ended January 25, 2026), NVIDIA reported approximately 36,000 full-time employees globally, operates in more than 50 countries, and ships its products through a hybrid model that combines direct sales to large cloud and enterprise customers with a deep channel of original equipment manufacturers (OEMs), original design manufacturers (ODMs), and value-added resellers.
NVIDIA’s business is now overwhelmingly oriented toward data-center accelerated computing. In fiscal 2026, the Data Center segment contributed approximately $172 billion of the company’s roughly $200 billion in total revenue (≈86%), driven principally by hyperscale cloud providers, sovereign AI initiatives, and enterprise AI factories deploying the Hopper (H100/H200) and the new Blackwell (B100, B200, GB200 NVL72) generations of GPUs together with NVIDIA-designed networking (NVLink, Quantum InfiniBand, and Spectrum-X Ethernet). The remaining ~14% of revenue is split across Gaming (GeForce RTX, including the RTX 50 “Blackwell” series), Professional Visualization (RTX workstation and Omniverse), Automotive (DRIVE Orin and DRIVE Thor), and OEM/Other.
What differentiates NVIDIA from a conventional semiconductor vendor is the depth of its software stack. CUDA — first released in 2006 — is now a mature parallel-programming platform with more than 5 million registered developers, an installed base of hundreds of millions of CUDA-capable GPUs, and a library ecosystem that spans deep learning (cuDNN, TensorRT, Triton Inference Server), high-performance computing (cuBLAS, cuFFT, RAPIDS), robotics (Isaac), digital twins (Omniverse), drug discovery (BioNeMo), and full-stack inference reference designs (NIM microservices). This software gravity, combined with two decades of compiler, library, and tools investment, is the single most important structural moat in the business and the reason that even credible silicon competitors have struggled to dent NVIDIA’s accelerator share, which remained above 80% of merchant AI accelerator units shipped in calendar 2025.
The company makes money primarily by selling discrete GPUs and full server-rack systems to a concentrated set of large customers. Average selling prices on the latest Blackwell GB200 NVL72 rack-scale systems run into the millions of dollars per rack, with gross margins on Data Center products running in the high-70% range, allowing NVIDIA to compound revenue and operating profit at a pace that has no real precedent at this scale of business. In fiscal 2026, GAAP gross margin was approximately 73% and GAAP operating margin was approximately 62%, with the company generating more than $98 billion in free cash flow.
Geographically, NVIDIA’s reported billings are heavily weighted toward the United States (~50% of revenue, reflecting U.S.-domiciled hyperscalers), with Singapore (a hub for global cloud customers’ procurement entities, not end-demand), Taiwan, China (despite ongoing U.S. export controls), and the rest of the world making up the balance. End-customer demand is more globally distributed than billings suggest, with sovereign AI projects in the Middle East, Europe, India, Japan, and Southeast Asia becoming a material new buyer class in fiscal 2026.
At the start of June 2026, NVIDIA’s market capitalization stood at approximately $3.8 trillion, making it the largest or second-largest company in the world by market value (alternating with Microsoft and Apple). The stock is a component of the S&P 500 and Dow Jones Industrial Average, and one of the most heavily-owned positions in both passive index funds and active growth portfolios globally.
The investment debate on NVIDIA is no longer about whether AI infrastructure spend is real — that question has been answered emphatically by three consecutive years of capacity-constrained demand — but rather about three forward-looking issues: (1) the durability and shape of the AI capex cycle, particularly whether 2027–2028 will see digestion as inference workloads commoditize; (2) the rate at which custom silicon (Google TPU, AWS Trainium, Microsoft Maia, Meta MTIA) displaces NVIDIA in inference workloads at the largest buyers; and (3) margin sustainability as Blackwell ramps, Rubin succeeds it, and a growing share of the system bill-of-materials goes to memory (HBM3e/HBM4 from SK Hynix, Micron, and Samsung) rather than to NVIDIA’s own value-add.
Brief history. NVIDIA was founded April 5, 1993, by Jensen Huang, Chris Malachowsky, and Curtis Priem at a Denny’s restaurant in San Jose, California. After a near-death experience with the NV1 in 1995, the 1999 release of the GeForce 256 — marketed as “the world’s first GPU” — established NVIDIA as the consumer 3D graphics leader. The pivotal strategic decision was the 2006 launch of CUDA, a general-purpose parallel-computing platform whose value compounded silently for years until Alex Krizhevsky’s 2012 AlexNet — trained on two consumer GeForce GTX 580 cards — ignited the deep-learning revolution. The 2020 acquisition of Mellanox added InfiniBand networking. The November 2022 ChatGPT moment transformed the business: fiscal 2024 grew 126% to $60.9B, fiscal 2025 grew 114% to $130.5B, and fiscal 2026 grew 53% to $200B. Over this three-year stretch, market cap grew from under $400B to over $3.8 trillion.
Colette Kress — EVP & CFO
CFO since 2013; ex-Cisco SVP/CFO Business Services Division; 13 years at Microsoft (CFO Server & Tools).
CFO since 2013; ex-Cisco SVP/CFO Business Services Division; 13 years at Microsoft (CFO Server & Tools).
Debora Shoquist — EVP, Operations
Owns global supply chain through HBM/CoWoS constraints; ex-JDSU, 26 years at HP.
Owns global supply chain through HBM/CoWoS constraints; ex-JDSU, 26 years at HP.
Business mix & divisional economics
| Division | FY26A rev | FY31E rev | CAGR | Est. op margin |
|---|---|---|---|---|
| Data Center | $171.9B | $409.8B | +19% | ~66% |
| Gaming | $15.0B | $22.5B | +8% | ~34% |
| Pro Visualization | $2.4B | $4.2B | +12% | ~38% |
| Auto & Robotics | $2.0B | $11.7B | +42% | ~12% |
| OEM & Other | $8.6B | $20.2B | +19% | ~22% |
Products & Services
NVIDIA’s product portfolio spans five major platforms — Data Center, Gaming, Professional Visualization, Automotive, and Networking/Software — but the economic center of gravity has decisively shifted toward Data Center accelerated computing, where the company sells silicon, full systems, networking, and software as a vertically integrated stack.
Data Center GPUs. The flagship product line is the Hopper architecture (H100, H200) and the new Blackwell architecture (B100, B200, GB200 Grace Blackwell Superchip, and the GB200 NVL72 rack-scale system). H100, launched in volume in late calendar 2022, remains in active production at depreciating but still-elevated pricing as customers fill out clusters and deploy inference workloads. H200, an HBM3e refresh of Hopper launched in mid-2024, extends Hopper’s life into 2026. Blackwell, the architecture announced at GTC 2024 and shipping in volume from Q4 calendar 2024, more than doubles training throughput per dollar versus H100 and is the primary growth driver for fiscal 2026 and fiscal 2027 revenue. The GB200 NVL72 system — a single liquid-cooled rack containing 36 Grace CPUs and 72 Blackwell GPUs interconnected by fifth-generation NVLink — is sold at multi-million-dollar ASPs and represents a fundamental shift in how NVIDIA monetizes silicon: the customer increasingly buys an integrated AI “factory” rather than a GPU, with rack-scale ASPs roughly 5–7x the equivalent loose-GPU configuration.
The next-generation Rubin platform, announced at GTC March 2026, is slated for sampling in late calendar 2026 and volume in 2027, with successor cadence (Rubin Ultra, then a third platform codenamed “Feynman”) on a roughly annual rhythm. The acceleration of NVIDIA’s product cadence from a two-year to an effective one-year rhythm is itself a competitive moat: each new generation resets the price-performance bar before custom silicon designs can catch up.
Networking. Through the Mellanox acquisition, NVIDIA controls two of the three credible AI fabric technologies: NVLink (proprietary GPU-to-GPU interconnect, 1.8 TB/s in the Blackwell generation) and Quantum-X InfiniBand (the dominant fabric for the largest training clusters). NVIDIA also offers Spectrum-X Ethernet, an AI-optimized Ethernet platform that competes directly with Broadcom’s Tomahawk and Arista’s offerings. Networking is a roughly $24B business inside the Data Center segment (FY26A) and grew faster than compute revenue in fiscal 2026 as customers built larger clusters with higher fabric content per GPU.
CUDA and AI software. NVIDIA’s software stack — CUDA, cuDNN, TensorRT, Triton, Megatron, NeMo, RAPIDS, Isaac, Omniverse, and the NIM inference microservices — has historically been distributed largely for free, with monetization captured in hardware ASPs. Beginning in fiscal 2025, NVIDIA began monetizing AI Enterprise (a curated, supported version of the stack) at approximately $4,500 per GPU per year, with NIM microservices priced on a similar per-GPU basis. Software revenue is not yet broken out separately but management commentary suggests it is approaching a multi-billion-dollar annual run rate and growing rapidly. The strategic significance is much larger than the near-term revenue: software monetization, if it scales, validates a recurring-revenue layer on top of the hardware business and provides additional margin upside.
Gaming. GeForce remains a roughly $15 billion annual business and the platform on which NVIDIA’s brand was built. The current generation, the RTX 50-series (“Blackwell” gaming derivatives), launched in early calendar 2026 with substantial AI-enhanced rendering features (DLSS 4 with multi-frame generation, RTX Neural Shaders). Gaming gross margins are lower than data center but still well above peer averages, and the segment generates significant cash flow.
Professional Visualization. RTX workstation products (formerly Quadro) and the Omniverse digital twin platform serve the architecture, engineering, manufacturing, and media/entertainment markets. The segment generates approximately $2.4 billion in annual revenue and serves as an early-stage on-ramp for industrial AI applications.
Automotive and Robotics. The DRIVE platform (DRIVE Orin in production, DRIVE Thor ramping) powers ADAS and autonomous driving in vehicles from Mercedes-Benz, Volvo, Polestar, Lucid, Hyundai, BYD, Li Auto, XPeng, and others. Annual revenue is approximately $2.0 billion (FY26A) and growing rapidly as Thor begins to ramp; we model the segment growing to $11.7B by FY31E. The Isaac robotics platform and Jetson edge-AI modules round out the embedded portfolio. Management has framed physical AI and robotics as the next major application platform, but commercialization at scale is still in early innings.
Pricing and deal sizes. Data Center deal sizes range from a few hundred thousand dollars for a single DGX appliance to multi-billion-dollar sovereign AI build-outs. Hyperscaler quarterly purchases have routinely run in the $5–10 billion range per customer. Gaming GPUs retail in the $300–2,500 range. Automotive contracts are multi-year design wins with revenue recognized over vehicle production cycles.
Customers & Go-to-Market
NVIDIA’s customer base is highly concentrated at the top: the four largest U.S. hyperscalers — Microsoft, Meta, Alphabet, and Amazon — together with Oracle and the Tier-1 GPU-as-a-service neoclouds (CoreWeave, Lambda, Crusoe, Nebius) account for an estimated 55–65% of fiscal 2026 Data Center revenue. The 10-K discloses that one customer represented approximately 19% of fiscal 2026 revenue and that three customers each exceeded 10%. Customer concentration is a structural risk that NVIDIA partially offsets by cultivating a broad base of secondary buyers.
Sovereign AI customers are an entirely new buyer class that emerged in fiscal 2025–2026. Multi-billion-dollar commitments include Saudi Arabia (Humain), the UAE (G42), France (Mistral and the broader Iliad/Free build-out), Japan (a SoftBank-led consortium with KDDI and NTT participation), Singapore, India, South Korea, Norway, Sweden, and Indonesia. These customers tend to buy mid-size clusters (4,000–32,000 GPUs) on multi-year payment plans backed by sovereign credit. We estimate sovereign AI bookings of $80–120B cumulative across FY27–31E, materially diversifying NVDA’s customer base.
Enterprise AI customers buy through NVIDIA’s DGX SuperPOD reference architectures or via systems integrators (Dell, HPE, Supermicro, Lenovo, Cisco). The enterprise wave is earlier in adoption than hyperscaler buying but is expected to compound over the next 3–5 years.
Tier-2 GPU clouds increasingly buy directly to fill capacity for AI startups that cannot get committed allocations from the hyperscalers. Gaming customers purchase GeForce GPUs through retail channels (Best Buy, Amazon) and OEM-integrated systems from ASUS, MSI, Gigabyte, Lenovo, HP, and Dell.
Go-to-market motion. NVIDIA blends three channels. First, a direct enterprise/strategic sales force, expanded materially in fiscal 2025–2026, calls on hyperscaler CTOs and sovereign decision-makers and increasingly negotiates multi-quarter supply commitments. Second, a deep OEM/ODM channel manufactures and integrates NVIDIA reference designs (HGX boards, MGX modular reference designs, DGX appliances, GB200 NVL72 racks). Third, a software-led developer motion — through GTC conferences, university programs, the Inception startup network, and the Deep Learning Institute — seeds the next generation of CUDA-native workloads.
Sales cycles vary materially by segment. Hyperscaler purchase commitments are negotiated 6–18 months ahead of delivery; sovereign deals take 6–12 months to structure; enterprise deals run 3–9 months; channel/retail is immediate. NVIDIA does not publish a backlog figure but management commentary has repeatedly described Data Center demand as “significantly exceeding supply” for the entirety of fiscal 2024–2026, with allocation rather than demand generation being the principal commercial activity.
Reference customers the company routinely cites include the xAI Colossus cluster (built with 100,000+ H100/H200 GPUs in 19 days), Meta’s “Llama factories” in Louisiana and Ohio, Microsoft’s Azure-hosted OpenAI training clusters, Tesla’s Dojo-adjacent training fleet, and the Stargate joint venture between OpenAI, Oracle, SoftBank, and MGX (Abu Dhabi), which is expected to consume tens of billions of dollars of NVIDIA capacity over its multi-year build-out.
Industry Overview
NVIDIA operates at the intersection of three industry verticals: semiconductors, data-center infrastructure, and AI software/platforms. The most relevant categorization for the company’s economics today is the merchant AI accelerator and AI infrastructure market, which has emerged as the fastest-growing major segment of the global technology economy.
Market size and growth. Worldwide spending on AI infrastructure — encompassing accelerators, server systems, networking, storage, and data-center facilities — totaled approximately $410 billion in calendar 2025 according to triangulated estimates from IDC, Dell’Oro Group, and Omdia, up from roughly $250 billion in 2024 and under $100 billion in 2022. Consensus sell-side forecasts project the market reaching $700–900 billion by 2027 and $1.2–1.7 trillion by 2030, implying a 25–30% CAGR over the next five years. Within this, the merchant AI accelerator market (GPUs and AI ASICs sold by chip vendors) is approximately $250 billion in 2025 and is projected to reach $500–700 billion by 2030. The hyperscaler AI capex cycle — Microsoft, Meta, Alphabet, Amazon, and Oracle together — exceeded $400 billion in calendar 2025 and is guided to grow further in 2026.
Growth drivers. The fundamental drivers are (1) the scaling laws of large language and multimodal models, which continue to deliver measurable capability gains with each order of magnitude increase in training compute; (2) the emergence of inference-time compute scaling (chain-of-thought, agentic reasoning), which makes inference itself a compute-intensive workload rather than a marginal one; (3) the proliferation of AI agents in enterprise workflows; (4) sovereign AI as a category, where national governments are building domestically-controlled AI infrastructure for security, language, and economic-policy reasons; and (5) the application of AI to physical systems — robotics, autonomous vehicles, drug discovery, and digital twins — which is expected to drive a second wave of demand from 2027 onward.
Industry structure. The accelerated computing market is structurally consolidated at the silicon layer. NVIDIA holds an estimated 80–85% share of merchant AI accelerator units shipped in 2025, with AMD a distant second (10–14%) and Intel marginal in the AI accelerator category. Custom silicon designed by hyperscalers and produced by foundries — Google TPU (designed with Broadcom), AWS Trainium/Inferentia (designed with Marvell), Microsoft Maia, Meta MTIA — has grown into an estimated $40–50 billion category in 2025 and is the most credible long-term competitive threat to NVIDIA’s share. Below the accelerator layer, the industry depends on a small number of critical suppliers: TSMC for advanced-node manufacturing (N5/N4/N3 today, N2 for Rubin), SK Hynix/Micron/Samsung for HBM3e and HBM4 memory, ASE/Amkor and TSMC’s CoWoS for advanced packaging, and a concentrated set of substrate and optical-component suppliers.
Barriers to entry. Entering the merchant AI accelerator market at scale requires four things in combination, each of which is itself a multi-year endeavor: (i) a competitive silicon design at leading-edge process nodes; (ii) HBM and CoWoS capacity allocation, which is constrained for years out; (iii) a credible software ecosystem with framework support, libraries, and developer mindshare; and (iv) a system-level reference architecture, including networking, that customers can integrate at hyperscale. NVIDIA’s two-decade head start on (iii) is the most difficult to replicate; AMD has made progress with ROCm but remains a smaller ecosystem, and the hyperscaler custom-silicon teams have viable software stacks for their internal workloads only.
Regulatory environment. The semiconductor industry is now meaningfully shaped by U.S.–China geopolitical considerations. The U.S. Department of Commerce’s Bureau of Industry and Security has progressively tightened export controls on advanced AI chips and chipmaking equipment to China since October 2022, with major updates in October 2023, December 2024, and January 2026. NVIDIA has responded with a series of compliant China-only SKUs (A800, H800, H20, B20, B30, CW30) whose performance has been deliberately limited to comply with successive thresholds. The CHIPS and Science Act has incentivized U.S. fab investment by TSMC, Samsung, Intel, and Micron, modestly diversifying advanced-node manufacturing capacity but with leading-edge production still concentrated in Taiwan. EU AI Act regulations, in effect since 2024, affect downstream applications but not the silicon layer directly.
Industry dynamics. The industry exhibits classic Moore’s Law-like cost-curve declines at the workload level — the cost to train a frontier model has fallen by roughly 75% per year on a token-equivalent basis — even as the total dollar spend rises. This is the central tension in the bull/bear debate on NVIDIA: aggregate compute demand is growing faster than per-unit cost declines, so dollar spend continues to rise, but if the curve crosses — i.e., if cost-per-token declines outpace demand-side growth — the AI accelerator market could enter a digestion phase. Most participants believe such a digestion phase is more likely in 2027–2028 than in 2026.
Competitive Landscape
NVIDIA’s competitive landscape can be organized into three tiers: direct merchant accelerator competitors, hyperscaler custom silicon, and adjacent threats in networking, software, and emerging architectures.
Tier 1 — Merchant accelerator competitors.
Advanced Micro Devices (AMD) is NVIDIA’s only credible direct merchant competitor. Its MI300X (Instinct), launched in late 2023, gained meaningful traction with Microsoft Azure, Meta, and Oracle for specific inference workloads where HBM capacity per GPU was a competitive advantage. The MI325X (2024), MI355X (2025), and the new MI400-series (2026) have continued the cadence. AMD’s data-center GPU revenue is expected to exceed $12–15 billion in calendar 2026, up from approximately $5 billion in 2024, representing a roughly 8–12% share of merchant accelerator dollars. The ROCm software stack has improved materially but remains a generation behind CUDA in framework support, library maturity, and developer mindshare. AMD’s strategic position is strongest in inference, where workload heterogeneity rewards a credible second source.
Intel. Intel’s Gaudi 3 (Habana) accelerator, launched in 2024, has not achieved meaningful share against the H100/MI300X duopoly. Intel’s stated strategy has shifted toward a multi-pronged AI approach combining Gaudi, Xeon CPUs with embedded AI acceleration (AMX), and a forthcoming Falcon Shores GPU. Intel’s accelerator business is sub-$1 billion in 2025 and not yet a meaningful competitive threat at the high end, though Xeon CPUs continue to be present in essentially every AI server as the host CPU.
Tier 2 — Hyperscaler custom silicon.
Google TPU, designed by Google with Broadcom on packaging and Marvell on networking, is the most mature hyperscaler custom-silicon program. TPU v5p, v5e, and the new Trillium (v6) and Ironwood (v7, 2026) chips power the majority of Google’s internal training and a large share of Google Cloud’s external AI inference workloads, including Anthropic Claude training under a multi-billion-dollar TPU commitment. Google does not sell TPUs commercially but offers them via Google Cloud, where they compete directly with NVIDIA-on-GCP. TPU is the most credible long-term threat to NVIDIA’s training share at the largest buyers.
AWS Trainium 2 (production 2024) and the new Trainium 3 (2026), designed in conjunction with Annapurna Labs and Marvell, target both training and inference workloads. Anthropic has committed to large-scale Trainium 3 capacity as part of its multi-billion-dollar AWS partnership. AWS continues to be NVIDIA’s largest single customer by reported revenue but is simultaneously its most strategically ambitious in-house silicon competitor.
Microsoft Maia (announced late 2023, ramping 2025–2026) is designed for OpenAI workloads on Azure. Maia is earlier-stage than TPU or Trainium but Microsoft’s commitment is significant, and Maia 200 (2027) is reportedly designed to handle a meaningful share of Azure’s GPT-class inference workload. Meta MTIA v2/v3 target Meta’s recommendation, ranking, and Llama inference workloads — a meaningful share of Meta’s internal compute, though MTIA has not displaced its NVIDIA training fleet.
Tier 3 — Adjacent and emerging competitors.
Broadcom is the silent giant in AI: its networking ASICs (Tomahawk 6 Ethernet switch silicon, Jericho 4 routing silicon) power most of the Ethernet AI fabrics in the world, and its custom-silicon design business is the silicon partner behind Google TPU and several other hyperscaler programs. Broadcom is a partner, supplier, and competitor to NVIDIA simultaneously, and is the principal beneficiary of any shift in AI fabric share from InfiniBand to Ethernet.
A long tail of well-funded AI silicon startups targets specific workloads: Cerebras (wafer-scale training), Groq (LPU inference for latency-sensitive workloads), SambaNova (dataflow training/inference), Tenstorrent (RISC-V plus AI accelerators). Collectively these vendors represent <3% of merchant accelerator revenue but capture mindshare for specific use cases.
Huawei (Ascend) 910C and the upcoming 910D/920 are the dominant AI accelerators within China for buyers that cannot procure even the export-compliant NVIDIA SKUs. Huawei’s domestic share has grown materially in 2025–2026 and represents a structural ceiling on NVIDIA’s China revenue irrespective of export-control resolution.
Competitive advantages — NVIDIA’s moat. NVIDIA’s defensible advantages are (1) the CUDA software ecosystem; (2) the vertically integrated NVLink/InfiniBand fabric; (3) annual product cadence at leading-edge nodes; (4) rack-scale reference architectures (GB200 NVL72) that capture more value per deployed GPU; (5) a developer community of 5+ million; and (6) a customer-engagement model that has elevated Jensen Huang to a peer of hyperscaler CEOs and heads of state.
Competitive vulnerabilities. NVIDIA is exposed where (1) inference economics matter more than peak performance — i.e., where the marginal cost per token at scale, not per-GPU performance, drives buying; (2) the workload is owned by a hyperscaler with its own silicon; (3) export controls or geopolitical considerations make NVIDIA procurement difficult; and (4) a buyer values open-standard Ethernet fabrics over InfiniBand. Each of these vulnerabilities is the principal long-thesis pushback on the stock.
Market Opportunity
NVIDIA’s serviceable opportunity can be sized at three levels: (1) the merchant AI accelerator market, (2) the broader AI infrastructure stack (accelerators + CPUs + networking + memory + storage + facilities), and (3) the long-tail enterprise and physical-AI software opportunity.
TAM sizing. At the silicon layer, the merchant AI accelerator TAM is estimated at $250 billion in 2025 and is projected to reach $500–700 billion by 2030. NVIDIA’s effective serviceable share, given export-control-driven exclusion from a meaningful share of China demand, is approximately $200–230 billion of the 2025 TAM. Including networking, NVIDIA’s full silicon-and-networking SAM is roughly $290 billion in 2025 and grows to $750 billion–$1 trillion by 2030 in consensus scenarios.
The broader AI infrastructure stack TAM — i.e., the dollars spent by hyperscalers, neoclouds, sovereigns, and enterprises on AI factories — is estimated at $410 billion in 2025 and is projected to reach $1.2–1.7 trillion by 2030. NVIDIA captures roughly 35–45% of this broader TAM today (compute + networking + a sliver of software), with the balance going to CPUs (Intel, AMD), memory (Hynix, Samsung, Micron), storage, facilities (utilities, real estate, cooling), and integration services.
The enterprise AI software and platform opportunity, captured via NVIDIA AI Enterprise and NIM microservices, is sized by management at several hundred billion dollars on a multi-year view, though the near-term recurring software revenue is sub-$10B. Physical AI (robotics, autonomous vehicles, industrial digital twins) is a separate multi-hundred-billion-dollar opportunity that is largely pre-revenue today.
Penetration strategy. NVIDIA’s playbook is to (1) lead each new generation of accelerated computing by 12–24 months, (2) bundle silicon with networking and software, (3) move up the stack into rack-scale systems and full AI factories, (4) monetize the software ecosystem incrementally via AI Enterprise and NIM, and (5) build new buyer classes — sovereigns, enterprises, robotics OEMs — to diversify away from hyperscaler concentration. The most important strategic vector is sovereign AI, which has emerged as a structurally less-price-sensitive buyer class than hyperscalers and which extends the duration of the current capex cycle.
Share outlook. A reasonable base case is that NVIDIA’s share of merchant AI accelerator units settles between 70% and 80% over the next 3–5 years (down modestly from 80–85% today) as AMD scales and hyperscaler custom silicon gains in inference. Even at the lower end of that range, the dollar TAM growth more than compensates: 70% share of a $600B 2030 TAM is roughly 2.0–2.5x the dollar revenue that 85% share of today’s $250B TAM represents.
Implication for the investment case. The TAM expansion alone — independent of any share-shift assumptions — supports our base-case forecast that NVDA revenue grows from $200B (FY26A) to $468B (FY31E). The bull case is principally about NVDA holding 80%+ share in a faster-growing TAM ($700B+ by 2030). The bear case is principally about NVDA losing share to custom silicon in a digesting TAM (digesting hyperscaler capex with custom-silicon share gain). In all three scenarios, the dollar opportunity is large enough to support the investment case at current multiples; the debate is about durability and pace.
8 · Risks to the Target
We organize risks into four categories: company-specific, industry/market, financial, and macroeconomic. The principal near-term risk is hyperscaler AI capex digestion in 2027–28; the principal long-term risk is custom silicon displacing NVDA in inference at the largest buyers.
Company-specific risks.
Customer concentration. Top-4 hyperscalers contribute an estimated 55% of Data Center revenue in FY26A. One customer disclosed at 19% of total revenue per the FY26 10-K. Multi-year pause or pivot to custom silicon by any one customer could compress quarterly results. Sovereign and enterprise growth partially offsets concentration, but it remains a multi-year structural reality.
Hyperscaler custom silicon. Google TPU, AWS Trainium, Microsoft Maia, and Meta MTIA are each multi-billion-dollar in-house alternatives, growing in inference at their parent companies. A successful generation prioritized over NVIDIA buy commitments would directly compress NVDA’s incremental dollar growth. We view this as a real but slow-moving threat; CUDA portability and rack-scale value capture provide multi-year offset.
Product execution risk. Annual cadence requires flawless execution on Rubin (sampling CY2026) and Rubin Ultra (2027). Any slip on CoWoS-L packaging yields, HBM4 ramp, or TSMC N2 node maturity gives competitors an opening. Complexity of NVL72 and planned NVL576 rack systems amplifies execution risk at the system level.
HBM and CoWoS supply. NVIDIA depends on a small base of critical suppliers — TSMC (foundry), SK Hynix/Micron/Samsung (HBM), TSMC/ASE (CoWoS). HBM and CoWoS capacity is binding through FY27. A natural disaster, major yield issue, or supply allocation shift constrains revenue regardless of demand. Capacity expansion is multi-year.
Key person risk. Jensen Huang is unusually integral to NVIDIA’s identity, strategy, and customer relationships. Any extended absence or departure would be a meaningful overhang on the stock. Deep bench (Kress, Shoquist, Puri, Catanzaro, Buck) mitigates operational risk but not the symbolic weight.
Industry / market risks.
AI capex digestion (2027–28). Hyperscaler AI capex grew >70% in 2025 to over $400B. A digestion phase — flat or declining capex in 2027–28 — is the principal bear case. Incremental sovereign and enterprise demand partially offsets, but a cyclical capex pause still compresses NVDA’s growth rate from 24% (FY28E) toward low-single-digits in our bear scenario.
Inference commoditization. Inference workloads are more price-sensitive than training and more amenable to specialized architectures (Groq LPU, Cerebras, hyperscaler ASICs). If inference economics shift decisively to non-NVIDIA accelerators, NVDA’s share at the largest buyers compresses materially.
Ethernet vs. InfiniBand fabric shift. A broader shift toward open Ethernet AI fabrics (Ultra Ethernet Consortium, Broadcom Tomahawk, Arista) erodes NVDA’s networking attach rates and dilutes the rack-level ASP captured per GPU. Spectrum-X partially hedges, but InfiniBand is the higher-margin franchise.
Technology disruption. A material architectural breakthrough — analog AI, photonic computing, neuromorphic, or extreme-sparsity software approaches — could disrupt the GPU-centric paradigm. Probability low on a 3-year view but non-zero on a 5–10 year view.
Financial risks.
Margin sustainability. Data Center gross margins are in the high-70% range, above historical semiconductor norms. As Blackwell ramps fully and memory content per system rises, gross margins moderate. Management has guided to mid-70s medium-term; faster-than-expected compression weighs on earnings.
Inventory and supply commitment risk. Long lead-times on Blackwell and Rubin force NVIDIA to commit substantial wafer, HBM, and CoWoS capacity in advance. Demand softening could result in inventory write-downs or supply-commitment penalties.
Macroeconomic risks.
Interest rate / discount rate sensitivity. NVDA trades at a premium multiple driven by long-duration cash flow expectations. Sustained rise in long real rates compresses the multiple even with unchanged fundamentals.
Geopolitical and export controls. Successive U.S. export controls have restricted advanced AI chip sales to China (historically >20% of revenue). Recent China-compliant SKUs (H20, B30, CW30) generate limited revenue at lower margin; Huawei domestic substitution is structural. Further tightening could zero China data-center revenue.
Power availability. AI factories are constrained by electricity availability as much as silicon supply. Slower grid build-out in U.S., Europe, parts of Asia could delay customer capex deployment and extend project timelines.
View triggers. Our view would shift to Cautious if (i) the 18-month Scenario Fair Value falls below the current share price, OR (ii) custom-silicon share at top-4 hyperscalers exceeds 30% of inference workloads by FY28E. Our view would strengthen if (i) software ARR exceeds $15B exit run-rate FY27E, AND (ii) sovereign AI bookings exceed $50B cumulative by end of FY27E.
Catalysts to watch
- Rubin platform sampling (CY2026 H2) — first independent ASP / performance data points
- Q3 FY27 earnings (Aug 2026) — Blackwell GB200 run-rate + sovereign AI bookings
- Hyperscaler 2027 capex guidance (Jan–Feb 2027) — single largest re-rating catalyst
- Explicit AI Enterprise / NIM software ARR ≥$10B run-rate disclosure
- China export-control resolution (any direction) — removes multi-quarter overhang
Upcoming events
- 2026-08-27 — Q2 FY27 earnings (approx.) (First full quarter of Blackwell GB200 NVL72 at peak production cadence)
- 2026-11-19 — Q3 FY27 earnings (approx.) (Rubin sampling commentary expected)