Dark Stone Capital | February 2026

There's a popular narrative circulating right now that NVIDIA's dominance is about to crack. Google has its TPUs. Amazon has Trainium. Meta is reportedly shopping for alternatives. Custom silicon is the future, the bears say, and NVIDIA's 75% gross margins are living on borrowed time.

They're wrong — or at least, they're early by several years. And in markets, early is the same as wrong.

Here's the case for why NVIDIA's competitive position is more durable than the consensus appreciates, why the margin structure is defensible through 2027, and why the data center buildout cycle is nowhere near peaking.

The CUDA Ecosystem: 18 Years of Compounding Lock-In

NVIDIA's moat isn't silicon. It's software.

CUDA — Compute Unified Device Architecture — launched in 2007 as a speculative bet on general-purpose GPU computing. There was no AI market. Deep learning was a fringe academic pursuit. But Jensen Huang committed to building out developer tools, documentation, and university partnerships for a market that didn't exist yet. That 18-year head start is now the single most defensible asset in the semiconductor industry.

The numbers tell the story. Over 4 million developers worldwide build on CUDA. The CUDA Toolkit has been downloaded more than 450 million times. There are over 3,000 GPU-accelerated applications optimized for NVIDIA hardware, and every major AI framework — PyTorch, TensorFlow, JAX — is optimized for CUDA first. The ecosystem includes specialized libraries like cuDNN for deep learning primitives, TensorRT for inference optimization, and NCCL for multi-GPU communication. Each of these represents thousands of engineering hours of performance tuning that competitors simply cannot replicate overnight.

This creates a textbook flywheel effect: more developers attract better software optimization, which drives superior real-world performance, which sells more hardware, which funds more R&D into the software stack. It's a self-reinforcing cycle that widens the gap with every rotation.

The switching costs are enormous and underappreciated by analysts who focus exclusively on hardware specs. Rewriting an existing CUDA codebase for AMD's ROCm or another platform can take months of engineering time and cost hundreds of thousands of dollars per project. For large enterprises with years of accumulated CUDA-optimized workloads, migration isn't a weekend project — it's an organizational transformation. SemiAnalysis put it bluntly in late 2024: the CUDA moat has yet to be crossed by AMD.

This is why NVIDIA commands 80-95% of the AI accelerator market depending on the segment and methodology. It's not because the hardware is always dramatically faster on paper. It's because the total cost of ownership — factoring in developer productivity, software maturity, library support, and deployment friction — overwhelmingly favors the NVIDIA stack.

Competition Analysis: Real Threats and Paper Tigers

Let's be honest about the competitive landscape rather than dismissive. There are real challengers, but the threat timeline matters enormously for investment decisions.

AMD (ROCm / Instinct MI350-MI500 roadmap)

AMD is the most credible GPU competitor. The MI300X was competitive on paper — 192GB HBM3 memory versus H100's 80GB — and AMD has landed real deployments at Meta, Microsoft, and xAI. ROCm 7.0 delivered meaningful improvements with day-zero support for major models. The OpenAI strategic partnership (6GW) and Oracle's 50,000-unit MI450 order signal genuine traction.

But context matters. AMD's data center revenue hit $4.3 billion in Q3 2025. That same quarter, NVIDIA posted $51.2 billion in data center revenue. That's a 12:1 gap. Third-party benchmarks still show NVIDIA delivering 10-30% better real-world training performance on many workloads — not from hardware superiority, but from software maturity. AMD's gross margins sit around 52%, versus NVIDIA's 73-75%. They're competing on price because they have to, not because they want to.

AMD's roadmap is aggressive — MI450 "Helios" in Q3 2026, MI500 in 2027 — but each new generation has to close both a hardware and software gap simultaneously. The risk for AMD investors isn't that they fail; it's that they succeed just enough to maintain a permanent #2 position in a market NVIDIA continues to expand.

Google TPUs (Ironwood / TPU v7)

Google is the most technically advanced custom silicon player, with a decade of TPU development behind it. Ironwood (TPU v7) features 192GB HBM3e and 7.4 TB/s bandwidth, competing directly with Blackwell-class architectures. Google reports that over 75% of Gemini model computations now run on internal TPUs. The reported Meta deal — billions of dollars in TPU capacity starting mid-2026 — was the first real signal that Google's chips could become a platform, not just an internal tool.

But TPUs remain single-purpose and Google Cloud-locked. They can't run HPC simulations, general-purpose scientific computing, or workloads requiring flexible execution models. NVIDIA chips run across every cloud, on-premise, at the edge, and on local workstations. That flexibility is a structural advantage that custom silicon can't replicate by definition. Google's $185 billion capex plan for 2026 is massive, but it's a bet on vertical integration within their own ecosystem, not a general-purpose threat to NVIDIA's TAM.

Amazon Trainium / Microsoft Maia

AWS's Trainium2 is in production with Anthropic's Project Rainier (500,000+ chips), and Trainium3 was announced in late 2024. These are real deployments at real scale. But Trainium is optimized for AWS's ecosystem and specific customer workloads, not general-purpose AI development.

Microsoft's story is more cautionary. Maia 100 remained in internal testing for over two years. The next-generation Maia 200, originally scheduled for 2025, was delayed to 2026 and reportedly suffered significant talent attrition — 20% of the team departed. Microsoft just unveiled Maia 200 in late January 2026, claiming strong benchmarks, but it remains to be seen if it can scale beyond internal workloads. Meanwhile, 70% of Azure AI still runs on NVIDIA hardware.

The Custom Silicon Verdict

JPMorgan projects custom chips will account for 45% of the AI chip market by 2028, up from 37% in 2024. That sounds bearish for NVIDIA until you realize two things: first, the total market is expanding so rapidly that NVIDIA's absolute revenue can grow even as its relative share contracts; and second, custom silicon primarily displaces NVIDIA in inference-heavy internal workloads at hyperscalers — not in the broader enterprise, sovereign AI, or startup ecosystems where CUDA lock-in is strongest.

As Bank of America's Vivek Arya put it, NVIDIA has "managed to consistently expand the size of the market." The pie is growing faster than any individual slice is shifting.

The Margin Structure: Why 73-75% Gross Margins Are Defensible

NVIDIA's gross margins are the most misunderstood number in semiconductor investing. Bears see 73-75% and assume mean reversion is inevitable. But these margins aren't a temporary anomaly — they reflect the structural economics of selling an integrated platform, not just a chip.

Here's the quarterly trajectory:

• FY25 Q3 (Oct 2024): 75.0% non-GAAP gross margin
• FY26 Q1 (Apr 2025): 61.0% (distorted by $4.5B H20 China export charge; 71.3% adjusted)
• FY26 Q2 (Jul 2025): 72.7%
• FY26 Q3 (Oct 2025): 73.6%
• FY26 Q4 guidance: 75.0%

Management has consistently guided toward mid-70% gross margins as the normalized target, and the trajectory since the H20 disruption confirms they're on track.

Compare this to the competition:

• AMD: ~52% gross margin (Q3 2025)
• Intel: ~30% gross margin (declining from historical peaks)
• NVIDIA: 73-75% and expanding

The 20+ point margin gap between NVIDIA and AMD isn't closing — it's structural. NVIDIA doesn't just sell GPUs. It sells NVLink interconnects, BlueField DPUs, Grace CPUs, and increasingly, complete rack-scale systems like the NVL72. The Blackwell generation accelerated the shift from selling individual chips to selling integrated AI supercomputers. This system-level approach captures more value per dollar of customer spend and creates integration barriers that make it harder for competitors to substitute individual components.

NVIDIA's pricing power comes from two sources: CUDA ecosystem lock-in (customers can't easily switch even if cheaper hardware exists) and performance-per-watt leadership (in power-constrained data centers, energy efficiency directly translates to revenue capacity). As long as both hold — and there's no indication either is eroding — the margin structure is sustainable.

The real risk to margins isn't competition. It's NVIDIA's own product cadence. Jensen Huang calls himself "the chief revenue destroyer," deliberately accelerating the obsolescence cycle to an annual architecture release cadence: Blackwell in 2025, Vera Rubin in Q3 2026 with HBM4 support, and Rubin Ultra in H2 2027. Each generation resets the performance bar and justifies premium pricing, but transition periods can temporarily compress margins as seen in Q1 FY26.

Data Center Buildout: The $1.15 Trillion Wave Through 2027

The capex cycle supporting NVIDIA's revenue is not peaking. It's accelerating.

Goldman Sachs projects total hyperscaler capex from 2025-2027 will reach $1.15 trillion — more than double the $477 billion spent from 2022-2024. CreditSights estimates the top five hyperscalers (Amazon, Alphabet, Microsoft, Meta, Oracle) will spend approximately $602 billion in 2026 alone, up 36% from ~$443 billion in 2025. Roughly 75% of that spend — approximately $450 billion — is directly tied to AI infrastructure: servers, GPUs, data centers, and related equipment.

The individual commitments are staggering:

• Amazon: $125 billion capex guidance for 2025, added 3.8 GW of capacity in 12 months
• Google/Alphabet: $185 billion capex plan for 2026 (record-breaking)
• Microsoft: ~$80 billion in 2025, with fiscal 2026 guidance suggesting further acceleration
• Meta: $64-72 billion in 2025, with $30+ billion incremental growth expected in 2026

And this doesn't capture the full picture. OpenAI alone has announced $1.4 trillion in data center ambitions through the Stargate Project. The "neocloud" providers — CoreWeave, Nebius, Lambda, Crusoe — are adding billions more. Sovereign AI buildouts are accelerating globally, with countries recognizing AI infrastructure as essential national capability.

The key insight for NVIDIA investors: hyperscaler capex estimates have been systematically too low for two consecutive years. At the start of both 2024 and 2025, consensus implied ~20% annual capex growth. Actual growth exceeded 50% both years. Analyst estimates for 2025 started at $280 billion and climbed to over $405 billion through successive revisions. If this pattern holds, 2026 numbers currently penciled at ~$600 billion could be revised higher.

Why the buildout extends through 2027 and beyond:

The transition from training to inference is creating a second wave of demand. Training built the initial GPU footprint. Now, as AI agents and reasoning models scale to production, inference token generation has surged 10x in just one year. Reasoning models like OpenAI's o-series require massive memory bandwidth and inter-chip communication during inference — exactly where NVIDIA's NVLink interconnect technology provides structural advantage.

Power constraints are actually extending the buildout timeline, not shortening it. Goldman Sachs estimates U.S. data centers face an 11+ GW capacity shortfall today, with the cumulative gap expected to exceed 40 GW by 2028. Building a data center takes 18-30 months from concept to commissioning. Securing grid connections can take several years. This means capacity additions planned today won't come online until 2027-2028, sustaining demand visibility well beyond the current planning horizon.

Each architecture generation forces re-architecture. Modern AI racks now exceed 100 kW per rack, with peak densities projected above that. Power, cooling, memory bandwidth, and networking standards must all scale in tandem. This isn't a one-time buildout — it's a recurring investment cycle, with hyperscalers re-architecting data centers roughly every one to two years as new chip generations arrive.

Risk Factors

No investment thesis is complete without honest risk assessment.

Valuation compression. At ~$3.4 trillion market cap, NVIDIA trades at a premium that prices in continued exceptional growth. Revenue growth has decelerated from +100% to +62% YoY — still extraordinary, but the rate of change matters for momentum-driven valuations. A further deceleration to +25-40% (Goldman's estimate for late 2026) could compress multiples even as absolute earnings grow.

Custom silicon adoption. If hyperscaler custom chips capture inference workloads faster than expected, NVIDIA's most price-sensitive market segment could face pressure sooner than 2028. The Meta-Google TPU deal bears watching as a leading indicator.

China export restrictions. NVIDIA lost meaningful China revenue (~13% of total in 2025, down from 26% in FY2022). Further restrictions or retaliatory measures remain a geopolitical wildcard.

DeepSeek-style efficiency shocks. A breakthrough in model efficiency that dramatically reduces compute requirements per unit of AI output could slow the capex cycle. The January 2025 DeepSeek moment demonstrated this risk is real, though so far efficiency gains have been absorbed by expanding use cases rather than reducing total compute demand.

Antitrust scrutiny. CUDA's tight hardware integration has attracted regulatory attention, with concerns about illegal tying and foreclosure of competition. The Run:ai acquisition faced scrutiny as a potential "killer acquisition." Material regulatory action could constrain NVIDIA's ecosystem strategy.

Positioning

NVIDIA's moat is wider than a single product cycle. It's an 18-year software ecosystem with 4+ million developers, a system-level integration advantage that competitors can't replicate with better specs alone, and a margin structure defended by genuine switching costs rather than temporary scarcity.

The data center buildout is a multi-year supercycle with $1.15 trillion in projected spending through 2027, and historical estimates have consistently proven too conservative. Custom silicon will capture a growing slice of inference workloads at hyperscalers, but the total addressable market is expanding faster than NVIDIA's share is contracting.

At current levels, $NVDA is priced for excellence — and the company keeps delivering it. The question isn't whether NVIDIA's moat is real. It's whether the market has already priced in the full duration and magnitude of the AI infrastructure cycle. Based on the systematic underestimation of capex spending we've seen for two consecutive years, I'd argue it hasn't.

The infrastructure play isn't over. We're still in the early innings.

Disclaimer: This analysis is for informational purposes only and does not constitute investment advice. Dark Stone Capital may hold positions in securities mentioned. Always conduct your own due diligence before making investment decisions.