This is a sample of CapRelay's research. CapRelay provides similar research and more for every US-listed company.

Start free trial
NVIDIA -

Bull Case

NVDA | Market Cap: $5.3T (06/04/26)
Industry:
Semiconductors

Thesis Summary

NVIDIA is the essential infrastructure layer of a multi-decade industrial revolution. The business is not just growing — it is accelerating. Revenue grew 85% year-over-year in Q1 FY27 to a record $82B, and Q2 FY27 guidance of $91B implies continued sequential momentum. What makes NVIDIA's position unusual is that demand is being driven by multiple compounding forces simultaneously: the shift from CPUs to accelerated computing, the replacement of classical ML with generative AI in core hyperscaler workloads, the explosion of agentic and reasoning AI, and the emergence of physical AI and robotics. Each of these is in early innings.

Key thesis points:

  • Reasoning AI is a structural step-function in compute demand. Agentic and reasoning models require orders of magnitude more compute per task than one-shot inference, driving a sustained increase in token generation demand that plays directly to NVIDIA's strengths.
  • Vera CPU opens a $200B TAM NVIDIA has never addressed. The world's first CPU purpose-built for agentic AI enters production in Q3 FY27, and management has visibility to nearly $20B in standalone CPU revenue in FY27 — entirely incremental to the existing Blackwell and Rubin GPU business.
  • NVLink networking is becoming as important as the GPU itself. Networking grew 263% year-over-year in Q4 FY26, and Spectrum-X Ethernet has surpassed a $10B annualized revenue run rate. NVLink Fusion now extends this networking ecosystem to third-party silicon, potentially expanding the addressable market further.
  • The ACIE segment (AI native clouds, enterprise, sovereign) is growing faster than hyperscale and is more defensible. This segment is fragmented, requires a full-stack solution, and is almost entirely served by NVIDIA. It is structurally less exposed to custom ASIC substitution.
  • Rubin is on track to be NVIDIA's most successful architecture yet. Every frontier model company is committed to Vera Rubin from day one, unlike the Blackwell transition. Management has visibility to $1T in Blackwell and Rubin revenue through end of calendar 2027.
  • Gross margins have normalized. After a one-time $4.5B H20 charge depressed Q1 FY26 margins to 61%, non-GAAP gross margin recovered to 75.2% in Q4 FY26 and held at 75.0% in Q1 FY27. Mid-70s is the target and the current run rate.

Reasoning and Agentic AI Are Structural Compute Multipliers

The shift from one-shot inference to reasoning AI is the single most important near-term demand driver, and it is still early.

  • Reasoning models (DeepSeek R1, OpenAI O3/GPT-5.5, Claude Code) require 100x to 1,000x more compute per task than a one-shot query, because the model must read context, break down problems, spawn sub-agents, and use tools.
  • Token generation has gone from a modest workload to a revenue-generating activity. H100 cloud rental pricing is up 20% YTD through Q1 FY27, and A100 pricing is up ~15% — the opposite of what you'd expect if demand were slowing.
  • Microsoft processed over 100 trillion tokens in Q1 FY26, a 5x increase year-over-year. Anthropic grew from $1B to $7B in annualized revenue in a single year. OpenAI's weekly users grew to 800M. These are not signs of digestion.
  • Agentic systems spawn sub-agents, each requiring inference. As the world moves from hundreds of thousands of agents today toward potentially billions, the compute demand per unit of economic activity rises significantly. This is the secular thesis in its most concrete form.

NVLink 72 is the direct architectural response to this shift. Reasoning models require enormous memory bandwidth to maintain context across long token chains. NVLink 72 connects 72 GPUs and 36 CPUs in a single rack, delivering 130 TB/s of bandwidth — equivalent to the world's entire peak internet traffic. For reasoning models like DeepSeek R1, GB300 NVL72 delivers 10x higher performance per watt and 10x lower cost per token versus H200. This generational leap is why inference performance now directly translates to customer revenue, and why NVIDIA is winning share in inference, not losing it.

Vera CPU: A Entirely New $200B Growth Driver

Perhaps the most underappreciated near-term driver is the Vera CPU. This is not an incremental product — it opens a market NVIDIA has never served.

  • Management has visibility to nearly $20B in standalone Vera CPU revenue in FY27. This is entirely additive to the Rubin GPU and networking TAM.
  • Vera is architecturally distinct from incumbent x86 server CPUs. It supports LPDDR5 memory (optimized for bandwidth-hungry AI workloads) and delivers up to 1.5x faster performance per core, 2x performance per watt, and 4x rack density versus x86 alternatives.
  • The use case is agentic orchestration: agents run on CPUs, sub-agents run on GPUs. Tool use, IO, memory management, and browser/compiler interactions are all CPU-bound. As the number of agents scales from hundreds of thousands to billions, so does the demand for Vera.
  • Every major hyperscaler and system maker is partnering with NVIDIA to deploy Vera, and Jensen Huang expects NVIDIA to become the world's leading CPU supplier.
  • Vera has four deployment modes: as part of Vera Rubin, as a standalone CPU, as a Vera+CX9 storage solution, and as a Vera+CX9 security/confidential computing solution — each addressing different enterprise needs.

ACIE: The More Defensible, Faster-Growing Second Segment

NVIDIA's new reporting framework segments its data center business into Hyperscale ($38B in Q1 FY27, ~50% of data center) and ACIE — AI cloud natives, industrial, and enterprise ($37B in Q1 FY27, ~50%). The ACIE segment grew 31% quarter-over-quarter, faster than Hyperscale's 12%.

  • AI native clouds, regional cloud providers, sovereign AI, and enterprises cannot build custom ASICs. They need a full-stack, pre-integrated solution that works out of the box and can serve the broadest possible range of workloads.
  • The ACIE segment is inherently diverse: 250,000+ companies globally, each with different requirements. NVIDIA's vertically integrated but openly deployable platform is the only solution that can address this market at scale.
  • Sovereign AI alone exceeded $30B in FY26, more than tripling year-over-year, and is deployed across nearly 40 countries representing $50T in GDP. Management expects sovereign AI to grow at least in line with overall AI infrastructure spending.
  • The number of partner data centers exceeding 10MW has nearly doubled in one year, now surpassing 80 sites.
  • The custom ASIC threat that investors focus on is almost entirely a hyperscale concern. Hyperscalers represent only ~50% of data center revenue. The other ~50% — ACIE — is structurally NVIDIA's.

The Rubin Transition: Better Positioned Than Blackwell

Vera Rubin enters production in Q3 FY27 and is set up for the fastest ramp in NVIDIA's history.

  • Unlike the Blackwell launch, every frontier model company is committed to Vera Rubin from day one. Blackwell did not start with that level of universal commitment.
  • Rubin delivers up to 35x higher inference throughput and up to 10x greater AI factory revenue versus Blackwell, with a 10x reduction in inference token cost.
  • The manufacturing transition is lower-risk than Blackwell's: Rubin uses a modular, cable-free tray design that improves resiliency and serviceability, and the supply chain has already mastered rack-scale architecture from the Blackwell ramp.
  • Google's XGS bare metal instances can support up to 960,000 Rubin GPUs across multiple sites. Purchase orders and customer commitments are already in place.
  • NVIDIA has visibility to $1T in Blackwell and Rubin revenue through end of calendar 2027, with management explicitly noting this excludes standalone Vera CPUs, which represent the second-largest upside vector.

Networking: A Durable Moat Builder

Networking has evolved from a Mellanox acquisition into a structurally critical, high-margin business that reinforces NVIDIA's data center lock-in.

  • Networking revenue in Q4 FY26 was $11B, up 263% year-over-year. For full FY26, networking exceeded $31B, up more than 10x from FY21.
  • Spectrum-X Ethernet exceeded a $10B annualized revenue run rate in Q2 FY26, and InfiniBand grew more than 4x year-over-year in Q1 FY27.
  • NVLink Fusion — announced in Q1 FY26 — allows third-party CPUs and ASICs to connect to NVIDIA's platform via NVLink. This is strategically significant: rather than competing against the custom ASIC trend, NVIDIA is absorbing it. Fujitsu, MediaTek, Qualcomm, and Arm have all announced NVLink integrations.
  • Spectrum XGS connects multiple data centers into gigascale AI super factories. Microsoft's Farweave — the world's most significant AI data center — is an early adopter.
  • Each GB200 NVL72 rack ships with 9 NVLink switches. As racks scale to millions of Rubin GPUs, the switch content per rack grows as well, making networking a structural co-traveler to GPU revenue growth.

Gross Margins: Normalized, Not Impaired

The FY26 gross margin story is simpler than it looks.

  • The full-year FY26 non-GAAP gross margin of 71.1% was almost entirely a function of the one-time $4.5B H20 charge in Q1 FY26, which compressed that quarter's margin to 61%. Excluding the charge, Q1 FY26 would have been 71.3%.
  • Margins recovered to 73.6% in Q3 FY26, then to 75.2% in Q4 FY26, and held at 75.0% in Q1 FY27 — exactly the mid-70s target management has consistently guided to.
  • Management's stated path to sustaining mid-70s margins is through continued improvement in cost structure, manufacturing cycle time, and mix — the same levers used to recover margins through the Blackwell ramp.
  • Input cost pressures (memory, yields, supply chain premiums) are real, but NVIDIA has the supply chain relationships, lead time management, and architectural performance advantages to offset them.

China: A Headwind That Is Already Fully Absorbed

The permanent closure of the China data center compute market is a real impairment, but it is already embedded in the numbers and the outlook.

  • H20 revenue was running at $7–8B per quarter before the ban. NVIDIA's Q2 FY27 guidance of $91B explicitly excludes any China data center compute revenue — meaning the current trajectory is already net of China.
  • The non-China business grew 120% year-over-year in Q1 FY27, accelerating even as China was foreclosed.
  • H200 licenses have been approved for some Chinese customers, and any revenue recognition from H200 would be entirely additive to current guidance.
  • NVIDIA's global sovereign AI pipeline, the ACIE segment, and the Rubin ramp collectively represent a far larger growth opportunity than the lost China revenue.

Free Cash Flow and Capital Returns

NVIDIA generated $49B in free cash flow in Q1 FY27, up from $35B in Q4 FY26. The company is now generating more free cash flow per quarter than most companies generate in a year.

  • NVIDIA returned $20B to shareholders in Q1 FY27, increasing its quarterly dividend from $0.01 to $0.25 per share (25x) and adding an $80B share repurchase authorization.
  • Management's stated intention is to return roughly 50% of free cash flow to shareholders in FY27, with the remainder deployed in R&D, supply chain, and ecosystem investments.
  • Diluted shares outstanding have declined from 24.9B in FY24 to 24.5B in FY26, and continued buybacks will further concentrate ownership.
  • The $17.5B in FY26 ecosystem investments in OpenAI, Anthropic, xAI, and others are not financial speculation — they are the mechanism by which NVIDIA ensures CUDA remains the universal platform for every major AI model, protecting the moat at its source.
Using data as of 2026-05-20