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

Start free trial
NVIDIA -

Bear Case

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

Thesis Summary

NVIDIA is an extraordinary business that has earned its position. The bearish case is not that NVIDIA is a bad company — it is that the company faces structural challenges that will make sustaining the current pace of growth increasingly difficult. Three core concerns:

  • Hyperscaler custom ASICs are moving from threat to reality. The largest customers, who represent over half of data center revenue, are systematically building capability to bypass NVIDIA for stable inference workloads. This is happening right now, not in some hypothetical future.
  • China is a permanent hole in the addressable market. NVIDIA has lost a market it estimates at nearly $50B annually, with no viable replacement product. Huawei is filling the void and building an alternative ecosystem that could challenge CUDA's global universality over time.
  • The business is structurally dependent on sustained, parabolic CapEx growth. NVIDIA's growth requires hyperscaler CapEx to keep accelerating toward $1T annually. The consequences of even a temporary pause are severe given the concentration of revenue and $145B in locked-in supply commitments.

The Custom ASIC Threat Is Accelerating, Not Stalling

The most important long-term structural threat to NVIDIA is hyperscalers systematically reducing their reliance on merchant GPUs for inference workloads:

  • Google (TPU), Amazon (Trainium/Inferentia), Microsoft (Maia), and Meta are all scaling internal ASIC programs. Broadcom, the primary ASIC design partner for several of these hyperscalers, has signaled strong AI ASIC revenue growth.
  • Hyperscalers collectively represent slightly over 50% of NVIDIA's data center revenue. Even partial substitution on inference — where model architectures are more stable and optimization is more tractable — would be material.
  • Inference is explicitly the fastest-growing workload. NVIDIA itself notes that reasoning and agentic AI are driving inference demand exponential. This is the workload where ASICs have the clearest advantage: stable model architectures, predictable access patterns, and optimizable pipelines.
  • NVIDIA's NVLink Fusion announcement — enabling third-party ASICs to connect to NVIDIA's NVLink fabric — can be read as an acknowledgment that some customers will use ASICs for inference and NVIDIA is trying to capture the networking and training revenue around the edges.

The counterargument that ASICs are narrow and CUDA is universal is compelling for training and post-training workloads. But as inference scales, the economic incentive to route stable, high-volume inference traffic to purpose-built chips is enormous. The question is not whether this happens, but how fast.

China Is a Permanent and Growing Problem

NVIDIA has been effectively foreclosed from the Chinese data center compute market, which it estimates will grow to nearly $50B annually:

  • Prior to the April 2025 H20 export ban, H20 revenue was running at approximately $7–8B per quarter. NVIDIA estimates it lost $8B of H20 revenue in Q2 FY26 alone.
  • The $4.5B inventory and purchase obligation charge in Q1 FY26 illustrates the financial severity of these controls and the risk of recurrence with future products.
  • NVIDIA has explicitly stated it has no viable replacement product. Management noted that the new export control limits make it impossible to further reduce Hopper for productive use, and no new China-compliant product exists.
  • Huawei's Ascend chips are scaling domestically with the benefit of IPO funding and government support. NVIDIA's own filings acknowledge that the export controls "helped our competitors build larger developer and customer ecosystems to challenge us worldwide."
  • China has approximately 50% of the world's AI researchers. If a generation of Chinese AI developers trains on and optimizes for Huawei's stack, the resulting models and tooling may be natively incompatible with CUDA, gradually eroding NVIDIA's claim to run every AI model globally.
  • A new complication: China's antitrust authority issued a preliminary finding in September 2025 that NVIDIA's compliance with U.S. export controls constituted unfair discrimination against Chinese customers, potentially creating financial penalties or operational restrictions on top of the export control losses.

The H200 license approvals so far have generated zero revenue, and management has explicitly excluded China data center compute revenue from forward guidance with no expected timeline for resolution.

Growth Requires Parabolic CapEx to Keep Accelerating

NVIDIA's growth trajectory requires not just sustained CapEx, but accelerating CapEx. Management acknowledges visibility to $1T in Blackwell and Rubin revenue through end of calendar 2027, and total supply commitments reached $145B as of Q1 FY27:

  • Analyst forecasts cited by management project hyperscaler CapEx approaching $1T in 2027, roughly doubling from current levels. NVIDIA's guidance is premised on this acceleration continuing.
  • If hyperscaler CapEx growth slows — even temporarily — the impact on NVIDIA would be disproportionate given the revenue concentration. Two customers accounted for 22% and 14% of FY26 total revenue, respectively. A change in buying behavior at either would have an outsized effect.
  • The historical precedent exists: AWS grew 29% in FY22 and slowed to 13% in FY23 as enterprises optimized IT spending. An analogous pause in AI infrastructure CapEx — driven by monetization challenges, balance sheet constraints, or demand normalization — would create a severe air pocket.
  • The $145B in supply commitments that NVIDIA has secured creates a potential inventory risk that mirrors the $4.5B H20 charge and the $2.17B gaming inventory provisions of FY23. NVIDIA has been here before.
  • Management's own signal is worth noting: in the Q3 FY26 call, Colette Kress acknowledged that "input costs are on the rise" and that NVIDIA is "working to hold gross margins in the mid-seventies." This is not the language of a company with pricing power to spare.

Gross Margins Face Structural Pressure From Complexity

The business model shift from selling HGX systems to delivering rack-scale Blackwell solutions has permanently changed the cost structure:

  • Non-GAAP gross margins fell from 75.0% in FY25 to 71.1% in FY26, primarily because of the H20 charge and Blackwell ramp complexity. Excluding the H20 charge, Q4 FY26 recovered to 75.2%.
  • The Blackwell NVL72 rack contains 1.2 million components across 350 manufacturing sites. Each new architecture generation — Blackwell Ultra, Rubin — requires managing this complexity afresh. NVIDIA is perpetually in a new architecture ramp, which is structurally a low-margin period.
  • Management's guidance of "mid-70s" for FY27 is aspirational, not guaranteed. Rising input costs (memory, yield management on leading-edge silicon, TSMC pricing) are explicitly acknowledged. NVIDIA's fabless model passes TSMC cost inflation directly through to COGS.
  • The annual product cadence that is central to NVIDIA's competitive moat also means the company is in some form of ramp complexity in nearly every quarter going forward. There is no stable plateau.

The Ecosystem Investment Program Introduces New Risk

NVIDIA deployed $17.5B in ecosystem investments in FY26 alone, including stakes in OpenAI, Anthropic, xAI, and Mistral, plus $3.5B in land, power, and shell guarantees to early-stage companies:

  • NVIDIA's largest customers are increasingly also its portfolio companies. This creates a conflict-of-interest dynamic: if NVIDIA is offering favorable commercial terms or prioritizing supply to investees, it distorts the arm's-length economics of its customer relationships.
  • The Groq IP licensing deal required "significant, nonrefundable payments" with uncertain returns, illustrating that not all technology bets will pay off.
  • Management has disclosed that NVIDIA has been asked to provide customer financing for data center build-outs. At $145B in supply commitments and growing, NVIDIA's balance sheet is increasingly being deployed to guarantee demand that may not materialize.
  • These investments are framed as ecosystem expansion, but they also represent a bet that the current frontier AI model companies — OpenAI, Anthropic, xAI — will be the long-term winners of the AI application layer. If the model layer commoditizes or consolidates differently, the ecosystem investments will prove expensive and dilutive.
Using data as of 2026-05-20