Customer Concentration and Counterparty Credit Risk in Frontier AI Labs

NVIDIA's growth increasingly depends on a small number of AI labs whose spending outpaces their own balance sheets:

• Customer concentration jumped in FY26: one direct customer accounted for 22% of total revenue and another 14%, up from a top-three group of 12%, 11%, and 11% in FY25.
• Management has invested nearly $50B in Frontier AI labs and disclosed it expects demand from AI labs where NVIDIA leverages its balance sheet to represent roughly a quarter of its business in FY28.
• NVIDIA has committed up to $105B in cumulative payment obligations tied to the SB Energy/OpenAI Portsmouth campus (a 4.25 GW initial commitment), plus discretion for additional exposure to another ~3.8 GW, and disclosed maximum gross guarantee exposure of $108.5B as of Q2 FY27.
• NVIDIA is also providing "selective credit enhancement" for nearly 2 GW of compute to another unnamed frontier lab.
• A new risk factor confirms NVIDIA has been asked to offer direct financing arrangements to customers for datacenter buildouts — a shift that would introduce genuine credit exposure to companies that are not yet cash-flow self-sufficient.
• Management explicitly addressed "circular financing" concerns on the FY27Q2 call, arguing risk is "limited" because compute is fungible — but this framing has not been tested through a demand downturn.

Competition from Hyperscaler Custom Silicon (ASICs)

NVIDIA's largest customers are simultaneously its potential future competitors:

• Google (TPU), Amazon (Trainium/Inferentia), and Microsoft (Maia) are all building internal AI accelerators, with Broadcom and Marvell supplying ASIC design capability.
• Hyperscalers represent roughly half of data center revenue, giving them both the capital and workload visibility to justify continued investment in custom silicon.
• Even NVIDIA's own equity partners are pursuing competing chips: OpenAI has discussed its "Jalapeno" chip, claiming performance advantages versus Blackwell, despite NVIDIA's multibillion-dollar investment and infrastructure partnership with the company.
• NVIDIA's counter-argument — that ASICs are workload-specific while NVIDIA's architecture is fungible across training, post-training, and inference — has held so far, but custom ASIC programs remain a long-term structural threat, particularly for stable inference workloads.

China Market Foreclosure and Ecosystem Erosion

U.S. export controls have eliminated a market NVIDIA estimates at ~$50B annually and are actively strengthening rivals:

• NVIDIA recorded a $4.5B charge in Q1 FY26 for excess H20 inventory and purchase obligations after new export controls ended its last compliant China Data Center compute product.
• China Data Center revenue fell from ~19% of Data Center revenue in FY23 to low-single-digit percentages by FY26; revenue from customers headquartered outside the U.S. fell to 31% of total revenue in FY26 from 53% in FY25.
• NVIDIA now states directly that the foreclosure "helped our competitors build larger developer and customer ecosystems to challenge us worldwide," referring to Huawei's Ascend platform gaining share domestically.
• China's antitrust authority issued a preliminary finding in September 2025 that NVIDIA's export-control compliance discriminates against Chinese customers, tied to Mellanox acquisition approval terms — creating a new avenue for fines or restrictions.

Margin Pressure from Memory Costs and Architecture Ramp Complexity

Gross margins are proving less stable than the "mid-70s" target suggests:

• On the FY27Q2 call, management reset gross margin guidance downward, citing memory pricing increases that "exceeded our prior expectations," guiding Q3 FY27 margins to 74% before bottoming at 71-72% in Q4 FY27 and settling at 72-73% in FY28.
• The H20 charge alone dragged Q1 FY26 non-GAAP gross margin to 61.0%, and prior architecture transitions have caused margin drag too — Q2 FY25 margins were hit by "inventory provisions for low-yielding Blackwell material."
• NVIDIA has moved to a one-year product cadence for Data Center solutions, which the company itself acknowledges "may magnify the challenges associated with managing supply and demand," raising the odds of recurring transition-related charges.

Capital Allocation Risk in Ecosystem Investments

NVIDIA is committing large sums to AI labs and adjacent technology bets with uncertain financial payback:

• Ecosystem investments totaled $17.5B in FY26 alone, on top of $3.5B in land, power, and shell guarantees provided to early-stage companies.
• A new intellectual property licensing arrangement with Groq required "significant, nonrefundable payments," and NVIDIA has disclosed it may not recover these costs or generate an adequate return.
• Equity stakes in OpenAI, Anthropic, xAI, and Mistral are framed as ecosystem-expanding rather than financial, but several of these same partners are pursuing independent silicon strategies that could reduce their reliance on NVIDIA over time.

Supply-Constrained Growth and Execution Risk

NVIDIA's growth is now explicitly capped by its own ability to deliver, not by demand:

• Management guided to ~70% revenue growth for FY28 while stating that unconstrained demand growth is closer to 100%, meaning any supply chain misstep directly caps revenue upside and disappoints commitments already made to customers.
• Total future purchase commitments reached $366B as of Q2 FY27, including $279B for supply and capacity — predominantly memory procurement — locking in enormous forward financial exposure regardless of whether end demand fully materializes.
• Inventory nearly doubled sequentially to $32B in Q2 FY27 in preparation for the Vera Rubin launch, and a new risk factor acknowledges customers' own inability to secure sufficient data center capacity or energy could constrain NVIDIA's growth independent of chip supply.

Sustainability of Hyperscaler and NeoCloud CapEx Growth

A large and growing share of NVIDIA's revenue now depends on a "revenue-sharing" financing model with new NeoCloud partners whose own economics are unproven at scale:

• Top-5 hyperscaler CapEx is expected to reach ~$800B in 2026 and $1.3T in 2027 (per FY27Q2 commentary), with cloud industry backlog exceeding $2T — figures that assume continued acceleration in customer spending that has not been tested through a downturn.
• NeoCloud capacity is expected to reach 8 GW by year-end 2026, up from ~3 GW at the end of 2025, funded partly through a structure where NVIDIA provides take-or-pay revenue guarantees to lenders in exchange for a share of rental revenue above a floor — a new, unproven exposure to NeoCloud utilization and pricing risk.
• Management itself concedes this data center segment (ACIE) is growing off a smaller, harder-to-underwrite base of smaller, capital-constrained operators rather than investment-grade hyperscalers.