NVIDIA Corporation NVDA Moat

Key data

The moat

A frontier lab picks a platform once and then spends two years writing kernels, schedulers and inference servers against it. Leaving means rewriting that work, requalifying a cluster, retraining staff, and accepting worse tokens per dollar during the transition. The behavior that proves it is not the purchase, it is the reorder: AWS, already running a vast installed base, committed to an additional 2 million GPUs plus Vera CPUs running into the second quarter of fiscal 2029. Classify this as proprietary technology first, with developer switching costs second, and with high confidence, because the audit trail is in measured third-party systems rather than in adjectives.

The swap test. One vendor ships the GPU, the CPU, the NVLink switch, the DPU, both InfiniBand and Ethernet scale-out, and the software, co-designed as one rack; 6,000 accelerated applications target it, 7.5 million developers write to its toolchain, and over 78% of the TOP500 list runs on it. Substitute AMD, Broadcom or Intel and the sentence collapses at the second clause.

What that produces is price. Gross margin ran 56.9%, 72.7%, 75.0%, 71.1% across four years while revenue went from $26.97 billion to $215.94 billion, and management is putting through an executed price increase effective in the first quarter against a memory cost shock.

Widening or narrowing

The ecosystem sequences widen cleanly. Supported applications: 2,800, 3,500, 4,400, 6,000. TOP500 presence: over 70%, over 75%, over 75%, over 78%.

The profit pool sits in the accelerator and the rack around it, and NVIDIA still holds the fat slice: Data Center revenue of $193.74 billion in fiscal 2026 against Broadcom's expected $58 billion of fiscal 2026 AI semiconductor revenue and AMD's $6.7 billion data center quarter. But the pool is being contested from inside the customer base, not outside it.

The overrated case, stated plainly. Third-party share of AI accelerator revenue has fallen on three successive calendar readings, roughly 87% to 80 to 90% to 75 to 81%. ROIC fell from 75.3% to 62.9%. Gross margin gave back 390 basis points and is guided to bottom at 71% to 72% before settling at 72% to 73%, so the input side now has pricing power over NVIDIA. Customers A, B and C were 12%, 11% and 11% of fiscal 2025 revenue, and every one of the hyperscalers named in the competition section designs its own accelerator. Verdict: narrowing.

What breaks it, and who

Custom silicon at the buyers is the live threat. Broadcom's XPU shipments rose over 3.5 times year over year and were 73% of its AI revenue in its third quarter, with AI semiconductor revenue of $16.7 billion in that quarter alone. Google TPU is estimated near 8% of 2026 accelerator revenue and AWS Trainium and Inferentia near 5%.

The merchant alternative is real for the first time in a decade. AMD data center revenue went $5.8 billion then $6.7 billion in consecutive 2026 quarters, up 107% year over year in the second, with multi-gigawatt MI450 commitments announced by Meta and Anthropic. Two quarters is not a sequence, so treat the slope as provisional.

Inputs and policy both bite. Memory price increases have exceeded the company's prior expectations and are headed higher, which is an input shock, not a competitive loss. China fell from 16.9% to 13.1% to 9.1% of revenue over three years while the United States went 44.3% to 46.9% to 69.3%, and government licenses remain the gate.

What would change the read over the next 12 to 18 months: whether Rubin ships in volume in the second half of fiscal 2027 at its claimed 10 times cost-per-token reduction, whether gross margin lands at 72% to 73% in fiscal 2028 as guided, and whether third-party share stabilizes above 75%.

Closing thoughts

The moat is real and rests on measured third-party evidence rather than claims, and its direction is narrowing from an extreme level rather than eroding. The checkable item is gross margin against the 72% to 73% fiscal 2028 guide, because that single line separates a memory cost problem, which passes, from a pricing problem, which does not. Customer concentration near one third of revenue across three buyers who all build their own accelerators is the structural weakness, and no software ecosystem fully offsets it. The moat strengthens if gross margin recovers to 73% or better in fiscal 2028 with third-party accelerator share holding above 75%, and weakens if margin settles below 71% while that share slips toward the 65 to 70% floor analysts describe.