Appearance
The Nvidia margin paradox: why FY27 datacentre profit may peak before revenue does
When Nvidia reported $82bn of datacentre revenue in Q1 FY27, 85% above the prior year, the headline was easy to write — and almost no one wrote the harder one underneath it. Adjusted gross margin came in at 73.4%, down 280 basis points sequentially. The decline was explained on the call as a "transitional quarter" reflecting Hopper sell-through into a Blackwell-dominated mix. What the call did not say, and what the 10-Q has to spell out by mid-June, is the longer arc behind that single quarter's compression.
The arc matters because Nvidia is the closest thing the public market has to a barometer of the AI capex cycle. If the dollar-per-GPU economics are decelerating even as units are still accelerating, the conclusions for everyone downstream — the hyperscalers buying the chips, the ETFs that own them, the engineers who optimise for them — change materially. So it's worth being precise about what's actually happening.
Three forces, one direction
The first force is the Hopper-to-Blackwell transition itself. Hopper SKUs (H100, H200) launched in 2022 into an underprovisioned market and held pricing for nearly three years — extraordinary by semiconductor standards. Blackwell, by contrast, launches into a market where its predecessor still has a multi-year depreciation tail on hyperscaler books. The result is that the H100 spot price fell from roughly $32k to under $24k between January and April 2026, and the Blackwell launch ASP is being calibrated to that market, not the 2024 one. Per-unit margin on Hopper compresses as the secondary market deepens; per-unit margin on Blackwell starts lower than Hopper's peak ever did.
The second force is hyperscaler ASIC encroachment. Google's TPU v6e, Amazon's Trainium2, Microsoft's Maia 200, and Meta's MTIA v2 are all in volume production. Critically, the four hyperscalers together represent something like 41% of Nvidia's datacentre revenue this fiscal year. Each one is now running internal workloads — primarily inference — at materially lower per-token cost on custom silicon. They are not replacing Nvidia for frontier training; they are replacing Nvidia for the workloads where the marginal Nvidia GPU was most profitable for Nvidia.
Figure 1 — Illustrative
Nvidia datacentre revenue vs. gross-margin trajectory, FY24–FY28E
Source: Nvidia 10-K, 10-Q filings (FY24–FY26 actuals); author's model for FY27–FY28. Forecast assumes Blackwell ASP discipline of ~$36k blended and ASIC share of hyperscaler datacentre spend reaching 28% by end-FY28. Illustrative only.
The third force is the one nobody on the sell-side wants to spend much time on: CoWoS-L allocation. TSMC's CoWoS-L is the advanced-packaging process that Blackwell, MI355X, and the next-generation TPU all depend on. TSMC is doubling capacity twice — once in 2026 and once in 2027 — but the reported allocation queue means Nvidia's share of incremental CoWoS-L wafers is forecast to fall from roughly 64% in 2025 to closer to 51% by late 2027 as AMD and the hyperscaler ASIC roadmaps scale. That doesn't reduce Nvidia's volumes in absolute terms. It does mean Nvidia stops being the price-setter at the packaging layer.
The numbers that decide it
Three line items in the next four 10-Qs will resolve the question of whether the margin compression is transitional or structural:
Datacentre product gross margin disclosure. Nvidia stopped breaking this out in FY25, but the Q&A pressure will force a colour comment. Anything below 72% on a sustained basis is the structural-compression read. Above 74% and the transition argument holds.
Customer concentration in the 10-K. The FY26 10-K already showed the four largest customers as 41% of revenue. If the FY27 filing shows that number drifting upward — even as ASIC deployments scale — the read is that hyperscalers are buying Nvidia for the highest-end frontier-training jobs only, which is the most price-sensitive use case. If concentration falls, it means a broader enterprise base is absorbing more Hopper-vintage product at lower price points. Both outcomes compress margin, in different ways.
Inventory days outstanding. Inventory was 134 days at end-Q1 FY27, up from 95 a year earlier. Some of that is healthy Blackwell pre-build. Some of it isn't. The split between "Blackwell ramp inventory" and "Hopper write-down risk" is the single most important number management will be asked about, and the answer determines whether next year's gross margin starts with a 7 or a 6.
The Hopper price collapse is not a Nvidia problem. It's a Nvidia valuation problem. The chips still work; the moat still exists; the question is what price the moat clears at when three structurally important buyers stop bidding it up.
What follows from this
If the structural read is right — and the evidence is currently 60/40 in its favour — three things follow for the rest of the market in roughly this order.
For the hyperscalers. The shift from Nvidia-only to Nvidia-plus-ASIC is good for hyperscaler unit economics and bad for hyperscaler differentiation. Once every cloud has comparable-cost inference silicon, the basis of competition moves up the stack to model serving, fine-tuning, and developer-facing API quality. AWS Bedrock, Google Vertex, and Azure AI Foundry start to look more like each other on price; they start to compete on the things wrapped around the price. That's healthier for customers and harder for any individual hyperscaler to monetise.
For the picks-and-shovels names. The CoWoS-L allocation story is the cleanest "picks-and-shovels" thesis left in the AI stack. Companies that supply into TSMC's packaging lines — Lam Research, Applied Materials, Onto Innovation, and the optical-interconnect names — capture some of the margin compression that Nvidia is shedding. This is the durable bet inside the broader hype cycle: when miners' margins compress, equipment suppliers' margins often don't.
For the model layer. Lower per-token inference cost across all clouds means the model layer's monetisation gets harder. OpenAI, Anthropic, and the open-weight labs all benefit from cheaper compute, but their pricing leverage falls roughly in proportion. The companies that capture the surplus are the ones built directly on top of cheap inference — vertical-specific applications that can afford to burn tokens because tokens stopped being expensive. The next twelve months of seed and Series A funding will lean very hard into that thesis.
The Nvidia margin question is not really a Nvidia question. It's the question of where the AI profit pool sits in 2028. As of this week, the answer is starting to look meaningfully different from the answer this time last year — and the move is happening quietly, in the second page of a 10-Q, while the headline number does its job of distracting.
Compounding all of this, the next two earnings cycles will arrive against a Federal Reserve still likely to hold rates above 4.25%, which raises the discount rate applied to terminal-year cash flows precisely as those cash flows become more uncertain. The compression in the multiple, if and when it comes, will look like the market suddenly noticing what the filings have been hinting at for two quarters. By then it will be too late to call it a surprise.