Nvidia's Blackwell and the AI chip demand curve, explained

01 What Blackwell actually is

Blackwell is a GPU microarchitecture developed by Nvidia as the successor to the Hopper and Ada Lovelace microarchitectures. That one-line definition undersells the scale of the change. Blackwell dies pair two large dies into a single package connected at very high bandwidth, so the pair of processors behaves — to software — as one enormous GPU. The architecture ships to cloud customers under the B100 and B200 names, and in a fully integrated configuration called GB200 that mounts the GPUs alongside Arm-based CPU cores (a Grace processor) in the same rack-level system. The companion networking and switching fabric, branded NVLink and NVSwitch, is what lets many of these packages behave as one accelerator for very large models.

It helps to separate three things that get conflated in coverage: an architecture (the internal design — Blackwell), the products built from it (B200 GPUs, GB200 systems), and the market story around them (sold-out allocation, multibillion-dollar orders). The measured facts are the first two: the architecture exists, products are shipping, and customers have disclosed large purchases. The market story — that demand will stay ahead of supply through the decade, for instance — is interpretation, and this article keeps the two visibly apart.

02 Why demand outran supply

The measurable core of the demand story is in Figure 1: data center revenue of $15.0 billion in the fiscal year ended January 2023, $47.5 billion in the year ended January 2024, and $115.2 billion in the year ended January 2025, per Nvidia's own 10-K filings. When revenue more than doubles twice in a row, the binding question stops being “who wants chips” and becomes “who can get them.” Through 2024 and into 2025, that was the actual constraint: Blackwell-class output was allocated before it existed, with customers publicly committing to quantities measured in hundreds of thousands of units, and Nvidia's own communications describing “full” demand — everything the fabs can produce is spoken for, in every configuration, for multiple quarters ahead.

Why this happened is a chain of dependencies rather than a single cause. Training frontier models requires enormous parallel compute, and inference at consumer scale — the part of the market that grew fastest once chatbots went mainstream — requires just as much, continuously. Manufacturing sits upstream at TSMC, whose advanced packaging capacity (the step that stitches two Blackwell dies into one package) became the true bottleneck rather than raw wafer output. Add a purchasing pattern peculiar to this market: buyers order ahead because being short of compute now is costlier than being long of compute later, which converts even mild uncertainty into over-ordering. Interpretation, clearly labeled: some of that is durable structural demand, and some of it is the buffer every buyer builds in when shortage is the alternative — the two are hard to tell apart from the outside.

03 The rack-scale shift: GB200 NVL72

The most consequential product of this generation is not a chip but a room-sized unit of sale. GB200 NVL72 is a rack that integrates 36 Grace CPUs and 72 Blackwell GPUs, connected with NVLink switching so the whole rack addresses memory and passes intermediate results as a single machine. For trillion-parameter models, that changes the unit of deployment: instead of networking thousands of separate servers, a datacenter installs integrated racks and treats the rack itself as the compute building block. Nvidia has framed this as selling AI factories rather than components, and the phrase is less marketing than it sounds — the company has publicly discussed pricing and delivery per rack.

The shift moves value and complexity up the stack. A rack-scale product integrates liquid cooling, power distribution, and switching that customers previously specified themselves, which raises what Nvidia captures per installation and deepens the practical switching cost of its ecosystem. It also raises the stakes for buyers: a full NVL72 rack is a six-figure-kilowatt load with its own plumbing, so a misjudged purchase is far more expensive to unwind than an over-ordered tray of PCIe cards. Measured fact: customers including the major cloud providers have committed to deploying these systems at scale. Interpretation: rack-scale integration is Nvidia's answer to the perennial question of whether faster competitors can undercut it — because competing with an integrated rack means competing with a supply chain, not just a chip.

04 Competition: AMD, Google TPUs, and custom silicon

Nvidia's share of the AI accelerator market is dominant but contested from three directions. AMD's Instinct line — most recently the MI300X and its successors — offers high memory capacity per part and, crucially, a software path (ROCm) that has improved enough that some large labs run training and inference on it. Google's tensor processing units (TPUs) are the deepest in-house alternative: purpose-built chips deployed inside Google Cloud for years and used both for Google's own models and by outside customers, including prominent AI labs that have disclosed TPU-based training runs. Microsoft, Meta, Amazon, and OpenAI have all disclosed custom accelerator programs of their own.

Two realities temper the competitive picture, and both are structural rather than temporary. First, CUDA — Nvidia's software platform, mature since 2006 — remains where nearly all AI research code is written first, and a model that runs on a competing accelerator must still be ported, tuned, and validated there; the chip is the cheaper part of that migration. Second, the custom-silicon programs belong to exactly the companies whose demand defines the shortage, so their chips largely reallocate internal demand rather than coming to market as rivals. Figure 2 shows the concentration that results: data center work now accounts for such a large majority of Nvidia's revenue that “AI chip competition” moves Nvidia's total far less than it moves any individual customer's options.

05 Power, cooling, and datacenter constraints

By 2026 the scarce input for AI is increasingly electricity rather than silicon. A single GB200 NVL72 rack draws on the order of 120 kilowatts, an order of magnitude beyond the air-cooled racks of the previous decade, which pushes essentially all Blackwell-scale deployments to liquid cooling — direct-to-chip cold plates at minimum, and in dense designs immersion. That requirement ripples outward: facilities need liquid-ready mechanical plants, water or closed-loop coolant supply, and floor structures rated for racks weighing well over a tonne each. New builds can be designed around this; retrofits of buildings standing in 2019 largely cannot, which is why AI capacity is arriving as new construction concentrated where power is cheap and grid connections can actually be secured.

The grid itself is the longer pole in the tent. Interconnection queues for new large loads stretch for years in the United States, and utilities and hyperscalers have signed agreements to restart dormant generators and co-locate generation with datacenters to bridge the gap. Nvidia's own roadmap acknowledges the constraint rather than denying it — successive generations are marketed on performance-per-watt precisely because watts are now the binding budget. The measured facts here are the engineering ones: rack densities, cooling retrofits, multi-year grid queues. The interpretation — that compute growth through the decade is gated more by power infrastructure than by chip fabrication — is widely held in the industry but remains a projection, not a settled outcome.

06 What Blackwell means for the 2026 buildout

The 2026 buildout is the year Blackwell becomes the default substrate of new AI capacity. The architecture behind it, as Nvidia's successor to Hopper, moves volume production from early allocation to broad availability during the year, and its follow-on (Blackwell Ultra, then the Rubin generation announced for later) layers on top of an installed base that is already enormous. For buyers, that changes the calculus from “can we get any” to “what mix”: established clouds diversify across Blackwell generations and their own silicon, while second-tier providers and sovereign AI projects — often the marginal buyer — take allocation as they can get it. The AI-factory framing matures too, with buyers increasingly contracting for delivered, operational compute capacity rather than chips.

Two second-order effects deserve attention. First, the installed base compounds software: every year of CUDA-first development deepens the ecosystem moat that Section 4 described, which is why “waiting for a competitor to mature” is a strategy whose cost rises over time rather than falling. Second, the buildout concentrates physical risk: more of the world's AI capacity sits in a modest number of new facilities that share the same cooling technology, the same power interconnection process, and the same vendor. That concentration is efficient, and it is also why 2026 discussions of AI capacity increasingly sound like discussions of critical infrastructure — because that is what it has become, functionally, for the companies that depend on it.

07 Risks: concentration, export rules, and depreciation

The bull case rests on assumptions that deserve explicit naming. Customer concentration: a handful of hyperscalers account for the large majority of data center purchases, so a pause in capex by two or three of them would be visible in Nvidia's results within quarters — Figure 2's 88 percent data center share is a measure of that exposure. Export controls: US rules restrict sales of top-tier accelerators to China and have already forced Nvidia to ship modified, lower-performance variants there; tightening rules shrink a real market overnight, and the policy direction across administrations has been one-way. Depreciation and the resale question: whether today's accelerators retain value as newer generations arrive is genuinely uncertain — buyers are amortizing these systems over several years, and if generational jumps keep being this large, older fleets could lose earning power faster than their balance-sheet schedules assume.

None of these risks negates what is measured: extraordinary, filing-verified revenue growth, and a product cycle (Blackwell and its successors) that competitors have not yet matched at scale. But the two layers of this article — the audited numbers and the forward narrative — are different kinds of claims, and the gap between them is where investment and policy mistakes live. The disciplined read of 2026: demand is real and physically constrained; supply is expanding as fast as packaging, power, and construction allow; and every extrapolation past the next fiscal year is a forecast wearing the costume of a trend line. When the buyers themselves hedge — building their own chips, signing multi-vendor deals, questioning depreciation schedules — that is signal worth weighing more heavily than any single quarter's results.