NVIDIA went from “Blackwell is delayed” to “Blackwell is sold out through 2026” in about six months, and the speed of that turnaround tells you almost everything about where the AI buildout stands. The Blackwell generation (GB200, then the refreshed GB300 with B300 Ultra parts) is the first serious attempt to sell compute by the rack rather than by the card, and the Rubin generation behind it is already reshaping supplier calendars. If you only follow the headlines, you'll miss the part that actually matters: the bottleneck has moved from the GPU die to power, memory and packaging.
What Blackwell actually is
Blackwell isn't a single chip. The B100, B200 and GB200/GB300 NVL72 configurations pair two GPU dies into a single GPU complex using a very high bandwidth on-package link, then hang that off NVIDIA's NVLink fabric so 72 GPUs in a rack behave, for many workloads, like one enormous accelerator. The headline numbers everyone quotes: around 2.5 times the FP4 dense throughput of Hopper on the same footprint, and rack-scale power draws in the 100 to 130 kW range for a full NVL72 deployment. Take the precise FLOPS with a pinch of salt; the real story is tokens per dollar per megawatt, and there Blackwell is a genuine step, not a marketing refresh.
The GB300 NVL72, with its HBM3E memory bump and faster NVLink, pushed shipments hard through 2025 and into 2026. Reportedly, NVIDIA was allocating racks to hyperscalers well ahead of general availability, which is why cloud capacity kept appearing in odd increments. Supply was never the die itself; it was everything around it.
The rack-scale economics
Buying an NVL72 rack is not like buying 72 discrete GPUs. You are buying a fixed, liquid-cooled, NVLink domain with its own networking, power delivery and failure modes. That changes the math in three ways:
- Density per megawatt goes up sharply. FP4 inference at rack scale is dramatically more token-efficient per watt than Hopper, which is the metric hyperscalers actually optimize for right now.
- Utilization economics flip. Training runs and large inference fleets want the whole NVLink domain; slicing it across tenants is awkward, so customers commit to big blocks.
- Operational complexity goes up. Liquid cooling, 130 kW-class power drops, and NVLink fabric management are real facilities problems, not just IT problems.
The practical consequence: Blackwell revenue lands in fewer, bigger customer contracts. Microsoft, Google, Amazon, Oracle and a handful of neoclouds are the bulk of the orders. That concentration is fine when the hyperscalers are flush, and slightly terrifying when capex scrutiny arrives.
Rubin: the next turn of the crank
NVIDIA has been unusually open about Rubin, the successor architecture. The public roadmap points to a ~2026–2027 window, with Rubin GPUs paired with a new Vera CPU, HBM4 memory, and NVLink 6 for multi-rack scaling. A denser NVL144-class rack configuration is expected, and NVIDIA has signaled an annual or near-annual architectural cadence: Hopper, Blackwell, Rubin, then Feynman after that.
The details worth watching are the ones that sound boring. HBM4 doubles memory bandwidth per stack roughly, which matters far more than raw compute for inference. NVLink 6 keeps the rack-scale fabric coherent at greater scale. And Rubin's package is expected to lean even harder on TSMC's CoWoS packaging and advanced nodes, which is why packaging capacity, not wafer starts, keeps dictating delivery dates.
Roadmaps slip. Blackwell slipped a little. Assume Rubin slips too, but don't let a quarter of timing fool you into thinking the direction changes. It doesn't. NVIDIA has effectively turned GPU architecture into an annual iPhone launch, with supply chains that plan a year or more ahead.
Where the bottlenecks actually are
Strip away the keynotes and three constraints decide who ships what:
- CoWoS interposer capacity. TSMC's advanced packaging lines are the single hardest constraint on Blackwell volumes. Expansions are underway, but packaging yields on giant packages are brutal.
- HBM supply. HBM3E is sold out in meaningful quantities well into 2026, and HBM4 qualification is a race between SK Hynix, Samsung and Micron.
- Power and facilities. You can deliver the rack, but can the datacenter feed it? Grid interconnect queues and cooling retrofits are the quiet limiters on deployment speed.
What this means if you actually buy or rent GPUs
If you're a practitioner, the Blackwell story is mostly a pricing and capacity story. Spot instances of Hopper hardware will stay cheap for a while, because installed base is huge. Blackwell-class capacity will stay scarce and pricey, and the neoclouds that grabbed allocations early will rent them at premium rates. For inference at scale, the FP4 economics are genuinely compelling; for most mid-size teams, the decision is simpler: rent, don't buy, and renegotiate quarterly.
One uncomfortable truth: the model quality race and the hardware race are now coupled so tightly that a roadmap miss at NVIDIA propagates into every lab's training plans. If you're betting a product on training runs sized to Rubin-class clusters, build slack into your timeline. The industry has been running on just-in-time silicon miracles for two years, and miracles don't always show up on schedule.
How to read the spec sheet without getting fooled
Vendors quote dense FP4 numbers because they're enormous, but almost nobody runs dense FP4 in production. Sparse marketing figures, sparsity-aware benchmarks and inflated tokens-per-second claims all inflate the gap between slide and reality. When evaluating Blackwell for your workload, benchmark with your actual model, your actual batch sizes and your actual accuracy floor. A 70B model at FP4 that collapses on your retrieval benchmarks is worthless no matter how pretty the TFLOPS chart looks.
Also watch the interconnect fine print. NVLink 5 domains scale to 72 GPUs, but crossing rack boundaries still leans on InfiniBand or Ethernet fabrics (Spectrum-X is NVIDIA's pitch here). Cluster-scale efficiency lives or dies on that fabric: all-reduce bandwidth, congestion behavior under load, and collective library tuning. Two clusters with identical GPUs can differ by 30% or more on real training throughput depending on the network. Always ask for the full topology, not the GPU count.
The software moat around the silicon
Blackwell's advantage isn't only hardware. CUDA, the libraries on top of it (cuDNN, NCCL, TensorRT), and the tooling that NVIDIA has spent two decades polishing mean that a mediocre workload on NVIDIA hardware often beats a heroic workload on competing silicon. Migrating off CUDA is a rewrite, not a config change — ask anyone who has attempted it on AMD's stack or on custom ASICs. That lock-in is uncomfortable for customers and wonderful for NVIDIA's margins, and it's a big part of why the Blackwell ramp stayed ahead of schedule.
There is real competition brewing, though. AMD's Instinct MI300/MI350 family is the strongest general-purpose alternative, with big memory capacity and genuinely competitive inference price-performance for some workloads. Google's TPU, Amazon's Trainium and Microsoft's Maia are credible inside their own clouds. Cerebras and Groq push exotic architectures for latency-sensitive inference. None of them matches the full CUDA ecosystem yet, but “good enough on someone else's cloud” erodes NVIDIA's pricing power at the margins, and NVIDIA knows it — hence the aggressive cadence and the rack-scale lock-in strategy.
The supply situation, quarter by quarter
Blackwell's ramp has been a masterclass in managing a constrained launch. Early 2025 saw the infamous design-flow hiccup and a whisper campaign about delays, which turned out to be a temporary rework of the GB200's cabling and thermal design. By mid-2025, shipments were ramping hard; by late 2025 and into 2026, the constraint shifted decisively to NVLink rack assembly and CoWoS interposer output, not the GPUs themselves. Reported figures floating around suggest NVIDIA was on pace to ship well over a million Blackwell-class GPUs through the cycle, with rack-scale units commanding enormous average selling prices.
Allocation remains everything. If you're a hyperscaler with a multi-billion-dollar order, you get racks; if you're a startup, you get what's left, often via neoclouds paying premiums. Expect this pattern to persist through Rubin: the first gigawatt-scale clusters will be nominated years ahead, and everyone else queues behind them.
The energy elephant in the room
Rack-scale AI has a power problem dressed up as a performance feature. An NVL72 deployment pulls on the order of 100+ kW per rack, and gigawatt-scale clusters now being planned would each draw as much electricity as a mid-size city. Grid interconnection queues in several US regions run years long; cooling retrofits require facilities that barely exist. That's why every credible AI plan for the late 2020s now includes power procurement, on-site generation and dedicated substations as first-class line items. The GPUs are the hard part to build; the electrons are the hard part to find.
Beyond Rubin: Feynman and the cadence problem
The part of the roadmap that gets whispered about is Feynman, the architecture expected after Rubin, reportedly built around a 1.6nm-class process and yet another memory transition. Nobody outside NVIDIA and its co-design partners knows the real specs yet, and honestly, anyone claiming precise numbers this far out is selling noise. What matters is the pattern: NVIDIA has committed to roughly an annual architecture cadence, and that forces TSMC, SK Hynix, ASML and the packaging houses to live on a treadmill they've never had to keep pace with before.
There's a real question of whether that cadence is sustainable or just an aggressive marketing posture. Fabs can't retool in twelve months; a new node takes years and a new package design takes 18 to 24 months of co-engineering. The likeliest reality is a hybrid: full-architecture refreshes every other year with meaningful refreshes (new memory, new interconnect, new rack config) in between. GB300 was exactly that kind of in-stride upgrade, and Rubin Ultra will probably be similar.
For buyers, the cadence cuts both ways. It's great for pushing performance up the curve every funding round, but it also means depreciation math gets weird — a rack bought in early 2026 may look modest by late 2027. If you're signing three-year GPU leases, price in that embarrassment now, because it's coming whether you plan for it or not.
The bottom line
Blackwell won because it delivered rack-scale economics soon enough to matter, and Rubin is the forcing function that will keep the entire supply chain — TSMC, SK Hynix, ASML, power utilities — flat out through 2027. NVIDIA isn't just selling GPUs anymore; it's selling a compute economy with an annual model year. Whether that's healthy or a little unhinged is a fair question, but if you want frontier inference at sane prices, you're shopping in their store for the foreseeable future.