At the end of 2025, Nvidia CEO Jensen Huang confirmed a headache for every company wanting to buy Blackwell: the B200 and GB200 NVL72 racks are sold out through mid-2026.
Among the backlog of orders from global cloud providers, just the largest few alone exceed 3.6 million units.
Where the supply chain is stuck
Blackwell's problem isn't a single link — it's the entire chain under strain simultaneously:
CoWoS packaging: TSMC has essentially locked up its relevant capacity for 2026-2027 for Nvidia, leaving other customers waiting.
HBM3e memory: SK Hynix and Micron's entire 2025 capacity is full, and they are planning a 20% price increase in early 2026. Expansion plans at SK Hynix and Samsung are racing to catch up.
Rack assembly: OEMs like Dell, Foxconn, Inventec, and Wistron previously encountered technical issues (mainly engineering challenges with liquid-cooled racks), which have only recently been resolved, allowing production to ramp up.
When these links are jammed, even a finished chip design cannot ship. The good news is that the OEM issues are largely resolved; the bad news is that the first two links remain bottlenecks.
Next generation: Blackwell Ultra (GB300)
Nvidia isn't letting buyers wait indefinitely.
Blackwell Ultra (B300/GB300) is on the way, scheduled for launch in the first half of 2026. Key upgrade: HBM3e memory per GPU increases from 192GB on the B200 to 288GB, which is much friendlier for inference on very large models.
Forecast: GB300 series rack shipments could more than double year-over-year in 2026.
Looking further ahead, the Rubin architecture is expected to enter the market in the second half of 2026, as previewed at GTC.
What it means for AI companies
Companies unable to get the latest GPUs have several paths:
Use H100/H200: Nvidia is discussing restarting H200 production, as Chinese companies like Alibaba and ByteDance have placed orders for over 2 million H200 units. Hopper architecture parts are somewhat easier to obtain than Blackwell.
Wait in line: Super-buyers like Microsoft, Google, and AWS have priority queues, but wait times are still measured in quarters.
Switch to competitors: AMD MI300X is still available, and companies are trying it — but the large-scale software ecosystem still favors CUDA's maturity.
Develop in-house: Google TPU, AWS Trainium, and Microsoft Azure Maia are all paths, but the build-out cycle is longer.
What this situation reveals
The growth in demand for AI infrastructure has already outpaced the global supply chain's ability to respond, and this gap cannot be closed in the short term. It's not that Nvidia doesn't want to sell; it's that the supply chain is only so large.
From TSMC to HBM suppliers, every link is either a monopoly or an oligopoly, with no room for rapid expansion.
The current reality: companies that want the latest Blackwell hardware either signed long-term procurement agreements early, or they have to wait.
Sources: CocoLoop, Nvidia's Blackwell Dynasty: B200 and GB200 Sold Out Through Mid-2026 (FinancialContent/TokenRing); NVIDIA's Blackwell Ultra GB300 AI Servers to Lead the AI Infrastructure Race (WCCFTech); Nvidia server makers solve Blackwell technical issues (Data Center Dynamics)