NVIDIA's Blackwell Ultra B300 Is Ramping This Quarter and the Hyperscaler Order Book Is Already Looking at 2027

NVIDIA's Blackwell Ultra B300 is in full production this quarter and the supply data coming out of the cloud providers is the cleanest read on AI demand right now. The B300 began shipping to hyperscalers in January 2026 and the rack-level GB300 NVL72 platform has been ramping through Q1 and into Q2. Capacity is opening up at the major cloud providers in late April. Spot instance availability is now possible for short bursts on AWS, Azure, and GCP. Dedicated reservation lead times for on-premise GB300 systems are still running 12 to 20 weeks, though that is meaningfully shorter than the 36-week wait the prior generation faced last fall.

The chip itself is a meaningful step up from the B200. Each GPU carries 288 GB of HBM3e memory and runs higher FP4 throughput, which is the precision that matters most for inference workloads on the latest large language models. The training side benefits from the same memory bump but is also limited by the rack-level interconnect, which Blackwell Ultra widens at the NVLink layer. Microsoft, Amazon, Meta, and Google are all building data centers around the new platform, and CoreWeave, Oracle, and a handful of newer GPU clouds are filling out the second tier of demand.

The financial picture is large and getting larger. NVIDIA's Q1 fiscal 2026 results, reported in late February, showed data center revenue around $33 billion, with Blackwell representing the majority of that mix by April. Bank of America's late April note pegged GB300 shipments for calendar 2026 at a 129 percent rise over the prior generation. Credible third-party estimates put Blackwell Ultra shipments at up to 60,000 racks for the year. At an average rack price north of $3 million, the rack-level total alone clears $180 billion before the rest of the data center capex picture is added on top.

The hyperscaler capex commitment is the larger number behind the chip story. Combined 2026 capex guidance from Microsoft, Google, Meta, and Amazon sits at roughly $385 billion, with the bulk pointed at AI infrastructure. Tonight's reports from Microsoft, Google, and Meta are likely to push that number up. Each of the four has confirmed multi-quarter purchase commitments for Blackwell and Blackwell Ultra. Several have begun publicly disclosing the rack count and power footprint they expect online by year end.

The supply tightness is the part that gets less attention but matters most. The wait list for B200 and GB200 systems is still running into mid-2026 with a backlog approaching 3.6 million units. The B300 is the unlock for hyperscalers that need to bring AI capacity online this calendar year, and the production ramp at TSMC and the packaging suppliers has been the gating factor. CoWoS-L packaging capacity expanded twice over the last year. Memory partners SK Hynix, Micron, and Samsung are all running at full output for HBM3e. Foundry capacity at TSMC's N4P remains the binding constraint, which is why pricing has held up.

The customer view is split into two camps. The first is the major hyperscaler buying directly from NVIDIA in long-tail commitments. The second is the AI startup buying time on the cloud GPU clouds, where the per-hour pricing has loosened in April for the first time in a year. CoreWeave dropped public hourly pricing for B200 instances by 18 percent earlier this month. The Together AI and Fireworks AI public pricing for inference models running on Blackwell hardware has come down by 22 to 30 percent since January. The supply ramp is reaching the cost curve.

The Rubin transition is the next shadow on the calendar. NVIDIA disclosed at GTC 2026 that Rubin, the next architecture, will begin sampling to hyperscalers in late 2026 and ramp in 2027. Hyperscaler procurement teams have already begun balancing how much Blackwell to commit to versus how much budget to hold back for the Rubin cycle. The right answer for most of them is that they need both, and the capex story does not slow down until at least the second half of 2027. Power constraints, not chip supply, become the binding factor by then. That story is already showing up in utility filings across Virginia, Texas, and the Pacific Northwest.

The takeaway for late April is straightforward. The B300 ramp is real, the supply is improving but still tight, and the hyperscaler order book continues to reach further into the calendar. The next forward read is tonight's earnings from Microsoft, Google, and Meta. Watch the capex guidance and the language on AI infrastructure spending. The chip story has become the data center story, and the data center story is rapidly becoming a power story. Every quarter that the B300 supply tightens at the rack level, the bottleneck moves further upstream into transformer manufacturing, transmission lines, and grid interconnect queues. The companies that own land near substations and water rights near data center corridors are looking at a real estate cycle that has not been priced into the public REIT market yet.