The cost of serving AI in production is becoming the number that decides what you can ship. NVIDIA built Vera Rubin to change that, and teams are claiming the first allocations now. Early access is open at Axe Compute, ahead of the general-availability queue.
The Inference Bill Is Outrunning the Budget
Picture where most AI teams sit by late 2026. Usually, a reasoning product or agent is in production, usage is climbing, and the inference bill is climbing faster. Meanwhile, context windows keep growing, and each one holds GPU memory longer. As a result, the capacity that would fix the unit economics sits either with a hyperscaler or three quarters out. The workload is working. The math is the problem.
Vera Rubin solves exactly that problem. Specifically, NVIDIA reports up to 10 times lower cost per generated token than Blackwell. That kind of shift decides whether an agent product runs at a margin or at a loss. Consequently, the teams that reach those economics first, and get them into production while competitors wait in line, set the pace for everyone else.
Reservations for NVIDIA Vera Rubin NVL72 capacity are open at Axe Compute. Axe plans the first United States deployment wave for August to September 2026. Afterward, a European rollout follows in Q1 2027. NVIDIA launched the Rubin platform at CES on 5 January 2026. In May, it confirmed that production shipments begin this fall. Early access now comes down to queue position, and the queue is forming.
One Vera Rubin NVL72 rack delivers 3.6 exaFLOPS of NVFP4 inference, carries 20.7 TB of HBM4, and moves 260 TB/s across its NVLink 6 fabric. Notably, that performance draws an estimated 200 kilowatts or more per rack, which is why where and how you deploy it matters as much as when.
NVFP4 inference per rack
HBM4 bandwidth per GPU
NVLink 6 fabric per rack
first US wave
Scaling AI Is Now a Compute Decision
Vera Rubin is the first NVIDIA platform built for where AI is heading: long-running, context-heavy, agentic work rather than one-shot prompts. Specifically, the models gaining ground plan across steps, hold memory, and reason over long horizons. These workloads lean on memory bandwidth and interconnect more than raw compute. Vera Rubin answers exactly that. In particular, each GPU nearly triples Blackwell’s memory bandwidth, and a single NVL72 rack serves 3.6 exaFLOPS of inference. In short, the platform that serves the most tokens per watt and per dollar sets the ceiling on what your AI can profitably do. Indeed, this generation raises that ceiling.
The leap arrives scarce. The largest cloud platforms hold the earliest supply. As a result, the advantage goes to teams that reserve Vera Rubin early. They also need a partner who can vet it and stand it up while the gains still count. Being first is necessary. However, being first with the right partner is what turns the reservation into results. That window is open now, and it spans quarters, not weeks.
Axe Compute Handles the Compute So You Can Scale
Vera Rubin is a decision before it is a deployment. Which configuration fits the workload, when to move from Blackwell, whether to take capacity on-demand or build a dedicated cluster: these choices decide whether the investment pays off. Therefore, Axe is the partner that helps you get them right.
We vet and source the hardware and match it to the workload, across any GPU type and configuration, with no lock-in. As a result, Rubin goes where inference needs it, and Blackwell stays where it still earns its place. Take capacity on-demand, and it arrives vetted and ready to run. For a dedicated cluster, Axe designs, builds, and operates the full stack, power, cooling, and facilities included. You also pay as operating expense with zero CapEx. Either way, you get an advisor who has stood up this hardware before.
Indeed, this infrastructure already exists. Axe Compute (NASDAQ: AGPU) runs 400,000+ GPUs in 93 countries at 99% uptime, with transparent flat-rate pricing and zero egress fees. In a generation moving this fast, a partner who can vet the hardware and build to spec is what stands between capacity you can count on and a bet you hope pays off.
Rubin Is Shipping, and the First Allocations Are Being Claimed
The Rubin platform is shipping on schedule. NVIDIA launched it at CES on 5 January 2026 with six new chips: the Vera CPU, the Rubin GPU, the NVLink 6 switch, and the networking silicon around them. By late May, the company confirmed full production across 350-plus factories in 30 countries. Subsequently, production shipments begin in the fall, and partner products follow in the second half of 2026.
Axe Compute has secured early-access allocations for enterprise deployments in the United States. The first wave lands in August to September 2026, and Europe follows in Q1 2027. The team processes reservations in the order commitments arrive, and it confirms each allocation window within one business day.
The Bigger Jump Is Memory Bandwidth
One Vera Rubin NVL72 rack behaves as a single accelerator. It combines 72 Rubin GPUs with 36 Vera CPUs in one NVLink 6 domain, each CPU built on 88 custom Olympus Arm cores. In total, every GPU contributes 50 petaFLOPS of NVFP4 inference, which puts the full rack at 3.6 exaFLOPS. By contrast, the Rubin GPU carries 336 billion transistors, against 208 billion on Blackwell.
Memory is the bigger jump, and it matters most. Specifically, each Rubin GPU pairs 288 GB of HBM4 with up to 22 TB/s of bandwidth, nearly triple Blackwell. Across the rack, that reaches 20.7 TB and roughly 1.6 PB/s. Reasoning models and agent loops hold GPU memory for the full length of a task. Consequently, memory throughput, more than raw compute, now sets the ceiling on tokens served per rack. We quantified that pattern in the agentic AI compute analysis.
What NVIDIA Reports
NVIDIA’s platform claims are specific, and they are NVIDIA’s: up to 10 times lower cost per generated token than Blackwell, mixture-of-experts training that needs 4 times fewer GPUs, and 10 times the agent throughput of Grace Blackwell. In addition, the rack runs fully liquid-cooled, and its modular, cable-free trays cut assembly and service time by up to 18 times.
Table 1: Vera Rubin NVL72 at a glance (NVIDIA published specifications)
| Dimension | Vera Rubin NVL72 |
|---|---|
| Rack inference compute | 3.6 exaFLOPS NVFP4 |
| GPUs | 72 Rubin GPUs · 50 PFLOPS NVFP4 each · 336B transistors |
| CPUs | 36 Vera CPUs · 88 custom Olympus Arm cores each |
| Memory | 20.7 TB HBM4 per rack · up to 22 TB/s per GPU |
| Scale-up fabric | NVLink 6 · 3.6 TB/s per GPU · 260 TB/s per rack |
| Scale-out network | 1.6 Tb/s per GPU via ConnectX-9 SuperNICs |
| Cooling and service | Fully liquid-cooled · modular cable-free trays · up to 18x faster servicing |
How Rubin Stacks Up Against B200, B300, and GB300
Vera Rubin is the direct successor to the rack that leads NVIDIA’s current line-up, the GB300 NVL72. The generational step is measurable. For example, NVFP4 inference rises to 3.6 exaFLOPS against 1.44 exaFLOPS FP4 sparse on GB300. Similarly, NVLink bandwidth doubles from 130 to 260 TB/s per rack. Moreover, the move from HBM3e to HBM4 lifts rack memory bandwidth from 576 TB/s to roughly 1.6 PB/s, while capacity holds close (20.7 TB against 20 TB). For the standalone B200 and B300 GPUs and the full Blackwell decision framework, see our NVIDIA Blackwell GPU comparison.
Table 2: GB300 NVL72 vs Vera Rubin NVL72 (NVIDIA published specifications)
| Dimension | GB300 NVL72 | Vera Rubin NVL72 |
|---|---|---|
| FP4 rack inference | 1.44 EF sparse (1.08 EF dense) | 3.6 EF NVFP4 |
| Rack GPU memory | 20 TB HBM3e | 20.7 TB HBM4 |
| Rack memory bandwidth | Up to 576 TB/s | Roughly 1.6 PB/s |
| NVLink fabric | 130 TB/s (1.8 TB/s per GPU) | 260 TB/s (3.6 TB/s per GPU) |
| Transistors per GPU | 208 billion (Blackwell) | 336 billion |
| Cost per token (NVIDIA claim) | Baseline | Up to 10x lower |
Where Rubin Fits First
The workloads that fit Rubin first are the ones straining Blackwell today. Sustained agentic and reasoning inference gains the most, because loop-based workloads lean on memory bandwidth and interconnect rather than raw compute. Long-context serving and mixture-of-experts models follow the same logic. In addition, Vera Rubin NVL72 is the first rack-scale platform with full NVIDIA Confidential Computing, which opens bare-metal deployment to teams working with regulated or proprietary data.
Teams already running Blackwell do not need to choose. B200 and B300 capacity is widely available, and it remains the right platform for training and fine-tuning that needs GPUs now. Rubin, meanwhile, is the expansion path for the inference fleet, and capacity moves to the new architecture as workloads call for it, without renegotiating the deployment.
The Infrastructure Behind the Performance
Rubin’s density is part of what makes it fast. For example, a Vera Rubin NVL72 rack draws an estimated 190 to 230 kilowatts, against roughly 120 to 130 for a Blackwell rack and about 40 for a Hopper rack a generation ago. The 2027 follow-on, Rubin Ultra, targets a 600-kilowatt rack with 800-volt DC distribution behind it. Axe’s live data centers already handle liquid cooling, purpose-built power delivery, and facilities designed around the rack. Take on-demand capacity, and it is simply there; commission a dedicated build, and Axe designs and operates that stack to your specification.
Early Reservation Converts a Queue Into a Date
NVIDIA allocates supply for a new architecture before general availability, and it commits the first Vera Rubin systems to the largest cloud platforms. As a result, enterprise customers usually meet the queue elsewhere: allocation reviews each quarter, delivery dates that shift. We documented those procurement cycles in The 52-Week Wait. Early reservation replaces that uncertainty with a confirmed allocation window.
The economics reward the same decision. Rubin’s efficiency gains land per token. Consequently, a team that moves inference onto the platform in 2026 banks the advantage from the first month. At scale, a large reduction in token cost changes what an agent program can afford to run. By contrast, waiting for general availability defers that by two quarters or more.
You Set the Spec, Axe Handles the Rest
Early access centers on client-specified capacity. First, you define the deployment: region, GPU configuration, and RoCE or InfiniBand interconnect. Then Axe Compute sources and configures the hardware, subject to availability and cost, and the deployment scales with the workload. The order of decisions is the point. The client chooses what the business needs, and the infrastructure follows.
“Clients should not have to design their AI roadmap around whatever GPUs happen to be available. Offering early access to Vera Rubin reflects Axe Compute’s commitment to giving businesses choice with a reliable partner. The client specifies region and needs, and we can deliver Rubin capacity against it, first starting in the United States”
Kyle Okamoto, President, Axe Compute (quote drafted, awaiting approval)
The process is short. Submit your requirements through the Vera Rubin early-access reservation form, and the team scopes the configuration and confirms an allocation window within one business day. The team processes reservations in the order commitments arrive. In addition, commitments made before 31 July 2026 hold launch pricing at 10 to 12% below market. The first US deployments follow in August to September.
The reservation window closes 31 July 2026, and the first racks power on within weeks. Scope a configuration now and hold your window. Finally, you can settle the remaining deployment details with the team once Axe confirms the allocation.
First-wave Vera Rubin NVL72 capacity. Reserve your allocation.
Frequently Asked Questions
What is the NVIDIA Vera Rubin NVL72?
The rack-scale system of the NVIDIA Rubin platform, the successor to Blackwell. It combines 72 Rubin GPUs and 36 Vera CPUs in one NVLink 6 domain, delivering 3.6 exaFLOPS of NVFP4 inference, 20.7 TB of HBM4, and 260 TB/s of fabric bandwidth. NVIDIA launched the platform at CES on 5 January 2026.
When will Vera Rubin be available?
NVIDIA reports the platform ramping into full production, with shipments in fall 2026 and partner products in the second half of the year. At Axe Compute, the first US deployment wave lands in August to September 2026, and Europe follows in Q1 2027.
How does Vera Rubin compare with Blackwell?
NVIDIA reports up to 10 times lower cost per generated token and MoE training that needs 4 times fewer GPUs. The Rubin GPU carries 336 billion transistors against 208 billion on Blackwell, and its HBM4 memory delivers up to 22 TB/s per GPU, nearly triple the prior generation.
How does Vera Rubin early access work at Axe Compute?
You define region, GPU configuration, and RoCE or InfiniBand interconnect. Axe Compute then sources and configures the deployment, subject to availability and cost, and confirms each allocation window within one business day. You submit reservations through the early-access reservation form.
What is the deadline for launch pricing?
Reservations made before 31 July 2026 receive launch pricing at 10 to 12% below market, processed in the order commitments arrive.
Should teams deploy Blackwell now or reserve Vera Rubin capacity?
Both paths stay open. B200 and B300 capacity is widely available and remains right for training and fine-tuning that needs GPUs now. The Vera Rubin NVL72 lists 3.6 exaFLOPS of NVFP4 inference per rack against 1.44 exaFLOPS FP4 sparse on GB300, so teams with sustained inference growth reserve Vera Rubin for the 2026 wave while running current workloads on Blackwell.
Why is Vera Rubin suited to inference?
Memory bandwidth, rather than raw compute, increasingly sets the limit on inference. Each Rubin GPU pairs 288 GB of HBM4 with up to 22 TB/s of bandwidth, and NVLink 6 provides 3.6 TB/s per GPU for communication-heavy reasoning and MoE inference.
About Axe Compute
Axe Compute (NASDAQ: AGPU) provides bare-metal GPU infrastructure across 200+ locations in 93 countries. The platform operates 400,000+ GPUs with 48-hour provisioning, zero egress fees, no virtualisation overhead, and 99% uptime, at transparent flat-rate pricing.
Sources
- NVIDIA press release, 5 Jan 2026 (CES). Rubin platform launch, NVL72 configuration, NVLink 6, cost-per-token and MoE claims. nvidianews.nvidia.com
- NVIDIA Technical Blog, Inside the NVIDIA Rubin Platform. Per-GPU compute, HBM4 capacity and bandwidth, transistor counts, cooling. developer.nvidia.com/blog/inside-the-nvidia-rubin-platform-six-new-chips-one-ai-supercomputer/
- NVIDIA press release, 31 May 2026 (GTC Taipei). Full production across 350+ factories, shipments from fall 2026, 10x agent throughput vs Grace Blackwell. nvidianews.nvidia.com
- NVIDIA GB300 NVL72 product page. 72 Blackwell Ultra GPUs, 36 Grace CPUs, 1,440 PFLOPS FP4 sparse, 20 TB memory at up to 576 TB/s, 130 TB/s NVLink. nvidia.com/en-us/data-center/gb300-nvl72/
- SemiAnalysis and Data Center Dynamics. Rubin NVL72 rack power estimates, Rubin Ultra 600 kW and 800 VDC. newsletter.semianalysis.com · datacenterdynamics.com