NVIDIA Launches Blackwell Ultra B300 Cloud Instances with 2.5x Inference Throughput Gain

NVIDIA has announced the general availability of its Blackwell Ultra B300 GPU instances across the three major hyperscalers — AWS, Microsoft Azure, and Google Cloud — marking a significant generational leap in cloud-based AI inference capability. On AWS, the new p6e.48xlarge instance type ships with 8x NVIDIA B300 GPUs and is priced at approximately $98.00 per hour on-demand, a figure that is virtually identical to the incumbent p5.48xlarge (8x H100 SXM5) at $98.32 per hour. Azure is offering the ND B300 v6-series, with the nd192adb300v6 SKU (8x B300) launching at roughly $96.80 per hour in East US and West Europe regions, compared to the ND H100 v5-series nd96asr_v4 at $98.00 per hour. Google Cloud is bringing the B300 to market via its A4 Ultra machine family, with the a4u-megagpu-8g configuration priced at approximately $97.50 per hour in us-central1, versus the A3 Mega (a3-megagpu-8g with 8x H100) at $98.00 per hour. Oracle Cloud Infrastructure and IBM Cloud have not yet announced B300 availability, meaning enterprises tied to those platforms face a temporary but material performance gap, a caveat worth flagging in any multi-cloud roadmap.

The headline technical improvement is a 2.5x inference throughput gain over the H100 SXM generation. A single B300 instance can process up to 15,000 tokens per second for a 70B-parameter model such as Meta Llama 3.1 70B or Mistral 70B, compared to approximately 6,000 tokens per second on an equivalent H100 instance. This means enterprises serving the same peak inference load can do so with 60% fewer GPU instances, dropping from, say, 10 p5.48xlarge nodes to just 4 p6e.48xlarge nodes for equivalent throughput. Because the hourly on-demand rates are essentially at parity, the cost-per-token reduction lands in the 35–45% range without any commitment or architectural change. For organisations running Reserved Instances or Committed Use Discounts at the H100 tier — typically 1-year or 3-year terms at 30–40% discounts — the calculus shifts slightly, since existing commitments may need to be run to term before migration is economically optimal. Reserved p5.48xlarge on AWS runs approximately $63.90 per hour on a 1-year no-upfront term; the equivalent p6e.48xlarge 1-year reserved rate has not yet been formally published but is expected around $63.00–$64.50 per hour based on historical GPU generation pricing patterns.

The customer segments with the most immediate upside are enterprises running large-scale LLM inference for production chatbots, autonomous coding assistants (GitHub Copilot-style deployments), high-volume document processing pipelines, and real-time recommendation systems built on transformer architectures. For a company processing 10 billion tokens per day across a fleet of 20 H100 instances, the migration to B300 effectively compresses that fleet to 8 instances at equivalent or lower total cost, freeing 12 instance-hours per hour — worth roughly $1,176 per hour or approximately $10.3 million annually at on-demand rates. The competitive pressure on AWS, Azure, and Google Cloud is now focused on allocation availability and regional reach rather than price, since all three have matched pricing almost exactly. The near-term risk for customers is GPU quota and capacity constraints — early access in GA does not guarantee unconstrained scale, and enterprises should expect waitlists in certain regions through mid-2026. Lock-in is a secondary concern: B300 instances use the same CUDA stack and NVLink interconnects as H100, meaning application-layer portability is high, but the instance families themselves differ enough that infrastructure-as-code templates and autoscaling policies require updates.

For CIOs and FinOps leads evaluating a migration, the recommended approach is to begin with a targeted inference workload audit within the next 30 days. Any workload currently consuming more than 5 H100-equivalent GPU-hours per day is a strong B300 migration candidate. Organisations should run a side-by-side benchmark of their specific model and sequence length distribution before committing to fleet-wide migration, as the 2.5x throughput figure is measured at optimal batch sizes and may vary for low-latency, single-request inference patterns. Teams holding H100 Reserved Instances with more than 12 months remaining should model the break-even point between running out existing commitments versus paying the early-exit or partial overlap cost to switch. For net-new inference infrastructure decisions being made in Q1–Q2 2026, B300 should be the default GPU selection on AWS, Azure, and Google Cloud, and procurement teams should request quota increases now to avoid allocation delays in Q3 2026.

At TCOIQ, we recommend starting with the TCO Calculator at tcoiq.com/tco.html to model the 35–45% cost-per-token reduction against your actual token volume and current GPU fleet composition — inputting your p5.48xlarge or ND H100 v5 hourly spend alongside your Reserved Instance commitments will surface a precise break-even timeline. The Inventory Builder at tcoiq.com/inventory.html can scan your existing cloud accounts to identify every H100-class instance currently in your environment, flag commitment expiry dates, and rank workloads by migration readiness. TCOIQ's AI Migration Assessment then maps those workloads against B300 compatibility, batch size optimisation potential, and regional availability, producing a prioritised migration roadmap with projected savings. For organisations planning a broader infrastructure modernisation, the Landing Zone Assessment ensures your networking, IAM, and autoscaling configurations are B300-ready before you scale. The single most valuable next step is to load your current GPU inventory into the TCOIQ Inventory Builder today — it takes under 15 minutes via read-only cloud role, and the output will immediately tell you which H100 workloads should migrate in the next 60 days versus which should run to commitment expiry.