NVIDIA Blackwell Ultra and B300: what the next GPU generation demands

Updated December 11, 2025

December 2025 Update: B300 delivering 15 PFLOPS FP4, 288GB HBM3e (12-high stacks), 8TB/s bandwidth, 1,400W TDP. GB300 NVL72 rack achieving 1.1 EXAFLOPS—exascale in single node. DGX B200 delivering 3x training, 15x inference vs Hopper. Systems shipping H2 2025. Requires liquid cooling, 800Gbps networking, power densities beyond most existing facilities.

The NVIDIA Blackwell Ultra GPU delivers 15 petaflops of dense FP4 compute, 50% more memory than the B200, and 1.5 times faster performance.¹ A single GB300 NVL72 rack achieves 1.1 exaflops of FP4 compute, operating as an exascale supercomputer in a single node.² Equipped with eight NVIDIA Blackwell GPUs, the DGX B200 delivers 3x the training performance and 15x the inference performance of previous-generation Hopper systems.³ The infrastructure requirements for Blackwell differ substantially from anything organizations deployed previously, demanding liquid cooling, 800-gigabit networking, and power densities that most existing facilities cannot support.

NVIDIA announced the B300 at GTC 2025, with systems shipping in the second half of 2025.⁴ The timing creates planning urgency for organizations that need to prepare facilities, secure power, and build operational capabilities before hardware arrives. Understanding Blackwell's infrastructure requirements now determines whether organizations can deploy effectively when systems become available.

Blackwell Ultra specifications

The Blackwell Ultra GPU features a dual-reticle design with 208 billion transistors and 160 streaming multiprocessors across two dies connected using NVIDIA's High-Bandwidth Interface.⁵ The B200 contained 208 billion transistors compared to 80 billion on the H100.⁶ The transistor count reflects the architectural complexity required for AI workloads at frontier scale.

The B300 delivers 288 gigabytes of HBM3e memory per GPU, achieved through 12-high memory stacks instead of the B200's 8-high configuration.⁷ Memory bandwidth reaches 8 terabytes per second.⁸ The memory capacity enables processing of models that previously required multi-GPU configurations on a single GPU.

Power requirements increase significantly. Each B300 chip draws 1,400 watts at the heart of the GB300.⁹ The B200 consumed 1,000 watts, up from the H100's 700 watts.¹⁰ The progression from 700 to 1,000 to 1,400 watts per GPU across three generations demonstrates the power trajectory organizations must plan for.

Dense FP4 performance reaches 14 petaflops on the B300 compared to 9 petaflops on the B200, representing a 55.6% improvement.¹¹ The FP4 compute capability reduces memory footprint by approximately 1.8 times compared to FP8 while maintaining nearly equivalent accuracy.¹² The lower-precision capability addresses inference workloads where reduced precision improves throughput without sacrificing quality.

Performance versus Hopper

Verified performance data shows up to 11 to 15 times faster LLM throughput per GPU compared to Hopper generation.¹³ The HGX B200 delivers up to 15x inference and 3x training improvements versus HGX H100, with 12x energy and cost reduction.¹⁴ The GB200 NVL72 cluster offers 4x faster training and 30x faster real-time inference versus H100 clusters.¹⁵

The B200 delivers 20 petaflops of AI performance from a single GPU. A single H100 had a maximum of 4 petaflops in AI calculations.¹⁶ The 5x improvement per GPU changes the economics of large-scale deployments. Organizations can achieve equivalent capability with fewer GPUs or substantially more capability with equivalent GPU counts.

Memory improvements complement compute gains. The B200 features 192 gigabytes of HBM3e compared to H100's 80 gigabytes of HBM3.¹⁷ Memory bandwidth reaches 8 terabytes per second, 2.4x faster than H100's 3.35 terabytes per second.¹⁸ The memory capacity enables single-GPU processing of models that previously required complex multi-GPU configurations.

For inference workloads, Blackwell delivers 25x lower energy per inference than the H100.¹⁹ A single B200 replaces 5x H100 nodes for Llama 3 inference, cutting costs and carbon footprints.²⁰ The efficiency gains compound across large deployments where inference dominates compute demand.

Architectural differences from Hopper

Hopper targets a broad mix of high-performance computing and AI workloads with focus on traditional precision in FP64 and FP32.²¹ Blackwell optimizes explicitly for large-scale generative AI tasks.²² The architectural focus reflects NVIDIA's assessment that AI workloads, particularly inference, will dominate GPU demand.

Blackwell introduces fifth-generation tensor cores with ultra-low-precision modes supporting 4-bit and 6-bit operations.²³ The low-precision capabilities accelerate inference workloads where quantized models maintain acceptable quality. Training workloads that require higher precision benefit less from the architectural changes.

NVLink connectivity increases dramatically. Each Blackwell GPU has 18 fifth-generation NVLink connections, 18 times more than available on the H100.²⁴ Each connection offers 50 gigabytes per second of bidirectional bandwidth.²⁵ The expanded interconnect enables the GB300 NVL72's architecture where 72 GPUs operate as a unified compute fabric.

For pure HPC numeric tasks including matrix algebra, fluid dynamics, and molecular dynamics with double precision, Hopper's strengths in FP64 per-watt, large shared memory, and well-provisioned caches for FP32 maintain advantage.²⁶ Organizations with traditional HPC workloads should not assume Blackwell improves all use cases equally.

GB300 NVL72 rack architecture

The liquid-cooled GB300 NVL72 rack integrates 36 Grace Blackwell Superchips, interconnected through NVLink 5 and NVLink Switching.²⁷ The rack contains 72 B300 GPUs, each with 288 gigabytes of HBM3e memory.²⁸ With each GPU interconnected via 1.8 terabytes per second of NVLink bandwidth, the system operates as a single exascale node.²⁹

The GB300 NVL72 enables 50x higher AI factory output, combining 10x better latency and 5x higher throughput per megawatt relative to Hopper platforms.³⁰ The efficiency gains demonstrate why liquid cooling requirements represent investment rather than overhead.

The DGX B300 system provides 2.3 terabytes of HBM3e memory with eight ConnectX-8 SuperNICs for 800-gigabit networking.³¹ The networking requirements match the compute capability. Undersized network fabrics create bottlenecks that waste GPU capacity.

Put eight NV72L racks together to form the full Blackwell Ultra DGX SuperPOD: 288 Grace CPUs, 576 Blackwell Ultra GPUs, 300 terabytes of HBM3e memory, and 11.5 exaflops of FP4 compute.³² The scale represents what frontier AI labs deploy for training the largest models.

Infrastructure requirements

Power and cooling requirements exceed what most existing facilities provide. The 4U HGX B300 system uses Supermicro's DLC-2 technology to capture up to 98% of heat through liquid cooling.³³ Air cooling cannot dissipate the thermal output. Organizations planning Blackwell deployments must implement liquid cooling infrastructure.

The 2-OU OCP liquid-cooled HGX B300 system enables up to 144 GPUs per rack for hyperscale and cloud providers.³⁴ A single ORV3 rack supports up to 18 nodes with 144 GPUs total, scaling with Quantum-X800 InfiniBand switches and 1.8-megawatt in-row coolant distribution units.³⁵ Eight HGX B300 compute racks, three Quantum-X800 InfiniBand networking racks, and two in-row CDUs form a SuperCluster scalable unit with 1,152 GPUs.³⁶

Networking requires 800-gigabit connectivity. Both the 2-OU OCP and 4U platforms double compute fabric network throughput to 800 gigabits per second via integrated ConnectX-8 SuperNICs.³⁷ The ConnectX-8 SuperNIC's I/O module hosts two ConnectX-8 devices for 800 gigabits per second of network connectivity per GPU.³⁸ Organizations with 400-gigabit infrastructure face upgrade requirements.

Hyperscaler and enterprise availability

Google Cloud became the first hyperscaler to announce preview availability of B200-based offerings.³⁹ AWS, Google Cloud, Microsoft Azure, and Oracle Cloud Infrastructure are among the first cloud providers to offer Blackwell-powered instances.⁴⁰ The hyperscaler availability provides cloud-based access for organizations not ready to deploy on-premises infrastructure.

HPE shipped its first NVIDIA Blackwell family solution, the GB200 NVL72, in February 2025.⁴¹ Global system makers Cisco, Dell, HPE, Lenovo, and Supermicro offer NVIDIA-Certified RTX PRO Servers with Blackwell.⁴² The vendor ecosystem matured rapidly from announcement to production availability.

Pegatron and 5C successfully deployed liquid-cooled racks based on HGX B200 with in-row CDU integration at a Maryland data center alongside air-cooled systems.⁴³ The deployment demonstrates production-ready infrastructure for organizations building their own AI factories.

Supply constraints affect availability. Demand from hyperscalers and AI labs overwhelms production capacity.⁴⁴ Major hyperscalers and AI companies order numerous nodes while smaller organizations can afford only limited quantities.⁴⁵ NVIDIA faces a backlog of Blackwell chips, partly due to design issues in early production.⁴⁶ Getting large clusters operational typically takes three additional months beyond initial delivery.⁴⁷

Deployment recommendations

Organizations should determine whether Blackwell's capabilities justify infrastructure investments. For inference-dominated workloads, Blackwell's efficiency gains prove compelling. For training workloads requiring FP64 precision, Hopper may remain appropriate.

Organizations can continue training large models on H100 or H200 GPUs while using B200 or B300 for inference and deployment tasks where Blackwell provides the largest throughput and latency gains.⁴⁸ The hybrid approach optimizes infrastructure investment across workload types.

Pricing reflects capability improvements. Early listings suggest B200 192GB SXM at $45,000 to $50,000 per GPU.⁴⁹ Complete 8x B200 server systems can exceed $500,000.⁵⁰ The capital requirements favor organizations with clear AI revenue models or strategic mandates.

The B200 suits model inference at scale, scientific computing, FP64 workloads, and multi-GPU systems with 4 to 8 GPUs.⁵¹ The B300 proves best for LLM training with higher throughput and NVLink fabric, model inference at scale, and supercomputers.⁵² The distinction helps organizations choose appropriate configurations.

Infrastructure investment decisions should account for Blackwell's liquid cooling, 800-gigabit networking, and power requirements. Organizations with existing air-cooled facilities face retrofit costs or new construction. Those without 800-gigabit network infrastructure need fabric upgrades. Facilities without adequate power density cannot host Blackwell systems regardless of other preparations.

The infrastructure gap between Hopper and Blackwell requirements exceeds any previous NVIDIA generation transition. Organizations that begin planning now position themselves for deployment when systems become available. Those that delay will find facility constraints limit their AI capabilities regardless of GPU budget.

Key takeaways

For infrastructure architects: - B300: 15 PFLOPS FP4, 288GB HBM3e (12-high stacks), 8TB/s memory bandwidth, 1,400W TDP per GPU - GB300 NVL72: 72 GPUs, 1.1 exaflops FP4, 1.8TB/s NVLink bandwidth per GPU; DGX SuperPOD: 576 GPUs, 11.5 exaflops - Power progression: H100 (700W) → B200 (1,000W) → B300 (1,400W); infrastructure gap exceeds any previous generation transition

For procurement teams: - B200 192GB SXM: $45,000-$50,000 per GPU; complete 8x B200 server systems exceed $500,000 - Supply constraints persist; demand from hyperscalers overwhelms production capacity with 3+ month deployment lag after delivery - HPE shipped first GB200 NVL72 February 2025; Cisco, Dell, HPE, Lenovo, and Supermicro offer Blackwell systems

For operations teams: - Liquid cooling mandatory: Supermicro DLC-2 captures 98% of heat; air cooling cannot dissipate thermal output - 800Gbps networking required via ConnectX-8 SuperNICs; 400Gbps infrastructure requires upgrade - ORV3 rack supports up to 144 GPUs (18 nodes); SuperCluster unit = 8 compute racks, 3 networking racks, 2 CDUs, 1,152 GPUs

For finance teams: - 30x faster real-time inference vs H100 clusters; 25x lower energy per inference; single B200 replaces 5x H100 nodes for Llama 3 - HGX B200: 15x inference and 3x training improvement over HGX H100 with 12x energy/cost reduction - Hybrid approach recommended: H100/H200 for training, B200/B300 for inference workloads

References

1. Tom's Hardware. "Nvidia announces Blackwell Ultra B300 — 1.5X faster than B200 with 288GB HBM3e and 15 PFLOPS dense FP4." March 2025. https://www.tomshardware.com/pc-components/gpus/nvidia-announces-blackwell-ultra-b300-1-5x-faster-than-b200-with-288gb-hbm3e-and-15-pflops-dense-fp4

2. WCCFTech. "NVIDIA Blackwell Ultra B300 Unleashing In 2H 2025 - 1.1 Exaflops of FP4 Compute." 2025. https://wccftech.com/nvidia-blackwell-ultra-b300-2h-2025-1-1-exaflops-fp4-ai-288-gb-hbm3e-memory-50-percent-faster-than-gb200/

3. NVIDIA. "DGX B200: The Foundation for Your AI Factory." 2025. https://www.nvidia.com/en-us/data-center/dgx-b200/

4. Tom's Hardware. "Nvidia announces Blackwell Ultra B300."

5. NVIDIA Developer Blog. "Inside NVIDIA Blackwell Ultra: The Chip Powering the AI Factory Era." 2025. https://developer.nvidia.com/blog/inside-nvidia-blackwell-ultra-the-chip-powering-the-ai-factory-era/

6. Exxact Corp. "Comparing Blackwell vs Hopper | B200 & B100 vs H200 & H100." 2025. https://www.exxactcorp.com/blog/hpc/comparing-nvidia-tensor-core-gpus

7. Tom's Hardware. "Nvidia announces Blackwell Ultra B300."

8. Glenn Klockwood. "NVIDIA B300." 2025. https://www.glennklockwood.com/garden/processors/B300

9. TechPowerUp. "NVIDIA GB300 Blackwell Ultra Will Feature 288 GB HBM3E Memory, 1400 W TDP." 2025. https://www.techpowerup.com/330154/nvidia-gb300-blackwell-ultra-will-feature-288-gb-hbm3e-memory-1400-w-tdp

10. ARC Compute. "Should You Wait for NVIDIA B300 or Go with H200 or B200 Now?" 2025. https://www.arccompute.io/arc-blog/should-you-wait-for-nvidia-b300-or-go-with-h200-or-b200-now

11. Verda. "NVIDIA B300 vs. B200: Complete GPU comparison to date." 2025. https://verda.com/blog/nvidia-b300-vs-b200-complete-gpu-comparison-to-date

12. NVIDIA Developer Blog. "Inside NVIDIA Blackwell Ultra."

13. Intuition Labs. "Blackwell vs Hopper: A Deep Dive GPU Architecture Comparison." 2025. https://intuitionlabs.ai/articles/blackwell-vs-hopper-gpu-architecture-comparison

14. NexGen Cloud. "NVIDIA Blackwell vs NVIDIA Hopper: A Detailed Comparison." 2025. https://www.nexgencloud.com/blog/performance-benchmarks/nvidia-blackwell-vs-nvidia-hopper-a-detailed-comparison

15. NexGen Cloud. "NVIDIA Blackwell vs NVIDIA Hopper."

16. Exxact Corp. "Comparing Blackwell vs Hopper."

17. Exxact Corp. "Comparing Blackwell vs Hopper."

18. Exxact Corp. "Comparing Blackwell vs Hopper."

19. Northflank. "B200 vs H200: Best GPU for LLMs, vision models, and scalable training." 2025. https://northflank.com/blog/b200-vs-h200

20. Northflank. "B200 vs H200."

21. Hyperbolic. "Hopper or Blackwell? Choosing the Right GPU Architecture." 2025. https://www.hyperbolic.ai/blog/comparing-nvidia-s-hopper-and-blackwell-architectures-for-ai-ml-practitioners

22. Hyperbolic. "Hopper or Blackwell?"

23. Hyperbolic. "Hopper or Blackwell?"

24. Continuum Labs. "Hopper versus Blackwell." 2025. https://training.continuumlabs.ai/infrastructure/servers-and-chips/hopper-versus-blackwell

25. Continuum Labs. "Hopper versus Blackwell."

26. Hyperbolic. "Hopper or Blackwell?"

27. WCCFTech. "NVIDIA Blackwell Ultra B300 Unleashing In 2H 2025."

28. Introl. "NVIDIA GB300 NVL72 Deployment Guide: Blackwell Ultra Specs & Engineering." 2025. https://introl.com/blog/why-nvidia-gb300-nvl72-blackwell-ultra-matters

29. Introl. "NVIDIA GB300 NVL72 Deployment Guide."

30. NVIDIA Developer Blog. "Inside NVIDIA Blackwell Ultra."

31. NVIDIA Newsroom. "NVIDIA Blackwell Ultra DGX SuperPOD Delivers Out-of-the-Box Performance." 2025. https://nvidianews.nvidia.com/_gallery/download_pdf/67d9c8093d6332de871510d6/

32. NVIDIA Newsroom. "NVIDIA Blackwell Ultra DGX SuperPOD."

33. AI Journal. "Supermicro Expands NVIDIA Blackwell Portfolio with New 4U and 2-OU Liquid-Cooled NVIDIA HGX B300 Solutions." 2025. https://aijourn.com/supermicro-expands-nvidia-blackwell-portfolio-with-new-4u-and-2-ou-ocp-liquid-cooled-nvidia-hgx-b300-solutions-ready-for-high-volume-shipment/

34. AI Journal. "Supermicro Expands NVIDIA Blackwell Portfolio."

35. AI Journal. "Supermicro Expands NVIDIA Blackwell Portfolio."

36. AI Journal. "Supermicro Expands NVIDIA Blackwell Portfolio."

37. AI Journal. "Supermicro Expands NVIDIA Blackwell Portfolio."

38. AI Journal. "Supermicro Expands NVIDIA Blackwell Portfolio."

39. Google Cloud Blog. "Introducing A4 VMs powered by NVIDIA B200 GPU aka Blackwell." 2025. https://cloud.google.com/blog/products/compute/introducing-a4-vms-powered-by-nvidia-b200-gpu-aka-blackwell

40. NVIDIA Newsroom. "NVIDIA Blackwell Platform Arrives to Power a New Era of Computing." 2024. https://nvidianews.nvidia.com/news/nvidia-blackwell-platform-arrives-to-power-a-new-era-of-computing

41. HPE Newsroom. "HPE announces shipment of its first NVIDIA Grace Blackwell system." February 2025. https://www.hpe.com/us/en/newsroom/press-release/2025/02/hpe-announces-shipment-of-its-first-nvidia-grace-blackwell-system.html

42. NVIDIA Newsroom. "NVIDIA RTX PRO Servers With Blackwell Coming to World's Most Popular Enterprise Systems." 2025. https://nvidianews.nvidia.com/news/nvidia-rtx-pro-servers-with-blackwell-coming-to-worlds-most-popular-enterprise-systems

43. PR Newswire. "PEGATRON Strengthens AI Infrastructure Collaboration with Together AI and 5C for NVIDIA GB300 NVL72 and NVIDIA HGX B200 Liquid-Cooled Rack Deployment." November 2025. https://www.prnewswire.com/news-releases/pegatron-strengthens-ai-infrastructure-collaboration-with-together-ai-and-5c-for-nvidia-gb300-nvl72-and-nvidia-hgx-b200-liquid-cooled-rack-deployment-in-us-data-centers-302617887.html

44. DEV Community. "What Is NVIDIA Blackwell Ultra? 2025 GPU Supply Crunch Guide." 2025. https://dev.to/banelygalan/what-is-nvidia-blackwell-ultra-2025-gpu-supply-crunch-guide-22ci

45. DEV Community. "What Is NVIDIA Blackwell Ultra?"

46. DEV Community. "What Is NVIDIA Blackwell Ultra?"

47. Data Center Dynamics. "Nvidia delays Blackwell GPUs, impacting hyperscaler data center plans." 2024. https://www.datacenterdynamics.com/en/news/nvidia-delays-blackwell-gpus-impacting-hyperscaler-data-center-plans-report/

48. Hyperbolic. "Hopper or Blackwell?"

49. Northflank. "How much does an NVIDIA B200 GPU cost?" 2025. https://northflank.com/blog/how-much-does-an-nvidia-b200-gpu-cost

50. Northflank. "How much does an NVIDIA B200 GPU cost?"

51. Verda. "NVIDIA B300 vs. B200."

52. Verda. "NVIDIA B300 vs. B200."

SEO Elements

Squarespace Excerpt (159 characters): B300 delivers 15 PFLOPS FP4, 288GB HBM3e at 1,400W per GPU. GB300 NVL72 reaches 1.1 exaflops. Infrastructure requirements for Blackwell Ultra deployment.

SEO Title (55 characters): NVIDIA Blackwell Ultra B300: Infrastructure Requirements

SEO Description (155 characters): B300 delivers 1.5x B200 performance with 288GB HBM3e. GB300 NVL72 reaches 1.1 exaflops. Analysis of power, cooling, and networking requirements for Blackwell.

URL Slugs: - Primary: nvidia-blackwell-ultra-b300-infrastructure-requirements - Alt 1: gb300-nvl72-deployment-specifications-2025 - Alt 2: blackwell-vs-hopper-gpu-architecture-comparison - Alt 3: nvidia-b300-b200-enterprise-deployment-guide