NVIDIA Blackwell Ultra

NVIDIA Blackwell Ultra is a mid-cycle refresh of NVIDIA's Blackwell data-center GPU architecture, announced at NVIDIA's GTC conference on March 18, 2025, and built for what NVIDIA calls the "age of AI reasoning." It centers on the B300 GPU and the GB300 (Grace Blackwell Ultra) superchip, each B300 carrying up to 288 GB of HBM3e memory (50 percent more than the 192 GB B200) and up to 15 petaFLOPS of dense NVFP4 compute, which NVIDIA describes as 1.5x the dense FP4 performance of the original B200. Delivered chiefly through the rack-scale GB300 NVL72 system (72 Blackwell Ultra GPUs and 36 Grace CPUs) and the HGX B300 server board, it began shipping from partners in the second half of 2025: the first GB300 NVL72 rack reached a cloud customer in July 2025, and NVIDIA reported that production shipments of Blackwell Ultra platforms began in its fiscal quarter ended July 27, 2025. ^{[1][2][3][30][55]}

Blackwell Ultra is not a new microarchitecture, but the B300 is a different chip from the B200. NVIDIA told Tom's Hardware at launch that Blackwell Ultra GPUs "are different chips than Blackwell GPUs (GB200 and B200)," and NVIDIA's CUDA documentation lists the B300 as compute capability 10.3, against 10.0 for the B200. ^[6][13] It keeps Blackwell's dual-die design, the same fifth-generation NVLink interconnect, and the same NVFP4 numerical format as the base generation, while raising memory capacity, dense FP4 and attention-layer throughput, and power budget, and giving up most of the B200's FP64 and INT8 throughput. ^[3][11][12] In NVIDIA's cadence it occupies the slot between the original 2024 Blackwell launch and the next-generation Vera Rubin platform, which began production shipments in August 2026, mirroring the earlier mid-life update of the prior Hopper generation, when the H200 raised the H100's 80 GB of memory to 141 GB of HBM3e. ^[3][15][58] NVIDIA chief executive Jensen Huang positioned it directly at reasoning and agentic workloads, saying AI "has made a giant leap" and that "reasoning and agentic AI demand orders of magnitude more computing performance." ^[1][2]

What is Blackwell Ultra?

Blackwell Ultra is the "Ultra" mid-cycle upgrade of NVIDIA's Blackwell GPU line, optimized for large-scale, token-heavy inference rather than for a wholesale compute redesign. The core changes over base Blackwell are larger memory (up to 288 GB of HBM3e per GPU) and higher dense low-precision math (up to 15 petaFLOPS of NVFP4), both aimed at serving reasoning models that emit long chains of thought at test time. NVIDIA describes the platform as a "single versatile platform" that handles pretraining, post-training, and reasoning inference, and frames its purpose around "AI factories" judged by tokens per second per user and tokens per second per megawatt, claiming more output per megawatt than any previous NVIDIA platform. Jensen Huang said at launch, "We designed Blackwell Ultra for this moment," calling it "a single versatile platform that can easily and efficiently do pretraining, post-training and reasoning AI inference." ^[1][2][3]

When was Blackwell Ultra announced?

NVIDIA unveiled the Blackwell Ultra AI factory platform on March 18, 2025, during Jensen Huang's keynote at GTC in San Jose. The company framed the launch around inference for reasoning models, which generate far more output tokens per query than conventional chatbots and therefore stress both compute and memory during the generation (decode) phase. NVIDIA tied the platform to test-time scaling inference, which it called "the art of applying more compute during inference to improve accuracy," and claimed the GB300 NVL72 increases Blackwell's revenue opportunity for AI factories 50x compared with those built with NVIDIA Hopper. ^[1][2][6]

The platform announced at GTC comprised two principal system designs: the rack-scale GB300 NVL72 and the eight-GPU HGX B300 NVL16 board for conventional servers. NVIDIA said Blackwell Ultra products would be available from partners starting in the second half of 2025, and that the GB300 NVL72 would also be offered through NVIDIA DGX Cloud. ^[1]

What is the B300 GPU?

The B300 is the Blackwell Ultra GPU. Like the base B200, it is a single package containing two reticle-limit dies built on a TSMC 4-nanometer-class (4NP) process and joined by the NV-HBI interface so software treats them as one device. The full GPU implementation has 160 streaming multiprocessors (SMs) organized into eight graphics processing clusters, with 640 fifth-generation Tensor Cores, and it supports NVIDIA's NVFP4, FP6, and FP8 formats through a second-generation Transformer Engine; NVIDIA notes that the available SM count and HBM capacity vary by product. ^[3][5]

The headline figures for the B300 are up to 288 GB of HBM3e (eight 12-high stacks) delivering about 8 TB/s of bandwidth, and 15 petaFLOPS of dense NVFP4 tensor compute (about 20 petaFLOPS with sparsity) in its GB300 form. The doubling of the special-function-unit throughput used for exponentials (10.7 trillion exponentials per second, against 5 trillion for Blackwell and 4.5 trillion for Hopper) gives up to 2x faster attention-layer compute than base Blackwell, a targeted change because the attention layer's softmax can become a latency bottleneck when reasoning models work through long contexts. Each GPU connects to its host Grace CPU over NVLink-C2C at 900 GB/s and to peers over fifth-generation NVLink at 1.8 TB/s. ^[3]

Announcement and shipments

NVIDIA announced the B300 on March 18, 2025 and published its chip-level details in an August 22, 2025 technical blog post. ^[1][3] The first GB300 NVL72 rack shipped in July 2025, Supermicro began volume shipments of HGX B300 servers in September 2025, and Amazon Web Services made B300-based instances generally available on November 18, 2025. ^[20][30][33]

Dies, transistors and cores

The two dies hold 208 billion transistors between them, 2.6 times as many as a Hopper GPU, and the NV-HBI link between them carries 10 TB/s. Each SM contains 128 CUDA cores (20,480 across 160 SMs), four Tensor Cores, 256 KB of Tensor Memory for intermediate results, and special function units for transcendental math. ^[3][5] CUDA reports the B300 and GB300 as compute capability 10.3, a separate target from the 10.0 of the B200 and GB200. ^[13] The B300 is also distinct from the B30A, a single-die Blackwell chip that NVIDIA was reported in August 2025 to be developing for China with roughly half the B300's raw performance; NVIDIA said only that it evaluates "a variety of products for our roadmap." ^[17]

Memory

The B300 carries eight 12-high HBM3e stacks of 36 GB each on sixteen 512-bit memory controllers, an 8,192-bit interface. That gives a maximum of 288 GB and 8 TB/s per GPU, 3.6 times the memory of an H100 and the same bandwidth as the B200. ^[3][6] Shipping products expose a little less: NVIDIA's Blackwell Ultra datasheet lists 279 GB at 8 TB/s per GPU in the GB300 NVL72 and 270 GB at 7.7 TB/s per GPU in HGX B300 servers. ^[11] The GPU can be split into up to seven Multi-Instance GPU partitions, for example two instances of 140 GB, four of 70 GB or seven of 34 GB. ^[3]

Compute rates

NVIDIA's datasheet gives these dense Tensor Core rates for one Blackwell Ultra GPU in a GB300 NVL72: 15 petaFLOPS of FP4, 5 petaFLOPS of FP8 or FP6, and 2.5 petaFLOPS of FP16 or BF16; in HGX B300 servers, rated at up to 1,100 W per GPU rather than 1,400 W, they are 14, 4.5 and 2.25 petaFLOPS. Sparsity doubles most of these rates but adds only a third for FP4: NVIDIA rates the B300 at 20 petaFLOPS of sparse FP4 in GB300 systems, the same sparse figure it gives the B200 in GB200 systems, so Blackwell Ultra's FP4 gain is entirely in dense math. ^[3][11][12]

Interconnect and networking

Each B300 has eighteen fifth-generation NVLink links at 100 GB/s, 1.8 TB/s of bidirectional bandwidth, the same as the B200 and twice Hopper's NVLink 4; the fabric scales to 576 GPUs in one non-blocking NVLink domain. The host interface is PCIe Gen6 x16 at 256 GB/s, where NVIDIA's Blackwell datasheet lists PCIe Gen5 at 128 GB/s for the B200. ^[3][11][12] Blackwell Ultra platforms pair each GPU with an 800 Gb/s ConnectX-8 SuperNIC, twice the 400 Gb/s ConnectX-7 per GPU of GB200 systems. ^[1][11][14]

Where published figures differ

Quoted B300 figures depend on the configuration. NVIDIA's technical blog compares 15 petaFLOPS of dense NVFP4 for Blackwell Ultra with 10 for Blackwell; those are the GB300 and GB200 figures, and in eight-GPU HGX servers NVIDIA lists 14 and 9 (108 and 72 petaFLOPS per board, so 13.5 per HGX B300 GPU). ^{[3][11][12][18]} Memory is quoted as 288 GB (the maximum), 279 GB (GB300 NVL72) and 270 GB (HGX B300), while cloud providers list 263 GB (Oracle) and 268 GB (Verda) for their HGX-class B300 instances. ^{[3][11][63][68]} Power is listed by NVIDIA as up to 1,100 W for HGX B300, but Supermicro said in September 2025 that both GB300 and B300 "can be configured to use up to 1400W per GPU." ^[11][20]

How is Blackwell Ultra different from Blackwell (B200)?

The defining changes from base Blackwell are larger memory and higher dense low-precision math, both aimed at inference. NVIDIA's technical materials emphasize three deltas over the B200: a step up from 192 GB to 288 GB of HBM3e per GPU, an increase in dense NVFP4 tensor throughput from about 10 to 15 petaFLOPS (which NVIDIA describes as a roughly 1.5x increase), and up to 2x faster acceleration of the attention layer's softmax (exponential) operations, which can become a latency bottleneck in long-context generation. The added memory lets a single GPU hold larger models and far longer key-value caches, which is precisely the bottleneck for reasoning workloads that emit thousands of tokens. ^[3][5]

Importantly, several core attributes are unchanged from the base generation, even though the B300 is a different chip. Blackwell Ultra keeps the arrangement of two reticle-sized GPU dies fused by NVIDIA's NV-HBI die-to-die link (about 10 TB/s) into a single logical GPU totaling 208 billion transistors, the same fifth-generation NVLink at 1.8 TB/s of bidirectional bandwidth per GPU, and the same 4-bit NVFP4 format introduced with Blackwell. The principal trade-off for the higher performance is power: NVIDIA's datasheets rate the B300 as configurable up to 1,400 W in GB300 systems and 1,100 W in HGX B300 servers, against 1,200 W and 1,000 W for the B200 in GB200 and HGX B200 systems. ^{[3][5][11][12]}

The table compares the two GPUs on the same basis, using NVIDIA's per-GPU figures for each chip in its eight-GPU HGX server and in its Grace-based NVL72 rack.

Sources: NVIDIA Blackwell and Blackwell Ultra datasheets (October 2025), NVIDIA technical blog, CUDA compute capability table. All compute figures are dense; where NVIDIA quotes only sparse rates (FP8, FP16/BF16, TF32, INT8), the values shown are half, as its footnotes specify. NVIDIA's HGX page rates an eight-GPU HGX B300 board at 108 petaFLOPS of dense FP4 (13.5 per GPU) and 3 petaOPS of sparse INT8, slightly different from the datasheet's rounded per-GPU figures. ^{[3][11][12][13][18]}

Two rows moved the other way. NVIDIA lists FP64 at 1.3 teraFLOPS per GPU in the GB300 NVL72, against 40 teraFLOPS for the B200 in the GB200 NVL72, and dense INT8 at 165 trillion operations per second against 5,000. At rack level the GB300 NVL72 is rated at 100 teraFLOPS of FP64, compared with 2,880 teraFLOPS for the GB200 NVL72. FP8, FP16/BF16 and TF32 Tensor Core rates, and FP32 at 80 teraFLOPS, are unchanged, so the B300's compute gain over the B200 is confined to FP4. ^{[4][8][11][12]}

What is the GB300 Grace Blackwell Ultra Superchip?

The GB300 superchip pairs NVIDIA's Arm-based Grace CPU with Blackwell Ultra GPUs on a single coherent module, the Blackwell Ultra equivalent of the earlier GB200. In the configuration used inside the GB300 NVL72, each superchip board combines one Grace CPU with two B300 GPUs, joined to the CPU by the 900 GB/s NVLink-C2C coherent link so that CPU and GPU share a unified memory space. NVIDIA rates the superchip at up to 30 petaFLOPS of dense NVFP4 compute (40 petaFLOPS with sparsity) and about 1 TB of unified memory combining HBM3e and the Grace CPU's LPDDR5X, with ConnectX-8 SuperNICs providing 800 Gb/s of networking. ^[3] This tight CPU-GPU coupling, together with the large HBM3e pool, is what NVIDIA leans on for serving very large mixture-of-experts and reasoning models. ^[3][5][9]

What is the GB300 NVL72?

The GB300 NVL72 is the flagship Blackwell Ultra system: a single liquid-cooled rack that links 72 B300 GPUs and 36 Grace CPUs through a fifth-generation NVLink switch fabric so the whole rack behaves as one large accelerator. NVIDIA states the rack delivers about 1.1 exaFLOPS (1,080 petaFLOPS) of dense NVFP4 compute and 360 petaFLOPS of dense FP8 for training (NVIDIA's headline figures, 1,440 and 720 petaFLOPS, are with sparsity), with around 20 TB of HBM3e GPU memory at up to 576 TB/s, about 37 TB of total fast memory including 17 TB of LPDDR5X on the Grace CPUs, and 130 TB/s of aggregate NVLink bandwidth. Networking uses NVIDIA ConnectX-8 SuperNICs providing 800 Gb/s per GPU. NVIDIA reports the GB300 NVL72 delivers "1.5x more AI performance than the NVIDIA GB200 NVL72." ^{[1][3][4][11]}

Against Hopper systems, NVIDIA claims 50x higher "AI factory output," combining 10x better tokens per second per user with 5x higher throughput per megawatt. NVIDIA also added power-smoothing features, including energy storage, to steady the rack's draw when its GPUs ramp up together. ^[3][4]

Sources: NVIDIA GB300 NVL72 and GB200 NVL72 product pages; NVIDIA Blackwell and Blackwell Ultra datasheets; NVIDIA DGX GB rack-scale user guide. NVIDIA quotes FP8 and FP16/BF16 rates with sparsity; the dense values are half of those, per NVIDIA's footnotes. With sparsity, NVIDIA rates both racks at 1,440 petaFLOPS of FP4. ^{[4][8][11][12][14]}

HGX B300 and DGX B300

For customers using standard eight-GPU servers rather than full racks, NVIDIA offers the HGX B300 baseboard, launched as the HGX B300 NVL16, which carries eight Blackwell Ultra GPUs in the familiar HGX form factor. The NVL16 label counted the sixteen GPU dies joined by NVLink rather than the eight GPU packages; NVIDIA's HGX page now calls the product simply HGX B300. The board also moved the eight ConnectX-8 SuperNICs onto the GPU baseboard, where their built-in PCIe switches replace chips used in earlier HGX designs, and server makers had to adapt their tray designs and cabling. ^[1][18][19] NVIDIA characterized the HGX B300 as delivering 11x faster large-language-model inference, 7x more compute, and 4x larger memory than the Hopper generation. ^[1] It rates an eight-GPU HGX B300 board at 108 petaFLOPS of dense FP4 (144 petaFLOPS with sparsity), 2.1 TB of GPU memory and 14.4 TB/s of total NVLink bandwidth, against 72 petaFLOPS of dense FP4 and 1.4 TB for the HGX B200; the board's FP64 rating falls from 296 teraFLOPS to 10. NVIDIA's HGX page lists both boards as shipping. ^[18]

The same silicon underpins the DGX B300 system, which packs eight Blackwell Ultra GPUs into a 10U chassis with two Intel Xeon 6776P host CPUs, eight 800 Gb/s ConnectX-8 SuperNICs and two BlueField-3 DPUs. NVIDIA rates it at 108 petaFLOPS of dense FP4 inference (144 with sparsity) and 36 petaFLOPS of dense FP8 for training (72 with sparsity), and lists power consumption of about 14.5 kW. ^[9][21][22] NVIDIA's sources disagree on its memory: the product page and the February 2026 datasheet list 2.1 TB of GPU memory, while the DGX B300 user guide and the March 2025 launch announcement give 2.3 TB (the user guide's "8 x 288 GB"). The lower figure matches the 270 GB per GPU that NVIDIA lists for HGX B300 systems. ^{[9][11][21][22][23]} The first DGX SuperPOD built from DGX B300 systems was Eli Lilly's, announced on October 28, 2025 with 1,016 Blackwell Ultra GPUs and inaugurated in Indianapolis in February 2026 after four months of assembly. ^[28][29]

Eight GB300 NVL72 racks combine into a Blackwell Ultra DGX SuperPOD of 576 GPUs and 288 Grace CPUs. Tom's Hardware's launch report gave it 300 TB of HBM3e and 11.5 exaFLOPS of FP4, but by NVIDIA's per-rack figures 300 TB is the combined fast memory of eight racks (about 37 TB each, including the Grace CPUs' LPDDR5X), not their HBM3e (about 160 TB), and 11.5 exaFLOPS is the sparse FP4 rating of 576 GPUs at 20 petaFLOPS each; dense FP4 is about 8.6 exaFLOPS. NVIDIA's own announcement gave each DGX GB300 rack 38 TB of fast memory. ^[4][6][23]

DGX Station

The DGX Station puts a single Blackwell Ultra GPU on a desk. NVIDIA announced it on March 18, 2025 alongside the smaller DGX Spark, which uses the GB10 Grace Blackwell chip rather than Blackwell Ultra. The DGX Station's GB300 Grace Blackwell Ultra Desktop Superchip pairs the GPU with a 72-core Grace CPU over NVLink-C2C and adds an 800 Gb/s ConnectX-8 SuperNIC; NVIDIA announced it with 784 GB of coherent memory and said partners would offer it later in 2025. ^[24][25] Orders opened through partners only at GTC in March 2026, and the shipping specification is smaller than first announced: 252 GB of HBM3e at 7.1 TB/s plus 496 GB of LPDDR5X, 748 GB in total, where NVIDIA's 2025 specifications had called for a fully enabled B300 with 288 GB at 8 TB/s. ServeTheHome noted that the revised figures are consistent with seven of the GPU's eight HBM3e stacks enabled. NVIDIA rates the system at 20 petaFLOPS of sparse FP4 (15 dense) and 1,600 W. ^[25][26] Partners set the prices: ServeTheHome heard of configurations in the $100,000 to $125,000 range in March 2026, and in August 2026 Exxact listed one starting at $94,930. ^[26][27]

When did Blackwell Ultra ship, and who adopted it?

NVIDIA guided that Blackwell Ultra systems would ship from partners in the second half of 2025, and the rollout broadly tracked that schedule. On August 27, 2025, NVIDIA said it "began production shipments of GB300" in its second fiscal quarter and that its rack production, converted to the GB300 in late July and early August, was running at about 1,000 racks per week. ^[54][55] Dell had delivered the first production GB300 NVL72 to CoreWeave in July 2025, but DigiTimes reported at the time that large-scale GB300 shipments would begin only in September 2025, the month Supermicro announced volume shipments of HGX B300 and GB300 NVL72 systems. ^[20][30][44]

The company named a wide hardware ecosystem at launch, including server makers Cisco, Dell, HPE, Lenovo, and Supermicro alongside Aivres, ASRock Rack, ASUS, Eviden, Foxconn, GIGABYTE, Inventec, Pegatron, QCT, Wistron, and Wiwynn, plus cloud providers AWS, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure, CoreWeave, Crusoe, Lambda, Nebius, Nscale, Yotta, and YTL. ^[1] In July 2026 Wistron opened a plant in Fort Worth, Texas, with one manufacturing cell building GB300 Grace Blackwell Ultra superchips and a second slated for Vera Rubin. ^[16]

By the second half of 2025 the platform was appearing in benchmarks and live deployments. In the September 2025 round of MLPerf Inference, NVIDIA reported that the GB300 NVL72 delivered 45 percent more DeepSeek-R1 inference throughput per GPU than the GB200 NVL72 in the offline scenario. ^[7][46] On October 9, 2025, Microsoft Azure brought up what it and NVIDIA called the first large-scale GB300 NVL72 cluster: 4,608 Blackwell Ultra GPUs in 64 racks, linked by Quantum-X800 InfiniBand and built for OpenAI's inference workloads. ^[10][32]

The rows mix deliveries, generally available cloud instances and multiyear capacity contracts whose deliveries run into 2027. The xAI figure comes from Musk's posts on X, as reported by Tom's Hardware, and has not been independently verified. ^[36][43]

MLPerf results

Blackwell Ultra systems appeared in every MLPerf data-center round from September 2025 to September 2026, first as NVIDIA's own GB300 NVL72 rack and later in submissions from many NVIDIA partners. ^[45][49][51]

In the September 2025 debut, NVIDIA said its platform set per-GPU records on all four benchmarks added that round (DeepSeek-R1, Llama 3.1 405B, Llama 3.1 8B and Whisper). It put the GB300 NVL72 at about 5x a Hopper H200 system on DeepSeek-R1, though the H200 figures were not verified by MLCommons, and the Llama 3.1 405B interactive record that round belonged to a GB200 NVL72. ^[7][46]

The November 2025 training round brought the first FP4 training submissions in MLPerf, run on both GB200 and GB300 systems. Its headline 10-minute Llama 3.1 405B record came from 5,120 Blackwell GPUs that MLCommons' system files identify as GB200 NVL72 racks, not Blackwell Ultra. ^[47][48] In April 2026, NVIDIA said Blackwell Ultra delivered up to 2.7x more DeepSeek-R1 throughput in the server scenario than at its debut six months earlier, crediting TensorRT-LLM updates; the faster result was a submission by partner Nebius, compared with NVIDIA's own 72-GPU rack entry from 2025. ^[9][46][49]

The September 2026 round was the first with a successor in the field: NVIDIA submitted Vera Rubin NVL72 in the preview category and claimed up to 3.7x the GB300 NVL72's throughput on Qwen3-VL and up to 2.5x on DeepSeek-R1. The same round included a 512-accelerator system from Crusoe, built on AMD Instinct MI355X GPUs, that broke the 288-accelerator record NVIDIA's GB300 racks had set in April 2026. ^{[49][52][53][78]}

Sales and revenue

NVIDIA does not report revenue by product, but its earnings materials tracked the Blackwell Ultra ramp:

In the same August 2026 commentary NVIDIA said its Blackwell architecture "remains the vast majority of our revenue." On the call that day, it said it "commenced production shipments of Vera Rubin earlier this month" and expected Vera Rubin to account for about 20% of Data Center revenue in its fiscal third quarter. ^[57][58]

How much has Blackwell Ultra capacity cost?

NVIDIA's product pages give no prices for B300 systems, which are sold through partners. ^[9] Cloud list prices give a dated record:

AWS describes its Capacity Block rates as reservation prices "updated regularly based on trends in supply and demand"; its rate for a p6-b300.48xlarge in Oregon was 38 percent higher in October 2026 than in December 2025. Between July and October 2026, Verda's B300 rate rose from $8.25 to $9.91 per hour and RunPod's from $7.39 to $8.99. For comparison, in October 2026 the same providers listed B200 capacity at $113.93 per instance-hour (AWS on demand, p6-b200.48xlarge), $14.00 per GPU-hour (Oracle), $8.50 (Nebius), $7.99 (RunPod) and $7.87 (Verda). ^{[60][61][62][63][64][65][66][67][68]}

Place in NVIDIA's roadmap

Blackwell Ultra was the 2025 step in the one-year cadence NVIDIA announced in 2024. At Computex on June 2, 2024, Huang presented a roadmap of new chips arriving on "a one-year rhythm" and revealed Rubin as Blackwell's successor, and coverage of the keynote reported a Blackwell Ultra GPU due in 2025. ^[69][70]

At GTC 2025, NVIDIA scheduled Vera Rubin systems for the second half of 2026. It declared Rubin in full production at CES on January 5, 2026, with partner products due in the second half of the year, and at GTC on March 16, 2026 it described seven Vera Rubin platform chips as in full production. ^[72][74][75] By July 21, 2026, Vera Rubin NVL72 racks were running at CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure and Nebius, and NVIDIA said on August 26, 2026 that production shipments had begun that month. ^[58][76] NVIDIA's Vera Rubin NVL72 page lists 288 GB of HBM4 at 19.2 TB/s per Rubin GPU, the same capacity as the B300's maximum at more than twice the bandwidth. ^[77]

Why does Blackwell Ultra matter?

Blackwell Ultra marked NVIDIA's pivot to an annual product rhythm in which a mid-cycle "Ultra" refresh extends a microarchitecture before the next full generation arrives. Its specific bias toward memory capacity and attention throughput, rather than a wholesale compute redesign, reflected the rise of reasoning models and inference-time scaling, where serving cost is dominated by long output sequences and large key-value caches rather than by raw matrix-multiply peaks. By raising per-GPU HBM3e to 288 GB and dense FP4 to 15 petaFLOPS while keeping the rack-scale NVLink fabric of base Blackwell, NVIDIA aimed to lower the cost per token of these workloads and to bridge the gap to the Vera Rubin platform, the next-generation architecture built around HBM4 memory, which began shipping in August 2026. ^{[2][3][4][58][69]} The cost of that focus was double precision: about 1.3 teraFLOPS of FP64 per B300 in GB300 systems, against 40 for the B200 in GB200 systems. ^[11][12]

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12NVIDIA Newsroom, "NVIDIA Blackwell Ultra AI Factory Platform Paves Way for Age of AI Reasoning," March 18, 2025. nvidianews.nvidia.com/...y-for-age-of-ai-reasoning
2. ^1 ^2 ^3 ^4 ^5NVIDIA Investor Relations, "NVIDIA Blackwell Ultra AI Factory Platform Paves Way for Age of AI Reasoning," March 18, 2025. investor.nvidia.com/...default
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22NVIDIA Technical Blog, "Inside NVIDIA Blackwell Ultra: The Chip Powering the AI Factory Era," August 22, 2025 (updated September 24, 2025). developer.nvidia.com/...owering-the-ai-factory-era
4. ^1 ^2 ^3 ^4 ^5 ^6NVIDIA, "GB300 NVL72" product page, accessed October 16, 2026. nvidia.com/...gb300-nvl72
5. ^1 ^2 ^3 ^4 ^5TweakTown, "NVIDIA details Blackwell Ultra GB300: dual-die design, 208B transistors, up to 288GB HBM3E," August 26, 2025. tweaktown.com/...08b-transistors-up-to-288gb-hbm3e
6. ^1 ^2 ^3 ^4Tom's Hardware, "Nvidia announces Blackwell Ultra B300, 1.5X faster than B200 with 288GB HBM3e and 15 PFLOPS dense FP4," March 19, 2025. tomshardware.com/...-hbm3e-and-15-pflops-dense-fp4
7. ^1 ^2Tom's Hardware, "Nvidia claims software and hardware upgrades allow Blackwell Ultra GB300 to dominate MLPerf benchmarks," September 10, 2025. tomshardware.com/...throughput-increase-over-gb200
8. ^1 ^2NVIDIA, "GB200 NVL72" product page, accessed October 16, 2026. nvidia.com/...gb200-nvl72
9. ^1 ^2 ^3 ^4 ^5NVIDIA, "DGX B300" product page, accessed October 16, 2026. nvidia.com/...dgx-b300
10. ^Tom's Hardware, "Microsoft deploys world's first 'supercomputer-scale' GB300 NVL72 Azure cluster," October 10, 2025. tomshardware.com/...le-of-1-44-pflops-of-inference
11. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15NVIDIA, "NVIDIA Blackwell Ultra Datasheet," October 2025. resources.nvidia.com/...blackwell-ultra-datasheet
12. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9NVIDIA, "NVIDIA Blackwell Datasheet," October 2025. dam-cdn.nvd.orangelogic.com/...18b2852i557jem5.pdf
13. ^1 ^2 ^3NVIDIA Developer, "CUDA GPU Compute Capability," accessed October 16, 2026. developer.nvidia.com/cuda-gpus
14. ^1 ^2NVIDIA, "Hardware," NVIDIA DGX GB Rack Scale Systems User Guide, accessed October 16, 2026. docs.nvidia.com/...hardware
15. ^NVIDIA Newsroom, "NVIDIA Supercharges Hopper, the World's Leading AI Computing Platform," November 13, 2023. nvidianews.nvidia.com/...ing-ai-computing-platform
16. ^1 ^2NVIDIA Blog, "Built in Fort Worth: Wistron Opens Advanced Manufacturing Plant to Produce NVIDIA AI Systems," July 21, 2026. blogs.nvidia.com/...wistron-manufacturing-texas
17. ^The Register, "To heck with export controls! Nvidia reportedly plotting cut-down B300 for Chinese market," August 19, 2025. theregister.com/...nvidia_cut_down_b300_for_china
18. ^1 ^2 ^3 ^4NVIDIA, "NVIDIA HGX Platform" product page, accessed October 16, 2026. nvidia.com/...hgx
19. ^ServeTheHome, "The NVIDIA HGX B300 NVL16 is Massively Different," April 15, 2025. servethehome.com/...0-nvl16-is-massively-different
20. ^1 ^2 ^3 ^4Supermicro, "Supermicro Begins Volume Shipments of NVIDIA Blackwell Ultra Systems and Rack Plug-and-Play Data Center-Scale Solutions," September 11, 2025. ir.supermicro.com/...default
21. ^1 ^2NVIDIA, "NVIDIA DGX B300 Datasheet," February 2026. resources.nvidia.com/...dgx-b300-datasheet
22. ^1 ^2NVIDIA, "Introduction to NVIDIA DGX B300 System," NVIDIA DGX B300 User Guide, accessed October 16, 2026. docs.nvidia.com/...introduction-to-dgxb300
23. ^1 ^2NVIDIA Newsroom, "NVIDIA Blackwell Ultra DGX SuperPOD Delivers Out-of-the-Box AI Supercomputer for Enterprises to Build AI Factories," March 18, 2025. nvidianews.nvidia.com/...upercomputer-ai-factories
24. ^NVIDIA Newsroom, "NVIDIA Announces DGX Spark and DGX Station Personal AI Computers," March 18, 2025. nvidianews.nvidia.com/...ion-personal-ai-computers
25. ^1 ^2NVIDIA, "DGX Station" product page, accessed October 16, 2026. nvidia.com/...dgx-station
26. ^1 ^2ServeTheHome, "NVIDIA DGX Station Systems Available At Last GB300 and GB200 Workstations For Your Desktop," March 20, 2026. servethehome.com
27. ^Tom's Hardware, "Nvidia's GB300-powered DGX Station desktop tower listed for nearly $100,000 online," August 23, 2026. tomshardware.com/...mere-mortals-with-lots-of-cash
28. ^1 ^2NVIDIA Blog, "Lilly Deploys World's Largest, Most Powerful AI Factory for Drug Discovery Using NVIDIA Blackwell-Based DGX SuperPOD," October 28, 2025. blogs.nvidia.com/...-nvidia-blackwell-dgx-superpod
29. ^1 ^2NVIDIA Blog, "Now Live: Lilly AI Factory for Pharmaceutical Discovery and Development," February 26, 2026. blogs.nvidia.com/...lilly-ai-factory-live
30. ^1 ^2 ^3 ^4Dell Technologies, "Dell Delivers Market's First NVIDIA GB300 NVL72 to CoreWeave," July 3, 2025. dell.com/...-first-nvidia-gb300-nvl72-to-coreweave
31. ^CoreWeave Docs, "August 19, 2025 - CoreWeave Launches GB300 NVL72-Powered Cloud Instances," August 19, 2025. docs.coreweave.com/...gb300-nvl72
32. ^1 ^2NVIDIA Blog, "Microsoft Azure Unveils World's First NVIDIA GB300 NVL72 Supercomputing Cluster for OpenAI," October 9, 2025. blogs.nvidia.com/...-supercomputing-cluster-openai
33. ^1 ^2Amazon Web Services, "Amazon EC2 P6-B300 instances with NVIDIA Blackwell Ultra GPUs are now available," November 18, 2025. aws.amazon.com/...a-blackwell-ultra-gpus-available
34. ^Amazon Web Services, "Amazon EC2 P6e-GB300 UltraServers accelerated by NVIDIA GB300 NVL72 are now generally available," December 2, 2025. aws.amazon.com/...-gb300-nvl72-generally-available
35. ^Google Cloud Blog, "Expanding our NVIDIA partnership: Now shipping A4X Max, Vertex AI Training, and more," October 29, 2025. cloud.google.com/...ax-vertex-ai-training-and-more
36. ^1 ^2Nscale, "Nscale Contracts Approximately 200,000 NVIDIA GB300 GPUs with Microsoft to Deliver NVIDIA AI Infrastructure Across Europe and the U.S.," October 15, 2025. nscale.com/...nscale-microsoft-2025
37. ^IREN, "IREN Secures $9.7bn AI Cloud Contract with Microsoft," Exhibit 99.1 to Form 8-K, November 3, 2025. sec.gov/...ef20058139_ex99-1
38. ^Lambda, "Lambda Announces Multibillion-Dollar Agreement With Microsoft to Deploy AI Infrastructure Powered by Tens of Thousands of NVIDIA GPUs," November 3, 2025. lambda.ai/...d-by-tens-of-thousands-of-nvidia-gpus
39. ^Data Center Dynamics, "Nebius launches Nvidia GPU cluster in London, UK," November 6, 2025. datacenterdynamics.com/...gpu-cluster-in-london-uk
40. ^NVIDIA Newsroom, "HUMAIN and NVIDIA Announce Strategic Partnership to Build AI Factories of the Future in Saudi Arabia," May 13, 2025. nvidianews.nvidia.com/...he-future-in-saudi-arabia
41. ^NVIDIA Newsroom, "Meta Builds AI Infrastructure With NVIDIA," February 17, 2026. nvidianews.nvidia.com/...nfrastructure-with-nvidia
42. ^CoreWeave, "CoreWeave Advances AI-Native Cloud Platform for the Next Phase of Production-Scale AI," March 16, 2026. coreweave.com/...next-phase-of-production-scale-ai
43. ^1 ^2Tom's Hardware, "Elon Musk's SpaceXAI to add another 660,000 AI GPUs this year, nearing a total of 1.44 million in operation," September 25, 2026. tomshardware.com/...-to-bring-systems-fully-online
44. ^Tom's Hardware, "Large-scale shipments of Nvidia GB300 servers tipped to start in September," July 18, 2025. tomshardware.com/...idespread-coolant-leak-reports
45. ^1 ^2MLCommons, "MLCommons Releases New MLPerf Inference v5.1 Benchmark Results," September 9, 2025. mlcommons.org/...mlperf-inference-v5-1-results
46. ^1 ^2 ^3 ^4NVIDIA Technical Blog, "NVIDIA Blackwell Ultra Sets New Inference Records in MLPerf Debut," September 9, 2025. developer.nvidia.com/...ce-records-in-mlperf-debut
47. ^1 ^2NVIDIA Technical Blog, "NVIDIA Blackwell Architecture Sweeps MLPerf Training v5.1 Benchmarks," November 12, 2025. developer.nvidia.com/...f-training-v5-1-benchmarks
48. ^1 ^2MLCommons, "training_results_v5.1" results repository (NVIDIA system descriptions), November 12, 2025. github.com/...training_results_v5.1
49. ^1 ^2 ^3 ^4NVIDIA Technical Blog, "NVIDIA Platform Delivers Lowest Token Cost Enabled by Extreme Co-Design," April 1, 2026. developer.nvidia.com/...abled-by-extreme-co-design
50. ^NVIDIA Blog, "Fastest, Largest, Strongest: NVIDIA Blackwell Sweeps MLPerf Training 6.0," June 16, 2026. blogs.nvidia.com/...blackwell-mlperf-training-6-0
51. ^MLCommons, "MLCommons Sets Participation Record with New MLPerf Inference v6.1 Benchmark Results," September 16, 2026. mlcommons.org/...mlperf-inference-v6-1-results
52. ^MLCommons, "Where the Industry Is Investing: A Look at MLPerf Inference v6.1," September 17, 2026. mlcommons.org/...chairs-mlperf-inference-v6-1
53. ^1 ^2NVIDIA Blog, "NVIDIA Vera Rubin NVL72 Delivers Leading Performance in MLPerf Inference v6.1 Debut," September 16, 2026. blogs.nvidia.com/...a-rubin-nvl72-mlperf-inference
54. ^1 ^2NVIDIA, "Q2 Fiscal 2026 Earnings Call" (corrected transcript), August 27, 2025. s201.q4cdn.com/...ll-27-August-2025-5_00-PM-ET.pdf
55. ^1 ^2 ^3NVIDIA, "Form 10-Q for the quarterly period ended July 27, 2025," August 27, 2025. s201.q4cdn.com/...-c226-4d05-8f74-aaca89a21b33.pdf
56. ^NVIDIA, "Q3 Fiscal 2026 Earnings Call" (corrected transcript), November 19, 2025. s201.q4cdn.com/...-19-November-2025-5_00-PM-ET.pdf
57. ^1 ^2NVIDIA, "CFO Commentary on Second Quarter Fiscal 2027 Results," August 26, 2026. s201.q4cdn.com/...Q2FY27-CFO-Commentary.pdf
58. ^1 ^2 ^3 ^4 ^5NVIDIA, "Q2 Fiscal 2027 Earnings Call" (transcript), August 26, 2026. s201.q4cdn.com/...ll-26-August-2026-5_00-PM-ET.pdf
59. ^NVIDIA, "CFO Commentary on Fourth Quarter and Fiscal 2025 Results," February 26, 2025. s201.q4cdn.com/...Q4FY25-CFO-Commentary.pdf
60. ^1 ^2Amazon Web Services, "Amazon EC2 Capacity Blocks for ML Pricing," archived December 24, 2025. web.archive.org/...pricing
61. ^1 ^2Amazon Web Services, "Amazon EC2 Capacity Blocks for ML Pricing," accessed October 16, 2026. aws.amazon.com/...pricing
62. ^1 ^2Amazon Web Services, EC2 On-Demand price list for US West (Oregon), Linux, published October 12, 2026. b0.p.awsstatic.com/...Linux
63. ^1 ^2 ^3Oracle, "Cloud Price List: Compute," accessed October 16, 2026. oracle.com/...pricing
64. ^1 ^2Nebius, "NVIDIA GPU Pricing," accessed October 16, 2026. nebius.com/prices
65. ^1 ^2RunPod, "Pricing," archived July 2, 2026. web.archive.org/...pricing
66. ^1 ^2RunPod, "Pricing," accessed October 16, 2026. runpod.io/pricing
67. ^1 ^2Verda, "Pricing," archived July 1, 2026. web.archive.org/...pricing
68. ^1 ^2 ^3Verda, "Pricing," accessed October 16, 2026. verda.com/pricing
69. ^1 ^2 ^3NVIDIA Blog, "'Accelerate Everything,' NVIDIA CEO Says Ahead of COMPUTEX," June 2, 2024. blogs.nvidia.com/...computex-2024-jensen-huang
70. ^1 ^2The Register, "AMD, Intel, Nvidia detail their datacenter roadmaps," June 5, 2024. theregister.com/...chipmakers_computex_roadmaps
71. ^NVIDIA Newsroom, "NVIDIA Blackwell Platform Arrives to Power a New Era of Computing," March 18, 2024. nvidianews.nvidia.com/...er-a-new-era-of-computing
72. ^1 ^2CNBC, "Nvidia announces Blackwell Ultra and Vera Rubin AI chips," March 18, 2025. cnbc.com/...ackwell-ultra-and-vera-rubin-ai-chips-
73. ^Tom's Hardware, "Nvidia announces Rubin GPUs in 2026, Rubin Ultra in 2027, Feynman also added to roadmap," March 19, 2025. tomshardware.com/...bin-ultra-in-2027-feynam-after
74. ^1 ^2NVIDIA Newsroom, "NVIDIA Kicks Off the Next Generation of AI With Rubin: Six New Chips, One Incredible AI Supercomputer," January 5, 2026. nvidianews.nvidia.com/...platform-ai-supercomputer
75. ^NVIDIA Investor Relations, "NVIDIA Vera Rubin Opens Agentic AI Frontier," March 16, 2026. investor.nvidia.com/...default
76. ^NVIDIA Blog, "NVIDIA Vera Rubin Driving Performance Per Watt, Lower Token Costs for Partners Worldwide," July 21, 2026. blogs.nvidia.com/...vera-rubin
77. ^1 ^2NVIDIA, "Vera Rubin NVL72" product page, accessed October 16, 2026. nvidia.com/...vera-rubin-nvl72
78. ^StorageReview, "MLPerf Inference v6.1: 5.7x Per-Accelerator Gains, a 512-GPU Run, and Vera Rubin's First Peer-Reviewed Numbers," September 16, 2026. storagereview.com/...s-first-peer-reviewed-numbers