NVIDIA B200

The NVIDIA B200 is a data center GPU based on the NVIDIA Blackwell microarchitecture, announced by Jensen Huang at NVIDIA's GTC conference in San Jose on March 18, 2024.^[1] It is the main accelerator of the Blackwell generation and the successor to the NVIDIA H100 and NVIDIA H200. Each B200 joins two reticle-limited dies, made on a custom TSMC 4NP process and linked by a 10 TB/s chip-to-chip interface, into one GPU with 208 billion transistors.^[1][2] The package carries 192 GB of HBM3E in eight stacks; NVIDIA's shipping specifications expose 180 GB at 7.7 TB/s per GPU on the eight-GPU HGX B200 board and 186 GB at 8 TB/s when the chip sits in a GB200 Grace Blackwell Superchip.^[2][3][4] In the HGX B200 configuration, rated at up to 1,000 W per GPU, the B200's dense tensor rates are 9,000 TFLOPS in FP4, 4,500 TFLOPS in FP8 and 2,250 TFLOPS in BF16; NVIDIA's headline figures double these by assuming 2:4 structured sparsity.^[5][2][6]

NVIDIA sells the B200 on the HGX B200 baseboard for x86 servers, in its own DGX B200 system, and inside the GB200 superchip, which pairs one Grace CPU with two B200 GPUs.^[1] The largest configuration is the GB200 NVL72, a liquid-cooled rack that links 72 B200 GPUs and 36 Grace CPUs into one NVLink domain rated at 720 PFLOPS of dense FP4 (1,440 PFLOPS sparse).^[7] NVIDIA pitched the platform as able to run real-time generative AI on trillion-parameter language models "at up to 25x less cost and energy consumption than its predecessor."^[1] Blackwell was the first NVIDIA architecture with FP4 Tensor Core math, and the B200 introduced a second-generation Transformer Engine, fifth-generation NVLink and a dedicated decompression engine.^[1][2][8]

A design flaw that hurt manufacturing yields delayed the volume ramp into NVIDIA's fiscal quarter ending January 2025, in which the company booked $11.0 billion of Blackwell revenue and called it the fastest product ramp in its history.^[9][10] Nvidia followed the B200 with the Blackwell Ultra B300, announced in March 2025 and first delivered to a customer in July 2025, and with the Vera Rubin platform, which NVIDIA began shipping in production in August 2026 and whose NVL72 systems CoreWeave made available to cloud customers in September 2026.^{[11][12][114][13]} As of October 2026, NVIDIA's HGX product page still lists the HGX B200 as shipping alongside the HGX B300, and the GetDeploying price tracker listed 37 cloud providers offering the B200.^[5][14]

What is the NVIDIA B200?

The B200 is the first NVIDIA GPU built from two reticle-limited dies that software sees as a single device. In NVIDIA's words, Blackwell GPUs are "packed with 208 billion transistors" and "manufactured using a custom-built 4NP TSMC process with two-reticle limit GPU dies connected by 10 TB/second chip-to-chip link into a single, unified GPU."^[1] Chips and Cheese, which tested a B200 cloud instance in 2025, called it "Nvidia's first chiplet GPU" and found that "both dies appear to software as a single GPU."^[15] It is the building block of the Blackwell systems that NVIDIA markets as "AI factory" infrastructure, from the eight-GPU HGX B200 baseboard and the DGX B200 server up to the rack-scale GB200 NVL72, and it was designed to train and serve trillion-parameter large language models.^[1][16][11]

Compared with the H100, the B200 has 2.6 times as many transistors (208 billion versus 80 billion), 2.25 times the usable memory in its HGX form (180 GB versus 80 GB), 2.3 times the memory bandwidth (7.7 TB/s versus 3.35 TB/s) and about 2.3 times the dense FP8 throughput (4,500 TFLOPS versus 1,979 TFLOPS). It also adds FP6 and FP4 Tensor Core formats that Hopper does not have.^[17][2][8] At rack scale, NVIDIA says the GB200 NVL72 provides "up to a 30x performance increase compared to the same number of NVIDIA H100 Tensor Core GPUs" for LLM inference while reducing cost and energy consumption by up to 25 times.^[1]

Background and announcement

NVIDIA announced the Blackwell architecture and the B200 on March 18, 2024, in Huang's keynote at the SAP Center in San Jose, the opening of the company's first in-person GTC in five years.^[18][20] Huang framed the launch around the growth of AI models: "Hopper is fantastic, but we need bigger GPUs."^[19] NVIDIA said Blackwell-based processors offered 20 petaflops of AI performance against 4 petaflops for the H100, and its keynote recap described Blackwell as delivering 2.5 times its predecessor's FP8 training performance per chip and 5 times its inference performance with FP4.^[19][20] Those launch numbers are sparse rates: the 20 petaflops is the sparse FP4 rate of a B200 running in the GB200 configuration (10 petaflops dense), and the H100's 4 petaflops is its sparse FP8 rate (about 2 petaflops dense). The HGX B200, limited to 1,000 W, is rated at 18 PFLOPS of sparse FP4 (9 PFLOPS dense).^[21][2][17] In the launch press release Huang said: "Generative AI is the defining technology of our time. Blackwell is the engine to power this new industrial revolution."^[1]

The architecture is named after David Harold Blackwell, a mathematician who worked in game theory and statistics and who in 1965 became the first Black scholar inducted into the National Academy of Sciences; the Rao-Blackwell theorem of statistics also carries his name.^[1][22] NVIDIA has long named its data center GPU architectures after scientists, and Blackwell follows Hopper, named after the computer scientist Grace Hopper.

At the same event NVIDIA introduced the GB200 Grace Blackwell Superchip, the GB200 NVL72 rack, the HGX B200 server board and a lower-power HGX B100 designed to fit existing HGX H100 infrastructure.^[1][21] It said Blackwell products would be available from partners "later this year" and named Amazon Web Services, Dell Technologies, Google, Meta, Microsoft, OpenAI, Oracle, Tesla and xAI among the organizations expected to adopt Blackwell.^[1] Amazon Web Services, Google Cloud, Microsoft Azure and Oracle Cloud Infrastructure were to be among the first cloud providers with Blackwell instances, along with NVIDIA Cloud Partner companies Applied Digital, CoreWeave, Crusoe, IBM Cloud, Lambda and Nebius.^[1] Cisco, Dell, Hewlett Packard Enterprise, Lenovo and Supermicro were expected to ship Blackwell servers, as were Aivres, ASRock Rack, ASUS, Eviden, Foxconn, GIGABYTE, Inventec, Pegatron, QCT, Wistron, Wiwynn and ZT Systems.^[1]

The launch specifications differed from what eventually shipped. NVIDIA's March 2024 figures, as tabulated by Tom's Hardware, gave the B200 192 GB of HBM3E at 8 TB/s, 40 TFLOPS of FP64 and a 1,000 W rating, and gave the eight-GPU HGX B200 1,536 GB of memory.^[21] By the October 2025 datasheet, the HGX B200 was specified at 180 GB and 7.7 TB/s per GPU, 37 TFLOPS of FP64 and 1.4 TB of total memory, while the GB200 version was listed at 186 GB and 8 TB/s.^[2] The 192 GB figure remains the physical memory on the package: eight 8-high HBM3E stacks of 24 GB each.^[23][3][4]

The day after the keynote, Huang told CNBC that Blackwell would cost $30,000 to $40,000 per unit and estimated that NVIDIA had spent about $10 billion on research and development for it (see Pricing and availability).^[24]

Architecture overview

Blackwell keeps Hopper's programming model and adds features aimed at large transformer models. NVIDIA's launch materials grouped them into six technologies: the two-die chip itself, a second-generation Transformer Engine with 4-bit floating point, fifth-generation NVLink, a reliability, availability and serviceability (RAS) engine, confidential computing under the name Secure AI, and a decompression engine.^[1][2]

Dual-die design

The B200 replaces Hopper's monolithic die with two compute dies, each at the largest size a lithography reticle allows, placed side by side in one package.^[1] Each die holds 104 billion transistors.^[23] NVIDIA calls the die-to-die link the NVIDIA High-Bandwidth Interface (NV-HBI); it carries 10 TB/s, enough that the pair works as "one fully coherent chip" with a single address space.^[25][26][2] Chips and Cheese measured the cost of the split in December 2025. The B200's L2 cache, 126 MB in total, is divided into two partitions that "almost certainly correspond to its two dies," and the cross-partition latency penalty was "higher than its predecessors, but only slightly so."^[15] For comparison, the H100 has 50 MB of L2 cache and the A100 40 MB.^[15]

The two-die layout gives NVIDIA 208 billion transistors in one product, against 80 billion in the H100 and H200 and about 54 billion in the A100.^[26][27] Each die is already "the largest die possible within the limits of reticle size," in NVIDIA's words, so NVIDIA reached 208 billion transistors by packaging two dies together and presenting them to software as one GPU.^[25][15]

Silicon and packaging

The B200 was the first commercial product built with TSMC's CoWoS-L packaging, according to TechInsights' teardown of an HGX B200 board.^[3] CoWoS-L joins the chiplets through a redistribution-layer (RDL) interposer with embedded local silicon interconnect (LSI) bridges, which carry the roughly 10 TB/s of die-to-die traffic. The bridges must be placed precisely, and this packaging was at the center of the 2024 yield problems described below.^[9] TechInsights identified the GPU silicon as GB100 and found SK Hynix HBM3E in the units it examined.^[3] Micron said in March 2025 that its 8-high, 24 GB HBM3E stacks were available for the HGX B200 and GB200 NVL72 platforms.^[28]

Each die physically contains 80 streaming multiprocessors (SMs), of which 74 are enabled, for 148 active SMs per GPU.^[15] Every SM has 128 FP32 CUDA cores, 128 INT32 cores, 64 FP64 cores and four fifth-generation Tensor Cores, which works out to 18,944 FP32 CUDA cores and 592 Tensor Cores across the 148 active SMs.^[29][15] Each SM also has a 256 KB L1 data cache, part of which can be set aside as shared memory, a 256 KB register file and 256 KB of tensor memory (TMEM).^[29] Launch-day reports put the B200 at 160 SMs and 20,480 CUDA cores; those numbers describe the full physical configuration of both dies rather than what is enabled in shipping parts.^[23][30]

Tensor Cores and precision formats

Blackwell's fifth-generation Tensor Cores handle FP64, FP32/TF32, FP16/BF16, INT8/FP8, FP6 and FP4 data.^[8] The H100 and H200 Tensor Cores stop at FP8 and INT8, so FP6 and FP4 are new with Blackwell.^[17] On the HGX B200, FP4 runs at twice the FP8 rate: 9,000 versus 4,500 dense TFLOPS per GPU.^[5]

The low-precision formats depend on microscaling, in which a small block of values shares a scale factor that recovers some of the dynamic range lost at 4 or 6 bits. NVIDIA describes Blackwell's Tensor Cores as introducing "new microscaling formats," including the community-defined formats of the Open Compute Project (OCP).^[2][25] The OCP's MXFP4 format attaches one power-of-two (E8M0) scale to each block of 32 values, while NVIDIA's own NVFP4 uses a finer block of 16 values with an FP8 (E4M3) scale plus a second FP32 scale for the whole tensor.^[8] NVFP4 stores about 4.5 bits per value, which cuts model memory by roughly 3.5 times relative to FP16 and 1.8 times relative to FP8, and NVIDIA reported 1% or less accuracy loss on key language-modeling tasks when DeepSeek-R1-0528 was quantized from FP8 to NVFP4.^[8] The 8-bit MXFP8 format is supported for training on Blackwell in NVIDIA's Transformer Engine library, and CUDA 12.8 extended cuBLAS with microscaled 4-bit and 8-bit floating-point matrix multiplication for compute capability 10.0.^[31][32]

Second-generation Transformer Engine

The Transformer Engine first appeared in the H100, where it managed FP8 training and inference. NVIDIA said Blackwell's second-generation version, backed by new micro-tensor scaling and dynamic range management algorithms, would "double the compute and model sizes with new 4-bit floating point AI inference capabilities."^[1] In software, the engine is implemented in TensorRT-LLM, NeMo and NVIDIA's open-source Transformer Engine library, which apply per-tensor or per-block scaling so that values fit the narrower formats.^[1][31]

FP4 was aimed first at inference: NVIDIA's first MLPerf submission with a B200, in Inference v4.1 in August 2024, already used the Blackwell FP4 Transformer Engine.^[109] NVIDIA later extended NVFP4 to training: its MLPerf Training v5.1 entries, published in November 2025, were the first FP4-precision submissions in MLPerf Training, and NVIDIA credited NVFP4 recipes and software changes for a gain of up to 1.4 times over its previous-round results on the same GB200 NVL72 hardware.^[33]

Decompression engine

Blackwell adds a fixed-function decompression engine that handles Snappy, LZ4 and Deflate (including Gzip) streams, which takes that work off both the CPU and the GPU's SMs.^[34][35] NVIDIA's nvCOMP documentation rates it at up to 600 GB/s, while NVIDIA's Blackwell technical brief gives up to 800 GB/s.^[34][25] NVIDIA's nvCOMP library uses it automatically, and the engine is present on the B200, B300, GB200 and GB300.^[34] In the GB200 NVL72, NVIDIA combines the engine with high-bandwidth memory and fast access to Grace CPU memory and projects 18 times faster database queries than CPUs, with five times better total cost of ownership; its technical blog names Intel Sapphire Rapids as the CPU baseline.^[2][36]

Confidential computing and RAS

Blackwell includes NVIDIA Confidential Computing, which uses hardware-based security to protect models and data from unauthorized access. NVIDIA describes Blackwell as the industry's first TEE-I/O capable GPU, with inline protection over NVLink when it is paired with TEE-I/O capable hosts.^[2] The launch press release aimed the feature at privacy-sensitive fields such as healthcare and finance.^[1] A dedicated RAS engine, with AI-based predictive management, monitors thousands of data points across hardware and software to catch likely faults early and reduce downtime.^[2][1]

B200 specifications

The tables below list NVIDIA's per-GPU figures. Dense rates are given first; NVIDIA's own tables quote sparse values, which assume 2:4 structured sparsity and are twice the dense rate.^[2][5] The same silicon runs in two main configurations: the HGX B200 board (also used in the DGX B200) and the GB200 superchip (see GB200 Grace Blackwell Superchip).

NVIDIA's widely quoted "20 petaflops" for the B200 is the sparse FP4 rate of the 1,200 W GB200 configuration. In NVIDIA's datasheet, Tensor Core figures are sparse values unless they are printed as a sparse and dense pair, and dense performance is half the sparse figure.^[2][19]

Sources disagree on the host interface. NVIDIA's August 2025 Blackwell Ultra comparison lists Blackwell with PCIe Gen 6 at 256 GB/s, but its October 2025 Blackwell datasheet, its Blackwell technical brief and Lenovo's HGX B200 product guide specify PCIe Gen5 at 128 GB/s for shipping B200 products; the technical brief reserves PCIe Gen6 for the HGX B300.^{[26][2][25][6]} The DGX B200 product page also lists 64 TB/s of total HBM3E bandwidth for eight GPUs, or 8 TB/s each, while the datasheet gives the HGX B200 62 TB/s in total.^[16][2]

Memory subsystem

The B200 has eight HBM3E stacks around its two dies on a combined 8,192-bit interface.^[23] TechInsights describes each stack as eight memory dies stacked in a true 3D configuration; like other high-bandwidth memory, the DRAM layers are joined by through-silicon vias.^[3] Each stack holds 24 GB, so the package carries 192 GB, of which NVIDIA exposes 180 GB per GPU on the HGX B200 and DGX B200 and 186 GB per GPU in GB200 systems.^[2][4] NVIDIA's technical brief describes the HGX B200 as having "up to 192 GB" of HBM3E at 7.7 TB/s, reflecting the physical capacity.^[25] At 8 TB/s across 8,192 bits, each pin runs at just under 8 Gb/s.

Memory is the most visible upgrade over Hopper. The H100 SXM carries 80 GB of HBM3 at 3.35 TB/s, and the H200 SXM 141 GB of HBM3E at 4.8 TB/s.^[17][39] The HGX B200's 180 GB and 7.7 TB/s are 2.25 and 2.3 times the H100's figures.^[2]

Capacity matters most for inference with long context windows or many concurrent users, because the KV cache grows with both. A 70-billion-parameter Llama model needs about 140 GB for its BF16 weights, which fits on one B200 but has to be split across two 80 GB H100s with tensor parallelism.^[40] For Llama 3.1 405B, the BF16 weights take about 810 GB: more than the 640 GB of an eight-GPU H100 server, but within the 1,440 GB of an eight-GPU B200 server, which can also hold the roughly 405 GB FP8 version with room left for KV cache.^[41][17] Fitting a model on fewer GPUs reduces the need for model parallelism and the communication overhead and latency that come with it.

Performance

How fast is the B200 at FP4?

FP4 (4-bit floating point) is the headline new precision in Blackwell. The HGX B200 delivers 9,000 TFLOPS of dense FP4 per GPU and 18,000 TFLOPS with 2:4 structured sparsity, a pattern in which two of every four weights are zero.^[5][30] The H100 and H200 have no FP4 Tensor Core support.^[17][39]

FP4 is used mostly for inference. At 4 bits, weights take half the memory and bandwidth of FP8, which allows larger batch sizes and lower per-token latency. NVIDIA's datasheet projects that an HGX B200 serves the 1.8-trillion-parameter GPT-MoE model at 58 output tokens per second per GPU, against 3.5 per GPU when eight HGX H100 servers run the same model, which NVIDIA presents as a 15-fold gain. The projection assumes a 50 ms token-to-token latency, a 5 s first-token latency, 32,768-token inputs and roughly 1,000-token outputs. It also projects 12 times lower energy use and total cost of ownership, comparing 100 racks of HGX H100 with 8 racks of HGX B200 for the same output.^[2]

FP8 performance

FP8 arrived with the H100 and remains the main training precision on Blackwell. The HGX B200 delivers 4,500 TFLOPS of dense FP8 per GPU (9,000 TFLOPS sparse), against 1,979 TFLOPS dense (3,958 TFLOPS sparse) on the H100 SXM, about 2.3 times as much per chip.^[5][17] In the 1,200 W GB200 configuration the dense rate is 5,000 TFLOPS, the basis of NVIDIA's "2.5x" FP8 training claim at launch.^[2][20] Higher FP8 throughput shortens training for a given model and batch configuration, although large runs are also limited by communication between GPUs.

BF16 and FP16

BF16 and FP16 remain the default formats for many training and fine-tuning jobs because they balance range and precision. The HGX B200 runs dense BF16 at 2,250 TFLOPS, against about 989 TFLOPS on the H100 SXM (1,979 TFLOPS with sparsity), again about 2.3 times higher.^[5][17]

How does the B200 compare to the H100, H200 and B100?

The table compares dense Tensor Core rates unless noted. B100 figures are NVIDIA's March 2024 launch specifications.

Sources: NVIDIA's H100 and H200 product pages, Tom's Hardware's table of the March 2024 launch specifications, and NVIDIA's HGX page and Blackwell datasheet.^{[17][39][21][5][2][26]}

The gap is larger for inference than for training. NVIDIA's HGX comparisons claim 3 times faster training and 15 times faster real-time inference than the HGX H100, and its GB200 NVL72 comparisons claim 4 times and 30 times.^[2] The inference advantage comes from FP4, which Hopper lacks, from larger memory, which allows bigger batches and longer contexts, and from higher bandwidth.

MLPerf results, 2024-2026

MLPerf, the benchmark suite run by MLCommons, has tracked the B200 since August 2024. The table summarizes selected results; details follow. Per-GPU comparisons are NVIDIA's or other submitters' own calculations, not an official MLPerf metric.

The B200's first MLPerf appearance came in Inference v4.1, published on August 28, 2024, when NVIDIA entered a single B200 in the preview category. Running Llama 2 70B with the FP4 Transformer Engine, it produced 10,756 tokens/s in the server scenario and 11,264 tokens/s offline, which NVIDIA put at 4 times and 3.7 times the per-GPU throughput of its eight-GPU H100 submissions.^[108][109]

In MLPerf Training v4.1, the B200's first training round, NVIDIA submitted eight-GPU Blackwell systems running at 1,000 W per GPU in the preview category. By NVIDIA's per-GPU normalization, they ran GPT-3 175B pretraining about 2 times as fast and Llama 2 70B LoRA fine-tuning 2.2 times as fast as H100 submissions, and the GPT-3 run needed only 64 Blackwell GPUs where Hopper needed 256 for its best per-GPU result.^[42]

MLPerf Inference v5.0 compared eight-GPU B200 and H200 systems in the data center closed division:^[43][102]

Source: MLCommons results for NVIDIA's DGX B200 and H200 entries (5.0-0056 and 5.0-0060); speedups are ratios of those figures.^[102]

The B200 system also posted 3.1 times the H200's throughput on the Llama 2 70B Interactive test. On Llama 3.1 405B, GB200 NVL72 delivered up to 3.4 times the per-GPU performance of an eight-GPU H200 system and up to 30 times its system-level throughput, a comparison that reflects nine times as many GPUs as well as a larger NVLink domain.^[43] In the server scenario the 72-GPU rack produced 8,850 tokens/s, against 291 tokens/s for the eight-GPU H200 system.^[102]

In MLPerf Training v5.0, a 2,496-GPU GB200 NVL72 cluster submitted jointly by CoreWeave, NVIDIA and IBM completed the new Llama 3.1 405B pretraining benchmark in 27.33 minutes, 2.1 times as fast as NVIDIA's larger 2,560-GPU H100 submission; a 2,048-GPU Blackwell run took 32.6 minutes, 2.15 times as fast as NVIDIA's 2,048-GPU H100 result.^[44][106] NVIDIA used NVFP4 across its DeepSeek-R1 and Llama submissions in Inference v5.1, where the B300 debuted.^[45] In the same round an eight-GPU DGX B200 served DeepSeek-R1 at 33,379 tokens/s offline and 18,592 tokens/s in the server scenario.^[107]

In November 2025 NVIDIA made the first FP4-precision submissions in MLPerf Training. It reported that GB200 NVL72 trained Llama 3.1 405B up to 3.2 times as fast as optimized Hopper FP8 submissions at the same GPU count, and up to 1.4 times as fast as its own previous-round results at similar scale; GB300 NVL72 was 1.9 times as fast as GB200 NVL72 had been in the previous round.^[33] NVIDIA also estimated that GB200 NVL72 gives almost twice the training performance per dollar of the H100, a figure it derived from public rental prices and that MLCommons did not verify.^[33] Nebius's HGX B200 entries in the same round showed how training time fell with GPU count:^[46][105]

MLPerf Inference v6.0, published on April 1, 2026, included benchmarks based on DeepSeek-R1, the 120B variant of OpenAI's gpt-oss (gpt-oss-120B) and the multimodal Qwen3-VL-235B model from Alibaba's Qwen family. Nebius's eight-GPU HGX B200 reached 51,693 tokens/s on DeepSeek-R1 in the server scenario and 87,444 tokens/s on gpt-oss-120B, results Nebius listed as first place; in the MLCommons tables they were the highest of any eight-GPU B200 system in those two tests.^[47][50] NVIDIA's 72-GPU GB200 NVL72 entry produced 879,542 tokens/s offline on gpt-oss-120B, about 12,200 per GPU, while the round's only eight-GPU H200 entry, a Red Hat system running the llm-d serving stack rather than TensorRT-LLM, produced 28,680 tokens/s, about 3,600 per GPU.^[50]

In MLPerf Training v6.0, published on June 16, 2026, a Microsoft Azure submission of 8,192 GB200 GPUs trained Llama 3.1 405B in 7.07 minutes, and 1,024 GB200 GPUs in an NVIDIA submission trained Llama 3.1 8B in 4.46 minutes; NVIDIA reported a 1.6-times gain for GB300 NVL72 over GB200 NVL72 on DeepSeek V3.^[48][104] In MLPerf Inference v6.1, published on September 16, 2026, Nebius's HGX B200 improved to 55,869 tokens/s on DeepSeek-R1 and 89,856 tokens/s on gpt-oss-120B in server mode.^[49] CoreWeave's eight-GPU B200 system posted the round's highest B200 figures in both tests, 56,121 and 90,205 tokens/s.^[103] AMD said its eight-GPU Instinct MI355X led the submitted B200 and B300 results on GPT-OSS-120B in its own comparisons, and NVIDIA submitted preview results for Vera Rubin NVL72 claiming up to 2.5 times the GB300 NVL72's throughput on DeepSeek-R1.^[51] The MLCommons tables support AMD's claim for the offline scenario, where its system reached 121,818 tokens/s against at most 114,569 for an eight-GPU B300 system and 91,487 for a B200 system; in the server scenario an HPE B300 system edged AMD's entry, 113,685 to 113,241 tokens/s.^[103]

InferenceMAX and InferenceX

SemiAnalysis released InferenceMAX v1, an open benchmark of inference throughput and cost across models and serving stacks, in October 2025; later results appear under the name InferenceX. NVIDIA wrote that the B200 "sets the pace with 60,000 tokens per second per GPU and 1,000 tokens per second per user on gpt-oss," and that software optimizations brought its cost to two cents per million tokens, a fivefold reduction in two months.^[52] On the dense Llama 3.3 70B model, NVIDIA cited more than 10,000 tokens per second per GPU at 50 tokens per second per user, four times the H200's per-GPU throughput.^[52] NVIDIA also claimed that a $5 million GB200 NVL72 system could generate $75 million in token revenue.^[52]

Citing InferenceX data as of April 2026, NVIDIA's DGX B200 page puts Blackwell at about $0.02 per million tokens on GPT-OSS-120B with TensorRT-LLM at 55 tokens per second per user, about 4.5 times cheaper than Hopper systems at $0.09 per million tokens running vLLM.^[16]

DeepSeek-R1 optimization

NVIDIA published a series of TensorRT-LLM optimizations for DeepSeek-R1 on the B200 in 2025. On March 18, 2025, it said a single DGX B200 could serve the 671-billion-parameter model at more than 250 tokens per second per user, or at a maximum throughput above 30,000 tokens per second, and that it had raised DeepSeek-R1 throughput about 36 times since January 2025, a roughly 32-fold improvement in cost per token.^[53] The same post said DGX B200 systems running FP4 delivered more than three times the inference throughput of DGX H200 systems on DeepSeek-R1, Llama 3.1 405B and Llama 3.3 70B.^[53] TensorRT-LLM's engineering blog later traced minimum-latency throughput on eight B200 GPUs from a 67 tokens per second per user baseline to 253 before GTC 2025 and then to 368, using NVFP4 for the mixture-of-experts layers and three multi-token-prediction layers.^[54]

GB200 Grace Blackwell Superchip

The GB200 is a superchip board that combines one NVIDIA Grace CPU with two B200 GPUs; the "G" stands for Grace, NVIDIA's Arm-based server CPU.^[1] Grace has 72 Arm Neoverse V2 cores and up to 480 GB of LPDDR5X memory with up to 512 GB/s of bandwidth.^[7] It connects to the two GPUs over NVLink Chip-to-Chip (NVLink-C2C), which carries 900 GB/s of bidirectional bandwidth, against 128 GB/s for PCIe Gen5.^[36][2]

NVLink-C2C makes CPU and GPU memory coherent, so GPU kernels can read Grace memory directly and the CPU can read GPU memory without staging copies. NVIDIA describes the 72 GPUs of an NVL72 rack as working with about 30 TB of unified fast memory, counting both HBM3E and Grace's LPDDR5X.^[36][2] In practice this simplifies programming and reduces latency for workloads that alternate between CPU preprocessing and GPU computation.

Per superchip, NVIDIA lists 40 PFLOPS of sparse FP4 (20 PFLOPS dense), 20 PFLOPS of sparse FP8 (10 PFLOPS dense), 372 GB of HBM3E at 16 TB/s and 3.6 TB/s of NVLink bandwidth.^[7] Launch materials gave 384 GB of HBM3E and a power envelope of up to 2,700 W for the superchip.^[21] In this configuration each B200 is configurable up to 1,200 W and exposes 186 GB at 8 TB/s.^[2]

The GB200 ships in rack-scale GB200 NVL72 and NVL36 systems (the NVL36 puts 36 GPUs in one rack as nine dual-GB200 compute nodes) and in the GB200 NVL4, a board for MGX servers that joins four B200 GPUs and two Grace CPUs, with 1.8 TB of fast memory and 32 TB/s of memory bandwidth.^[36][2] NVIDIA projects the NVL4 at 1.3 to 2.3 times the performance of a four-way GH200 configuration across HPC and AI-for-science workloads, assuming 1,200 W per Blackwell GPU.^[2]

GB200 NVL72 rack

The GB200 NVL72 combines 36 GB200 superchips in one liquid-cooled rack: 72 B200 GPUs and 36 Grace CPUs in a single NVLink domain.^[36][2] NVIDIA says the rack acts "as a single, massive GPU" and calls it "an exascale computer in a single rack."^[2][7]

Physical configuration

According to NVIDIA's technical description, the NVL72 rack contains:^[36]

• 18 compute trays based on the MGX reference design, each with two Grace CPUs and four B200 GPUs (two GB200 superchips)
• 9 NVLink switch trays, each with two NVLink switch chips
• a copper cable cartridge that interconnects the GPUs, chosen for operational simplicity

The compute trays are liquid-cooled through cold plates on the CPUs and GPUs, while networking and storage devices and the rack's power supplies are air-cooled, and NVIDIA's DGX GB rack user guide puts rack power consumption at about 120 kW.^[111] Introl, a data center deployment firm, gives the same 120 kW figure and notes that NVIDIA specifies coolant inlet temperatures of 20 to 25 degrees Celsius.^[55] An air-cooled DGX B200 with eight GPUs, by comparison, is rated at about 14.3 kW, and NVIDIA's DGX SuperPOD reference design for DGX B200 uses racks drawing more than 25 kW.^[16][56] The NVL72's power density requires purpose-built facilities or significant retrofits with facility water loops.

NVLink fabric in NVL72

Each B200 has 18 fifth-generation NVLink links, for 1.8 TB/s of bidirectional bandwidth, and the nine switch trays connect all 18 links of every one of the 72 GPUs, giving the rack 130 TB/s of total NVLink bandwidth.^[36][37] Any GPU reaches any other through a single switch hop, so the all-reduce and all-to-all operations used in tensor parallelism and expert parallelism stay on NVLink instead of crossing InfiniBand. That keeps latency predictable and avoids the bandwidth limits of multi-hop network switches. The switch chips include SHARP engines that perform reductions and multicast inside the network, and Blackwell's switches add FP8 support for SHARP, which NVIDIA says improves bandwidth efficiency fourfold.^[37][2]

Fifth-generation NVLink can extend a single domain to 576 GPUs.^[1][2] NVIDIA's planned Vera Rubin Ultra NVL576, which belongs to a later NVLink generation (Rubin GPUs use sixth-generation NVLink), is designed to join eight racks of 72 Rubin Ultra GPUs into one 576-GPU NVLink domain over copper and direct optical connections.^[57][5]

GB200 NVL72 performance

Source: NVIDIA's GB200 NVL72 product page and Blackwell datasheet; NVIDIA lists sparse Tensor Core figures, and dense values are half.^[7][2]

NVIDIA's launch recap described the rack as delivering 720 petaflops of AI training performance and 1.4 exaflops of AI inference performance, which correspond to the sparse FP8 and sparse FP4 rows above (360 and 720 petaflops dense).^[20] The datasheet's feature list gives up to 13.5 TB of HBM3E, while the product page gives 13.4 TB, which matches 72 GPUs at 186 GB each.^[2][7]

NVIDIA's comparisons with Hopper are projections:^[2][36]

The 30-fold inference figure reflects both the FP4 advantage over FP8 and the NVLink domain, which lets a trillion-parameter model be sharded across the rack's 13.4 TB of GPU memory without spilling onto a slower scale-out network.^[2][36]

DGX B200 and HGX B200

Alongside the rack systems, NVIDIA and its partners sell eight-GPU servers built on SXM B200 modules without a Grace CPU, following the DGX and HGX product structure of earlier generations.

HGX B200

The HGX B200 is a baseboard with eight B200 GPUs linked through two NVLink switch chips, which give each GPU 1.8 TB/s of NVLink bandwidth and the board 14.4 TB/s in total.^[16][5] It carries 1.4 TB of HBM3E with 62 TB/s of aggregate bandwidth and 0.8 TB/s of networking bandwidth, and it connects to x86 host processors from Intel or AMD.^{[2][5][1][58]} Each GPU is configurable up to 1,000 W.^[2] Per board, NVIDIA lists 144 PFLOPS of sparse FP4 (72 PFLOPS dense) and 72 PFLOPS of sparse FP8 (36 PFLOPS dense).^[5] Communication between servers uses InfiniBand or Spectrum-X Ethernet.

The HGX B200 is the GPU subsystem that OEMs build into their own chassis. Examples include Lenovo's air-cooled ThinkSystem SR680a V3 and liquid-cooled SR780a V3, Dell's 4U liquid-cooled PowerEdge XE9685L with AMD EPYC 9005 processors, and HPE's ProLiant Compute XD685, which offers B200 in a 5U direct-liquid-cooled chassis.^[6][58][59]

Cloud providers that want standard x86 AI servers with Blackwell use HGX B200. Lambda made multi-node HGX B200 clusters of 16 to 512 GPUs available on demand through its 1-Click Clusters on March 18, 2025.^[60] AWS made P6-B200 instances, with eight GPUs, 1,440 GB of GPU memory and up to 3.2 Tb/s of EFAv4 networking, generally available on May 15, 2025, and CoreWeave made HGX B200 instances generally available on May 29, 2025.^[41][61] RunPod, Nebius and other GPU clouds also rent B200 capacity by the GPU-hour.^[62]

DGX B200

The DGX B200 is NVIDIA's own eight-GPU server built on the HGX B200 board and the successor to the DGX H100.^[16] It packages the GPUs with two Intel Xeon Platinum 8570 processors, system memory, NVMe storage and networking in a 10U chassis, as a turnkey option for organizations that want NVIDIA-validated hardware rather than sourcing parts from several vendors.^[16][63]

Sources: NVIDIA's DGX B200 product page and user guide.^[16][63]

NVIDIA projects that the DGX B200 delivers 3 times the training performance and 15 times the inference performance of the DGX H100.^[16] The NVL72's larger multiples reflect its 72-GPU NVLink domain, which NVIDIA uses to "overcome communication bottlenecks"; an eight-GPU server has to send traffic for large models over the scale-out network once a model spans more than one node.^[2]

The DGX B200 is also the building block of NVIDIA DGX SuperPOD. NVIDIA's reference architecture groups 32 DGX B200 systems (256 GPUs) into a scalable unit, centers its design on four units with 128 systems, and scales beyond 64 units and 2,000 systems over 400 Gb/s NDR InfiniBand.^[56]

The successor system is the DGX B300, with eight Blackwell Ultra GPUs and 2.1 TB of GPU memory in the same 10U size. NVIDIA says it raises dense FP4 performance by 1.5 times and attention performance by 2 times over the DGX B200 at about 14 kW.^[64]

B200 variants and product lineup

NVIDIA announced three Blackwell data center products in March 2024: the B100, the B200 and the GB200 superchip pairing B200 GPUs with Grace.^[1][21] The B200 and GB200 became the volume products; the B100 saw limited deployment.^[65]

B100 (HGX B100)

The NVIDIA B100 used the same two-die Blackwell silicon at a 700 W limit, so that HGX B100 boards could drop into the power and air-cooling designs of existing HGX H100 servers.^[21] At launch NVIDIA rated it at 7 PFLOPS of dense FP4 (14 PFLOPS sparse) and 3.5 PFLOPS of dense FP8 (7 PFLOPS sparse), with the same 192 GB of HBM3E at 8 TB/s as the B200's launch specification.^[21]

In August 2024 TrendForce described the B100 as a transitional, lower-power product that would be replaced by the B200, a planned B200A and GB200 racks once existing cloud-provider orders were filled. It reported that the B200A would use TSMC's CoWoS-S packaging with four 12-high HBM3E stacks (144 GB) and a lower TDP than the B200, aimed at OEM and enterprise customers after the first half of 2025.^[66] By September 2025, Modal described B100s as hard to find and said the industry had largely skipped them for the B200.^[65]

B200 (HGX B200)

The B200 on the eight-GPU HGX baseboard runs at up to 1,000 W per GPU.^[2] It is offered air-cooled, as in NVIDIA's DGX B200 and Lenovo's SR680a V3, and with direct liquid cooling in servers such as Lenovo's SR780a V3 and Dell's XE9685L.^[16][6][58] This is the form factor that CoreWeave, Lambda, AWS, Google Cloud and other clouds rent as eight-GPU B200 instances.^{[61][60][41][67]}

GB200 (with Grace CPU)

In the GB200, the same silicon is configurable up to 1,200 W per GPU and exposes 186 GB of HBM3E at 8 TB/s.^[2] This is the highest-performance configuration of B200 silicon, used in liquid-cooled GB200 NVL72 and NVL36 racks and in the GB200 NVL4 board.^[36][2]

B300 (Blackwell Ultra)

The B300, marketed as Blackwell Ultra, is NVIDIA's mid-generation refresh. NVIDIA announced it at GTC on March 18, 2025, with the GB300 NVL72 rack and the HGX B300 NVL16 system, and said products would be available from partners starting in the second half of 2025.^[11] Dell delivered the first GB300 NVL72 rack to CoreWeave in July 2025.^[12]

The B300 is a different chip from the B200: CUDA lists it as compute capability 10.3, against 10.0 for the B200.^[38] Blackwell Ultra keeps the two-die, 208-billion-transistor design and the 10 TB/s NV-HBI link; its full implementation has 160 SMs with 640 Tensor Cores, and NVIDIA notes that the number of SMs available varies by product.^[26] Per GPU it delivers 15 PFLOPS of dense NVFP4, against 10 PFLOPS for Blackwell in the same NVIDIA comparison, while the sparse NVFP4 rate stays at 20 PFLOPS and FP8 stays at 5 PFLOPS dense. Maximum memory rises to 288 GB of HBM3E at the same 8 TB/s, attention-layer exponential throughput roughly doubles (10.7 versus 5 trillion exponentials per second), and maximum power rises from 1,200 W to 1,400 W.^[26] NVIDIA's Blackwell Ultra datasheet lists 279 GB per GPU in the GB300 NVL72 and rates each HGX B300 GPU at 14 PFLOPS of dense FP4 (18 sparse), against 9 PFLOPS dense (18 sparse) for the HGX B200.^[110][2] On NVIDIA's HGX page, the eight-GPU HGX B300 is rated at 108 PFLOPS of dense FP4 (144 sparse), or 13.5 PFLOPS per GPU, versus 72 PFLOPS dense (144 sparse) for the HGX B200. Across eight GPUs, FP64 falls from 296 TFLOPS to 10 TFLOPS and sparse INT8 from 72 POPS to 3 POPS; per GPU, NVIDIA's datasheets give dense INT8 of 4,500 TOPS for the HGX B200 and about 154 TOPS for the HGX B300.^[5][2][110] NVIDIA's technical brief gives each HGX B300 GPU 270 GB of usable HBM3E at 7.7 TB/s, a 1,100 W limit and a PCIe Gen6 host interface, against 1,000 W and PCIe Gen5 for the HGX B200.^[25]

NVIDIA says the GB300 NVL72 provides 1.5 times the AI performance of the GB200 NVL72.^[11] In MLPerf Inference v5.1, GB300 NVL72 delivered 45% higher per-GPU DeepSeek-R1 throughput than GB200 NVL72 in the offline scenario and 25% higher in the server scenario.^[45] The B300's larger memory matters most for inference: a 70-billion-parameter model's BF16 weights take about 140 GB, which leaves about 40 GB of an HGX B200's 180 GB for KV cache and activations, against about 130 GB of an HGX B300's 270 GB.^[25]

NVLink and NVSwitch

Fifth-generation NVLink

NVLink 5 doubles per-GPU bandwidth over the fourth generation used in Hopper. Each GPU still has 18 links, but each link carries 100 GB/s instead of 50 GB/s, for 1.8 TB/s per GPU against 900 GB/s on the H100.^[37][17] NVIDIA says that is more than 14 times the bandwidth of PCIe Gen5.^[36]

Inside the NVL72, GPUs connect to the switch trays through a copper cable cartridge rather than optical modules, which NVIDIA chose for operational simplicity.^[36] The same generation can extend a single NVLink domain to 576 GPUs.^[1]

Fourth-generation NVSwitch

The NVSwitch chip used with NVLink 5, which NVIDIA calls the NVLink Switch, provides any-to-any connectivity in the NVL72. Eighteen switch chips in nine trays give 130 TB/s of total switch bandwidth across the rack's 72 GPUs, and an HGX B200 board uses two of the same chips.^[36][37][16] The switches include engines for NVIDIA's Scalable Hierarchical Aggregation and Reduction Protocol (SHARP), which perform reductions and multicast inside the network instead of routing all data back to individual GPUs. That reduces the bandwidth needed for gradient synchronization during training, and Blackwell's switches add FP8 support for SHARP.^[37][2]

Software support

CUDA and compute stack

The B200 has compute capability 10.0, the sm_100 compilation target.^[38][68] CUDA 12.8, released in January 2025, was the first toolkit to support Blackwell across NVIDIA's compilers, libraries, profilers and other developer tools, and its compiler added the SM_100, SM_101 and SM_120 Blackwell targets.^[32][68] In that release cuBLAS gained microscaled 4-bit and 8-bit floating-point matrix multiplication for compute capability 10.0, and CUTLASS 3.8 added Blackwell support.^[32]

The CUDA programming model is unchanged from Hopper. Applications that include PTX code run on Blackwell after just-in-time compilation by the driver, while binaries that contain only compiled code for older GPUs must be rebuilt; code that uses architecture-specific features, such as Hopper's compute_90a target, is not portable between generations.^[69] Framework users reach Blackwell's formats through NVIDIA libraries such as cuBLAS, cuDNN and Transformer Engine, which PyTorch and JAX call rather than requiring custom kernels.^[31][32]

TensorRT-LLM and Dynamo

TensorRT-LLM, the large language model inference library in NVIDIA's TensorRT family, added Blackwell support in version 0.17, with optimizations for Blackwell instructions, its memory hierarchy and FP4.^[53] Its Blackwell work includes:^[54][53]

• NVFP4 inference kernels for the B200's Tensor Cores
• NVFP4 mixture-of-experts layers with mixed-precision recipes that trade some speed for accuracy
• multi-token prediction layers for DeepSeek-R1, with relaxed acceptance to raise speculative acceptance rates
• tensor and expert parallelism over NVLink, with a custom one-shot all-reduce kernel

NVIDIA Dynamo, introduced at GTC on March 18, 2025, is open-source inference-serving software that sits above engines such as TensorRT-LLM, vLLM and SGLang. It uses disaggregated serving, running the prefill (processing) and decode (generation) phases of LLM inference on different GPUs so that each can be tuned on its own, and it is designed to coordinate inference across thousands of GPUs, including multi-rack GB200 NVL72 deployments.^[70]

Transformer Engine library

NVIDIA's open-source Transformer Engine library provides building blocks for mixed-precision training and inference of transformer models. It supports FP8 training on Hopper, Ada, Blackwell and Rubin GPUs and MXFP8 and NVFP4 training on Blackwell, offers PyTorch and JAX APIs, and is integrated into frameworks including Megatron-LM, NeMo Megatron Bridge, Hugging Face Accelerate and Lightning.^[31]

NIM microservices

NVIDIA NIM, introduced at GTC 2024 alongside Blackwell, is a set of inference microservices that package foundation models in prebuilt containers for deployment in any cloud or data center.^[19][2] NIM is part of NVIDIA AI Enterprise, the software platform NVIDIA sells for its data center GPUs.^[2] NVIDIA publishes NIM containers for frontier models and widely used open models, including the Llama, Mistral and DeepSeek families, for deployment on Blackwell hardware without writing serving code.

Collective communication

NCCL, NVIDIA's collective communications library, carries all-reduce and other collective operations over NVLink within a server or NVL72 rack and over InfiniBand or Ethernet between them. The SHARP engines in NVLink switches can take over part of that reduction work.^[37]

Pricing and availability

How much does a B200 cost?

NVIDIA does not publish list prices for its data center GPUs. On March 19, 2024, the day after the launch, Huang told CNBC that Blackwell would cost $30,000 to $40,000 per unit, a range CNBC noted was similar to analysts' estimates of $25,000 to $40,000 per chip for the H100. Huang later added that the cost covered more than the chip, because NVIDIA designs data centers and integrates its systems into other companies' facilities. He estimated NVIDIA's research and development spending on Blackwell at about $10 billion.^[24]

In May 2024, HSBC analysts, as cited by a Barron's writer and reported by Tom's Hardware, estimated the GB200 superchip's average selling price at $60,000 to $70,000 and the B100's at $30,000 to $35,000, and put GB200 NVL36 and NVL72 racks at about $1.8 million and $3 million.^[71] In October 2024 the reseller Broadberry listed the DGX B200 at $515,410.43.^[72] A July 2025 pricing survey by Modal gave $30,000 to $40,000 for a standalone B200 SXM module, $60,000 to $70,000 for a GB200 superchip and about $515,000 for a DGX B200.^[73]

Epoch AI estimated in December 2025 that a B200 costs about $6,400 to manufacture. Its breakdown put HBM at about $2,900, nearly half the total, CoWoS-L packaging at about $1,100, packaging yield losses at about $1,000 (within a wide range of $420 to $1,700) and the two logic dies at about $900. Against reported selling prices of $30,000 to $40,000, that implies chip-level gross margins of around 80%, though Epoch noted that NVIDIA's realized Blackwell margins are lower because most revenue comes from bundled servers and racks.^[4]

With the H100 estimated by analysts at $25,000 to $40,000 per chip in March 2024, the B200's per-chip price sits in a similar range while delivering more than twice the dense throughput per GPU, which improves price-performance for workloads that can use the extra throughput.^[24][5][17]

Cloud rental pricing

Rental rates have moved with supply and demand, so the figures below are dated observations rather than current prices.

Sources: Modal, July 14, 2025; Thunder Compute, reviewed October 9, 2026.^[73][62]

In July 2025 Lambda's three-year reserved rate of about $2.99 per GPU-hour was the lowest in Modal's survey, and on-demand rates at AWS ($14.24) and Google Cloud ($18.53) were roughly four to five times Lambda's $3.79.^[73] Price indexes recorded a rise in September 2026. Ornn's OCPI-B200 index, a volume-weighted average of transacted neocloud rental prices that is listed on Bloomberg, settled at $7.92 per GPU-hour on October 15, 2026; it had closed September at $8.09, up 33.3% from $6.07 at the end of August.^[74] GetDeploying's tracker, which compares 37 providers, put the median on-demand price across the 20 with listed on-demand configurations at $7.99 per GPU-hour on October 15, 2026, with listings from $3.75 to $14.24 and a rise of about 19% over the previous 12 months.^[14]

Availability and supply

NVIDIA began shipping Blackwell in late 2024, after the delayed ramp described in Production timeline and issues. In September 2024, Tom's Hardware reported that server deliveries would start in early December, with Microsoft first.^[75] OpenAI received one of the first engineering builds of the DGX B200 in early October 2024 and posted a photo of its team with the system on X.^[76] In October 2024, after hosting Huang, CFO Colette Kress and other executives for a three-day roadshow, Morgan Stanley wrote that "the Blackwell ramp is progressing on schedule with that product sold out for the next 12 months."^[77] In November 2024 NVIDIA said Blackwell production shipments would begin in its fiscal fourth quarter and that demand was "expected to exceed supply for several quarters in fiscal 2026."^[78]

NVIDIA reported $11.0 billion of Blackwell revenue in the quarter ending January 26, 2025, calling it the fastest product ramp in the company's history. Large cloud service providers accounted for about half of its data center revenue that quarter, and H200 sales were still growing.^[10] At GTC in March 2025, NVIDIA said the four largest cloud providers had already bought 3.6 million Blackwell GPUs, up from 1.3 million Hopper GPUs, under a new convention that counts each Blackwell package as two GPUs.^[79]

The major clouds brought Blackwell online through 2025. CoreWeave announced on February 4, 2025, that it was the first cloud provider to make GB200 NVL72 instances generally available, in clusters of up to 110,000 GPUs on Quantum-2 InfiniBand.^[80] Microsoft Azure made its ND GB200 v6 virtual machines generally available on March 18, 2025.^[81] Google Cloud's A4 virtual machines, based on HGX B200, were generally available by GTC in March 2025, and its GB200 NVL72-based A4X machines followed on May 29, 2025; Google said it was the first hyperscaler to offer both.^[67][82] AWS made P6-B200 instances generally available on May 15, 2025, initially through EC2 Capacity Blocks in its US West (Oregon) region, and launched P6e-GB200 UltraServers on July 9, 2025.^[41][83] Oracle began taking orders in September 2024 for OCI Superclusters of up to 131,072 Blackwell GPUs, which it rated at 2.4 zettaFLOPS of peak performance, a sparse-precision figure (see Oracle Cloud Infrastructure).^[84]

Packaging and memory, more than wafer output, limited supply. In December 2025 DigiTimes reported, via Wccftech, that NVIDIA had reserved more than half of TSMC's CoWoS capacity for 2026, about 800,000 to 850,000 wafers.^[85] Memory makers were similarly committed: Micron said it had completed price and volume agreements for its entire calendar 2026 HBM supply, and SK Hynix said its HBM supply discussions for the following year were finished.^[86]

In October 2025 Huang visited TSMC's fab in Phoenix, Arizona, to sign the first Blackwell wafer produced in the United States; NVIDIA said the milestone showed that Blackwell had reached volume production.^[87] At its GTC event in Washington, D.C., the same month, Huang projected $500 billion in combined GPU revenue from Blackwell and the following Rubin generation and said NVIDIA had shipped 6 million Blackwell units over the previous four quarters.^[88] On its November 2025 earnings call NVIDIA described the figure as visibility to $0.5 trillion of Blackwell and Rubin revenue from the start of calendar 2025 through the end of calendar 2026.^[112] In November 2025 he described demand for Blackwell as "very strong" and said there were "no active discussions" about selling Blackwell chips to China, sales that the Trump administration had blocked under U.S. export controls.^[89] As of October 2026, NVIDIA's HGX page lists both the HGX B300 and the HGX B200 as shipping.^[5]

Major deployments

B200 adoption was led by hyperscale clouds, specialized GPU clouds and frontier model developers. The following deployments were announced or reported between 2024 and 2026.

NVIDIA's DGX SuperPOD reference designs use the DGX B200 as their building block, with 32-system scalable units for customers that build their own clusters (see DGX B200).^[56]

CoreWeave

CoreWeave was among the earliest and largest deployers of GB200 NVL72. It announced general availability of GB200 NVL72 instances on February 4, 2025, describing clusters that scale to 110,000 GPUs over NVIDIA Quantum-2 InfiniBand at 400 Gb/s per GPU.^[80] NVIDIA said on April 15, 2025, that thousands of Blackwell GPUs were live there: IBM was training its next Granite models, Cohere was building its North platform for AI agents, and Mistral AI was receiving its first thousand Blackwell GPUs.^[93] CoreWeave added HGX B200 instances on May 29, 2025, received the first GB300 NVL72 rack from Dell in July 2025, and on September 30, 2026, announced Vera Rubin NVL72 availability with Cognition as the first customer running production workloads on it.^[61][12][13]

Microsoft Azure

Microsoft committed early to deploying Blackwell at scale as part of its AI infrastructure spending. Tom's Hardware reported in September 2024 that the first Blackwell servers would go to Microsoft in early December.^[75] Azure made ND GB200 v6 virtual machines, built on GB200 NVL72, generally available on March 18, 2025.^[81] In March 2026, Microsoft said it was the first cloud to bring up a Vera Rubin NVL72 system for validation.^[94][95]

Oracle Cloud Infrastructure

Oracle announced on September 11, 2024, that OCI Superclusters could be ordered with up to 131,072 Blackwell GPUs, which it called the largest AI supercomputer in the cloud, with 2.4 zettaFLOPS of peak performance.^[84] Oracle did not state the precision; the figure works out to about 18 petaFLOPS per GPU, the B200's sparse FP4 rating in HGX systems (9 petaFLOPS dense).^[2] Oracle was among the cloud providers NVIDIA named at launch as early Blackwell adopters.^[1] By October 2026 Oracle listed an eight-GPU bare-metal B200 shape, BM.GPU.B200.8, at $14.00 per GPU-hour on demand.^[62]

Meta

Meta was a large buyer of the previous generation: Mark Zuckerberg said in January 2024 that Meta's computing infrastructure would include 350,000 H100 GPUs by the end of that year.^[117] Meta was also named among the expected Blackwell adopters at launch.^[1] On February 17, 2026, the companies announced a multiyear, multigenerational partnership covering NVIDIA CPUs, networking and millions of Blackwell and Rubin GPUs, including what they described as the first large-scale deployment of Grace CPUs on their own and Meta's use of NVIDIA confidential computing for AI features in WhatsApp.^[90]

Amazon Web Services

AWS offers the B200 through EC2 P6-B200 instances, generally available since May 15, 2025, with eight GPUs, 1,440 GB of GPU memory and up to 3.2 Tb/s of EFAv4 networking; AWS said they deliver up to twice the performance of its P5en instances for AI training and inference.^[41] Its P6e-GB200 UltraServers, generally available since July 9, 2025, are built on GB200 NVL72 and sold through EC2 Capacity Blocks, starting in the Dallas Local Zone.^[83] In both July 2025 and October 2026, AWS's on-demand B200 price was higher than those of the specialized GPU clouds in the same surveys.^[73][62]

xAI

In June 2024 Elon Musk wrote that xAI's next step after the 100,000-GPU H100 cluster it was building in Memphis, later named Colossus, would probably be a system of about 300,000 B200 GPUs with CX8 networking, targeted for the following summer. DatacenterDynamics noted Musk's history of missing such timelines.^[91]

Alpha Compute and second-wave clouds

Smaller GPU clouds brought B200 capacity online behind the hyperscalers in 2026. Alpha Compute said in April 2026 that its first large cluster, ALPHA-01, with 504 B200 GPUs at a hydro-powered site in Canada, had slipped from a March target because of hardware delivery delays and was due to be handed over to customers on May 8, 2026. The company also held a right of first refusal on an expansion of the same site to more than 1,000 B200 GPUs, targeted for August 2026, and was building a 576-GPU B300 cluster in Sweden.^[92] On May 12, 2026, it said an unnamed frontier AI laboratory had leased the 504-GPU cluster for two years for $32.2 million, and on May 21 it reported ALPHA-01 live for that first enterprise customer, with the Swedish cluster moved to a third-quarter 2026 target.^[115][116]

MLPerf and benchmark partners

Cloud providers have used MLPerf to market their B200 capacity. Nebius listed first-place results for its HGX B200 systems in several categories of MLPerf Training v5.1 and Inference v6.0, and second places for all four of its HGX B200 language-model entries in Inference v6.1, where CoreWeave's B200 system led each of those tests.^{[46][47][49][103]} CoreWeave described its 2,496-GPU GB200 NVL72 run as the largest MLPerf Training v5.0 submission on Blackwell.^[44]

Use cases

What is the B200 used for?

The B200 is built mainly for training and serving large language models, and it is also used for scientific computing and data analytics. The sections below describe each.

LLM training

For training, the B200 targets large transformer models, including mixture-of-experts architectures with trillions of parameters. Its dense FP8 throughput (4,500 TFLOPS per HGX GPU), memory (180 to 186 GB) and NVLink bandwidth (1.8 TB/s per GPU) address the three usual limits in distributed training: compute, memory for model state, and communication for gradient synchronization.^[5][2]

For frontier training runs that need thousands of GPUs, the NVL72's 72-GPU NVLink domain simplifies the network. Tensor-parallel and expert-parallel traffic stays on the 130 TB/s NVLink fabric, and InfiniBand carries only traffic between racks, which reduces the load on the scale-out network.^[7] In MLPerf Training v6.0, a Microsoft Azure GB200 NVL72 system trained Llama 3.1 405B on 8,192 GPUs in 7.07 minutes.^[48][104]

LLM inference

Inference is where the B200's lead over Hopper is largest, for several reasons:

• FP4 Tensor Cores double throughput over FP8 for models that tolerate 4-bit weights, and NVFP4 lost 1% or less accuracy on key tasks in NVIDIA's DeepSeek-R1 tests^[8]
• 180 to 186 GB of memory per GPU fits larger models on fewer GPUs, which reduces cross-GPU communication^[2]
• 7.7 to 8 TB/s of memory bandwidth speeds the reading of weights and KV cache during token generation^[2]
• NVFP4 storage at about 4.5 bits per value uses roughly 1.8 times less memory than FP8, leaving more room for KV cache and larger batches^[8]

NVIDIA's projection for real-time inference on a 1.8-trillion-parameter MoE model shows the GB200 NVL72 serving 30 times as many tokens per GPU as an HGX H100 system.^[2]

Reasoning model inference

With the growth of chain-of-thought and reasoning models such as DeepSeek-R1 and the OpenAI o-series, inference demand shifted toward long generated sequences. These models produce many internal tokens before answering, which raises the number of tokens served per query and the importance of per-user token speed. NVIDIA's DeepSeek-R1 work on the B200 targeted that case, reaching more than 250 tokens per second per user on one DGX B200 by March 2025 and 368 in later TensorRT-LLM work.^[53][54] NVIDIA said in March 2025 that its optimizations had raised DeepSeek-R1 throughput about 36 times since January 2025 and cut cost per token about 32 times.^[53]

HPC and scientific computing

The B200 keeps double-precision support for traditional HPC work such as molecular dynamics, climate simulation and computational fluid dynamics. Its FP64 rate is 37 TFLOPS in the HGX B200 and 40 TFLOPS in the GB200, for both vector and Tensor Core operations; the H100 SXM is rated at 34 TFLOPS for vector FP64 and 67 TFLOPS with FP64 Tensor Cores.^[2][17] Blackwell Ultra cut FP64 sharply, to 10 TFLOPS across an eight-GPU HGX B300 against 296 TFLOPS for the HGX B200, so the B200 and GB200 retain far more double-precision throughput than their successors.^[5]

NVIDIA positions the GB200 NVL4 for converged HPC and AI work, with projected gains over GH200 on workloads that include GraphCast training, OpenFold and GenSLM inference, HPL, GROMACS, LAMMPS and ICON.^[2] Large memory capacity also helps HPC data sets that previously had to be partitioned across several GPUs, and the decompression engine targets analytics pipelines that read compressed data.^[34]

Power and cooling

The B200 SXM module in HGX systems is configurable up to 1,000 W, 43% more than the H100 SXM's 700 W, and up to 1,200 W in GB200 systems.^[2][17] NVIDIA's DGX B200 draws about 14.3 kW at most through six 3.3 kW power supplies arranged for 5+1 redundancy. Its user guide notes that nvidia-smi reports a 1,000 W per-GPU limit with five supplies working and 800 W when only three or four are available.^[63]

Cooling depends on the configuration. HGX B200 servers are available air-cooled, including the DGX B200 and Lenovo's SR680a V3, and with direct liquid cooling, as in Lenovo's SR780a V3, Dell's XE9685L and HPE's XD685.^{[16][6][58][59]} The GB200 NVL72 liquid-cools its compute trays, uses air cooling for networking, storage and power supplies, and draws about 120 kW per rack.^[111][55]

NVIDIA's claim of 25 times better energy efficiency for the GB200 NVL72 compares one liquid-cooled NVL72 rack with 65 racks of air-cooled eight-way HGX H100 servers delivering the same GPT-MoE-1.8T real-time inference throughput, and the HGX B200's projected 12-fold energy saving compares 8 racks with 100 racks of HGX H100.^[2] For the wider context of data center power demand, see AI energy consumption.

Production timeline and issues

Initial delay (mid-2024)

In early August 2024, The Information reported that NVIDIA had told Microsoft about delays affecting the most advanced Blackwell models, and other outlets reported that volume shipments of chips such as the B200 could slip by three months or more, into early 2025. The Register tied the problem to the complexity of TSMC's CoWoS packaging, and Tom's Hardware reported that Bloomberg had corroborated the delay through a source at Microsoft.^[96][97]

On August 28, 2024, NVIDIA said it had "executed a change to the Blackwell GPU mask to improve production yield" and that the production ramp would begin in its fiscal fourth quarter, when it expected "to ship several billion dollars in Blackwell revenue." Reports attributed the problem to the CoWoS-L packaging that joins the two dies: an alleged mismatch in thermal expansion among the GPU chiplets, LSI bridges, RDL interposer and substrate caused warping, and NVIDIA reportedly redesigned the top metal layers and bumps of the GPU silicon. NVIDIA said no functional changes to the silicon were required.^[9]

In October 2024 Huang said: "We had a design flaw in Blackwell, it was functional, but the design flaw caused the yield to be low. It was 100% Nvidia's fault." He credited TSMC with helping NVIDIA recover and dismissed reports of tension between the companies as "fake news."^[98] NVIDIA's November 2024 financial commentary said it had "completed a successful mask change for Blackwell" that improved production yields.^[78]

Q4 2024 ramp

NVIDIA began ramping Blackwell production in its fiscal fourth quarter, which ended January 26, 2025, and delivered $11.0 billion of Blackwell revenue in that quarter, well above the "several billion dollars" it had forecast in August.^[9][10] Hopper continued to ship in large volumes as a bridge while Blackwell ramped; NVIDIA reported sequential growth in H200 sales in the same quarter.^[10]

Liquid cooling and rack issues

Early GB200 NVL72 racks ran into system-level problems beyond the chip yield issue. Engineers at NVIDIA's suppliers cited overheating, liquid-cooling leaks, software bugs and chip-to-chip connectivity problems, which emerged at the end of 2024 and delayed rack shipments.^[99][100] The Financial Times reported in May 2025 that Foxconn, Inventec, Dell and Wistron had made "a series of breakthroughs." NVIDIA's Taiwanese partners said at Computex 2025 that GB200 rack shipments had begun at the end of the first quarter of 2025 and that "production capacity was now being rapidly scaled up."^[99]

Ramp trajectory and volume production

By mid-2025 cloud providers were offering B200 and GB200 capacity at commercial scale (see Availability and supply). By March 2025 the four largest cloud providers had bought 3.6 million Blackwell GPUs under NVIDIA's two-GPUs-per-package count, compared with 1.3 million Hopper GPUs.^[79] In October 2025 NVIDIA presented the first Blackwell wafer made in the United States, at TSMC's Arizona fab, as a sign that Blackwell had reached volume production, and Huang said NVIDIA had shipped 6 million Blackwell units over four quarters.^[87][88] Blackwell Ultra systems began shipping in the second half of 2025, and in its third quarter of fiscal 2026, reported in November 2025, NVIDIA said GB300 had overtaken GB200 and contributed roughly two-thirds of Blackwell revenue.^[11][112] In the quarter ended July 26, 2026, Data Center revenue reached $89.0 billion, driven by Blackwell Ultra, and NVIDIA said the Blackwell architecture still made up the vast majority of its revenue; it began production shipments of Vera Rubin in August 2026.^[113][114] As of October 2026, both Blackwell generations remain in NVIDIA's shipping HGX lineup.^[5]

The delays kept Hopper as NVIDIA's main data center GPU through most of 2024 rather than seeing it displaced soon after the March announcement; NVIDIA said in November 2024 that it would ship both Hopper and Blackwell systems from its fiscal fourth quarter onward.^[78]

Comparison to predecessors and successors

B200 vs. H100

The H100 was NVIDIA's main data center GPU from its 2022 launch through 2024. It uses a single 80-billion-transistor die on TSMC's 4N process.^[26] The HGX B200 exceeds it on the main AI specifications: 208 versus 80 billion transistors, 180 versus 80 GB of memory, 7.7 versus 3.35 TB/s of bandwidth, 4,500 versus 1,979 dense FP8 TFLOPS, and FP4 support that the H100 lacks. Double precision is the exception: the H100 SXM's FP64 Tensor Core rate of 67 TFLOPS is higher than the HGX B200's 37 TFLOPS.^[2][17]

For organizations still running H100 clusters in 2025 and 2026, the upgrade case has been strongest for inference on large models. Training gains are smaller because large runs are limited by network communication as well as per-GPU throughput; in MLPerf Training v4.1, by NVIDIA's per-GPU comparison, the B200's gain over the H100 was about 2 to 2.2 times on the two language-model benchmarks and 1.4 to 2 times on the others.^[42]

B200 vs. H200

The H200 was a mid-cycle memory upgrade to the H100. It replaced 80 GB of HBM3 with 141 GB of HBM3E at 4.8 TB/s but kept the same compute die and 700 W limit.^[39] Against it, the HGX B200 offers 180 versus 141 GB, 7.7 versus 4.8 TB/s, 4,500 versus 1,979 dense FP8 TFLOPS, and native FP4.^[2][39] In MLPerf Inference v5.0, an eight-GPU B200 system served Llama 2 70B about three times as fast as an eight-GPU H200 system.^[43]

B200 vs. B100

The B100 used the same silicon in a 700 W envelope for HGX H100-compatible servers, with the B200's launch memory (192 GB at 8 TB/s) but lower throughput: 7 PFLOPS of dense FP4 (14 sparse) and 3.5 PFLOPS of dense FP8 (7 sparse) per GPU, against 9 and 4.5 PFLOPS dense for the B200.^[21] It saw limited deployment.^[65]

B200 vs. B300 (Blackwell Ultra)

The B300 is the B200's direct successor within the Blackwell generation and began shipping in the second half of 2025.^[11][12] It raises dense NVFP4 per GPU by about half (15 versus 10 PFLOPS in the GB300 and GB200 configurations, and 14 versus 9 PFLOPS in HGX systems by NVIDIA's datasheets), raises maximum memory from 192 to 288 GB of HBM3E, doubles attention-layer exponential throughput and raises maximum power from 1,200 W to 1,400 W, while keeping the same 8 TB/s bandwidth and 1.8 TB/s NVLink.^[26][110][2] In exchange it gives up most of the B200's FP64 and INT8 throughput.^[5]

For organizations that ordered B200 systems in 2024 and 2025, the B300 is an upgrade path rather than an immediate replacement. The B200 remains competitive for the workloads it was designed for, and the two coexist in NVIDIA's HGX lineup and in cloud catalogs as of October 2026.^[5][14]

B200 vs. Vera Rubin

NVIDIA Vera Rubin is the architecture generation after Blackwell, pairing the Vera CPU with the Rubin GPU in place of Grace and Blackwell; each Rubin GPU reportedly has two reticle-sized dies and eight HBM4 interfaces.^[88] NVIDIA said at GTC in March 2026 that Vera Rubin NVL72 was in full production and on track to ship in the second half of 2026, and Huang repeated at Computex in June 2026 that the platform had reached full production.^[57][101] Microsoft said in March 2026 that it was the first cloud to bring up a Vera Rubin NVL72 system for validation.^[94] On August 26, 2026, NVIDIA said it had begun production shipments of Vera Rubin earlier that month.^[114]

On September 30, 2026, CoreWeave announced Vera Rubin NVL72 availability, and its first customer, Cognition, reported up to 4.8 times the total token throughput of a GB200 NVL72 baseline on its SWE-2 inference workloads.^[13] NVIDIA's preview results in MLPerf Inference v6.1 claimed up to 2.5 times the GB300 NVL72's throughput on DeepSeek-R1.^[51] As of October 2026, HGX B200 systems remained on sale while Vera Rubin ramped, and B200 rental prices had risen in September 2026.^[5][74]

B200 vs. AMD Instinct

Chips and Cheese found that the B200's HBM3E gives it a memory bandwidth lead over AMD's Instinct MI300X, which uses HBM3 at up to 5.3 TB/s, while the MI300X "retains some surprising advantages" in its tests despite being an outgoing product.^[15] In MLPerf Inference v6.1, AMD said its eight-GPU Instinct MI355X led the submitted B200 and B300 results on GPT-OSS-120B in its own comparisons.^[51]

ELI5: what is the B200, simply?

A chip can only be made so large, because the machines that print it can expose only a limited area of silicon at once. NVIDIA built each of the B200's two dies at that limit and joined them with a link so fast (10 trillion bytes per second) that software treats the pair as one GPU. Together they hold 208 billion transistors, the tiny switches that do the computing. Around them sit 192 GB of very fast memory, of which 180 to 186 GB is usable, enough to hold a 70-billion-parameter AI model in 16-bit precision on a single chip. The B200's Tensor Cores can also do math with 4-bit numbers instead of 8- or 16-bit ones, which lets it run AI chatbots faster and more cheaply than the previous generation, as long as the model tolerates the lower precision.

See also

• NVIDIA Blackwell
• NVIDIA Blackwell Ultra
• NVIDIA B100
• NVIDIA GB200 NVL72
• NVIDIA GB300 NVL72
• NVIDIA DGX B300
• NVIDIA H100
• NVIDIA H200
• NVIDIA Hopper
• NVIDIA Vera Rubin
• NVIDIA Grace
• NVIDIA DGX SuperPOD
• NVLink
• HBM3e
• Tensor Core
• NVFP4
• CUDA
• TensorRT
• NVIDIA Dynamo
• NVIDIA NIM
• MLPerf
• Cloud AI GPU pricing comparison
• List of AI compute deals
• Jensen Huang
• TSMC

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28 ^29"NVIDIA Blackwell Platform Arrives to Power a New Era of Computing." NVIDIA Newsroom, 2024-03-18. nvidianews.nvidia.com/...er-a-new-era-of-computing
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28 ^29 ^30 ^31 ^32 ^33 ^34 ^35 ^36 ^37 ^38 ^39 ^40 ^41 ^42 ^43 ^44 ^45 ^46 ^47 ^48 ^49 ^50 ^51 ^52 ^53 ^54 ^55 ^56 ^57 ^58 ^59 ^60 ^61 ^62 ^63 ^64 ^65 ^66 ^67 ^68 ^69 ^70 ^71 ^72 ^73 ^74 ^75 ^76 ^77 ^78 ^79 ^80 ^81 ^82 ^83"NVIDIA Blackwell Datasheet (GB200 NVL72, GB200 NVL4, HGX B200)." NVIDIA, October 2025. dam-cdn.nvd.orangelogic.com/...18b2852i557jem5.pdf
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7"TechInsights' NVIDIA HGX B200 Analysis Reveals HBM3E Supplier and Advanced Packaging Innovation." TechInsights, 2025-04-14. techinsights.com/...-advanced-packaging-innovation
4. ^1 ^2 ^3 ^4Luke Emberson. "NVIDIA's B200 costs around $6,400 to produce, with memory accounting for half." Epoch AI, 2025-12-10. epochai.substack.com/...-b200-costs-around-6400-to
5. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23"NVIDIA HGX Platform." NVIDIA, accessed 2026-10-15. nvidia.com/...hgx
6. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13"ThinkSystem NVIDIA HGX B200 180GB 1000W GPU Product Guide (LP2226)." Lenovo Press, 2025-05-28, updated 2026-01-13. lenovopress.lenovo.com/...dia-b200-180gb-1000w-gpu
7. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10"NVIDIA GB200 NVL72." NVIDIA, accessed 2026-10-15. nvidia.com/...gb200-nvl72
8. ^1 ^2 ^3 ^4 ^5 ^6 ^7"Introducing NVFP4 for Efficient and Accurate Low-Precision Inference." NVIDIA Technical Blog, 2025-06-24. developer.nvidia.com/...te-low-precision-inference
9. ^1 ^2 ^3 ^4Anton Shilov. "Nvidia addresses significant Blackwell yield issues, production ramps in Q4." Tom's Hardware, 2024-08-29. tomshardware.com/...d-issues-production-ramp-in-q4
10. ^1 ^2 ^3 ^4"CFO Commentary on Fourth Quarter and Fiscal 2025 Results." NVIDIA (Form 8-K exhibit), 2025-02-26. sec.gov/...q4fy25cfocommentary
11. ^1 ^2 ^3 ^4 ^5 ^6"NVIDIA Blackwell Ultra AI Factory Platform Paves Way for Age of AI Reasoning." NVIDIA Newsroom, 2025-03-18. nvidianews.nvidia.com/...y-for-age-of-ai-reasoning
12. ^1 ^2 ^3 ^4"Dell Delivers Market's First NVIDIA GB300 NVL72 to CoreWeave." Dell Technologies, 2025-07-03. dell.com/...-first-nvidia-gb300-nvl72-to-coreweave
13. ^1 ^2 ^3"CoreWeave Delivers NVIDIA Vera Rubin NVL72 at Scale." CoreWeave, 2026-09-30. coreweave.com/...ion-scale-starting-with-cognition
14. ^1 ^2 ^3"Nvidia B200 Rental Prices: Compare 37 Providers." GetDeploying, 2026-10-15. getdeploying.com/...nvidia-b200
15. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9clamchowder and Neggles. "Nvidia's B200: Keeping the CUDA Juggernaut Rolling ft. Verda (formerly DataCrunch)." Chips and Cheese, 2025-12-15. old.chipsandcheese.com
16. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13"NVIDIA DGX B200." NVIDIA, accessed 2026-10-15. nvidia.com/...dgx-b200
17. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14"NVIDIA H100 Tensor Core GPU." NVIDIA, accessed 2026-10-15. nvidia.com/...h100
18. ^"NVIDIA GTC 2024: A Glimpse Into the Future of AI With Jensen Huang." NVIDIA Blog, 2024-03-14. blogs.nvidia.com/...gtc-2024-jensen-huang
19. ^1 ^2 ^3 ^4Kif Leswing. "Nvidia CEO Jensen Huang announces new AI chips: 'We need bigger GPUs'." CNBC, 2024-03-18. cnbc.com/...well-ai-chip-launching-later-this-year
20. ^1 ^2 ^3 ^4Brian Caulfield. "'We Created a Processor for the Generative AI Era,' NVIDIA CEO Says." NVIDIA Blog, 2024-03-18. blogs.nvidia.com/...2024-gtc-keynote
21. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Jarred Walton. "Nvidia's next-gen AI GPU revealed: Blackwell B200 GPU delivers up to 20 petaflops of compute and massive improvements over Hopper H100." Tom's Hardware, 2024-03-18. tomshardware.com/...-improvements-over-hopper-h100
22. ^"David Blackwell." Encyclopaedia Britannica. britannica.com/...David-Blackwell
23. ^1 ^2 ^3 ^4 ^5 ^6 ^7Hassan Mujtaba. "NVIDIA Blackwell GPU Architecture Official: 208 Billion Transistors, 5x AI Performance, 192 GB HBM3e Memory, 8 TB/s Bandwidth." Wccftech, 2024-03-18. wccftech.com/...ai-performance-192-gb-hbm3e-memory
24. ^1 ^2 ^3Kif Leswing. "Nvidia's Blackwell AI chip will cost more than $30,000, CEO says." CNBC, 2024-03-19. cnbc.com/...hip-will-cost-more-than-30000-ceo-says
25. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10"NVIDIA Blackwell Architecture Technical Brief: Built for the Age of AI Reasoning." NVIDIA, updated October 2025. resources.nvidia.com/en-us-blackwell-architecture
26. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9"Inside NVIDIA Blackwell Ultra: The Chip Powering the AI Factory Era." NVIDIA Technical Blog, 2025-08-22, updated 2025-09-24. developer.nvidia.com/...owering-the-ai-factory-era
27. ^"Nvidia Unveils Its Next-Generation 7nm Ampere A100 GPU for Data Centers, and It's Absolutely Massive." Tom's Hardware, 2020-05-14. tomshardware.com/...nvidia-ampere-A100-gpu-7nm
28. ^"Micron announces data center, edge innovations with Nvidia." TipRanks (The Fly), 2025-03-18. tipranks.com/...enter-edge-innovations-with-nvidia
29. ^1 ^2 ^3 ^4"B200 SM." Cornell Virtual Workshop: Understanding GPU Architecture, Cornell University Center for Advanced Computing. cvw.cac.cornell.edu/...b200_sm
30. ^1 ^2Glenn K. Lockwood. "NVIDIA B200." glennklockwood.com. glennklockwood.com/...B200
31. ^1 ^2 ^3 ^4"Transformer Engine." NVIDIA (GitHub repository), accessed 2026-10-15. github.com/...TransformerEngine
32. ^1 ^2 ^3 ^4"CUDA Toolkit 12.8 Delivers NVIDIA Blackwell Support." NVIDIA Technical Blog, 2025-01-31. developer.nvidia.com/...s-nvidia-blackwell-support
33. ^1 ^2 ^3 ^4"NVIDIA Blackwell Enables 3x Faster Training and Nearly 2x Training Performance Per Dollar than Previous-Gen Architecture." NVIDIA Technical Blog, 2025-12-11. developer.nvidia.com/...-previous-gen-architecture
34. ^1 ^2 ^3 ^4 ^5"Decompression Engine." nvCOMP documentation, NVIDIA, accessed 2026-10-15. docs.nvidia.com/...decompression_engine_faq
35. ^"Speeding Up Data Decompression with nvCOMP and the NVIDIA Blackwell Decompression Engine." NVIDIA Technical Blog, 2025-10-06. developer.nvidia.com/...kwell-decompression-engine
36. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13"NVIDIA GB200 NVL72 Delivers Trillion-Parameter LLM Training and Real-Time Inference." NVIDIA Technical Blog, 2024-03-18. developer.nvidia.com/...ng-and-real-time-inference
37. ^1 ^2 ^3 ^4 ^5 ^6 ^7"NVLink and NVLink Switch." NVIDIA, accessed 2026-10-15. nvidia.com/...nvlink
38. ^1 ^2 ^3"CUDA GPU Compute Capability." NVIDIA Developer, accessed 2026-10-15. developer.nvidia.com/cuda-gpus
39. ^1 ^2 ^3 ^4 ^5"NVIDIA H200 Tensor Core GPU." NVIDIA, accessed 2026-10-15. nvidia.com/...h200
40. ^"NVIDIA B200: 192GB Specs, Benchmarks, vs H100." Spheron Blog, 2026-05-14. spheron.network/...nvidia-b200-complete-guide
41. ^1 ^2 ^3 ^4 ^5 ^6"Amazon EC2 P6-B200 instances powered by NVIDIA B200 GPUs now generally available." Amazon Web Services, 2025-05-15. aws.amazon.com/...-b200-instances-nvidia-b200-gpus
42. ^1 ^2 ^3"NVIDIA Blackwell Doubles LLM Training Performance in MLPerf Training v4.1." NVIDIA Technical Blog, 2024-11-13. developer.nvidia.com/...ce-in-mlperf-training-v4-1
43. ^1 ^2 ^3 ^4"NVIDIA Blackwell Delivers Massive Performance Leaps in MLPerf Inference v5.0." NVIDIA Technical Blog, 2025-04-02. developer.nvidia.com/...s-in-mlperf-inference-v5-0
44. ^1 ^2 ^3 ^4Peter Salanki. "CoreWeave, NVIDIA, and IBM Set MLPerf Record with Largest NVIDIA GB200 Blackwell Cluster, Achieving Over 2x Faster Training." CoreWeave, 2025-06-04. coreweave.com/...achieving-over-2x-faster-training
45. ^1 ^2 ^3"NVIDIA Blackwell Ultra Sets New Inference Records in MLPerf Debut." NVIDIA Technical Blog, 2025-09-09. developer.nvidia.com/...ce-records-in-mlperf-debut
46. ^1 ^2 ^3"MLPerf Training v5.1: Leading results on NVIDIA Blackwell and Blackwell Ultra systems." Nebius, 2025-11-12. nebius.com/...mlperf-training-v5-1-results
47. ^1 ^2 ^3"MLPerf Inference v6.0: Top-tier AI performance on NVIDIA Blackwell and Blackwell Ultra." Nebius, 2026-04-01. nebius.com/...mlperf-inference-v6-0-results
48. ^1 ^2 ^3"NVIDIA Blackwell Tops MLPerf Training 6.0 with Industry-Leading Scale and Performance." NVIDIA Technical Blog, 2026-06-16. developer.nvidia.com/...ding-scale-and-performance
49. ^1 ^2 ^3"MLPerf Inference v6.1 Results: NVIDIA Vera Rubin NVL72." Nebius, 2026-09-16. nebius.com/...mlperf-inference-v6-1-results
50. ^1 ^2 ^3"MLPerf Inference v6.0 Datacenter results." MLCommons, 2026-04-01. github.com/...inference_results_v6.0
51. ^1 ^2 ^3"MLPerf Inference v6.1: 5.7x Per-Accelerator Gains, a 512-GPU Run, and Vera Rubin's First Peer-Reviewed Numbers." StorageReview, 2026-09-16. storagereview.com/...s-first-peer-reviewed-numbers
52. ^1 ^2 ^3Dion Harris. "NVIDIA Blackwell Raises Bar in New InferenceMAX Benchmarks, Delivering Unmatched Performance and Lowest Cost Per Token." NVIDIA Blog, 2025-10-09. blogs.nvidia.com/...inferencemax-benchmark-results
53. ^1 ^2 ^3 ^4 ^5 ^6"NVIDIA Blackwell Delivers World-Record DeepSeek-R1 Inference Performance." NVIDIA Technical Blog, 2025-03-18. developer.nvidia.com/...k-r1-inference-performance
54. ^1 ^2 ^3"Pushing Latency Boundaries: Optimizing DeepSeek-R1 Performance on NVIDIA B200 GPUs." NVIDIA TensorRT-LLM Tech Blog, last updated 2026-03-16. nvidia.github.io/...erformance_on_NVIDIA_B200_GPUs
55. ^1 ^2"GB200 NVL72 Deployment: Managing 72 GPUs in Liquid-Cooled Configurations." Introl, 2026-04-15. introl.com/...vl72-deployment-72-gpu-liquid-cooled
56. ^1 ^2 ^3"NVIDIA DGX SuperPOD: Next Generation Scalable Infrastructure for AI Leadership, Reference Architecture Featuring NVIDIA DGX B200." NVIDIA, accessed 2026-10-15. docs.nvidia.com/...abstract
57. ^1 ^2"NVIDIA Vera Rubin POD: Seven Chips, Five Rack-Scale Systems, One AI Supercomputer." NVIDIA Technical Blog, 2026-03-16. developer.nvidia.com/...stems-one-ai-supercomputer
58. ^1 ^2 ^3 ^4"PowerEdge XE9685L Spec Sheet." Dell Technologies, accessed 2026-10-15. delltechnologies.com/...dge-xe9685l-spec-sheet.pdf
59. ^1 ^2"HPE ProLiant Compute XD685." Hewlett Packard Enterprise, accessed 2026-10-15. buy.hpe.com/...1014862906
60. ^1 ^2 ^3"Lambda Announces General Availability of Multi-Node NVIDIA HGX B200 Accelerated Clusters On Demand." Business Wire, 2025-03-18. businesswire.com/...Accelerated-Clusters-On-Demand
61. ^1 ^2 ^3 ^4"B200 instances." CoreWeave Changelog, 2025-05-29. docs.coreweave.com/...b200-instances
62. ^1 ^2 ^3 ^4"NVIDIA B200 Pricing (October 2026)." Thunder Compute, published 2026-04-02, reviewed 2026-10-09. thundercompute.com/...nvidia-b200-pricing
63. ^1 ^2 ^3"Introduction to NVIDIA DGX B200 Systems." NVIDIA DGX B200 User Guide, accessed 2026-10-15. docs.nvidia.com/...introduction-to-dgxb200
64. ^"NVIDIA DGX B300." NVIDIA, accessed 2026-10-15. nvidia.com/...dgx-b300
65. ^1 ^2 ^3"NVIDIA Showdown: H100s and H200s vs B100s and B200s." Modal, 2025-09-05. modal.com/...h100-and-h200-vs-b100-and-b200
66. ^"NVIDIA's New B200A Targets OEM Customers; High-End GPU Shipments Expected to Grow 55% in 2025." TrendForce via I-Connect007, 2024-08-07. iconnect007.com/...ein
67. ^1 ^2 ^3"Google Cloud goes to NVIDIA GTC." Google Cloud Blog, 2025-03-19. cloud.google.com/...google-cloud-goes-to-nvidia-gtc
68. ^1 ^2"CUDA Toolkit 12.8 Release Notes." NVIDIA. docs.nvidia.com/...cuda-toolkit-release-notes
69. ^"Blackwell Architecture Compatibility." NVIDIA CUDA 12.8 documentation. docs.nvidia.com/...blackwell-compatibility-guide
70. ^"NVIDIA Dynamo Open-Source Library Accelerates and Scales AI Reasoning Models." NVIDIA, 2025-03-18. investor.nvidia.com/...default
71. ^Anton Shilov. "Nvidia's next-gen Blackwell AI Superchips could cost up to $70,000: fully-equipped server racks reportedly range up to $3,000,000 or more." Tom's Hardware, 2024-05-14. tomshardware.com/...nge-up-to-dollar3000000-report
72. ^Muhammad Zuhair. "NVIDIA 'Blackwell' DGX B200 Listed, Prices Start At A Half A Million Dollars For The Top of The Line AI Hardware." Wccftech, 2024-10-12. wccftech.com/...ollars-top-of-the-line-ai-hardware
73. ^1 ^2 ^3 ^4"How much does it cost to run NVIDIA B200 GPUs in 2025?" Modal, 2025-07-14. modal.com/...nvidia-b200-pricing
74. ^1 ^2"OCPI-B200: NVIDIA B200 Price Index." Ornn Data, accessed 2026-10-15. data.ornn.com/...b200
75. ^1 ^2Anton Shilov. "Nvidia will reportedly start Blackwell server deliveries in early December: the first servers will go to Microsoft." Tom's Hardware, 2024-09-26. tomshardware.com/...t-servers-will-go-to-microsoft
76. ^1 ^2Sarfraz Khan. "NVIDIA Delivers 'One of The First' Blackwell DGX B200 AI System To OpenAI." Wccftech, 2024-10-09. wccftech.com/...kwell-dgx-b200-ai-system-to-openai
77. ^"Nvidia meetings highlight magnitude, length of runway, says Morgan Stanley." TipRanks (The Fly), 2024-10-10. tipranks.com/...ngth-of-runway-says-morgan-stanley
78. ^1 ^2 ^3"CFO Commentary on Third Quarter Fiscal 2025 Results." NVIDIA (Form 8-K exhibit), 2024-11-20. sec.gov/...q3fy25cfocommentary
79. ^1 ^2Kif Leswing. "Nvidia's Huang says faster chips are the best way to reduce AI costs." CNBC, via NBC Los Angeles, 2025-03-19. nbclosangeles.com/...3658527
80. ^1 ^2 ^3"CoreWeave First Cloud Provider to Announce General Availability of NVIDIA GB200 NVL72 Instances." CoreWeave, 2025-02-04. coreweave.com/...y-of-nvidia-gb200-nvl72-instances
81. ^1 ^2 ^3"Accelerating the Intelligence Age with Azure AI Infrastructure and the GA of ND GB200 v6." Microsoft Tech Community, Azure High Performance Computing Blog, 2025-03-18. techcommunity.microsoft.com/...4394575
82. ^1 ^2"New A4X VMs powered by NVIDIA GB200 GPUs." Google Cloud Blog, 2025-02-20, updated 2025-05-29. cloud.google.com/...s-powered-by-nvidia-gb200-gpus
83. ^1 ^2 ^3"New Amazon EC2 P6e-GB200 UltraServers powered by NVIDIA Grace Blackwell GPUs for the highest AI performance." AWS News Blog, 2025-07-09. aws.amazon.com/...s-for-the-highest-ai-performance
84. ^1 ^2 ^3"Oracle Offers First Zettascale Cloud Computing Cluster." Oracle, 2024-09-11. oracle.com/...e-cloud-computing-cluster-2024-09-11
85. ^Muhammad Zuhair. "NVIDIA Alone Has TSMC's Advanced Packaging Lines Booked for Several Years Ahead." Wccftech, 2025-12-10. wccftech.com/...nes-booked-for-several-years-ahead
86. ^"Why GPU and HBM Supply Is Still Broken in 2026: CoWoS, 2nm, and What's Next." Fusion Worldwide, 2025-12-04. info.fusionww.com/...city-constraints-through-2027
87. ^1 ^2"The Engines of American-Made Intelligence: NVIDIA and TSMC Celebrate First NVIDIA Blackwell Wafer Produced in the US." NVIDIA Blog, 2025-10-17. blogs.nvidia.com/...tsmc-blackwell-manufacturing
88. ^1 ^2 ^3"NVIDIA Dismisses AI Bubble Concerns, Reportedly Projects $500B in GPU Sales from Blackwell and Rubin." TrendForce, 2025-10-29. trendforce.com/...u-sales-from-blackwell-and-rubin
89. ^Wen-Yee Lee. "Nvidia CEO Huang sees strong demand for Blackwell chips." Reuters, via Yahoo Finance, 2025-11-08. finance.yahoo.com/...o-huang-sees-strong-041748560
90. ^1 ^2"Meta Builds AI Infrastructure With NVIDIA." NVIDIA, 2026-02-17. investor.nvidia.com/...-Infrastructure-With-NVIDIA
91. ^1 ^2"Elon Musk claims 300,000 B200 GPU supercomputer for xAI by next summer; the usual caveats apply." DatacenterDynamics, 2024-06-03. datacenterdynamics.com/...-the-usual-caveats-apply
92. ^1 ^2"Alpha Compute AI GPU Launch Update; First Large-Scale NVIDIA Blackwell Cluster." GlobeNewswire, 2026-04-27. globenewswire.com/...cale-nvidia-blackwell-cluster
93. ^Ian Buck. "Thousands of NVIDIA Grace Blackwell GPUs Now Live at CoreWeave." NVIDIA Blog, 2025-04-15. blogs.nvidia.com/...ve-grace-blackwell-gb200-nvl72
94. ^1 ^2"Microsoft brings up first NVIDIA Vera Rubin NVL72 system." Converge Digest, 2026-03-15. convergedigest.com/...idia-vera-rubin-nvl72-system
95. ^"Microsoft claims to be first hyperscaler to 'power on' Nvidia Vera Rubin NVL72 system." DatacenterDynamics, March 2026. datacenterdynamics.com/...-vera-rubin-nvl72-system
96. ^Dan Robinson. "Nvidia delays Blackwell GPUs until 2025, say reports." The Register, 2024-08-05. theregister.com/...dia_delays_blackwell_gpus_until
97. ^Jowi Morales. "Nvidia Blackwell GPUs allegedly delayed due to design flaws: launch expected to be pushed back by three months or more." Tom's Hardware, 2024-08-03. tomshardware.com/...ly-delayed-due-to-design-flaws
98. ^Anton Shilov. "Nvidia's Jensen Huang admits AI chip design flaw was '100% Nvidia's fault': TSMC not to blame, now-fixed Blackwell chips are in production." Tom's Hardware, 2024-10-23. tomshardware.com/...ckwell-chips-are-in-production
99. ^1 ^2Stephen Warwick. "Nvidia GB200 production ramps up after suppliers tackle AI server overheating and liquid cooling leaks." Tom's Hardware, 2025-05-28. tomshardware.com/...ating-and-liquid-cooling-leaks
100. ^"Nvidia server makers solve Blackwell technical issues, ramp up shipments of GB200 racks: report." DatacenterDynamics, May 2025. datacenterdynamics.com/...ts-of-gb200-racks-report
101. ^"Jensen Huang opens Computex with Vera Rubin in production and a move into Windows PCs." The Next Web, 2026-06-01. thenextweb.com/...jensen-huang-computex-2026-keynote
102. ^1 ^2 ^3 ^4"MLPerf Inference v5.0 Datacenter results." MLCommons, 2025-04-02. github.com/...inference_results_v5.0
103. ^1 ^2 ^3 ^4"MLPerf Inference v6.1 Datacenter results." MLCommons, 2026-09-16. github.com/...inference_results_v6.1
104. ^1 ^2 ^3"MLPerf Training v6.0 results." MLCommons, 2026-06-16. github.com/...training_results_v6.0
105. ^1 ^2"MLPerf Training v5.1 results." MLCommons, 2025-11-12. github.com/...training_results_v5.1
106. ^1 ^2"MLPerf Training v5.0 results." MLCommons, 2025-06-04. github.com/...training_results_v5.0
107. ^1 ^2"MLPerf Inference v5.1 Datacenter results." MLCommons, 2025-09-09. github.com/...inference_results_v5.1
108. ^1 ^2"MLPerf Inference v4.1 Datacenter results." MLCommons, 2024-08-28. github.com/...inference_results_v4.1
109. ^1 ^2 ^3"NVIDIA Blackwell Platform Sets New LLM Inference Records in MLPerf Inference v4.1." NVIDIA Technical Blog, 2024-08-28. developer.nvidia.com/...s-in-mlperf-inference-v4-1
110. ^1 ^2 ^3"NVIDIA Blackwell Ultra Datasheet." NVIDIA, October 2025. resources.nvidia.com/...blackwell-ultra-datasheet
111. ^1 ^2"Hardware." NVIDIA DGX GB Rack Scale Systems User Guide, accessed 2026-10-16. docs.nvidia.com/...hardware
112. ^1 ^2"Q3 Fiscal 2026 Earnings Call (corrected transcript)." NVIDIA, 2025-11-19. s201.q4cdn.com/...-19-November-2025-5_00-PM-ET.pdf
113. ^"CFO Commentary on Second Quarter Fiscal 2027 Results." NVIDIA, 2026-08-26. s201.q4cdn.com/...Q2FY27-CFO-Commentary.pdf
114. ^1 ^2 ^3"Q2 Fiscal 2027 Earnings Call (transcript)." NVIDIA, 2026-08-26. s201.q4cdn.com/...ll-26-August-2026-5_00-PM-ET.pdf
115. ^1 ^2"Alpha Compute Closes $32.2 Million Revenue Contract with AI Lab Customer." GlobeNewswire, 2026-05-12. globenewswire.com/...Contract-with-AI-Lab-Customer
116. ^1 ^2"Alpha Compute Corp. Provides Mid-Q2 2026 Update." GlobeNewswire, 2026-05-21. globenewswire.com/...p-provides-mid-q2-2026-update
117. ^Jonathan Vanian. "Mark Zuckerberg indicates Meta is spending billions on Nvidia AI chips." CNBC, 2024-01-18. cnbc.com/...s-spending-billions-on-nvidia-ai-chips