# NVIDIA GB300 NVL72

> Source: https://aiwiki.ai/wiki/nvidia_gb300_nvl72
> Updated: 2026-10-15
> Fact-checked: 2026-10-15
> Categories: AI Compute, AI Hardware, AI Infrastructure, Data Centers, NVIDIA
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "NVIDIA GB300 NVL72." aiwiki.ai, 15 Oct 2026. https://aiwiki.ai/wiki/nvidia_gb300_nvl72
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

The NVIDIA GB300 NVL72 is a liquid-cooled, rack-scale AI computing system that integrates 72 [NVIDIA Blackwell Ultra](https://aiwiki.ai/wiki/nvidia_blackwell_ultra) (B300) GPUs and 36 Arm-based [NVIDIA Grace](https://aiwiki.ai/wiki/nvidia_grace) CPUs into a single NVLink fabric, delivering 1.1 exaFLOPS (1,080 petaFLOPS) of dense FP4 compute and about 20 TB of [HBM3E](https://aiwiki.ai/wiki/hbm3e) memory per rack.[1][3] NVIDIA announced the platform on March 18, 2025 at its GTC conference, positioning the system for reasoning AI, agentic workloads, and test-time scaling inference.[1] It is the direct successor to the [NVIDIA GB200 NVL72](https://aiwiki.ai/wiki/nvidia_gb200_nvl72), and NVIDIA states it provides "1.5x more dense FP4 Tensor Core FLOPS and 2x higher attention performance compared to NVIDIA Blackwell GPUs" along with 1.5x larger HBM3E memory.[3]

NVIDIA rates the rack at 1,080 petaFLOPS of dense FP4 and 1,440 petaFLOPS of sparse FP4. The GB200 NVL72 carries the same 1,440 petaFLOPS sparse FP4 rating, so the GB300's FP4 gain is in dense math, from 720 to 1,080 petaFLOPS.[3][22] Against Hopper-based platforms, NVIDIA claims up to 50x higher AI factory output, combining 10x better user responsiveness (tokens per second per user) with 5x higher throughput per megawatt.[3] Dell shipped the first GB300 NVL72 to [CoreWeave](https://aiwiki.ai/wiki/coreweave) in July 2025, CoreWeave, Microsoft Azure, Google Cloud and AWS brought GB300 NVL72 capacity online between August and December 2025, and NVIDIA said in November 2025 that GB300 had overtaken GB200 and contributed roughly two-thirds of its Blackwell revenue.[8][11][28][29][9][40]

## How does the GB300 NVL72 differ from the GB200 NVL72?

The GB300 NVL72 shares the GB200 NVL72's overall topology (72 GPUs, 36 Grace CPUs, a single 130 TB/s NVLink domain, liquid-cooled compute trays) but upgrades the GPU from the standard Blackwell B200 to Blackwell Ultra (B300).[3][22] The differences are concentrated in the components that matter most for reasoning and inference: memory capacity, dense FP4 throughput, attention-layer compute, and scale-out networking.

| Difference | GB200 NVL72 (B200) | GB300 NVL72 (B300) |
|---|---|---|
| GPU | Blackwell B200 | Blackwell Ultra B300 |
| CUDA compute capability | 10.0 | 10.3 |
| HBM3E per GPU | 192 GB maximum, 186 GB as listed (8-Hi stacks) | 288 GB maximum, 279 GB as listed (12-Hi stacks) |
| Total GPU memory | 13.4 TB | ~20 TB |
| Dense FP4 per GPU | 10 PFLOPS | 15 PFLOPS (1.5x) |
| Sparse FP4 per GPU | 20 PFLOPS | 20 PFLOPS |
| Rack dense FP4 | 720 PFLOPS | 1,080 PFLOPS (1.1 exaFLOPS) |
| Rack sparse FP4 | 1,440 PFLOPS | 1,440 PFLOPS |
| Rack FP64 | 2,880 TFLOPS | 100 TFLOPS |
| Attention-layer (softmax) throughput | 1x | Up to 2x |
| GPU power (configurable) | Up to 1,200 W | Up to 1,400 W |
| Scale-out NIC per GPU | ConnectX-7 (400 Gb/s) | ConnectX-8 (800 Gb/s) |
| NVLink fabric | 130 TB/s | 130 TB/s |

Sources: NVIDIA GB300 NVL72 and GB200 NVL72 product pages, NVIDIA technical blog, CUDA compute capability table, DGX GB rack documentation, Oracle Cloud shape listings.[2][3][21][22][23][31]

The headline upgrades are therefore the 50% larger HBM3E capacity per GPU, 50% more dense FP4 FLOPS, doubled attention-layer throughput, and doubled scale-out network bandwidth per GPU. The NVLink scale-up fabric (130 TB/s) is unchanged between the two generations, and double-precision (FP64) throughput fell sharply.[3][14][21][22]

## Background: from Blackwell to Blackwell Ultra

NVIDIA's [Blackwell](https://aiwiki.ai/wiki/nvidia_blackwell) architecture (the B100/B200 generation) was announced in March 2024 as the successor to Hopper. The original Blackwell GPUs introduced a dual-die design, FP4 [tensor core](https://aiwiki.ai/wiki/tensor_core)s, fifth-generation NVLink, and a tight integration with the Grace CPU through a high-speed chip-to-chip interconnect.[4] The GB200 NVL72, built around the standard B200 GPU, became the company's flagship rack-scale product for [hyperscaler](https://aiwiki.ai/wiki/hyperscaler)s and other large buyers.[22]

Blackwell Ultra, designated the B300, is a refined version of the same core architecture on the same TSMC 4NP process. NVIDIA did not redesign the architecture from scratch, but the B300 is a different chip from the B200: NVIDIA's CUDA documentation lists it as compute capability 10.3, against 10.0 for the B200.[2][23] The company made targeted changes to the components most relevant to inference and reasoning workloads: the attention-layer special function units, the memory stack configuration, and the dense FP4 tensor core throughput.[2] The result is a GPU rated at up to 1,400 W, against up to 1,200 W for the B200 in GB200 systems, that delivers 50% more dense FP4 compute.[2]

[Jensen Huang](https://aiwiki.ai/wiki/jensen_huang), NVIDIA's CEO, framed the timing at GTC 2025, saying "AI has made a giant leap" and that "reasoning and agentic AI demand orders of magnitude more computing performance."[1] He added that Blackwell Ultra is "a single versatile platform that can easily and efficiently do pretraining, post-training and reasoning AI inference."[1] The GB300 NVL72 was designed specifically to serve that demand, with the open-source [Dynamo](https://aiwiki.ai/wiki/nvidia_dynamo) inference framework announced alongside it and NVIDIA AI Enterprise software, including [NVIDIA NIM](https://aiwiki.ai/wiki/nvidia_nim) microservices, providing the software counterpart.[1]

## What is the B300 chip architecture?

### Dual-die design and transistor count

Like its B200 predecessor, the Blackwell Ultra B300 is not a single monolithic die. Two reticle-limited GPU dies are connected through NVIDIA's NV-HBI (High-Bandwidth Interface), a custom die-to-die interconnect delivering 10 TB/s of internal bandwidth.[2] The two dies operate as a single logical GPU, sharing a unified memory address space and appearing to software as one accelerator.[2][5]

The combined package contains 208 billion transistors, identical to the standard Blackwell count, manufactured on TSMC's 4NP process, a custom member of TSMC's 4-nanometer-class family.[2][5] The 208 billion transistor count exceeds the 185 billion in AMD's competing Instinct MI355X by approximately 12%.[35]

Each B300 GPU contains:

- Up to 160 Streaming Multiprocessors (SMs) distributed across the two dies; NVIDIA notes that the available SM count varies by product[2]
- 20,480 CUDA cores (128 per SM)[2]
- 640 fifth-generation Tensor Cores (4 per SM)[2]
- 256 KB of Tensor Memory (TMEM) per SM[2]

### Tensor cores and precision support

The fifth-generation Tensor Cores in the B300 support FP8, FP6, and NVFP4 (four-bit floating point) precision.[2] FP4 inference was the headline addition in the original Blackwell generation, but Blackwell Ultra doubled the attention-layer compute specifically to address a known bottleneck in transformer inference.[2]

Transformer attention layers stress exponentials, divisions and other transcendental operations executed by each SM's Special Function Units (SFUs), and the [softmax](https://aiwiki.ai/wiki/softmax) step can become a bottleneck when models work through long contexts and long chains of reasoning. Blackwell Ultra doubles SFU throughput for the key instructions used in attention, to 10.7 trillion exponentials per second against 5 trillion for Blackwell and 4.5 trillion for Hopper, giving up to 2x faster attention-layer compute than Blackwell GPUs.[2]

Dense NVFP4 throughput per GPU is 15 petaFLOPS (20 petaFLOPS sparse), compared to 10 petaFLOPS dense (also 20 petaFLOPS sparse) on the B200, a 50% increase in dense math. FP8 throughput is 5 petaFLOPS dense (10 petaFLOPS sparse), the same as the B200.[2]

### Memory subsystem: HBM3E at 288 GB

The most significant change in the B300 relative to the B200 is memory capacity. The B300 uses eight 12-Hi HBM3E stacks instead of the 8-Hi stacks in the B200.[5] This allows up to 288 GB of HBM3E per GPU, compared to 192 GB on the B200, a 50% increase per chip; in the GB300 NVL72, the capacity listed per GPU is 279 GB.[2][5][31]

The memory interface consists of sixteen 512-bit controllers (8,192-bit total bus width) with a peak bandwidth of 8 TB/s per GPU.[2] This is the same bandwidth figure as the B200, because the taller 12-Hi stacks add capacity rather than per-pin speed. NVIDIA says the 288 GB capacity lets models of more than 300 billion parameters reside entirely on one GPU without offloading, which avoids inter-GPU traffic for some serving configurations.[2] Each B300 can also be split into Multi-Instance GPU partitions, for example two instances of 140 GB, four of 70 GB, or seven of 34 GB.[2]

Across the 72-GPU GB300 NVL72 rack, total HBM3E memory is about 20 TB, with up to 576 TB/s of aggregate GPU memory bandwidth.[3]

### Host connectivity

Each B300 GPU connects to its paired Grace CPU through NVLink-C2C, a coherent chip-to-chip interconnect running at 900 GB/s.[2] The CPU and GPU share a unified memory address space over this link, allowing CPU-side LPDDR5X memory to be addressed directly from GPU kernels without explicit data transfers.

For host system connectivity in server configurations without a Grace CPU, the B300 provides a PCIe Gen 6 x16 interface delivering 256 GB/s bidirectional bandwidth, double that of PCIe Gen 5.[2]

## GB300 Grace Blackwell Ultra Superchip

NVIDIA packages the B300 GPU with the Grace CPU in a component called the GB300 Grace Blackwell Ultra Superchip. Each GB300 superchip contains two B300 GPUs and one Grace CPU, connected through NVLink-C2C.[2][3] This three-chip package is the building block of the GB300 NVL72 rack; NVIDIA rates it at up to 30 petaFLOPS of dense NVFP4 compute (40 petaFLOPS sparse) with about 1 TB of unified memory combining HBM3E and LPDDR5X.[2]

The Grace CPU in each superchip is based on Arm's Neoverse V2 architecture, the same core used in the GB200, with 72 cores per CPU.[3][22] In the GB300 NVL72 rack, 36 superchips provide 36 Grace CPUs with a combined 2,592 Neoverse V2 cores.[3]

Grace CPU memory in the NVL72 system totals approximately 17 TB of LPDDR5X with up to 14 TB/s of bandwidth.[3] The CPU memory is available to GPU kernels through the coherent NVLink-C2C interface, giving each pair of B300 GPUs access to a combined pool of CPU DRAM and GPU HBM3E.[2]

## What is inside the GB300 NVL72 rack?

### Physical configuration

The GB300 NVL72 is a pre-integrated 48U MGX rack containing:[36][21]

- 18 compute trays, each with two Grace CPUs and four Blackwell Ultra GPUs (36 GB300 superchips in all)[21][36]
- 72 Blackwell Ultra GPUs and 36 NVIDIA Grace CPUs in total[3]
- 9 NVLink switch trays, each housing two NVSwitch chips[21]
- Four single-port 800 Gb/s ConnectX-8 SuperNICs per compute tray, one per GPU[21]
- BlueField-3 DPUs for storage and management networking (two per compute tray in NVIDIA's DGX version)[21]
- Liquid-cooling manifolds and cold plates, with coolant distribution units installed in the rack or the row[21][32]
- Power shelves of 33 kW each (up to eight in Lenovo's configuration)[21][36]

The rack is shipped as a complete, pre-validated unit by NVIDIA's manufacturing partners, including Dell, HPE, [Lenovo](https://aiwiki.ai/wiki/lenovo), Supermicro, GIGABYTE, and others.[1] Dell announced on July 3, 2025 that it had become the first to ship the GB300 NVL72, built from its PowerEdge XE9712 servers, to CoreWeave.[8]

### NVLink fabric and NVSwitch

All 72 GPUs in the rack are connected through NVIDIA's fifth-generation [NVLink](https://aiwiki.ai/wiki/nvlink) Switch chips into a single non-blocking NVLink fabric. Each B300 GPU has eighteen fifth-generation NVLink links of 100 GB/s each, for 1.8 TB/s of bidirectional bandwidth per GPU.[2]

The NVSwitch chips inside the rack aggregate these links into a fabric that delivers 130 TB/s of total GPU-to-GPU bandwidth within the NVL72 domain.[3] Every GPU can communicate directly with every other GPU at full NVLink speed. This topology makes the 72-GPU rack function as a single logical compute unit for large model inference and training jobs.[1]

Each NVLink Switch also has engines for NVIDIA SHARP (Scalable Hierarchical Aggregation and Reduction Protocol), enabling in-network reductions and multicast acceleration that reduce communication overhead for [distributed training](https://aiwiki.ai/wiki/distributed_training) and inference.[24]

### Networking (scale-out)

For connecting multiple GB300 NVL72 racks or integrating with storage and front-end systems, each GPU has 800 Gb/s of network bandwidth through its own ConnectX-8 SuperNIC, using either NVIDIA Quantum-X800 [InfiniBand](https://aiwiki.ai/wiki/infiniband) or NVIDIA Spectrum-X [Ethernet](https://aiwiki.ai/wiki/ethernet).[1][21]

The ConnectX-8 SuperNIC combines 800 Gb/s networking with a 48-lane PCIe Gen 6 switch in a single device, reducing the need for separate PCIe switch chips on the data path between GPU and network.[14]

BlueField-3 Data Processing Units (DPUs) handle multi-tenant networking, GPU compute elasticity, accelerated data access and cybersecurity functions.[1]

## What are the GB300 NVL72 specifications?

### B300 GPU specifications

| Specification | Value |
|---|---|
| Architecture | Blackwell Ultra |
| Process node | TSMC 4NP |
| Transistors | 208 billion |
| Die configuration | Dual-die (NV-HBI, 10 TB/s) |
| Streaming Multiprocessors | Up to 160 |
| [CUDA](https://aiwiki.ai/wiki/cuda) cores | 20,480 |
| [Tensor Core](https://aiwiki.ai/wiki/tensor_core)s | 640 (5th generation) |
| CUDA compute capability | 10.3 |
| Precision support | FP4, FP6, FP8, FP16, BF16, TF32, FP64 |
| HBM3E capacity | 288 GB maximum (279 GB as listed in GB300 NVL72) |
| HBM3E bandwidth | 8 TB/s |
| HBM3E stacks | 8 stacks of 12-Hi |
| Memory bus width | 8,192-bit |
| FP4 compute | 15 petaFLOPS dense (20 sparse) |
| FP8 compute | 5 petaFLOPS dense (10 sparse) |
| FP16/BF16 compute | 2.5 petaFLOPS dense (5 sparse) |
| Power | Up to 1,400 W (configurable) |
| NVLink bandwidth | 1.8 TB/s (bidirectional) |
| NVLink version | 5th generation |
| PCIe interface | Gen 6 x16 (256 GB/s bidirectional) |
| NVLink-C2C bandwidth | 900 GB/s |

Sources: NVIDIA technical blog, GB300 NVL72 product page, CUDA compute capability table, Oracle Cloud shape listings.[2][3][23][31] The FP16/BF16 figures are NVIDIA's rack ratings divided by 72; NVIDIA quotes that rate with sparsity, and the dense value is half.[3]

### GB300 NVL72 rack specifications

| Specification | Value |
|---|---|
| Total GPUs | 72 (Blackwell Ultra B300) |
| Total Grace CPUs | 36 |
| CPU cores | 2,592 (Arm Neoverse V2) |
| GPU memory (HBM3E) | ~20 TB |
| GPU memory bandwidth | ~576 TB/s |
| CPU memory (LPDDR5X) | ~17 TB |
| CPU memory bandwidth | ~14 TB/s |
| Total fast memory | ~37 TB |
| NVLink fabric bandwidth | 130 TB/s |
| Network bandwidth per GPU | 800 Gb/s (ConnectX-8) |
| FP4 peak compute (dense) | 1,080 PFLOPS (~1.1 exaFLOPS) |
| FP4 peak compute (sparse) | 1,440 PFLOPS |
| FP8/FP6 peak compute | 720 PFLOPS sparse (360 dense) |
| FP16/BF16 peak compute | 360 PFLOPS sparse (180 dense) |
| TF32 peak compute | 180 PFLOPS sparse (90 dense) |
| INT8 peak compute | 24 POPS sparse |
| FP32 | 6 PFLOPS |
| FP64 / FP64 Tensor Core | 100 TFLOPS |
| Rack power draw | ~120 kW (NVIDIA DGX GB rack documentation) |
| Cooling | Liquid-cooled compute trays; networking and storage components air-cooled |
| Form factor | 48U MGX rack |
| Networking options | Quantum-X800 [InfiniBand](https://aiwiki.ai/wiki/infiniband) / Spectrum-X [Ethernet](https://aiwiki.ai/wiki/ethernet) |

Sources: NVIDIA GB300 NVL72 product page, whose footnotes state that Tensor Core figures are with sparsity unless marked otherwise; NVIDIA DGX GB rack user guide; Lenovo GB300 NVL72 datasheet.[3][21][36]

## How does the GB300 compare to earlier GPUs?

| Metric | H100 SXM5 | H200 SXM5 | B200 (in GB200 NVL72) | B300 (in GB300 NVL72) |
|---|---|---|---|---|
| Architecture | Hopper | Hopper | Blackwell | Blackwell Ultra |
| Process node | TSMC 4N | TSMC 4N | TSMC 4NP | TSMC 4NP |
| Transistors | 80B | 80B | 208B | 208B |
| HBM capacity | 80 GB | 141 GB | 192 GB maximum (186 GB listed) | 288 GB maximum (279 GB listed) |
| HBM bandwidth | 3.35 TB/s | 4.8 TB/s | 8 TB/s | 8 TB/s |
| FP8 dense compute | 1.98 PFLOPS | 1.98 PFLOPS | 5 PFLOPS | 5 PFLOPS |
| FP4 dense compute | N/A | N/A | 10 PFLOPS | 15 PFLOPS |
| Softmax exponentials per second | 4.5 trillion | 4.5 trillion | 5 trillion | 10.7 trillion |
| Power | Up to 700 W | Up to 700 W | Up to 1,200 W | Up to 1,400 W |

Sources: NVIDIA technical blog and product pages.[2][3][22] In eight-GPU HGX servers the B200 and B300 run at lower power limits and lower rated throughput than in the NVL72 racks; NVIDIA rates an eight-GPU HGX B200 board at 72 petaFLOPS of sparse FP8, or 4.5 petaFLOPS dense per GPU.[41]

The GB300 NVL72 as a full rack system versus Hopper-based platforms, according to NVIDIA:[3]

| Workload metric | GB300 NVL72 vs Hopper (NVIDIA claims) |
|---|---|
| AI factory output | Up to 50x higher |
| Throughput per megawatt | 5x higher |
| Video generation | 30x faster |
| [Tokens per second](https://aiwiki.ai/wiki/tokens_per_second) per user | 10x improvement |

## NVLink, NVSwitch, and networking details

NVLink 5 is the interconnect generation shipping with both Blackwell and Blackwell Ultra GPUs. Each GPU has 18 NVLink links of 100 GB/s each, for a total of 1.8 TB/s bidirectional per GPU.[2] Compared to NVLink 4 (used in Hopper), NVLink 5 doubles per-GPU bandwidth.[2]

Within the NVL72 rack, the NVLink Switch chips aggregate all GPU links. The NVL72 domain runs at 130 TB/s total GPU-to-GPU bandwidth, non-blocking.[2][3] NVLink Switch also includes SHARP engines for in-network reductions, which reduce the amount of data that must traverse the fabric during collective operations like all-reduce, commonly used in distributed training.[24]

For scale-out networking beyond a single rack, the ConnectX-8 SuperNIC provides 800 Gb/s per GPU.[1] This is double the networking bandwidth available on GB200 NVL72 systems, whose compute trays carry 400 Gb/s ConnectX-7 NICs, one per GPU.[21]

Multi-rack GB300 NVL72 deployments connect racks over Quantum-X800 InfiniBand or Spectrum-X Ethernet fabrics.[3] [Microsoft Azure](https://aiwiki.ai/wiki/azure)'s first at-scale deployment, announced on October 9, 2025, connected more than 4,600 Blackwell Ultra GPUs through Quantum-X800 InfiniBand for [OpenAI](https://aiwiki.ai/wiki/openai) workloads.[28][10]

## Why does the GB300 NVL72 require liquid cooling?

NVIDIA describes the GB300 NVL72 as a fully liquid-cooled, rack-scale architecture.[3] In NVIDIA's DGX GB racks, the compute trays are cooled by coolant running through manifolds to cold plates on the CPUs and GPUs, while networking and storage components are air-cooled by fans; NVIDIA puts rack power consumption at approximately 120 kW.[21] Lenovo describes its GB300 NVL72 as a hybrid design that relies mainly on warm-water liquid cooling, with supplemental airflow for the remaining heat.[36]

Each rack connects to a facility water loop through a Coolant Distribution Unit (CDU), installed in the rack or in the row.[32] Lenovo's datasheet specifies warm-water operation at 18 to 45 degrees C.[36] At 120 kW, a single GB300 NVL72 rack generates about 410,000 BTU/hr of heat.

The cooling system bill of materials for one NVL72 rack totals approximately $49,860, according to a Morgan Stanley estimate reported in November 2025. This includes cooling hardware across all 18 compute trays (about $2,260 per tray, $40,680 in all) and the 9 NVSwitch trays (about $1,020 per tray, $9,180 in all).[12] This cooling hardware cost is separate from the rack system price itself.

Data centers deploying GB300 NVL72 racks at scale need to plan power distribution at the row level, not per rack: at about 120 kW each, a row of 10 GB300 NVL72 racks draws about 1.2 megawatts.[21] Planning estimates published before NVIDIA's final specifications ran higher; Sunbird projected in July 2025 that GB300 racks could need up to about 150 kW.[18]

NVIDIA has cited liquid cooling as a significant efficiency improvement. In April 2025 it said the liquid-cooled GB200 NVL72 delivers "300x more water efficiency than traditional air-cooled architectures," and said liquid cooling lets facilities operate with warmer water, reducing or eliminating the need for mechanical chillers in many climates; it gave no separate figure for the GB300 NVL72.[20]

## What software runs on the GB300 NVL72?

### CUDA and cuDNN

The B300 GPU is fully [CUDA](https://aiwiki.ai/wiki/cuda)-compatible. NVIDIA's CUDA documentation lists the B300 and GB300 as compute capability 10.3, a separate target from the B200 and GB200 (10.0) and from Hopper (9.0).[23] NVIDIA says Blackwell Ultra is supported by its full-stack AI platform, which includes deep learning libraries such as [cuDNN](https://aiwiki.ai/wiki/cudnn) and cuBLAS and the NCCL collective communication library.[28]

### NVIDIA Dynamo

NVIDIA Dynamo is an open-source inference framework announced alongside the GB300 NVL72 at GTC 2025.[1] It is designed for distributed, disaggregated inference serving on large [GPU cluster](https://aiwiki.ai/wiki/gpu_cluster)s.[15] Dynamo splits the two phases of LLM inference, prefill (context processing) and decode (token generation), across separate pools of GPUs. This [disaggregated serving](https://aiwiki.ai/wiki/disaggregated_serving) approach allows each phase to be independently scaled and optimized.[1][15]

In NVIDIA's MLPerf Inference v5.1 submissions, disaggregated serving on the GB200 NVL72 delivered nearly 1.5x the per-GPU throughput of aggregated serving on a DGX B200 for the Llama 3.1 405B interactive scenario; NVIDIA noted that Dynamo provides disaggregated serving support.[6] For Mixture-of-Experts models like [DeepSeek-R1](https://aiwiki.ai/wiki/deepseek_r1), NVIDIA says the combination of GB300 NVL72 hardware and Dynamo improves throughput by up to 50x compared to Hopper-based systems.[15]

Dynamo supports major LLM serving frameworks as backends, including [NVIDIA TensorRT-LLM](https://aiwiki.ai/wiki/tensorrt_llm), [vLLM](https://aiwiki.ai/wiki/vllm), and [SGLang](https://aiwiki.ai/wiki/sglang).[15]

### TensorRT-LLM

NVIDIA TensorRT-LLM is the primary compiler and runtime for optimized LLM inference on NVIDIA hardware. In NVIDIA's first MLPerf submissions for the GB300 NVL72, most DeepSeek-R1 weights were quantized to NVFP4 with the TensorRT Model Optimizer library, and the key-value cache was quantized from BF16 to FP8 to reduce its memory footprint.[6]

### NCCL and collective communications

NVIDIA's NCCL (NVIDIA Collective Communications Library) handles multi-GPU and multi-node communication. On the GB300 NVL72, NCCL operations within the 72-GPU NVLink domain run over the 130 TB/s NVLink fabric.[3] Cross-rack NCCL operations use the ConnectX-8 network at 800 Gb/s per GPU.[1] NCCL is aware of the NVLink topology and routes intra-rack collectives through the NVSwitch fabric rather than the network interface.

## What is the GB300 NVL72 used for?

### Reasoning AI and test-time scaling

The primary design goal of the GB300 NVL72 is reasoning AI inference. Models like OpenAI's o1/o3, DeepSeek-R1, and similar "chain-of-thought" reasoning systems generate many more tokens per query than conventional LLMs, working through internal reasoning steps before answering. NVIDIA says this reasoning process can require over 100x the compute of a single inference pass on a traditional LLM for challenging queries.[37]

This token explosion makes memory capacity and memory bandwidth the primary bottlenecks, not raw FLOPS. The B300's 288 GB of HBM3E addresses capacity: more memory means larger [KV cache](https://aiwiki.ai/wiki/kv_cache)s for longer contexts, and the doubled attention throughput means the attention computation itself is less of a bottleneck in the decode phase.[2]

Test-time scaling, which NVIDIA describes as "the art of applying more compute during inference to improve accuracy," is the workload NVIDIA built Blackwell Ultra around.[1] The GB300 NVL72's exaFLOP-scale compute capacity is intended to make [test-time compute](https://aiwiki.ai/wiki/test_time_compute) economically viable at production scale.

### Large model training

The GB300 NVL72 is also used for training and fine-tuning [large language models](https://aiwiki.ai/wiki/large_language_model). The 130 TB/s NVLink fabric allows large [model parallelism](https://aiwiki.ai/wiki/model_parallelism) within a single rack.[3] For trillion-parameter model training, multiple racks are connected via InfiniBand or Ethernet.

In MLPerf Training v5.1 (November 2025), NVIDIA said the GB300 NVL72 delivered more than 4x the Llama 3.1 405B pretraining performance and nearly 5x the Llama 2 70B LoRA fine-tuning performance of Hopper with the same number of GPUs.[25] In the same round, Lambda's 72-GPU GB300 NVL72 cluster finished the Llama 2 70B LoRA fine-tuning benchmark in 1.26 minutes, 27% faster than a 72-GPU GB200 NVL72 result from the previous round.[19] In MLPerf Training v6.0 (June 2026), NVIDIA said the GB300 NVL72 trained up to 1.6x faster than the GB200 NVL72 at the same scale.[26] The increased per-GPU compute and memory capacity allows larger [batch size](https://aiwiki.ai/wiki/batch_size)s and reduces the communication overhead relative to compute time.

### Agentic AI workloads

[Agentic AI](https://aiwiki.ai/wiki/agentic_ai) systems use LLMs to plan and execute multi-step tasks, often calling external tools, running code, browsing web content, or invoking specialized models. These workloads require both high throughput (to handle many concurrent agent instances) and low latency (to minimize response time per step). The GB300 NVL72's per-GPU memory capacity, FP4 throughput, and attention acceleration are aimed at both requirements.

### Physical AI and video generation

NVIDIA also targets [physical AI](https://aiwiki.ai/wiki/physical_ai) applications including robotics simulation and [synthetic data](https://aiwiki.ai/wiki/synthetic_data) generation. For video generation using [diffusion model](https://aiwiki.ai/wiki/diffusion_model)s, NVIDIA claims 30x faster generation on GB300 NVL72 systems compared to Hopper: it says a single five-second video sequence processes 4 million tokens and takes nearly 90 seconds to generate on Hopper GPUs.[3] Faster generation supports synthetic dataset creation for [autonomous vehicle](https://aiwiki.ai/wiki/autonomous_vehicle) training and robotic manipulation.

## How fast is the GB300 NVL72 in MLPerf benchmarks?

The GB300 NVL72 made its official benchmark debut in MLPerf Inference v5.1, published on September 9, 2025. Key results, as reported by NVIDIA:[6]

**DeepSeek-R1 671B (per GPU):**
- GB300 NVL72: 5,842 tokens/second offline and 2,907 tokens/second in the server scenario[6]
- GB200 NVL72: 4,024 tokens/second offline and 2,327 in the server scenario[6]
- Improvement: 45% offline and 25% server per GPU over GB200 NVL72[6]
- Improvement vs Hopper: about 5x per GPU, against unverified results on a DGX H200 system (1,253 tokens/second offline)[6]

**Llama 3.1 405B:** NVIDIA's per-GPU records in the round were 224 tokens/second offline, 170 in the server scenario and 138 in the new interactive scenario; NVIDIA listed these as platform records rather than as GB300 NVL72 comparisons.[6]

The DeepSeek-R1 submissions quantized most of the model's weights to NVFP4 and its key-value cache to FP8.[6]

| Round | Results published | Headline GB300 NVL72 results |
|---|---|---|
| MLPerf Inference v5.1 | 2025-09-09 | Debut; DeepSeek-R1 per-GPU throughput 45% (offline) and 25% (server) above GB200 NVL72 [6] |
| MLPerf Training v5.1 | 2025-11-12 | Debut; more than 4x Hopper on Llama 3.1 405B pretraining at equal GPU count [25] |
| MLPerf Training v6.0 | 2026-06-16 | Up to 1.6x faster than GB200 NVL72 at the same scale; CoreWeave trained DeepSeek-V3 671B in 2.02 minutes on 8,192 GB300 GPUs [26] |
| MLPerf Inference v6.1 | 2026-09-16 | A 288-GPU, four-rack submission reached 99% scaling efficiency; Vera Rubin NVL72 previewed at up to 3.7x the GB300 NVL72 [27] |

In MLPerf Training v5.1, NVIDIA said its platform posted the fastest time to train on all seven benchmarks. Its record 10-minute Llama 3.1 405B result came from more than 5,000 Blackwell GPUs, which NVIDIA reported separately from its Blackwell Ultra results.[25]

## Who is buying and deploying the GB300 NVL72?

### ODM and server manufacturers

System makers named at GTC 2025 as delivering Blackwell Ultra servers include:[1]

- Cisco
- Dell Technologies (PowerEdge XE9712)[8]
- Hewlett Packard Enterprise
- Lenovo
- Supermicro
- Aivres
- ASRock Rack
- ASUS
- Eviden
- Foxconn
- GIGABYTE
- Inventec
- Pegatron
- QCT
- Wistron
- Wiwynn

### Cloud providers

The four largest cloud providers were named at GTC 2025 among the first to offer Blackwell Ultra instances:[1]

**[Amazon Web Services](https://aiwiki.ai/wiki/amazon_web_services):** AWS announced the general availability of Amazon EC2 P6e-GB300 UltraServers on December 2, 2025, at AWS re:Invent. They provide 1.5x the GPU memory and 1.5x the FP4 compute (without sparsity) of the prior P6e-GB200 UltraServers.[9] AWS made HGX B300-based EC2 P6-B300 instances generally available for single-node workloads on November 18, 2025, initially in US West (Oregon).[30]

**Microsoft Azure:** On October 9, 2025, Azure announced the NDv6 GB300 VM series and the first supercomputing-scale production cluster of GB300 NVL72 systems, with more than 4,600 Blackwell Ultra GPUs connected through Quantum-X800 InfiniBand for OpenAI's inference workloads.[28][10]

**[Google Cloud](https://aiwiki.ai/wiki/google_cloud):** Named at GTC 2025 among the first cloud providers to offer Blackwell Ultra.[1] On October 29, 2025, Google said its GB300 NVL72-based A4X Max instances were shipping in production, with customer access through a preview.[29]

**Oracle Cloud Infrastructure:** Named at GTC 2025 among the first cloud providers to offer Blackwell Ultra instances.[1] OCI lists a BM.GPU.GB300.4 bare-metal shape with four B300 GPUs of 279 GB from an NVL72 rack.[31]

**CoreWeave:** The first cloud provider to bring GB300 NVL72 instances into production. CoreWeave made GB300 NVL72-powered instances available on August 19, 2025, initially in the US-WEST-01A availability zone.[11] Dell had shipped CoreWeave's first GB300 NVL72 in July 2025.[8]

Additional GPU cloud providers named at GTC 2025 include Lambda, [Nebius](https://aiwiki.ai/wiki/nebius), Nscale, Crusoe, Yotta, and YTL.[1]

## What does the GB300 NVL72 cost, and when was it available?

NVIDIA has not published official list prices for the GB300 NVL72, which is sold through partners, and estimates vary widely. Press reports put a GB300 NVL72 rack at $6 million to $6.5 million, against $2.8 million to $3.4 million for a GB200 NVL72. In August 2026, Data Gravity reported purchase orders for small quantities at just under $5.0 million per GB300 NVL72 rack without an in-rack CDU, and about $32,000 more with a 250 kW CDU.[32]

Dated cloud prices give another reference. In October 2026, Oracle Cloud Infrastructure listed GB300 capacity at $18.00 per GPU-hour on a pay-as-you-go basis, against $16.00 for GB200.[31] AWS did not list a public rate for P6e-GB300 UltraServers; its general availability announcement directed customers to their sales representatives.[9]

Initial shipments began in July 2025 (Dell to CoreWeave).[8] NVIDIA said on its November 19, 2025 earnings call that GB300 had crossed over GB200 and contributed roughly two-thirds of total Blackwell revenue in its third fiscal quarter.[40] AWS P6e-GB300 instances became generally available in December 2025.[9]

For conventional servers, NVIDIA also sells the eight-GPU HGX B300 board, launched as the HGX B300 NVL16, which brings Blackwell Ultra to workloads that do not require the full rack-scale NVLink domain.[1][13]

## How does the GB300 compare with AMD Instinct MI355X?

| Specification | NVIDIA B300 (GB300 NVL72) | AMD Instinct MI355X |
|---|---|---|
| Architecture | Blackwell Ultra | CDNA 4 |
| Process node | TSMC 4NP | TSMC 3 nm (compute), 6 nm (I/O) |
| Transistors | 208 billion | 185 billion |
| HBM capacity | 288 GB HBM3E (279 GB listed) | 288 GB HBM3E |
| Memory bandwidth | 8 TB/s | 8 TB/s |
| FP4 compute | 15 PFLOPS dense (20 sparse) | 10.1 PFLOPS dense (MXFP4; AMD lists no sparse rate) |
| FP8 compute | 5 PFLOPS dense (10 sparse) | 5 PFLOPS dense (10.1 sparse) |
| FP16/BF16 compute | 2.5 PFLOPS dense (5 sparse) | 2.5 PFLOPS dense (5 sparse) |
| Power | Up to 1,400 W | 1,400 W typical board power |
| Scale-up interconnect | NVLink 5 (1.8 TB/s) | 7 Infinity Fabric links at 153 GB/s each |
| Rack-scale domain | 72 GPUs / 130 TB/s | No equivalent in this generation |
| Software ecosystem | [CUDA](https://aiwiki.ai/wiki/cuda), [TensorRT](https://aiwiki.ai/wiki/tensorrt), Dynamo | [ROCm](https://aiwiki.ai/wiki/rocm), MIOpen |

Sources: NVIDIA technical blog and GB300 NVL72 page; AMD Instinct MI355X specifications.[2][3][35]

On paper, the B300 leads the MI355X in dense FP4 compute (15 versus 10.1 petaFLOPS), while the two chips match on memory capacity, bandwidth, FP8 and FP16 throughput and board power.[2][35] In real-world inference, [SemiAnalysis](https://aiwiki.ai/wiki/semianalysis)'s [InferenceX](https://aiwiki.ai/wiki/inferencex) v2 analysis (February 2026) found that NVIDIA's B300 and GB300 NVL72 dominate when advanced techniques like disaggregated prefill, wide expert parallelism, and FP4 precision are used together. The MI355X was competitive when only a subset of these optimizations was enabled but fell behind when they were combined, which SemiAnalysis attributed to a lack of kernel and collective optimization in the ROCm software stack.[33]

The GB300 NVL72's primary competitive advantage is not per-GPU FLOPS but rather the 130 TB/s NVLink fabric connecting all 72 GPUs.[3] For very large model inference (400B+ parameters across a whole rack) and for agentic workloads requiring rapid KV cache sharing, the unified NVLink domain provides practical throughput advantages that peak FLOPS comparisons do not capture.

AMD disputes how such comparisons are framed. In March 2026, it argued that when the two run in the same operating modes, the MI355X "often delivers comparable or better results than GB300 NVL72," and said that on FP8 the MI355X delivered lower cost per token than the GB300 at a high-interactivity operating point (60 or more tokens per second per user).[34]

## What comes after the GB300 NVL72?

NVIDIA announced the Vera Rubin platform as the successor to Blackwell Ultra at GTC 2025, for arrival in the second half of 2026.[7][17] At CES on January 5, 2026, Jensen Huang said Vera Rubin was in full production; the Vera Rubin NVL72 pairs 72 Rubin GPUs with 36 Vera CPUs and follows the same rack-scale approach as the GB300 NVL72.[17][16]

Announced characteristics of Vera Rubin NVL72:

- Sixth-generation NVLink: NVIDIA's March 2026 technical blog gave 260 TB/s of rack scale-up bandwidth (3.6 TB/s per GPU), while its product page in October 2026 listed 216 TB/s (3 TB/s per GPU)[16][38]
- HBM4 memory replacing HBM3E: 288 GB per Rubin GPU at 19.2 TB/s, more than twice the B300's 8 TB/s, and 20.7 TB per rack[38]
- The Vera CPU replacing Grace, built from 88 custom NVIDIA Olympus Arm cores rather than Neoverse cores[17][38]
- A Rubin GPU with about 1.6 times Blackwell's transistor count, rated at 50 petaFLOPS of sparse NVFP4 inference and 35 petaFLOPS of dense NVFP4 training per GPU[17][38]
- Rack ratings of 3,600 petaFLOPS of sparse NVFP4 inference (2.5 times the GB300 NVL72's 1,440 sparse) and 2,520 petaFLOPS of dense NVFP4 training (2.3 times its 1,080 dense)[38][3]

NVIDIA said on August 26, 2026 that Vera Rubin was ramping into full production, with racks running at partners including CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure and Nebius.[39] In MLPerf Inference v6.1 in September 2026, NVIDIA's preview submission put the Vera Rubin NVL72 at up to 3.7x the GB300 NVL72's throughput on Qwen3-VL and up to 2.5x on DeepSeek-R1.[27]

NVIDIA has also outlined a longer roadmap built around Rubin Ultra: a Kyber rack that doubles the NVLink domain to 144 GPUs per rack, and a Vera Rubin Ultra NVL576 system that combines eight racks of 72 Rubin Ultra GPUs. SemiAnalysis reported at GTC 2025 that the Kyber rack was expected to launch with Rubin Ultra in 2027.[16][7]

## Power and infrastructure requirements

The GB300 NVL72 represents a significant step up in data center infrastructure requirements compared to prior GPU generations. Key infrastructure considerations:

**Power density:** NVIDIA's DGX GB rack documentation puts rack power consumption at approximately 120 kW, with each B300 GPU configurable up to 1,400 W alongside the Grace CPUs and switching equipment.[21][2]

**Liquid cooling infrastructure:** The compute trays' GPUs and CPUs are cooled by liquid through cold plates, while some networking and storage components remain air-cooled.[21] Data centers must provide a facility water loop with sufficient flow rate and heat rejection capacity; Lenovo specifies warm-water operation at 18 to 45 degrees C.[36]

**Power distribution:** At about 120 kW per rack, a row of 10 GB300 NVL72 racks draws about 1.2 MW.[21] The racks are fed by 33 kW power shelves, so power feeds must be redundant and capacity-planned at the row level, and floor loading must accommodate the weight of liquid-cooled racks.[21][36]

**Power smoothing:** GB300 systems use multiple power-shelf configurations to handle synchronous GPU load ramps, and NVIDIA says its power smoothing features, including energy storage and burn mechanisms, help stabilize power draw across training workloads.[2]

## See also

- [NVIDIA GB200 NVL72](https://aiwiki.ai/wiki/nvidia_gb200_nvl72)
- [NVIDIA Blackwell](https://aiwiki.ai/wiki/nvidia_blackwell)
- [NVIDIA H100](https://aiwiki.ai/wiki/nvidia_h100)
- [NVIDIA H200](https://aiwiki.ai/wiki/nvidia_h200)
- [NVLink](https://aiwiki.ai/wiki/nvlink)
- [Tensor Processing Unit](https://aiwiki.ai/wiki/tpu)
- [AMD Instinct MI300X](https://aiwiki.ai/wiki/amd_instinct_mi300x)
- [Large language model](https://aiwiki.ai/wiki/large_language_model)
- [Transformer architecture](https://aiwiki.ai/wiki/transformers)
- [Test-time compute](https://aiwiki.ai/wiki/test_time_compute)

## References

1. NVIDIA Newsroom. "NVIDIA Blackwell Ultra AI Factory Platform Paves Way for Age of AI Reasoning." March 18, 2025. https://nvidianews.nvidia.com/news/nvidia-blackwell-ultra-ai-factory-platform-paves-way-for-age-of-ai-reasoning
2. NVIDIA Technical Blog. "Inside NVIDIA Blackwell Ultra: The Chip Powering the AI Factory Era." August 22, 2025 (updated September 24, 2025). https://developer.nvidia.com/blog/inside-nvidia-blackwell-ultra-the-chip-powering-the-ai-factory-era/
3. NVIDIA. "GB300 NVL72 Product Page." Accessed October 2026. https://www.nvidia.com/en-us/data-center/gb300-nvl72/
4. NVIDIA. "Blackwell Architecture Overview." https://www.nvidia.com/en-us/data-center/technologies/blackwell-architecture/
5. TweakTown. "NVIDIA details Blackwell Ultra GB300: dual-die design, 208B transistors, up to 288GB HBM3E." August 26, 2025. https://www.tweaktown.com/news/107373/nvidia-details-blackwell-ultra-gb300-dual-die-design-208b-transistors-up-to-288gb-hbm3e/index.html
6. NVIDIA Technical Blog. "NVIDIA Blackwell Ultra Sets New Inference Records in MLPerf Debut." September 9, 2025. https://developer.nvidia.com/blog/nvidia-blackwell-ultra-sets-new-inference-records-in-mlperf-debut/
7. SemiAnalysis. "NVIDIA GTC 2025: Built For Reasoning, Vera Rubin, Kyber, CPO, Dynamo Inference, Jensen Math, Feynman." March 19, 2025. https://newsletter.semianalysis.com/p/nvidia-gtc-2025-built-for-reasoning-vera-rubin-kyber-cpo-dynamo-inference-jensen-math-feynman
8. Dell Technologies. "Dell Delivers Market's First NVIDIA GB300 NVL72 to CoreWeave." July 3, 2025. https://www.dell.com/en-us/blog/dell-delivers-market-s-first-nvidia-gb300-nvl72-to-coreweave/
9. AWS. "Amazon EC2 P6e-GB300 UltraServers accelerated by NVIDIA GB300 NVL72 are now generally available." December 2, 2025. https://aws.amazon.com/about-aws/whats-new/2025/12/amazon-ec2-p6e-gb300-ultraservers-nvidia-gb300-nvl72-generally-available/
10. Microsoft Azure Blog. "Microsoft Azure delivers the first large-scale cluster with NVIDIA GB300 NVL72 for OpenAI workloads." October 9, 2025. https://azure.microsoft.com/en-us/blog/microsoft-azure-delivers-the-first-large-scale-cluster-with-nvidia-gb300-nvl72-for-openai-workloads/
11. CoreWeave. "GB300 NVL72-Powered Cloud Instances Changelog." August 19, 2025. https://docs.coreweave.com/docs/changelog/release-notes/gb300-nvl72
12. Tom's Hardware. "Cooling system for a single Nvidia Blackwell Ultra NVL72 rack costs a staggering $50,000." November 5, 2025. https://www.tomshardware.com/pc-components/cooling/cooling-system-for-a-single-nvidia-blackwell-ultra-nvl72-rack-costs-a-staggering-usd50-000-set-to-increase-to-usd56-000-with-next-generation-nvl144-racks
13. ServeTheHome. "The NVIDIA HGX B300 NVL16 is Massively Different." 2025. https://www.servethehome.com/the-nvidia-hgx-b300-nvl16-is-massively-different/
14. NVIDIA Technical Blog. "NVIDIA ConnectX-8 SuperNICs Advance AI Platform Architecture with PCIe Gen6 Connectivity." May 18, 2025. https://developer.nvidia.com/blog/nvidia-connectx-8-supernics-advance-ai-platform-architecture-with-pcie-gen6-connectivity/
15. NVIDIA Developer. "Dynamo Inference Framework." https://developer.nvidia.com/dynamo
16. NVIDIA Technical Blog. "NVIDIA Vera Rubin POD: Seven Chips, Five Rack-Scale Systems, One AI Supercomputer." March 16, 2026. https://developer.nvidia.com/blog/nvidia-vera-rubin-pod-seven-chips-five-rack-scale-systems-one-ai-supercomputer/
17. Tom's Hardware. "Nvidia CEO confirms Vera Rubin NVL72 is now in production." January 6, 2026. https://www.tomshardware.com/tech-industry/artificial-intelligence/nvidia-ceo-confirms-vera-rubin-nvl72-is-now-in-production-jensen-huang-uses-ces-keynote-to-announce-the-milestone
18. Sunbird DCIM. "How Much Power Does a NVIDIA GB300 NVL72 Need?" July 10, 2025. https://www.sunbirddcim.com/blog/how-much-power-does-nvidia-gb300-nvl72-need
19. Lambda AI. "MLPerf Training v5.1: Lambda's NVIDIA GB300 NVL72 outperforms NVIDIA GB200 NVL72 by 27%." November 12, 2025. https://lambda.ai/blog/lambda-mlperf-training-benchmarks-v5.1
20. NVIDIA Blog. "NVIDIA Blackwell Platform Boosts Water Efficiency by Over 300x." April 22, 2025. https://blogs.nvidia.com/blog/blackwell-platform-water-efficiency-liquid-cooling-data-centers-ai-factories/
21. NVIDIA. "Hardware." NVIDIA DGX GB Rack Scale Systems User Guide. Accessed October 2026. https://docs.nvidia.com/dgx/dgxgb200-user-guide/hardware.html
22. NVIDIA. "GB200 NVL72 Product Page." Accessed October 2026. https://www.nvidia.com/en-us/data-center/gb200-nvl72/
23. NVIDIA Developer. "CUDA GPU Compute Capability." Accessed October 2026. https://developer.nvidia.com/cuda-gpus
24. NVIDIA. "NVLink and NVLink Switch." Accessed October 2026. https://www.nvidia.com/en-us/data-center/nvlink/
25. NVIDIA Blog. "NVIDIA Wins Every MLPerf Training v5.1 Benchmark." November 12, 2025. https://blogs.nvidia.com/blog/mlperf-training-benchmark-blackwell-ultra/
26. NVIDIA Blog. "Fastest, Largest, Strongest: NVIDIA Blackwell Sweeps MLPerf Training 6.0." June 16, 2026. https://blogs.nvidia.com/blog/blackwell-mlperf-training-6-0/
27. NVIDIA Blog. "NVIDIA Vera Rubin NVL72 Delivers Leading Performance in MLPerf Inference v6.1 Debut." September 16, 2026. https://blogs.nvidia.com/blog/vera-rubin-nvl72-mlperf-inference/
28. NVIDIA Blog. "Microsoft Azure Unveils World's First NVIDIA GB300 NVL72 Supercomputing Cluster for OpenAI." October 9, 2025. https://blogs.nvidia.com/blog/microsoft-azure-worlds-first-gb300-nvl72-supercomputing-cluster-openai/
29. Google Cloud Blog. "Expanding our NVIDIA partnership: Now shipping A4X Max, Vertex AI Training, and more." October 29, 2025. https://cloud.google.com/blog/products/compute/now-shipping-a4x-max-vertex-ai-training-and-more
30. AWS. "Amazon EC2 P6-B300 instances with NVIDIA Blackwell Ultra GPUs are now available." November 18, 2025. https://aws.amazon.com/about-aws/whats-new/2025/11/amazon-ec2-p6-b300-instances-nvidia-blackwell-ultra-gpus-available/
31. Oracle Cloud Infrastructure. "Compute Pricing: GPU." Accessed October 2026. https://www.oracle.com/cloud/compute/gpu/pricing/
32. Data Gravity. "How Much Does an NVIDIA NVL72 Cost?" August 31, 2026. https://www.datagravity.dev/p/how-much-does-an-nvidia-nvl72-cost
33. SemiAnalysis. "InferenceX v2: NVIDIA Blackwell Vs AMD vs Hopper." February 16, 2026. https://newsletter.semianalysis.com/p/inferencex-v2-nvidia-blackwell-vs
34. AMD. "The Many Aspects of Inference Performance." March 18, 2026. https://www.amd.com/en/developer/resources/technical-articles/2026/the-many-aspects-of-inference-performance.html
35. AMD. "AMD Instinct MI355X GPUs." Accessed October 2026. https://www.amd.com/en/products/accelerators/instinct/mi350/mi355x.html
36. Lenovo Press. "Lenovo NVIDIA GB300 NVL72 Datasheet." March 16, 2026. https://lenovopress.lenovo.com/DS0207
37. NVIDIA Blog. "How Scaling Laws Drive Smarter, More Powerful AI." February 12, 2025. https://blogs.nvidia.com/blog/ai-scaling-laws/
38. NVIDIA. "Vera Rubin NVL72 Product Page." Accessed October 2026. https://www.nvidia.com/en-us/data-center/vera-rubin-nvl72/
39. NVIDIA Newsroom. "NVIDIA Announces Financial Results for Second Quarter Fiscal 2027." August 26, 2026. https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-second-quarter-fiscal-2027
40. MarketBeat. "NVIDIA Q3 2026 Earnings Call Transcript." November 19, 2025. https://www.marketbeat.com/earnings/reports/2025-11-19-nvidia-co-stock/
41. NVIDIA. "NVIDIA HGX Platform." Accessed October 2026. https://www.nvidia.com/en-us/data-center/hgx/