GPU Power Consumption Trend From 700 W Chips to 142 kW Racks

NVIDIA’s August 2025 technical blog on Blackwell Ultra lists the Blackwell GPU at “Up to 1,200W” of maximum power, while the DGX B200 user guide lists 1,000 W as the maximum power per GPU in its power-cable table, and neither document mentions the other number. That gap is the first warning that a GPU power consumption trend read from chip ratings alone will mislead a cooling plan. This piece traces how the doubling of per-chip ratings since Hopper turned into rack loads more than three times larger. It uses rated and reference-design figures only and leaves out measured draw under training jobs, which none of these documents publish.

One Blackwell GPU carries a 1,000 W and a 1,200 W rating

The two numbers answer different questions. The blog’s comparison table labels its row “Max power (TGP)” and sets Hopper, Blackwell and Blackwell Ultra side by side as chips. The DGX B200 user guide gives a limit for one product, reported by the nvidia-smi tool, in a table of what happens when AC power cords are pulled.

That table lists 1,000 W per GPU with five of the six power supplies running, a state the guide describes as having no effect on the system. With three or four supplies left, the limit falls to 800 W.

Source: NVIDIA, Inside NVIDIA Blackwell Ultra (August 2025); NVIDIA DGX B200 User Guide, Power Specifications.

The reading that fits both documents is that “Up to” in the blog describes the envelope of the silicon across the products NVIDIA builds from it, and the DGX figure is one product’s setting inside that envelope. This is an inference. Neither document says which Blackwell product runs at the full 1,200 W, so a single Blackwell TDP figure does not exist in NVIDIA’s own documents; there is a ceiling, and there are system settings.

AMD’s product pages show the same mechanism with a cleaner pair. The MI350X and MI355X pages both list 256 compute units and 288 GB of memory, the same CDNA 4 architecture and the same launch date. They differ in peak engine clock and in Typical Board Power (TBP).

Power ratio = 1,400 W ÷ 1,000 W = 1.40.

Clock ratio = 2,400 MHz ÷ 2,200 MHz = 1.09.

A 9% higher clock comes with 40% more board power. The part number therefore does not fix the heat load. The configuration does, and the per-GPU limit belongs in the design record next to the model name.

H100 power showed the same looseness at 700 W. Lenovo’s product guide lists the H100 80GB SXM board at 700 W “Max Power Consumption” and lists a water-cooled feature code and an air-cooled one that the guide marks as withdrawn, so even at 700 W the rating alone did not decide between air and liquid.

Table 1. Published per-GPU power ratings, sorted by value (W)

* Labels differ by source; TGP, TBP, per-GPU limit and max power consumption are not one quantity.
* DGX B200 value with five of six power supplies; 800 W with three or four.

Source: NVIDIA, “Inside NVIDIA Blackwell Ultra” (2025-08-22); NVIDIA, “DGX B200 User Guide” (2026-09-08); AMD Instinct MI300X, MI325X, MI350X and MI355X product pages; Lenovo Press, “ThinkSystem NVIDIA H100 PCIe Gen5 GPUs” LP1732 (2026-02-19)

Eight-GPU boxes held airflow between 103 and 108 cfm per kilowatt

The eight-GPU DGX guides publish the same fields for each generation. The DGX H100/H200 guide gives 10.2 kW max, and the DGX B200 guide gives 14.3 kW max.

The DGX B300 guide is less tidy. It lists power consumption of 14.5 kW in one table and “15 kW system max.” in the next.

Source: NVIDIA DGX H100/H200, DGX B200 and DGX B300 User Guides, Introduction chapters.

The GPU share of each box follows from the per-GPU figures:

DGX H100: 8 × 700 W = 5,600 W, and 5,600 W ÷ 10,200 W = 55%.

DGX B200: 8 × 1,000 W = 8,000 W, and 8,000 W ÷ 14,300 W = 56%.

DGX B300 at the blog’s ceiling: 8 × 1,400 W = 11,200 W, and 11,200 W ÷ 14,500 W = 77%.

The B300 line is the one to distrust. At that ceiling, everything else in the box would have to fit in 3.3 kW, against 6.3 kW in the DGX B200. The B300 guide states no per-GPU limit, so a 1,400 W GPU should not be assumed for this air-cooled system.

Airflow tells a steadier story. Each guide lists an airflow figure, so the mean exhaust temperature rise can be estimated from ΔT = P ÷ (ρ × cp × V). The assumptions are air density of 1.2 kg/m³, specific heat of 1,005 J/(kg·K), and all rated power leaving in the listed airflow, with 1 cfm taken as 0.000471947 m³/s.

DGX H100: 1,105 cfm = 0.5215 m³/s, so ΔT = 10,200 W ÷ (1.2 × 1,005 × 0.5215) = 16.2 K.

DGX B200: 1,550 cfm = 0.7315 m³/s, so ΔT = 14,300 W ÷ (1.2 × 1,005 × 0.7315) = 16.2 K.

DGX B300: 1,500 cfm = 0.7079 m³/s, so ΔT = 14,500 W ÷ (1.2 × 1,005 × 0.7079) = 17.0 K.

These are estimates at rated power, not measurements, and the fan settings differ: the H100 airflow is quoted at 80% fan PWM and the B300 airflow at 70%. Within those limits, each kilowatt still needs about the same air. The box drew 42% more power from DGX H100 to DGX B300, and the airflow grew with it.

Table 2. Eight-GPU DGX systems, sorted by rated system power (kW, cfm, K, %)

* B300 GPU share assumes the 1,400 W chip ceiling; the DGX B300 guide states no per-GPU limit.
* cfm per kW and air rise calculated: 1.2 kg/m³, 1,005 J/(kg·K), all rated power in listed airflow.
* DGX B300 airflow is the power-supply version; the busbar version lists 1,350 cfm.
* DGX B300 guide also lists 15 kW system max; DGX H100 guide lists heat output of 38,557 BTU/hr, about 11.3 kW.

Source: NVIDIA, “DGX H100/H200 User Guide” (2026-01-26), “DGX B200 User Guide” (2026-09-08), “DGX B300 User Guide” (2026-10-05); calculations by this article

The strain appears one level up. Four DGX B200 systems in one rack would add up to 57.2 kW and 6,200 cfm.

The ASHRAE TC 9.9 white paper on liquid cooling states that a 40 to 50 kW rack could need up to 5000 cfm, and that a best-in-class floor tile delivers 1900 cfm.

That puts the four-box rack at 3.3 tiles of air, the point where the floor tile, rather than the server, sets the air limit.

Why did rack power grow 3.5 times when the chip only doubled?

The answer is GPU count per rack. NVIDIA’s blog describes the GB300 NVL72 as a liquid-cooled rack of 36 Grace Blackwell Superchips, each pairing one Grace CPU with two Blackwell Ultra GPUs, so one rack holds 72 GPUs. Vertiv’s reference architecture for this platform supports densities up to 142kW per rack.

Source: NVIDIA, Inside NVIDIA Blackwell Ultra; Vertiv press release, June 11, 2025.

Taking four DGX H100 systems per rack as the Hopper baseline, the growth splits into three factors:

Rack ratio = 142 kW ÷ (4 × 10.2 kW) = 142 kW ÷ 40.8 kW = 3.48.

Chip factor = 1,400 W ÷ 700 W = 2.0.

GPUs per rack factor = 72 ÷ 32 = 2.25.

Overhead factor = (142,000 W ÷ 100,800 W of GPU ceiling) ÷ (10,200 W ÷ 5,600 W) = 1.41 ÷ 1.82 = 0.77.

Check: 2.0 × 2.25 × 0.773 = 3.48, using the unrounded overhead factor.

This compares ratings and design densities, an upper bound rather than a measurement. GPU packing contributed slightly more to the rack figure than the chip rating did, and non-GPU power per GPU watt fell rather than rose.

Per GPU, the rack is not a different power regime from the box:

GB300 rack: 142,000 W ÷ 72 = 1,972 W per GPU. DGX B300: 14,500 W ÷ 8 = 1,813 W per GPU.

What separates an air-cooled box from a liquid-cooled rack is how many of those slots share one footprint. NVIDIA’s reference architecture for the previous generation points the same way: each scalable unit of eight DGX GB200 rack systems requires a Thermal Design Power (TDP) of 1.2 MW, a ceiling of 150 kW per rack when divided evenly.

Source: NVIDIA DGX SuperPOD Reference Architecture (DGX GB200), DGX SuperPOD Architecture page.

Not all of that rack heat goes to liquid. Lenovo’s GB300 NVL72 page lists 18kW of air-cooled power and 118kW of liquid-cooled power, and states that 85% of the rack is liquid cooled.

Liquid share from the two listed loads = 118 kW ÷ 136 kW = 86.8%, which does not match the stated 85%. Lenovo does not define the basis of the 85%.

Mass flow for the liquid share = 118,000 W ÷ (4,180 J/(kg·K) × 10 K) = 2.82 kg/s, about 169 L/min of water at 1 kg/L.

The 10 K rise is an assumption, and glycol mixtures need more flow for the same heat. How the loop’s temperature rise sets usable capacity matters as much as the rack rating.

The residual air is not small either. At the 16.2 K rise estimated for the DGX boxes:

Volume flow for the air share = 18,000 W ÷ (1.2 × 1,005 × 16.2) = 0.92 m³/s = 1,952 cfm.

That is one best-in-class floor tile per rack for the part of the load that is still air.

GPUs per rack, not watts per GPU, pushed cooling onto liquid

The doubling of chip ratings from Hopper to Blackwell Ultra was absorbed inside the eight-GPU box without changing how much air each kilowatt needed. What changed the cooling method was the decision to put 72 GPUs in one rack behind one NVLink domain, which multiplied GPUs per rack by 2.25 before any change in silicon.

Uptime Institute’s 2026 survey reports that the average of modal rack densities surpassed 11 kW for the first time. A 142 kW design density is the load of about 13 of those average racks in one position, and even the residual air in Lenovo’s design exceeds that average for a whole rack.

Liquid-first designs still plan for a large air share. Vertiv’s AI Reference Design #027 gives its cooling topology as 77% liquid (direct-to-chip) and 23% air, so a plan built on chip ratings alone would size the liquid loop and miss the air plant that still carries close to a quarter of that hall. The figure that belongs at the start of a design is the per-rack load split into liquid and air parts, taken from the integrator’s documents for the exact configuration, with the per-GPU limit recorded beside it. Chip ratings track the GPU power consumption trend well and size a room badly, and any real design built on these figures needs review by a qualified engineer against the vendor’s site documentation.

Whether a 1,400 W rating is a 1,400 W heat load

Does a cold plate see the rated watts all the time?

The labels say no. NVIDIA writes “Up to”, AMD lists the MI300X as “750W Peak”, and the DGX B200 guide lowers the per-GPU limit to 800 W when supplies drop out. NVIDIA’s blog adds that GB300 systems use power smoothing with energy storage to stabilize draw across training workloads. None of the documents publish a sustained average under a real job, which leaves the margin between rating and draw unmeasured.

Can a Blackwell Ultra system still run on air?

The DGX B300 guide lists an airflow of 1,500 cfm and an inlet air temperature range, so the eight-GPU box is specified as an air-cooled system. The difficulty sits at the rack, where four of them reach 58 kW. That is above the 40 to 50 kW case in the ASHRAE white paper. Whether a given room can deliver that air is a site question for qualified engineering review.

Did AMD’s parts follow the same climb?

AMD’s line went from the MI300X at 750W Peak to the MI355X at 1400W over about 18 months of launch dates, the same endpoint as Blackwell Ultra. The MI355X page lists its cooling as “Passive & Active” against “Passive OAM” for the MI350X. The page does not define these terms, so they say nothing certain about liquid.

References

• Inside NVIDIA Blackwell Ultra: The Chip Powering the AI Factory Era / NVIDIA Technical Blog / 2025-08-22, updated 2025-09-24 / https://developer.nvidia.com/blog/inside-nvidia-blackwell-ultra-the-chip-powering-the-ai-factory-era/
• Introduction to NVIDIA DGX H100/H200 Systems, NVIDIA DGX H100/H200 User Guide / NVIDIA / last updated 2026-01-26 / https://docs.nvidia.com/dgx/dgxh100-user-guide/introduction-to-dgxh100.html
• Introduction to NVIDIA DGX B200 Systems, NVIDIA DGX B200 User Guide / NVIDIA / last updated 2026-09-08 / https://docs.nvidia.com/dgx/dgxb200-user-guide/introduction-to-dgxb200.html
• Introduction to NVIDIA DGX B300 System, NVIDIA DGX B300 User Guide / NVIDIA / last updated 2026-10-05 / https://docs.nvidia.com/dgx/dgxb300-user-guide/introduction-to-dgxb300.html
• DGX SuperPOD Architecture, NVIDIA DGX SuperPOD Reference Architecture featuring DGX GB200 / NVIDIA / last updated 2025-11-19 / https://docs.nvidia.com/dgx-superpod/reference-architecture-scalable-infrastructure-gb200/latest/dgx-superpod-architecture.html
• AMD Instinct MI300X Accelerators / AMD / product page, launch date 2023-12-06 / https://www.amd.com/en/products/accelerators/instinct/mi300/mi300x.html
• AMD Instinct MI325X Accelerators / AMD / product page, launch date 2024-10-10 / https://www.amd.com/en/products/accelerators/instinct/mi300/mi325x.html
• AMD Instinct MI350X GPUs / AMD / product page, launch date 2025-06-12 / https://www.amd.com/en/products/accelerators/instinct/mi350/mi350x.html
• AMD Instinct MI355X GPUs / AMD / product page, launch date 2025-06-12 / https://www.amd.com/en/products/accelerators/instinct/mi350/mi355x.html
• ThinkSystem NVIDIA H100 PCIe Gen5 GPUs, Product Guide LP1732 / Lenovo Press / updated 2026-02-19 / https://lenovopress.lenovo.com/lp1732-thinksystem-nvidia-h100-pcie-gen5-gpu
• Lenovo NVIDIA GB300 NVL72 / Lenovo Press / undated product tour / https://lenovopress.lenovo.com/3dtours/gb300nvl72
• Vertiv develops energy-efficient cooling and power reference architecture for the NVIDIA GB300 NVL72 platform / Vertiv / 2025-06-11 / https://www.vertiv.com/en-us/about/news-and-events/corporate-news/vertiv-develops-energy-efficient-cooling-and-power-reference-architecture-for-the-nvidia-gb300-nvl72/
• Vertiv 360AI High Density Reference Design for AI, AI Reference Design #027 / Vertiv / 2025 / https://www.vertiv.com/493f45/globalassets/documents/brochures/vertiv-reference-design-027-ds-en-na-2025-sl-80148-web.pdf
• Emergence and Expansion of Liquid Cooling in Mainstream Data Centers / ASHRAE Technical Committee 9.9 / 2021 / https://www.ashrae.org/File%20Library/Technical%20Resources/Bookstore/Emergence-and-Expansion-of-Liquid-Cooling-in-Mainstream-Data-Centers_WP.pdf
• Uptime Institute Global Data Center Survey 2026, executive summary / Uptime Institute / July 2026 / https://uptimeinstitute.com/uptime_assets/77a6191bef6c85cb4ac9bfb169ddc313cae517030b48709697b1f80bc10edc90-uptime-institute-global-data-center-survey-results-2026.pdf