AI rack power by GPU generation

LIVING INDEX · NO. 2 IN THE SERIES

AI rack power by GPU generation

Living index of rack-scale power requirements across NVIDIA + AMD + adjacent AI accelerator generations. Categorical cooling requirement, 800V DC ecosystem readiness, integrator + power partners, per-row primary source. Quarterly cadence.

11Accelerator generations
7 kW → 500+ kWRack power range
QuarterlyUpdate cadence
v1.0Aug 2026

01Rack power comparison table

Note on undisclosed values. Where specific TDP or rack power is not publicly disclosed by the vendor, cell reads "Not publicly disclosed" rather than a fabricated estimate. Where an estimate is presented, it is attributed as such. NVIDIA Vera Rubin specifics have not been fully released as of August 2026; entries reflect what has been publicly announced.
Accelerator generation TDP per GPU Rack config Rack power Cooling requirement 800V DC readiness Integrator + power partners Primary source
NVIDIA H100 (Hopper)
Shipped 2022
700W SXM58-GPU HGX / 4-GPU/8-GPU DGX30-60 kW typical (HGX H100 8-GPU) / 10 kW (DGX H100 single node)Air (up to ~40 kW/rack); DLC required aboveShips in AC-powered racks; 800V DC racks introduced by ecosystem partners 2025+Foxconn, Wiwynn, Supermicro, Dell, HPE, Lenovo (multiple integrators)NVIDIA H100 product page
NVIDIA H200 (Hopper refresh)
Shipped 2024
700W SXM5 (same as H100)8-GPU HGX30-60 kW typical (HGX H200 8-GPU)Air-capable; DLC preferred at 8-GPU HGX densityShips in AC-powered racks; 800V DC via ecosystem partnersSame integrator ecosystem as H100NVIDIA H200 product page
NVIDIA B100 / B200 (Blackwell)
Shipped 2024-2025
700W (B100) / 1000W (B200 dense)8-GPU HGX / MGX modular60-120 kW (HGX B200 8-GPU)Liquid cooling standard (DLC or immersion); air-cooled variants limitedEcosystem transitioning to 800V DC rack designs for Blackwell densityFoxconn, Wiwynn, Quanta, Supermicro, Dell, Delta Electronics (integrators + power)NVIDIA HGX platform
NVIDIA GB200 NVL72 (Grace Blackwell)
Shipped 2025
2700W per GB200 superchip (2×B200 + Grace CPU)72-GPU NVL72 (full rack), 36-GPU NVL36 (subset)120 kW per NVL72 rack (public NVIDIA guidance)Liquid cooling required (DLC); air-cooled not a supported configurationNVIDIA reference design assumes 800V DC facility feed; OCP Mt Diablo rack fitsFoxconn, Wiwynn, Quanta, Delta (800V DC power shelf), Flex (800V DC rack), CoolIT (thermal)NVIDIA GB200 NVL72 + 800V DC ecosystem blog
NVIDIA Vera Rubin (announced)
Shipped 2026-2027 target
Not publicly disclosedNot publicly disclosed (successor to GB200 architecture)Estimated 500-600 kW per rack based on public commentary; not officially disclosedLiquid cooling required; NVIDIA guidance emphasises 800V DC + advanced cooling800V DC facility feed assumed; 1500V DC discussion emergingSame partner ecosystem as GB200; specific product roadmaps not publicly disclosedNVIDIA 800V HVDC architecture blog
AMD MI300X
Shipped 2023
750W per GPU8-GPU OAM Universal Baseboard30-60 kW typical (8-GPU UBB)Air-capable; DLC preferred at 8-GPU densityAC-powered racks primary; 800V DC via ecosystem partnersSupermicro, Dell, HPE, Lenovo (multiple integrators)AMD MI300X product page
AMD MI325X
Shipped 2024
1000W per GPU8-GPU OAM UBB60-90 kW typical (8-GPU UBB)DLC standard at 1000W densityAC-powered primary; 800V DC via ecosystem partnersSame integrator ecosystem as MI300XAMD MI325X product page
AMD MI350 series (announced)
Shipped 2025
Not publicly disclosedNot publicly disclosedNot publicly disclosedDLC + advanced cooling expectedNot publicly disclosedNot publicly disclosedAMD Instinct product line
Google TPU v5
Shipped 2023
Not publicly disclosed (Google internal)Google internal pod architectureNot publicly disclosed (Google internal)Liquid cooling (Google DC standard)Google internal DC architecture; not publicly detailed for TPUGoogle internal + ODM partners (Quanta et al)Google Cloud TPU
Cerebras WSE-3
Shipped 2024
~23 kW per wafer-scale engineCS-3 system (1 WSE-3 per system chassis)~23 kW per CS-3 systemLiquid cooling required (integrated Cerebras thermal)Standard AC input; not part of 800V DC ecosystemCerebras direct; system-level integrationCerebras CS-3 system page
Groq LPU
Shipped 2024
Not publicly disclosedGroqRack (server-scale)Not publicly disclosedAir-cooled variant available; liquid optionsStandard AC input; not part of 800V DC ecosystemGroq direct + system integratorsGroq product page

02What this tracks

What this map ISN'T

Not a benchmark. Not a performance comparison. Not investment advice.

Not exhaustive of all AI accelerators (Cerebras + Groq included as representative non-NVIDIA/non-AMD points; other silicon vendors deferred to next revision). Not a substitute for direct vendor engagement or published product datasheets. Cooling + 800V DC readiness columns reflect ECOSYSTEM CONFIGURATION rather than accelerator-inherent limits.

03Method + sourcing

How this index is built
  • Coverage. Includes accelerators with public product-page presence + verified generation. Each row cites a canonical vendor source.
  • TDP. Per-GPU thermal design power as publicly disclosed by the vendor. Blank or "not publicly disclosed" where vendor has not released.
  • Rack power. Reflects typical rack configurations (e.g. HGX 8-GPU, NVL72 full-rack). Range spans lower-density variants + higher-density configs.
  • Cooling requirement. Categorical: air-capable, DLC preferred, DLC required, liquid required. Reflects ecosystem-standard configurations, not necessarily accelerator-inherent thermal limits.
  • 800V DC readiness. Whether the accelerator generation ships in racks that assume 800V DC facility feed, OR is designed for AC-input via traditional PDUs. Ecosystem partners (Delta, Flex, CoolIT) determine actual deployment configurations more than the accelerator vendor.
  • Update cadence. Quarterly. Between-cycle updates when NVIDIA/AMD publish new accelerator generations or significant ecosystem announcements.

04Change log

DateVersionChange
2026-08-21v1.0Initial public release.
Definition A living index of accelerator generation rack power envelopes across the AI compute buildout, from the H100 baseline (~40 kW/rack typical) through Blackwell GB200 NVL72 (~120 kW/rack) and forward to Rubin and Rubin Ultra deployments where per-rack power crosses 500 kW. The table is refreshed as hyperscaler specifications and public disclosures update.