Skip to main content

    GPU Colocation: Power-Ready Racks for H100/B200 Clusters

    You own the GPUs. We find the rack: 40 kW air-cooled to 120 kW liquid-cooled positions in Dallas, Ashburn, the Pacific Northwest, the Bay Area, and secondary US markets. One request, quotes from multiple facilities within one business day.

    • $120–$350 per kW per month across the primary US markets, per published 2026 rate guides. See pricing by market
    • Liquid-cooled and rear-door heat exchanger racks for DGX H100, DGX B200, and GB200 NVL72
    • InfiniBand and 400G Ethernet cross-connects, contiguous rack positions
    • No fee to you; the facility pays GYGO

    Deploying H100s specifically? H100 colocation requirements

    Get quotes from multiple facilities within one business day

    Tell us your power requirement and region. We match it against partner facilities and send you comparable quotes, free of charge.

    No fee to you. Quotes come from the facilities; you contract directly with the one you choose.

    Colocation for GPU clusters vs standard colocation

    Standard enterprise colocation quotes 5–10 kW per rack and puts your cabinets wherever space is free. Uptime Institute’s 2025 global survey puts the average modal rack density, excluding the few high-density sites, at 7.5 kW, right in the middle of that band. A training cluster breaks both assumptions.

    A single DGX H100 pulls about 10.2 kW and a DGX B200 about 14.3 kW. Four H100 systems fill a 40 kW air-cooled rack. A GB200 NVL72 needs roughly 120 kW in one rack and has no air-cooled option, so a building without direct-to-chip liquid cooling can’t take the deployment at all.

    Placement matters as much as power. InfiniBand and 400G Ethernet links between racks have to stay inside cable-length limits, so the facility has to hand you contiguous rack positions instead of whatever is free across two halls.

    Most buildings can’t do either. In the same Uptime survey, 82% of operators said their densest rack sits below 30 kW, so a 120 kW liquid-cooled deployment rules out the large majority of the market before you start negotiating. That is why we start from your kW and your topology rather than from a list of buildings.

    What GYGO brokers

    Colocation is the core. If your plan also involves renting, buying, or monetizing GPUs, the same team handles it.

    Place (colocation)

    Rack space, power, cooling, and cross-connects for GPU clusters you own. This page.

    Rent

    Compare cloud and bare-metal GPU providers by workload, latency, or cost.

    Buy

    Source H100, H200, and B200 systems through value-added resellers.

    Invest

    Turn idle GPU hardware into revenue through leasing and compute programs.

    Power and density: air vs liquid

    The rack you need depends on the systems you are landing. Most facilities quote per kW, so the density decides both the building and the price.

    Air-cooled / rear-door
    up to 40 kW
    • About four DGX H100 systems (10.2 kW each) or two to three DGX B200 (14.3 kW each) per rack
    • Widely available in Dallas, Ashburn, and Pacific Northwest facilities
    • Indicative $4,800–$14,000 per rack per month at $120–$350/kW
    Direct-to-chip liquid
    100–120 kW
    • Required for GB200 NVL72 (about 120 kW per rack) and dense B200 deployments
    • Scarcer inventory; expect longer lead times. Liquid-ready 20–40 kW racks publish at $130–$400/kW including the premium
    • Indicative $14,400–$42,000 per rack per month at the air-cooled $120–$350/kW; no guide prices 100 kW+ separately

    GPU colocation cost by market

    Air-cooled $/kW/month (colocation MRC) as printed in published 2026 rate guides. Full table with liquid-cooled ranges, hidden costs, sources and contract terms on the GPU colocation pricing page.

    Dallas–Fort Worth
    $140–$240 /kW/mo

    Deep high-density inventory, competitive power, central US latency.

    Northern Virginia (Ashburn)
    $180–$300 /kW/mo

    The largest data center market in the world (CBRE, Q1 2026); capacity is tight and priced accordingly.

    Pacific Northwest
    $120–$200 /kW/mo

    Low-cost hydro power; strong value for training clusters that tolerate distance.

    Bay Area / Silicon Valley
    $200–$350 /kW/mo

    Premium for proximity to AI teams; limited liquid-cooled capacity.

    Secondary US markets
    $78–$170 /kW/mo

    Colorado and the Dakotas as published examples; longer lead times for 100 kW+ racks.

    Ranges as printed in published 2026 rate guides, read 13 October 2026: QuoteColo (AI colocation pricing, 23 March 2026; GPU colocation and liquid-cooled colocation guides, 2026), Inflect (1 April 2026), CBRE (26 February 2026) and Digital Realty's Q2 2026 filing (23 July 2026). Indicative only; the facility's quote governs.

    How it works

    1. 1

      Tell us your kW, region, and timeline

      One form. Power requirement, preferred region, and when the hardware lands. GPU model and interconnect help but are optional.

    2. 2

      We collect quotes from matching facilities

      We check partner data centers with the density and cooling you need and return comparable quotes within one business day.

    3. 3

      You compare and sign with the facility

      Pick the quote that fits. You contract directly with the operator. GYGO is paid by the facility, not by you.

    Request quotes

    GPU colocation FAQ

    Published 2026 rate guides put air-cooled GPU colocation at $120–$350 per kW per month across the four primary US markets, depending on market, rack density, and contract term; that is the colocation MRC, and cross-connects, remote hands and power billing typically add 25–30% on top. It puts a 40 kW air-cooled rack at roughly $4,800–$14,000 per month before those add-ons. Liquid-cooled 20–40 kW racks publish at $130–$400 per kW including the premium. See the pricing page for ranges by market and the sources.

    Air-cooled racks with rear-door heat exchangers top out around 40 kW. Direct-to-chip liquid cooling supports 100–120 kW per rack, which is what a full GB200 NVL72 or a dense B200 deployment needs. A single DGX H100 draws about 10.2 kW and a DGX B200 about 14.3 kW at full load.

    Standard enterprise racks are provisioned at 5–10 kW. GPU cluster colocation needs 40 kW+ per rack, liquid or rear-door cooling, InfiniBand or 400G Ethernet cross-connects between racks, and contiguous rack positions so the fabric stays within cable-length limits. Not every facility can do this, which is why we start from your kW and topology rather than from a list of buildings.

    You submit power, region, and timeline. We match the requirement against partner facilities with available capacity, collect comparable quotes, and send them to you within one business day. You contract directly with the facility you choose. GYGO is compensated by the facility, so there is no fee to you.

    Many high-density facilities quote from a single 40 kW rack, which fits roughly four DGX H100 systems, and QuoteColo's 2026 guide reports that most Tier 1 operators now gate small deployments behind 50–100 kW+ minimums, so a single rack usually lands with a mid-tier or regional facility. Below a rack, managed hosting or bare-metal rental is usually cheaper than a colocation contract. We can advise on either.

    It depends on the market and the cooling. QuoteColo's 2026 NVIDIA colocation guide puts air-cooled H100/H200 in power-slotted facilities at 2–6 weeks, liquid-cooled racks in prime markets (Ashburn, Santa Clara) at 3–6 months, and GB200 NVL72 racks at 3–6 months minimum, with lead times assuming hardware is in hand; its GPU colocation guide reports a 6–12 month queue in Ashburn. CBRE's H1 2026 North America report says ready-for-service fit-out dates are expanding from weeks to several months, and beyond six months where an air-cooled hall has to be converted to liquid cooling. Tell us your cluster delivery date up front and we quote lead times from the facilities themselves.
    GPU colocation pricing by market
    $/kW table, hidden costs, minimum terms.
    H100 colocation requirements
    Power, cooling, and interconnect for DGX H100 racks.
    Colocation vs cloud ROI
    When owning and colocating beats renting.
    All FAQs
    Rental, purchase, colocation, and investment questions.