H100 at $3.83 an hour: the GPU and colocation market this week

Market summary

GetDeploying's H100 on-demand median was $3.83 per GPU-hour in the week of October 26, 2026, across 39 providers, up 3.2% from the week before and up 13.3% over four weeks.¹ GetDeploying's B200 on-demand median was $7.99 per GPU-hour in the week of October 26, 2026, across 22 providers, unchanged from the week before and up 17.7% over four weeks.²

TrendForce (October 29, 2026): However, TrendForce notes that the new MacBook Pro’s premium starting price of US$2,899 for the 14-inch model, coupled with its initial availability alongside existing Mini-LED models, could temper near-term replacement demand.³

Xcel Energy rolls out $70B, 5-year capital plan focused on renewables. The spending plan includes 9.2 GW of renewables and 2.3 GW of storage.⁴

GPU pricing this week

GetDeploying's H100 on-demand median was $3.83 per GPU-hour in the week of October 26, 2026, across 39 providers, up 3.2% from the week before and up 13.3% over four weeks.¹ The H100 12-month reserved median was $3.21 per GPU-hour in the week of October 26, 2026.⁵

GetDeploying's H200 on-demand median was $4.88 per GPU-hour in the week of October 26, 2026, across 35 providers, down 2.4% from the week before and up 8.4% over four weeks.⁶ The H200 12-month reserved median was $3.68 per GPU-hour in the week of October 26, 2026.⁷

GetDeploying's B200 on-demand median was $7.99 per GPU-hour in the week of October 26, 2026, across 22 providers, unchanged from the week before and up 17.7% over four weeks.² The B200 12-month reserved median was $8.44 per GPU-hour in the week of October 26, 2026.⁸

GetDeploying's B300 on-demand median was $9.15 per GPU-hour in the week of October 26, 2026, across 17 providers, up 1.8% from the week before and up 6.1% over four weeks.⁹ The B300 12-month reserved median was $6.00 per GPU-hour in the week of October 26, 2026.¹⁰

GetDeploying's MI300X on-demand median was $2.92 per GPU-hour in the week of October 26, 2026, across 6 providers, down 33.9% from the week before and down 33.8% over four weeks.¹¹ The MI300X 12-month reserved median was $2.23 per GPU-hour in the week of October 26, 2026.¹²

As a second source: Price of Compute's on-demand median for H100 SXM was $3.99 per GPU-hour on November 2, 2026, across 9 providers.¹³ Price of Compute's on-demand median for H200 SXM was $6.28 per GPU-hour on November 2, 2026, across 3 providers.¹⁴ Price of Compute's on-demand median for B200 was $8.60 per GPU-hour on November 2, 2026, across 5 providers.¹⁵ Price of Compute's on-demand median for B300 was $6.60 per GPU-hour on November 2, 2026, across 3 providers.¹⁶

Colocation and power

Xcel Energy rolls out $70B, 5-year capital plan focused on renewables. The spending plan includes 9.2 GW of renewables and 2.3 GW of storage.⁴

AEP increases investment plan 30% on generation, transmission spending. American Electric Power’s contracted large loads jumped 22%, to 84 GW, mainly on a 10.4-GW data center in Ohio being developed by SB Energy, company officials said Wednesday.¹⁷

Chinese-linked solar developer loses appeal over Texas grid interconnection. Court of Appeals for the 5th Circuit said Texas has a “free lane” to enforce restrictions on foreign-owned companies’ access to the state’s electric grid in a ruling that could have implications for other states and market participants.¹⁸

Deals and announcements

What could 1,000x more reliable quantum computation unlock?¹⁹

Beyond the grid: the maintenance challenges putting the data center boom at risk.²⁰

Sponsored: One security strategy. Four different data center realities.²¹

Europe’s AI ambitions are hitting a power bottleneck.²²

Data center mooted in Glasgow by Scottish automotive firm.²³

Modular data center provider DXN loses neocloud customer contract.²⁴

What it means for buyers

My take: rental prices are still moving up, so if you have a renewal or a new cluster in the next quarter, get counter-quotes now and ask for multi-year terms with a price lock.¹

My take: memory and flash costs feed straight into server and GPU node prices. Budget for higher hardware quotes into next year.³

My take: power, not floor space, sets colocation timelines. Ask every facility for the energized date and committed kW per rack in writing.⁴