The Information Machine
Updated todayDay 9in the edition since 26 Oct 20264 sources

GPU rental prices rising despite oversupply forecasts

The gist

Latest: GPU Rental Prices Surge as Lead Times Hit 45 Weeks

GPU rental cost directly determines which AI teams can train and run models at scale. The combination of rising rental rates, multi-year lead times, and large upfront payment requirements concentrates compute access among well-capitalized firms.

When stories like this move, I write about them in an email each morning.

One free email each morning, around 8 AM US Eastern.

The full picture

GPU rental prices have surged across all major NVIDIA chip generations while supply chain constraints push lead times well above historical norms. FastGPU data covering 81 provider instances over 60 days shows the RTX PRO 6000 up 307%, the B300 up 52.8%, the H200 up 33.3%, the B200 up 27.8%, and the H100 up 30%. Over a longer horizon, a16z data shows the H200 up about 160% and the B200 up about 120% since the start of 2026, while the H100 is up about 35% and the A100 is roughly flat.

On the supply side, TrendForce reports top-end GPU lead times jumped 5 weeks to 45 weeks as NVIDIA transitions to its Rubin architecture, with substrates and capacitors running at more than three times balanced lead times and delaying complete server builds. Hyperscalers are expected to receive Rubin GPUs first, extending enterprise customers' dependence on existing generations. Spheron reports H100 and H200 lead times of 36 to 52 weeks, with TSMC's CoWoS packaging capacity fully allocated through at least mid-2027 and HBM supply running short. Consumer GPU production cuts of 30 to 40% have added pressure across the NVIDIA stack.

IREN co-CEO Dan Roberts said on CNBC that IREN's 2026 GPU capacity is largely sold out, new three-year contracts are priced at more than double last year's rates, and customers are prepaying 45 to 55% of GPU costs upfront. Roberts said the bottleneck is getting chips online fast enough, not finding buyers. Nebius's CRO said the company could theoretically sell out its entire 2027 capacity immediately at current prices.

Evan Conrad warned that within six months most AI startups will not be able to buy compute, and cloud providers are requiring upfront payments as high as 100% of contract value from non-investment-grade customers. Dean Ball of OpenAI responded: "This seems right to me, unfortunately, and it is essentially a species of interest-rate rise (the opportunity cost of capital is increasing)." The gap between chip rental cost of roughly $6 per hour and inference revenue of roughly $30 per hour was cited as evidence that GPU prices remain below equilibrium.

AI infrastructure stocks including IREN, CRWV, and NBIS fell over 8% on fears of oversupply and financing costs even as underlying compute rental rates rose. Milk Road AI characterized this as "the market pricing a demand scare that the people renting out the GPUs are not seeing."

How it developed
2 November 2026in the 3 Nov edition

TrendForce data from November 2026 found top-end GPU lead times jumped to 45 weeks as NVIDIA transitions to its Rubin architecture, with substrates and capacitors running at more than three times balanced lead times delaying server builds.

A separate Spheron analysis found TSMC's CoWoS packaging capacity fully allocated through at least mid-2027, HBM supply from SK Hynix and Micron running short, and consumer GPU production cuts of 30 to 40% adding pressure across the stack. Both reports extend the supply constraints behind elevated GPU rental rates into next-generation hardware components.

29 October 2026in the 30 Oct edition

Cloud providers are now requiring upfront payments as high as 100% of contract value from non-investment-grade customers, and Evan Conrad warned that most AI startups will be unable to buy compute within six months.

Dean Ball of OpenAI characterized this as equivalent to an interest-rate rise, with the opportunity cost of capital increasing. A reported gap between chip rental cost (roughly $6 per hour) and inference revenue (roughly $30 per hour) suggests prices remain below market-clearing levels.

26 October 2026

Korean media reported MLCC lead times stretching to 61 weeks as AI-driven component shortages spread beyond GPUs and memory

First in the edition · 26 October 2026
25 October 2026in the 26 Oct edition

GPU rental prices have risen sharply in 2026, contradicting expectations that compute would oversupply and prices would fall.

FastGPU data across 81 providers over 60 days shows the Nvidia B300 up 52.8% and the RTX PRO 6000 up 307%. A16z data shows the H200 up roughly 160% and the B200 up about 120% since January. Nebius's CRO said the company could sell out its entire 2027 capacity immediately at current prices, and Ramp spending data shows over a third of AI budgets at top AI-spending firms goes to raw GPU compute rather than API calls or subscriptions.

23 October 2026

IREN co-CEO Dan Roberts said on CNBC that 2026 GPU capacity is largely sold out and new contracts are priced at more than double last year's rates

20 October 2026

FastGPU data showed GPU rental prices rose across 81 provider instances over 60 days, with RTX PRO 6000 up 307% and B300 up 52.8%

17 October 2026

Milk Road AI reported Amazon CEO Andy Jassy said compute shortage extends beyond 2028; Morgan Stanley estimated a 76.5 GW US power shortfall by 2029

30 April 2026

US federal prosecutors indicted Supermicro's Taiwan general manager and two others for allegedly diverting $2.5B in Nvidia Blackwell-class servers to China; B300 servers in China reach ~$1M, nearly double US prices

6 April 2026

Spheron published analysis putting H100/H200 lead times at 36-52 weeks and TSMC CoWoS capacity as fully allocated through at least mid-2027

Sources

Get the daily email

When stories like this move, I write about them in an email each morning.

One free email each morning, around 8 AM US Eastern.

See the latest edition · Privacy