Question
What will the average on-demand rental price of an H100-equivalent GPU be on 2029-12-31, in 2026 US dollars per hour?
Baseline and Current Dynamics My median estimate is $1.75 per hour per H100-equivalent (H100e) on-demand in 2026 dollars, reflecting a substantial structural decline in unit compute costs offset by persistent constraints and sticky list prices. The middle half of my distribution runs from $1.20 to $2.70 per hour. As of late 2026, the on-demand price for an H100e hovers roughly between $3 and $4.50. Recent tracker data shows upward pressure, with the Silicon Data neocloud H100 index reaching $3.13 Silicon Data H100 Rental Price Index, Ornn at $3.29 H100 Price: $3.39/hr (Nov 12, 2026) | Rental Price Index, and medians from platforms like GetDeploying and AIMultiple near $3.90–$4.00 Nvidia H100 Rental Prices: Compare 63 Providers Cloud GPU Rental Price Index. The background research notes Uvik's median at $4.20 across 16 providers, alongside recent price hikes from AWS and Nebius. Because the forecast targets 2026 dollars, deflating nominal late-2029 prices by projected US CPI trims the final outcome by roughly 8–10%.
Supply Expansion and Generational Gains Substantial capacity build-outs and generational leaps in chip architecture will apply massive downward pressure on the normalized per-H100e price by 2029. By late 2029, the mainstream fleet will likely feature Rubin Ultra or Feynman architectures https://vrlatech.com/nvidia-gpu-roadmap-2026-2030/ Nvidia Vera Rubin NVL72 Full Production: What Changed, which could realistically deliver 5 to 15 times the training throughput of a baseline H100. Even if these advanced, multi-die systems command $10–$25 per hour, the effective cost per H100-equivalent of compute will fall significantly. These performance gains are backed by aggressive expansion in physical infrastructure, including Alphabet's expected 2027 capex surge and TSMC's planned scaling of N2 and CoWoS packaging https://investor.tsmc.com/english/encrypt/files/encrypt_file/reports/2026-10/eba4fac72873bf546b669bd2b5d973df5c93215d/TSMC%203Q26%20Transcript.pdf TSMC's CoWoS Capacity to Double by 2028—and Rivals .... Taken together, this massive supply expansion will likely outpace growing demand from inference workloads, making a substantial decline from current rental prices my central expectation.
Demand Shocks and Supply Bottlenecks Counteracting these cost declines are physical deployment bottlenecks and explosive compute demand. Translating fab output into usable capacity remains constrained by slow grid connections, equipment supply chains, and persistent high-bandwidth memory (HBM) shortages expected to last into 2030 Executive summary - Key Questions on Energy and AI SK Hynix says 2027 will be the 'worst year' for memory .... On the demand side, internal compute usage for inference is surging—OpenAI reports a 100-fold increase in internal coding inference over six months, while Anthropic notes a 250-fold jump in internal Claude use. If automated AI R&D capabilities compound over the next three years, willingness to pay could remain high enough to absorb expanding supply, pushing equilibrium rents closer to the marginal value of AI labor rather than just hardware depreciation curves. I regard that as an upside risk rather than the central case: inference growth should absorb substantial capacity, but is less likely to fully offset the combined expansion in physical supply and compute delivered per chip.
Measurement Nuance and Tail Risks The right tail (p90 of $4.50) is driven heavily by the resolution criteria: an arithmetic average across providers, rather than a volume-weighted transaction price, will be pulled up by notoriously sticky hyperscaler list prices, which historically run 1.7x to 6x higher than neocloud rates GPU Price Index 2026: NVIDIA H100, H200 and B200 Prices NVIDIA Blackwell & Vera Rubin GPU Price Hike 2027. Supply growth likely outpacing inference demand limits the weight I place on sustained scarcity pricing, even though sticky list prices leave substantial upside risk. Furthermore, relying on realized training throughput rather than peak FP4/FP8 FLOP marketing claims—such as NVIDIA’s MLPerf submission showing GB200 delivering roughly 2.2x the performance of H100 on pretraining NVIDIA Blackwell Delivers up to 2.6x Higher Performance ...—provides a conservative anchor for the normalization denominator. The left tail (p10 of $0.80), meanwhile, accounts for the risk of an AI financing bust or neocloud capacity glut, signaled by existing credit stress like high yields on CoreWeave debt and major players curbing neocloud partnerships Microsoft is renting less AI capacity, and the neoclouds that ... Nvidia now agrees to rent back unused GPU capacity from ....
Set against related questions, this estimate was slightly adjusted to reflect the massive supply expansion that will likely outpace the growing demand from inference workloads.
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited