AI Compute Cost Outlook: Quarterly Briefing

AI compute is entering a two-speed market: surplus Hopper capacity is driving sharp rental deflation, while Blackwell and HBM-constrained systems retain scarcity premiums. Over the next several quarters, aggregate cost per useful token should fall, but frontier training and tightly coupled inference will remain expensive because networking, memory, power, and deployment scarcity—not GPU silicon alone—set the marginal cost.

Hopper deflation, Blackwell premium

H100 rental pricing has compressed to roughly $1.49–$3.50 per GPU-hour in specialist and spot markets, versus as high as $10–$12 at hyperscaler on-demand rates.^[1][3][7] This is a structural repricing of depreciating Hopper fleets as new capacity comes online and utilization becomes more uneven.

Blackwell pricing remains materially firmer. B200 capacity generally trades around $4.50–$7.00 per GPU-hour in specialist channels, with hyperscaler rates near $14 in some configurations.^[1][3] B300 preemptible pricing has recently been listed at $9.50 per GPU-hour, while Nebius raised several on-demand GPU rates by approximately 17–21% in October.^[4][9]

The implication is a widening distinction between nominal accelerator price and effective compute cost. Blackwell’s higher throughput and memory bandwidth can justify the premium for frontier training and high-volume inference, but not for price-sensitive batch workloads.

Capex declines; infrastructure remains the bottleneck

Street pricing for individual H100s is approximately $25,000–$40,000, while B200 pricing is broadly $30,000–$50,000; complete eight-GPU systems range from roughly $250,000 for H100 to $500,000 for B200-class configurations.^[1][2] These prices are unlikely to fall rapidly while HBM, advanced packaging, optical networking, and rack-level power delivery remain constrained.

The market is therefore shifting from a chip shortage to a systems bottleneck. Volume buyers still face reported 36–52 week lead times for some configurations, even as secondary-market H100 availability expands.^[3]

Compute economics: declining unit cost, rising utilization discipline

For model developers, cost per token should continue declining through 2027 as newer accelerators improve throughput and quantization reduces memory requirements. However, the savings will accrue disproportionately to operators with high utilization, optimized software stacks, and access to reserved capacity.

Low-utilization buyers will continue paying hyperscaler scarcity rents. The observed spread—from roughly $1.50 spot H100 pricing to nearly $12 on demand—makes workload scheduling, reservation strategy, and cluster utilization as financially important as GPU selection.^[1][7]

AMD remains the principal price-disruption vector. MI300X rental indications around $1.20–$2.50 per hour and MI325X around $1.80–$3.20 imply meaningful discounting versus comparable NVIDIA capacity, though software compatibility and cluster availability limit substitution.^[8]

What to watch

• Hopper spot curves: sustained sub-$2 H100 pricing would accelerate write-down risk and force hyperscalers to discount older fleets.
• HBM and networking availability: easing constraints would compress Blackwell premiums; persistent shortages would support NVIDIA’s pricing power.
• Cloud contract mix: rising reserved-capacity adoption would signal durable demand, while widening spot/on-demand spreads would indicate utilization stress.

Members-only briefing synthesized by the AISA LLM layer (AISA Perplexity API). 2026-10-15.