Moore’s Law, AI Compute Costs, and the Semiconductor Pricing Bounce
H100 GPU rentals collapsed up to 75% in three years, vindicating a quiet Moore's Law for AI. A 10-40% bounce late in 2025 shows the supply-demand truce is fragile, not broken.
In December 2025, on-demand H100 rentals jumped roughly 10% in a single month, and by March 2026 one-year contract rates had climbed about 40% from their October 2025 low, according to pricing data from Silicon Data and SemiAnalysis. The bounce interrupted a three-year price collapse and reset how investors should read Moore’s Law, AI compute costs, and the semiconductor cycle.
Key facts at a glance
- H100 on-demand rates fell roughly 64% to 75% from their 2023 peak of $8 to $12 per GPU-hour to a 2026 range of $1.49 to $6.98 across major clouds, per IntuitionLabs.
- One-year H100 contract rates rose from about $1.70 per hour in October 2025 to about $2.35 by March 2026, a roughly 40% bounce, per SemiAnalysis.
- Introl pegged the December 2025 on-demand cut at 64% from the 2023 highs before the spike landed.
- Crypto Briefing reported Nvidia (NVDA) lifting reference rental rates around 20% into 2026.
What happened
The long-run direction is down. Silicon Data’s 2023-to-2025 dataset shows the on-demand H100 falling from a 2023 peak above $8 per hour to roughly $2 by mid-2025, a curve much closer to Moore’s Law than the AI bears expected.
Eugene Cheah at Latent Space called the floor in 2024 by counting newly built colocation capacity, and Spheron now lists H100 SXM from $1.03 per hour.
| Provider | Mid-2023 listed | June 2026 listed |
|---|---|---|
| AWS P5 | ~$7.50/hr | ~$6.88/hr |
| GCP A3 High | ~$11.00/hr | ~$3.00/hr |
| Lambda Cloud | ~$4.00/hr | ~$3.29/hr |
| Spheron on-demand | not offered | ~$2.50/hr |
Numbers compiled from Lambda’s rate card and CloudZero’s 2026 buy-and-rent comparison.
Why it matters
The bounce matters because it shows demand still outruns capacity at the margin, even after a brutal supply ramp. That keeps pricing power with chip suppliers like Nvidia (NVDA), TSMC (TSM), and the HBM stack inside every H100 from Micron (MU).
It also feeds the abundance argument from economist Alex Imas. “You actually have abundance. For abundance to generate negative economic growth, that’s really hard to get,” Imas told Phil Trammell in the interview.
He framed Alphabet’s (GOOGL) Gemini Omni as the demand side: the model takes “any kind of input, whether that’s text or audio or video,” he said later in the same conversation, and surfaces emergent capabilities Google never explicitly targeted.
For the cycle, the read is split. AI infrastructure names including Nvidia (NVDA), Broadcom (AVGO), and AMD (AMD) keep the demand tailwind. The capex partners TSMC (TSM) and Micron (MU) still benefit from a tight HBM and advanced-node bottleneck.
What this does not tell you
The three-year price collapse is the dominant signal; the late-2025 bounce is six months old. If colocation supply outpaces a single deployment cycle of Nvidia’s Blackwell and Rubin successors, prices can fall again quickly. None of the cited datasets resolve whether the spike reflects training demand pulling forward inference capacity or a genuine shortage. The remarks on Gemini Omni are framing from Alex Imas, not disclosed Alphabet (GOOGL) capex or H100 utilization data.
FAQ
Did H100 rental prices actually fall over three years?
Yes. Multiple datasets show on-demand H100 rentals fell 50% to 75% from 2023 to mid-2025 before the late-2025 bounce.
Is the recent price spike a reversal of Moore’s Law?
Not yet. The bounce is roughly 10% to 40% on a base that fell 64%, leaving the three-year trajectory aligned with Moore’s Law trends.
Disclaimer. This article is news analysis, not investment advice. References to listed tickers are descriptive context for the news, not a recommendation to buy or sell any security.
Past positioning, prices, and capacity figures are point-in-time snapshots from the cited sources and can change rapidly as the AI compute cycle evolves.
