The cleanest market read on the price of raw AI compute: what it costs to rent one NVIDIA H100 for an hour. It's the empirical test of [[The demand for intelligence is virtually infinite]] and a live gauge of the commodity-compute layer in [[AI creates deflationary abundance and inflationary capital demand simultaneously]]. **Current level: ~$2.35/GPU-hour** (1-year contract, March 2026) — **up ~40%** from the cycle low of **$1.70 in October 2025**, with on-demand capacity reported "sold out across all GPU types" ([SemiAnalysis](https://newsletter.semianalysis.com/p/the-great-gpu-shortage-rental-capacity)). **The shape is a V, not a slide** — this is the whole story: - **2023:** **$7–$12+/hr** at peak scarcity (often >$12 on-demand at hyperscalers) — the "forever above $4.70" era. - **Early 2024:** ~**$2.85/hr** as 300+ neoclouds flooded capacity ([Latent Space](https://www.latent.space/p/gpu-bubble)). - **Oct 2024:** **$2.99/hr** on-demand SXM after a 30% cut in two months ([SemiAnalysis](https://x.com/SemiAnalysis_/status/1843153111118201270)). - **2025:** drifts to ~**$2.2–$2.4** (marketplace); AWS cuts P5 on-demand ~30–44% in June. - **Oct 2025:** **$1.70/hr** — the cycle low, sitting right at the **~$1.65/hr** level below which the GPU never recoups its capex ([Latent Space](https://www.latent.space/p/gpu-bubble)). - **Dec 2025 → Mar 2026:** snaps back to **$2.35** (+40%) — the "Great GPU Shortage" round two, driven by **inference/agentic and media-generation** demand (Claude Code, multi-agent, Suno), *not* training. Blackwell (B200) capacity is itself sold out, so buyers bid up Hopper. **Two prices — don't confuse them:** - **Rental ($/GPU-hr):** cyclical, above. Breakeven rules of thumb: **>$2.85/hr** to beat a stock-market IRR; **<$1.65/hr** = expected loss on the hardware. - **Purchase (per card):** new list price held remarkably stable at **~$25k–$40k** since launch ($40–46k on eBay at the Apr 2023 peak). The *used* market round-tripped from a ~$50k mid-2024 scarcity spike down to **~$19k–$22k/GPU** by early 2026 — the real depreciation signal. Why it matters: the bear case ([[Should you diversify into real estate in the age of AGI?]]) calls a GPU a "melting ice cube" — a depreciating consumable whose output price collapses. The V-shape complicates that: cyclical shortages can re-price the chip *up* 40% in months. But the secular force is still down — Hopper gets obsoleted by Blackwell, then Rubin — so a rental spike is a capacity-lag, not the chip becoming durably scarce. The note that interprets all this: [[The demand for intelligence is virtually infinite]]. Where to find it: - [SemiAnalysis — GPU rental price index / "The Great GPU Shortage"](https://newsletter.semianalysis.com/p/the-great-gpu-shortage-rental-capacity) — the authoritative 1-yr-contract series (Dylan Patel). - [Silicon Data — H100 rental price over time](https://www.silicondata.com/blog/h100-rental-price-over-time) — dedicated tracker, Bloomberg ticker SDH100RT. - [Latent Space — "$2 H100s: How the GPU Rental Bubble Burst"](https://www.latent.space/p/gpu-bubble) — the breakeven math. - Marketplace live pricing: [RunPod](https://www.runpod.io), [Vast.ai](https://vast.ai), [Lambda](https://lambdalabs.com). Related: [[AI capex]] (the buildout flooding the commodity layer), [[Real interest rates (10-year TIPS yield)]] (the cost of the capital funding it), [[The demand for intelligence is virtually infinite]] (what the V-shape means).