The H100 resale market is crashing. Pricing data from 6 months.

I've been tracking H100 GPU resale prices on four broker sites since January. The trend is unmistakable: prices are falling fast.

H100 SXM resale price trend

| Month | Avg resale price (H100 SXM) | Change from Jan peak | |-------|---------------------------|---------------------| | Jan 2025 | $32,500 | Baseline | | Feb 2025 | $28,800 | -11% | | Mar 2025 | $26,200 | -19% | | Apr 2025 | $24,100 | -26% | | May 2025 | $22,400 | -31% | | Jun 2025 | $20,800 | -36% | | Jul 2025 | $19,500 | -40% |

Sources: GPU broker listings from 4 sites (aggregated), secondary market data, SemiAnalysis reporting, NVIDIA.

$32,500 in January. $19,500 in July. A 40% drop in six months.

For context, NVIDIA's list price for the H100 SXM is approximately $25,000-30,000. The resale market was at a premium in January (supply constrained). Now it's below list price.

Why prices are falling

Three forces are converging:

| Force | Impact | Timeline | |-------|--------|----------| | DeepSeek efficiency shock | Proved you need fewer GPUs for frontier models | Jan 2025 | | H200 availability | Next-gen GPU entering supply, makes H100 less desirable | Q2 2025 | | B200 pre-orders | Companies saving budgets for B200, selling H100 inventory | Ongoing |

Sources: NVIDIA product roadmap, CoreWeave, Lambda Labs, market reporting.

DeepSeek demonstrated that 2,048 H800s (the export-restricted version of H100) could train a frontier model. The narrative shifted from "you need 10,000+ H100s" to "you need 2,000 GPUs plus good algorithms." That reduced the urgency of H100 hoarding.

H200 GPUs started shipping in volume in Q2 2025. The H200 offers 1.4-1.9x better inference performance than H100 at a similar price point. Why buy a used H100 when you can get a new H200?

B200 GPUs (the next generation) are expected in late 2025. Companies are deferring purchases to wait for the 2.5x performance jump.

Cloud pricing impact

The hardware price drop is flowing through to cloud inference costs:

| Provider | H100 hourly rate (Jan 2025) | H100 hourly rate (Jul 2025) | Change | |----------|---------------------------|---------------------------|--------| | Lambda Labs | $2.49/hr | $1.99/hr | -20% | | CoreWeave | $2.06/hr | $1.79/hr | -13% | | AWS p5 (H100) | $32.77/hr (8-GPU) | $32.77/hr | 0% (AWS is slow to reprice) | | RunPod | $2.39/hr | $1.89/hr | -21% |

Smaller providers (Lambda, CoreWeave, RunPod) have already cut prices 13-21%. AWS hasn't moved yet, which is typical. They'll adjust in their next pricing cycle.

What this means for AI startups

| Budget | Jan 2025 (what you could get) | Jul 2025 (what you can get) | |--------|-------------------------------|---------------------------| | $100K | 3 H100 GPUs | 5 H100 GPUs | | $500K | 15 H100 GPUs | 25 H100 GPUs | | $1M/year cloud | ~400K GPU-hours | ~500K GPU-hours |

A startup with a $500K GPU budget can now get 67% more compute than they could six months ago. For small teams trying to fine-tune or run inference at scale, the economics are improving rapidly.

The bigger picture

| GPU generation | Release | Approximate price | Performance (inference) | |---------------|---------|-------------------|----------------------| | A100 (80GB) | 2020 | $8,000-12,000 (used) | 1x baseline | | H100 SXM | 2023 | $19,500 (current resale) | 3x | | H200 | 2025 | $25,000-30,000 (new) | 4.5x | | B200 | Late 2025 | $30,000-40,000 (expected) | 7.5x |

Sources: NVIDIA, broker pricing, SemiAnalysis estimates.

The performance-per-dollar curve is improving on both ends: new GPUs are more capable, and old GPUs are getting cheaper. The "AI compute crisis" narrative from 2023-2024 is being replaced by a buyer's market.

Good news if you're building. Less good if you're NVIDIA's stock price.

---

If you found this interesting, you might also like:

• The GPT-3 API waitlist is 6 months long. Here's what the early data looks like.
• The GPU shortage data: who has capacity and who's lying about it
• The DeepSeek effect: AI stock prices dropped $1 trillion in a day. The data.
• I counted every AI startup that raised money in Q1 2021. The numbers are strange.
• 5 charts that explain why GPU prices went insane in 2021

-- dataku