The compute crunch: what GPU price data shows, and the free credits that soften it

“The world’s compute is sold out.” That is how Evan Conrad of SF Compute put it in The end of startups, published October 22, 2026. He argues that within six months most AI startups won’t be able to buy compute, and that lenders now ask for large upfront payments, in one case “as high as 100% of the contract”. Keep in mind that SF Compute sells GPU capacity, and the post ends with a hiring note.

Here is what the public price data says, including the parts that point the other way, followed by the free credits that cut your bill today. Every figure links to its source; we read each one on October 23, 2026 (ET).

What the numbers show

Rental prices for the newest GPUs went up over the past month. Ornn’s Compute Price Index, which is built from executed compute trades and settled October 23, 2026, puts the H100 SXM at $3.17 per GPU-hour (+3.9% over one month), the H200 at $5.74 (+14.3%) and the B200 at $9.58 (+20.5%).

List prices are moving one way. In its October 2026 GPU rental market report, FastGPU found that in the 60 days to October 10, every confirmed list-price change for the H100 SXM, H200, B200 and B300 was a rise, with median steps of 14.3% to 17.7% (10.4% for the B300). Its price-change log holds 81 rises against 4 cuts. One example: RunPod’s on-demand B200 went from $6.79 to $7.99 per GPU-hour on October 8.

The squeeze started earlier. SemiAnalysis reported in April that one-year H100 rental contracts rose from a low of $1.70 per GPU-hour in October 2025 to $2.35 by March 2026, and that on-demand rental capacity was sold out across GPU types.

Conrad expects more of the same. “We expect GPU prices to rise,” he writes, adding that “you can buy a B300 for $6/hr and then make about $30/hr selling inference on it”, a margin he expects to shrink as rents go up. He doesn’t give the model or utilization behind that estimate.

What points the other way

• Older and cheaper cards are not in a crunch. FastGPU’s GPU Cloud Price Index, which tracks the cheapest rentable price for each GPU, stood at 99.1, 0.9% below its August base, with 8 of 18 tracked GPUs cheaper. The floor held while list prices climbed (FastGPU).
• Tightness eased this month. Bargo’s Compute Tightness Index reads 52.0 out of 100 (“Balanced”) as of October 22, down 5 points over 30 days. Spot H100, H200 and B200 listings sit 44%, 39% and 47% below on-demand prices. Bargo still describes capacity as “easing but remains far from abundant”.
• Token prices are not following GPU rents. On Ornn’s token index (a blend of posted per-token prices, settled October 22), Anthropic sits at $0.70 per million tokens, down 53.3% over one month, and OpenAI was roughly flat.

Our read: if you rent H200- or B200-class GPUs on demand, expect higher list prices and less spare capacity. If you mainly call model APIs, your per-token prices have not risen with GPU rents. Either way, free credits are the cheapest compute you can get, and most of the ones below need no application.

Free GPU and inference credits you can claim today

The compute crunch pack groups them. Sign-up credits and free tiers come first, followed by startup and research programs that have eligibility rules.

Larger programs (application required)

Where to look next

• Cloud & GPU credits and LLM API credits: every live offer in each category, with source quotes and the time each was last verified.
• GPU & inference pack: includes grants such as Hugging Face community GPUs and NVIDIA Inception.
• GPU & inference providers: RunPod, Vast.ai, Lambda, SF Compute, OpenRouter, Together AI, Fireworks and more, with links to each provider’s own pricing page and any free credits we’ve verified for it.

Prices move daily. Check the linked sources before you commit to a reservation, and each vendor’s own page before you count on a credit.