GPU prices jump 30%+, startups squeezed

GPU prices are rising again — not in the abstract, but in the very practical market where startups rent H100s by the hour or by the month. The important change is that this is no longer just a “can you get capacity?” problem. It is turning back into a “can you afford predictable capacity?” problem. SemiAnalysis flagged the shift in April when it launched a daily H100 spot and contract pricing index and framed the market as a fresh GPU shortage. ### What exactly got tighter? The market for rented Nvidia H100s got tighter across both spot access and longer contracts. SemiAnalysis now tracks daily, on-demand, 1-month, 3-month, 6-month, and 1-year H100 rental pricing — which tells you the stress is showing up across the whole curve, not just in last-minute overflow demand. That kind of tracking only becomes useful when pricing is moving enough, and unpredictably enough, that buyers need a benchmark. (api.semianalysis.com) ### Why do contracts matter so much? Because startups do not just need cheap GPUs — they need GPUs that will still be there next month when a training run slips, a customer pilot expands, or an inference workload spikes. Spot is great until it disappears. So the market is leaning harder on reserved and longer-duration deals. Crusoe’s pricing page makes the split visible: some newer high-end parts are listed as contact-sales only, while H100 on-demand is posted at $3.90 per GPU-hour and spot is not publicly listed there at all. (api.semianalysis.com) That is a hint that guaranteed supply is becoming more negotiated and less commodity-like. ### Why is this hitting startups first? Big model labs can pre-buy clusters, sign giant commitments, or spread workloads across multiple providers. Early-stage teams usually cannot. They rent. That means GPU inflation shows up directly in burn rate. A team that thought it had 12 months of runway can discover that the same experimentation plan now buys fewer training cycles, fewer retries, and less room for mistakes. The squeeze is financial, but it also changes behavior. (crusoe.ai) ### So what changes inside a startup? Teams get stricter. They do more preflight testing on smaller machines. They kill weak experiments earlier. They become more careful about data prep, checkpointing, and evaluation so they do not waste expensive runs. And they revisit whether a feature really needs live GPU inference or whether it can be batched, distilled, quantized, or pushed onto cheaper hardware. Basically, “move fast” starts to mean “waste less compute.” (newsletter.semianalysis.com) ### Is this just a short-term blip? Maybe not. The broader backdrop is still heavy AI infrastructure demand. SemiAnalysis has been writing all spring about shortages spanning GPUs, memory, and datacenter bottlenecks, while infrastructure spending keeps rising across the major cloud and model players. When upstream supply stays constrained, rental markets feel it fast — especially the secondary and neocloud layers that many startups rely on. (semianalysis.com) ### Why does a 30% move hurt so much? Because compute costs compound. If training gets 30% more expensive and inference assumptions were already tight, margins can flip from plausible to shaky in one planning cycle. It is like building a startup on rented office space and then finding out the landlord also controls your electricity, your internet, and whether you can enter the building next week. The price is one problem. The uncertainty is the bigger one. (newsletter.semianalysis.com) ### What should readers watch next? Watch whether this pressure stays concentrated in H100 rentals or spreads into H200, B200, and AMD alternatives. Also watch contract length. When buyers start accepting longer lockups just to secure supply, that usually means they think availability will stay tight for a while. The bottom line is simple: AI startups built on rented GPUs just got a harsher operating environment. Higher rental prices do not just make experiments cost more — they make strategy less forgiving. (api.semianalysis.com) (newsletter.semianalysis.com) (crusoe.ai)