Nvidia H200 Rental Prices Up 143% to 160% Since January, a16z Data Shows

Renting Nvidia's H200 costs sharply more than it did at the start of 2026. Figures circulated by venture firm a16z put the increase at about 160% from Jan. 1 through mid-October. A separate cut of the same data, tracked by Ornn Data, shows 143%.

The gap between those two numbers depends on how the tracking is done. The direction is the same in both.

Every chip is up, some far more than others

The Ornn Data tally shows these increases since Jan. 1:

• H200: up 143%
• B200: up 111%
• RTX PRO 6000: up 73%
• H100: up 54%
• A100: up 9%

Another tally of the same period puts the H200 at about 160%, the B200 at about 120%, the RTX PRO 6000 at about 65% and the H100 at about 35%. Separately, investor Chamath Palihapitiya said renting an H100 costs 45% more than it did in December.

The A100 is the odd one. It is a six-year-old chip, and it now rents for more than it did nine months ago. Hardware that was supposed to be sliding down a depreciation schedule is instead getting more expensive to use.

The path was not a straight line

Most of these chips hit their low in the first quarter. The H100 and B200 fell roughly 23% below their January levels, then rebounded in April and May.

The H100 stayed flat until early May before jumping within weeks. The H200 spiked in late May, gave back about a quarter of the gain within weeks, then resumed climbing into October.

What renters pay

Median on-demand H200 rates run about $4.27 to $4.94 per GPU-hour across tracked providers. Spot marketplaces are cheaper, with rates as low as about $1.47 and up to $3.88.

Spheron, a GPU marketplace, lists its own H200 at $4.80 per hour dedicated and $3.35 on spot capacity that can be reclaimed at any time. In its published comparison, AWS lists $4.98 per hour, CoreWeave $6.31, Azure $10.60 and Google Cloud $10.87. Spheron says those are on-demand rates from each provider's own pricing page, and that a listed rate is not the same as available capacity.

Some providers are raising prices outright. Nebius lifted its on-demand H200 rate from $4.50 to $5.40 per GPU-hour, a 20% increase, effective Oct. 1, 2026. In the same adjustment it raised H100 rates by 17%.

Two explanations for the squeeze

a16z points to token consumption. The firm says token demand on OpenRouter has risen about 27x since last November, from roughly 6 trillion to 165 trillion tokens per week. Open-weight models now handle about 72% of that traffic, up from 28% a year ago.

In a16z's framing, falling per-token prices are driving more consumption, and that is creating scarcity in the GPUs available to serve it.

Palihapitiya, the Social Capital founder, argues the constraint sits elsewhere. "Power is the binding constraint," he said. "This is why prices continue to climb so ferociously."

Those two explanations are not mutually exclusive. The charts, however, do not settle the question. They do not identify what is behind the demand, whether agentic workloads, enterprise adoption or something else. They also do not show supply, data center capacity or power availability.

Why older chips still have value

The H200 stands out partly on specifications. It carries 141 GB of HBM3e memory and 4.8 TB/s of bandwidth, about 1.76 times the memory capacity of the H100. That makes it a better fit for long-context inference and serving 70B to 100B parameter models at large batch sizes, according to Spheron.

Older chips are still competitive for running medium-sized models, which helps explain why their rental prices have not collapsed.

What to watch

The rental market has already shown it can reverse. The H200 gave back about a quarter of its late-May spike within weeks before climbing again. Nebius's October increase is the one provider-level price move in the data. Whether other providers follow, and whether the H200's climb holds into the winter, will show if this is a lasting shortage or another spike.