H100 Rental Prices "Up 45% Since December": What Silicon Data Actually Shows

A post on X on October 24, 2026 said the cost to rent an Nvidia H100 has jumped 45% since December, citing rental indices from Silicon Data, and investor Chamath Palihapitiya added that "power is the binding constraint." The direction matches what public data shows: H100 rent was about 2.00 dollars per GPU-hour in December 2025 and was reported near 2.81 dollars in mid-October 2026, roughly 40 percent higher. We could not reproduce exactly 45 percent from a public Silicon Data post, and the power explanation is an opinion, not an index finding.

This is a claimed-versus-verified read. It matters because the H100 is a 2022 chip, two Nvidia generations old, and its price is a clean gauge of whether compute demand still outruns supply. If you rent GPUs for fine-tuning, batch inference or evals, your bill follows this curve. We covered the same pattern for newer parts in our B300 rental record claim check, and for one provider in Nebius raising GPU rates 20 percent.

Update (October 25, 2026): the a16z chart behind the "H200 up 160%" follow-up

A follow-up claim is now circulating alongside the H100 number: that renting an Nvidia H200 costs about 160% more than at the start of the year, and that OpenRouter token demand rose roughly 27x, from 6T to 165T tokens a week, since last November. The source is a16z's Charts of the Week newsletter of October 23, 2026, in a section titled "GPUs, Still Not Obsolete." What that text actually says:

• Every GPU model in the Ornn Data series, "including the oldest A100 models," costs more to rent than at the beginning of the year.
• Against a straight-line six-year depreciation schedule, prices are "anywhere from 9% higher to 143% higher" than that schedule would imply. That is a gap to a depreciation line, not a rental-price rise since January.
• A100 and H100 spent more of the longer history below the depreciation line than above it, and a16z says the market "really took off" in April 2026.
• OpenRouter token demand began a steep climb in spring, with a brief late-summer plateau. a16z notes OpenRouter "captures only a small slice of the picture" and that the indexes are not perfect proxies.

What we could not find in the newsletter text: a 160% H200 figure, or the 6T-to-165T and 27x token numbers. Those appear in a social post summarizing the charts, so treat them as secondhand until you can read the chart values yourself. Public indexes also disagree on levels: Silicon Data's H200 index page showed about 3.48 dollars per GPU-hour as of October 24, a different method and baseline from Ornn Data's. The direction (rents up across generations) is supported; the size of the H200 jump depends on the index and start date.

Quick reference: claimed vs. verified

What the public Silicon Data numbers show

Silicon Data, which publishes daily GPU rental indices, posted its own account of the first leg of this climb in January. In its analysis of an H100 price spike, the company reported that its H100 rental index rose from 2.00 dollars per hour on December 9, 2025 to 2.20 dollars on January 6, 2026, a 10% move that it described as the largest short-term increase since mid-2025 and as isolated, since A100 and B200 rates were flat in the same window.

Our earlier B300 post cited the Silicon Data page on October 15 with H100 at 2.81 dollars per GPU-hour, up 5.6% over seven days.

Put the December baseline and the October reading side by side:

From 2.00 to 2.81 is a 40.5% increase. A 45% rise from the same base would land near 2.90 dollars. When we opened Silicon Data's H100 index page on October 24, 2026, it showed 2.88 dollars per GPU-hour for the neocloud series (ticker SDH100RT), about 44% above the December 9 base. That is close to 45%, though the same page carried a different figure in its FAQ text, and we could not confirm the exact base date the post used. We label the 45% as roughly consistent, not exactly verified.

Why one percentage can mislead: tiers and baselines

Silicon Data's own writing is a reminder that "the H100 price" is several prices.

• Neocloud versus hyperscaler. In its H100 price spike post the company noted rates as low as about 1.90 dollars an hour at specialist providers while AWS, Azure and Google Cloud can charge more than 4 dollars for comparable configurations.
• Thin listings. Rental indices sample listed offers, so a daily index can move because a few providers reprice.
• Start date. December 9 was a local low before a spike. Measuring from a different December date changes the percentage by several points.

If the 45% was computed on the neocloud series from the December base, it is a statement about the cheaper, more flexible tier, which is where startups and researchers rent. If a hyperscaler reserved rate is what you pay, your increase may be far smaller.

Is power really the constraint?

Chamath Palihapitiya's claim, relayed in the post, is that power constraints drive the climb. That is a reasonable hypothesis, and it fits other reporting we have covered: AWS saying capacity is sold out for years, the grid fights behind Virginia's data center power tax, and Google's 1 GW virtual power plant deal on the PJM grid.

But an index of rental prices measures price, not cause. Silicon Data's January analysis gave different drivers for the first spike: hyperscalers prioritizing reserved capacity and long-term contracts, which squeezed on-demand inventory; year-end training deadlines; and budget-use-it-or-lose-it spending. Those can coexist with a power story. Another candidate is demand from newer agent workloads that run on older chips because they are cheap per token. To test the power explanation you would need data on new data-center capacity delayed by grid connection, which an index does not provide.

Why would a 2022 chip get more expensive?

Three mechanisms are plausible, and none requires a new shortage of H100 silicon itself.

1. Substitution upward. When newer parts such as B200 and B300 are scarce or expensive (B300 neocloud rent was reported near 6.77 dollars in mid-October), buyers fall back on H100 and H200 for workloads that fit.
2. Inventory locked in term contracts. If more of the fleet is reserved, the spot remainder is thin and volatile, as Silicon Data argued in January.
3. Inference demand. Serving open-weight models at scale runs well on 80 GB parts, so demand is not limited to training.

The counterpoint is the neocloud shakeout thesis in Gartner's forecast for neoclouds: if capacity is being added fast, a price rise can be a temporary squeeze rather than a trend. Rising rent also makes new neocloud buildouts, such as the Nebius Madrid hub targeting 4 million H100-equivalents, look more profitable, which pulls more supply in.

What people are asking

Is this the same as a chip shortage? Not necessarily. Silicon Data noted in January that CoWoS packaging and HBM supply had eased compared with 2023 and 2024, yet the system still ran near capacity with little buffer. A market with no slack reacts sharply to small demand shocks, so a price rise does not by itself mean fewer chips are being made.

Will prices fall when Blackwell is plentiful? Silicon Data's January outlook expected older parts to settle into a mainstream workhorse role with modest declines. Ten months later that expectation has not held for the H100, which is a useful reminder that forecasts in this market age quickly.

Can an index be wrong? Yes. Rental indices sample listed offers, which can be thin, stale, or not available in size. A 5% weekly jump can come from a handful of providers repricing. That is why we recommend reading a second independent series before acting.

What the claim means for what you pay

• Check which index tier you are exposed to. Read the neocloud, hyperscaler, spot and term series separately.
• Lock rates when it helps. For most teams a plain term contract with the provider is the simplest hedge.
• Re-run cost per task. At 2.8 dollars an hour versus 2.0, a job that took ten GPU-hours costs 28 instead of 20 dollars. Faster batching, quantization and smaller models cut hours, which now saves more.
• Watch the chip-cost context. Upstream costs are also rising: see US chip producer prices up 27 percent on AI demand.
• Do not overreact to a single day. The index moved 5.6% in a week in October; weekly noise is large relative to the headline trend.

How to verify it yourself

Open Silicon Data's H100 index page, pick the series (neocloud or hyperscaler), and compare the earliest December 2025 value to the latest. Note the ticker (SDH100RT for the neocloud rate). Cross-check with a second, independent series where you can find one; two independent series rising together is stronger evidence than a single index.

A worked budget example

To see why the tier and the base date matter, take a team that runs 1,000 H100 GPU-hours a month for fine-tuning and evals. At the December 9 reading of 2.00 dollars, that is 2,000 dollars a month. At the October 24 neocloud reading of 2.88 dollars, the same usage costs 2,880 dollars, an extra 880 dollars a month or roughly 10,500 dollars a year. If the same team is on a term contract signed before the move, the increase may not show up until renewal, which is why the renewal date is worth putting on the calendar. If instead it buys on-demand capacity from a hyperscaler at more than 4 dollars an hour, the percentage change may be smaller even though the absolute bill is higher. In both cases the useful number is cost per finished job, not the headline percentage.

Bottom line

H100 rent is clearly higher than it was in December, by about 40 percent on public numbers, and the 45% claim is plausible but unconfirmed to the decimal. The cause is still debated; power is one hypothesis among several. For anyone budgeting GPU hours into 2027, the safe assumption is that older-generation capacity is not getting cheaper on its own.

Prices and index values are as published by Silicon Data and cited sources as of October 24, 2026 and change daily.

Related reading

• B300 GPU rental prices at a record: what the public indexes show
• Nebius raises GPU rental rates 20% in a second hike
• Nebius Madrid AI hub and 4 million H100-equivalents
• AWS capacity sold out through 2028: claimed vs verified
• US chip producer prices up 27 percent
• Virginia's data center power tax
• Gartner on the neocloud shakeout