Nvidia's H100 is now two architecture generations behind Blackwell and Vera Rubin, whose production shipments began in 2026. Yet the H100 remains a heavily traded rental product—and its prices tell conflicting stories depending on the market being measured.

Start with the index budgets are written against. Silicon Data's H100 Rental Index stood at $2.88 per GPU-hour on October 24. That is 44% above its $2.00 reading on December 9, 2025. It is also about 22% above the roughly $2.37 level where the index launched in May 2025, which is the first reason to handle the headline number carefully: the 44% measures a climb out of a trough. One more caution belongs with any index figure: Silicon Data's rental indices draw on every rental type and convert them to a common on-demand basis, so $2.88 is a normalized benchmark reading, not any single seller's on-demand quote.

The posted lists moved the same way. Nebius, the Amsterdam-based neocloud, repriced its on-demand table effective October 1. Its posted on-demand prices, in USD per GPU-hour, before and from October 1: H100 $3.85 to $4.50 (+16.9%); H200 $4.50 to $5.40 (+20.0%); B200 $7.15 to $8.50 (+18.9%); B300 $7.85 to $9.50 (+21.0%). The current rates are on Nebius's own pricing page; the before-and-after schedule rests on two independent tracker reads of that page on October 1 and October 8, with IntuitionLabs displaying the same $4.50 H100 rate from October 1. The increases were selective: the RTX PRO 6000 stayed at $1.80, and the L40S lines, at $1.55 and $1.82, did not move. Selective increases of that shape are what a provider posts when particular cards are drawing demand; a shortage of specific cards is one possible explanation, not the only one. Runpod moved at a smaller scale: its Secure Cloud H100 SXM rate is displayed at $3.99, up from $3.29 earlier in the autumn.

The layer underneath disagrees

Silicon Data also measures the spot layer, and its study of October 8 to 16 reads differently. Spot capacity is the interruptible kind, sold cheap because the provider can take it back. Across sellers, and excluding Nebius, the H100 spot median over those nine days was $2.33 per GPU-hour, with quotes from $1.38 to $4.49. Nebius had just moved its preemptible prices to a dynamic system that resets as often as every 15 minutes, with a ceiling one cent below on-demand. Its H100 quotes averaged $0.80, sat at the $0.79 floor for most of the window, in a single region, eu-north1, and were the cheapest quote in the market on all nine days. The newest card behaved differently: the B300 spot median was $4.95, and Nebius's B300 quotes rose through the week.

Silicon Data says Nebius supplied the quote data for the study and that the companies were discussing broader data sharing. Nebius did not review or approve the analysis. The reported figures are quoted prices, not evidence of completed rentals.

Nebius's H100 spot quote remained near its floor in one region during the nine-day sample, consistent with slack at that seller and location. The data do not measure available GPU quantities or realized utilization. Silicon Data's own broader reading goes further: it calls the pattern consistent with older GPUs in ample supply and the newest in strong demand. Nine days, though, is all the study claims to describe; Silicon Data itself says the window is not enough to measure volatility or how these prices respond to the rest of the market.

One more observation from the study cuts against reading too much into any single quote. More than half of the spot price series Silicon Data tracks did not change price once in the nine days, and three quarters changed on two days or fewer. Silicon Data's own possible explanations are mundane: prices set and revisited only occasionally, customer preference for predictable rates, operating costs, and thin spot volume. An unchanged quote, on that account, says something about how a seller prices, not about whether its GPUs are busy.

Both sets of numbers can hold at once because most capacity never touches either price. The bulk of GPU rental is sold on reserved contracts negotiated for months or years. On-demand lists are the visible price of flexibility; spot is the residual balancing layer. Rising posted on-demand prices say buyers who need guaranteed hours are paying more for them, which sits comfortably beside Amazon chief executive Andy Jassy's October 23 statement that AWS's large capacity blocks are mostly booked through 2027, a claim this site covered on October 24. Floored spot quotes say something narrower: at one seller, in one region, interruptible H100 hours were quoted at the floor for nine days in October.

The older-card escape valve

Price pressure at the top is pushing some work down the stack, which is where a small open-source project fits. gpu-unlock, posted to GitHub on October 18 by a developer working as Zagitta, is a Linux kernel module that lifts the FP16 and BF16 instruction issue-rate limit Nvidia applies to consumer Ampere cards, the RTX 30-series. In the project's own benchmark, a dense BF16 matrix multiplication on a 3090 Ti improves by about 80%. User reports on real workloads range from meaningful gains in prefill to none at all, and token generation speed is mostly unchanged. The tool is a week old, carries an experimental label, and installing it needs root access, a kernel change and a reboot.

What to quote

If you need one number for a budget, take it from the layer you will actually buy, on the date you will buy it, and say which layer it came from. A team renting on-demand H100s in late October faces posted rates about 17% higher at Nebius than in September. A team that can tolerate interruptions is in a different market, where one seller quoted H100 hours at $0.79 this month. Neither figure is the market; each is a layer of it.

The year-to-date roundup percentages now circulating for these chips are worse than unhelpful: depending on the roundup, H200 rentals are up about 160% since January, or about 143%, with H100 up about 54%. Those figures are excluded here because their bases are incompatible — different baskets, different start points, different rental types — so the disagreement describes the roundups, not the market. Even two live indices diverge: Ornn's H100 reading was $3.02 on October 24, the same day Silicon Data printed $2.88. Different coverage and normalization produce different numbers; the divergence does not invalidate either reading, but a rental claim without a named index, a date and a layer is not a claim at all.

Sources

  1. [1] Silicon Data, "GPU Spot Prices and Dynamic Pricing", October 21, 2026 (data window October 8–16, 2026) — spot medians and quote ranges, the $0.79 Nebius H100 floor in eu-north1, the nine-day limits, the stable-price explanations, the index methodology (indices draw on every rental type, converted to a common on-demand basis), the note that all figures are offered prices, not realized rentals, and the disclosure that Nebius provided the quote data (read first-hand)Read source
  2. [2] Startup Fortune, "H100 rental prices are spiking even as Wall Street worries about an AI bubble", October 25, 2026 — the $2.00 to $2.88 index readings and the +44% calculation, the Runpod September-to-October move, Jassy contextRead source
  3. [3] Nebius pricing page — current posted on-demand rates read first-hand on October 25, 2026: H100 $4.50, H200 $5.40, B200 $8.50, B300 $9.50, RTX PRO 6000 $1.80, L40S $1.55/$1.82 per GPU-hour; the page displays current rates only, with no before/after scheduleRead source
  4. [4] Nebius price page as read on October 1 and October 8, 2026, via FourWeekMBA's table of the page — the before/after on-demand schedule effective October 1 for H100, H200, B200, B300, RTX PRO 6000 and L40SRead source
  5. [5] IntuitionLabs, "H100 Rental Prices Compared" (revised October 2026) — Runpod Secure Cloud H100 SXM displayed at $3.99, up from $3.29; Nebius HGX H100 at $4.50 from October 1, 2026, previously $3.85; the May 2025 index launch level of about $2.37Read source
  6. [6] GPUQuant / Ornn Compute Price Index (read October 25, 2026) — H100 daily settlement of $3.02 on October 24, 2026, cited only to show that two live indices printed different H100 numbers on the same dayRead source
  7. [7] Zagitta, gpu-unlock repository (GitHub, posted October 18, 2026) — mechanism, supported dies, the project's own benchmarks, experimental statusRead source