Nebius (NASDAQ: NBIS) is raising on-demand hourly rates for four Nvidia GPU models by between 17% and 21%, effective October 1, 2026.

The price increases cover four of Nvidia’s most sought-after chips: the H100, H200, B200, and B300, along with Nebius’ CPU-only server offerings.

The October 1 round is not the first increase this year, with Nebius having already raised rates substantially since early May.

In May, the older H100 chip was priced at $2.95 per hour and the B300 at $6.10, but after this latest round those same chips will cost $4.50 and $9.50 respectively.

That represents cumulative increases of roughly 53% and 56% on those two chips alone over a span of just five months.

Notably, rising prices for the H100, a chip that predates Nvidia’s current Blackwell generation, may say more about the depth of the supply shortage than any premium on the newest hardware.

When a company is expanding capacity at speed, conventional logic suggests it should be working to fill that capacity, not charging customers more for it.

Management has indicated it could sell all of its 2027 capacity today on the terms of its most recent contracts, underscoring just how acute demand for AI compute has become.

The aggressive build-out has required substantial financing, including $5.75 billion in convertible notes sold in August and approximately $2.8 billion in new shares sold through June.

Shortage pricing can reverse quickly if supply catches up with demand, and Nebius and its rivals continue to add capacity at a rapid pace, which could put downward pressure on hourly rates.

The October 1 price hike is described as a strong signal of market tightness rather than a primary revenue driver, given that new rates apply to a smaller share of Nebius’ overall business than the headline numbers suggest.

With shares trading at around $243, Nebius carries a market valuation of over $65 billion, equivalent to more than 20 times sales at the midpoint of the company’s 2026 revenue guidance.

That valuation implies investors are already pricing in the assumption that current compute rates hold well into 2027, leaving little margin for error if the supply-demand balance shifts.