Nebius B300 Spot Hours Averaged Up to 63% of List Price

Nebius’s spot price for Nvidia’s B300 chip averaged $5.97 per GPU-hour in its us-north1 region between 8 and 21 October 2026, or 63% of the $9.50 on-demand list price, according to our analysis of Nebius’s own public price data. Over the same period its older H100 averaged $0.80 in eu-north1, 18% of a $4.50 list price.

Nebius chief executive Arkady Volozh said on 21 October that “the prices are still there, demand is still there,” according to a transcript posted on X by Fireside Alpha. Investor Gavin Baker said in the same exchange that the cost to rent an hour of GPU “has gone vertical” even as token prices fall. As we read it, Nebius’s data shows that split running through one cloud’s price list, chip by chip.

Business Pill · RENTING THE CHIPS BY THE HOUR

A one-minute explainer of neoclouds: companies that rent AI chips by the hour to anyone who needs compute. It teaches the general idea only and says nothing about any company in this story.

What Nebius’s Price Data Shows

Nebius publishes a live spot price for each chip in each region, and an API that returns up to 30 days of history without a login. We pulled the hourly averages for all 12 markets it lists on 21 October at about 09:30 UTC, and compared the average since 8 October with the on-demand list price on its price page.

Prices in US dollars per GPU-hour. Spot averages are the mean of the hourly average prices in Nebius’s public spot-price API from 8 October 12:00 UTC to 21 October 09:00 UTC (310 hours); on-demand list prices are from nebius.com/prices. The shares are our arithmetic.

• The newest chip sold closest to list. The B300 averaged 49% to 63% of its $9.50 list price across three regions, on our arithmetic.
• It hit the ceiling. In the hourly data, the B300 averaged $9.49, one cent under list, for 27 hourly averages in us-north1, 31 in eu-west2 and 12 in uk-south1.
• Older chips sold far below list. The H100 averaged 18% of list, the B200 about 26% to 27% and the H200 between 26% and 44%, depending on the region.
• The same chip, a different price. At 09:29 UTC on 21 October, an H200 hour cost $4.49 in eu-west1 and $0.886 in us-central1, about five times as much, on our arithmetic.

The prices move in both directions. Nebius’s spot-price page showed the cheapest B300 at $4.10 at 09:33 UTC, down 41.4% on the day before, and the cheapest B200 at $1.90, down 48.0%.

What Was Posted on X

Fireside Alpha posted this clip with a transcript: Gavin Baker says that as token prices have fallen, the cost to rent an hour of GPU has gone vertical, and Arkady Volozh replies that Nebius now raises prices every 12 minutes on its auctions, and that sometimes they go down. This piece quotes the transcript; the table above tests the GPU-hour side against Nebius’s own data.

In a second clip, Volozh argues that GPU capacity should be priced by a market mechanism, like internet advertising or taxi markets, and says Nebius launched it for preemptible capacity and saw prices go up. Nebius’s own blog dates its dynamic spot pricing to 8 October.

How Nebius’s Spot Market Works

Spot, or preemptible, capacity is spare GPU time that a customer can lose at short notice. Nebius said in a blog post on 22 September that it would price this capacity with an algorithm that reads customer demand and available capacity, and that dynamic pricing would take effect on 8 October. Its price history shows every market changing price from that day.

The spot price “always stays between a price floor and a price ceiling, where the ceiling is one cent below the current on-demand GPU-hour rate,” the blog says. It adds that the price “is currently updated up to every 15 minutes.” Customers can set a maximum price, and their machine stops if the spot price goes above it, or follow the market price as it moves.

What Volozh and Baker Said

Volozh and Baker spoke on 21 October in clips that Fireside Alpha posted on X; Nebius said on X that Volozh was speaking on the #NVIDIAGTC pregame about GPU pricing. Baker asked for a real-time view on compute prices. Volozh answered, as transcribed: “Yes, the prices are still there, demand is still there.”

Baker then separated the two prices that are often blurred. “And as the price of tokens has gone down and the mix has shifted to cheaper tokens. The cost to rent an hour of GPU has gone vertical,” he said, adding that “as you’re getting more tokens that are more useful out of a given GPU hour, that GPU hour becomes more valuable.”

Asked how often Nebius had raised prices, Volozh said: “Now we raise prices every 12 minutes on the auctions. Not only raising, by the way. Sometimes they go down.” In a later clip he said “The pricing should be market driven,” and described the preemptible system as launched “yesterday,” adding: “we got the first traction, the prices went up.” Nebius’s blog dates dynamic spot pricing to 8 October; the clips do not say what changed on 20 October.

Tokens Get Cheaper, Hours Do Not

A token is a small unit of text that an AI model reads or writes, and AI labs sell access by the million tokens. A GPU-hour is one chip rented for one hour, which is how clouds such as Nebius sell compute. Baker’s point is that the two prices can move apart: if each hour of a chip produces more tokens, the price per token can fall while the price of the hour holds or rises.

Two figures in our earlier coverage show the gap from the seller’s side. DeepSeek’s deployment notes say a single cluster earns “overall revenue of US$38 per GPU-hour (at valley pricing)”, against the $9.50 list price for a Nebius B300 hour. SpaceX’s chief financial officer said on 20 October that its compute pricing “is currently tracking slightly above the high end of the $30 to $50 per watt range.” These are different chips, contracts and units, so they are not directly comparable.

Update, 21 October: Nebius Publishes Its Pricing Approach

Nebius “today announced public live spot pricing for GPU compute, giving teams greater visibility and control over their compute costs,” according to its 21 October release. The release says “reserve auctions offer defined capacity blocks for short-term compute needs, while pay-as-you-go enables access to compute on demand without a fixed-term reservation.”

It adds that Nebius “is also piloting a capacity-rebate program that enables customers to recover value by releasing idle GPUs back into the spot capacity pool.”

Update, 21 October: Nebius Tells CNBC Compute Prices Are Rising

Nebius’s chief revenue officer, Marc Boroditsky, told CNBC on Wednesday that AI computing prices are rising more than the company expected. Asked whether he sees the price per megawatt continuing to rise ahead of expectations, he replied: “Yeah.” He said the auction Nebius ran on Tuesday showed continued price rises: “We’re seeing continued upward mobility in terms of the kind of price per megawatt.”

Boroditsky also told CNBC that if Nebius brought its 2027 capacity online, it would sell out, and that there could be “the same type of dynamic” in 2028. On debt, he said Nebius was not in CoreWeave’s position because “we’re not carrying the kind of debt that they are.”

CNBC frames the remarks against investor wariness about neocloud economics after CoreWeave’s cost of borrowing spiked, which it says has weighed on Iren and Nebius shares over the last month. In the same article, Crusoe CEO Chase Lochmiller told CNBC: “We’re seeing higher pricing today for A100s than we saw in 2020 or 2021. We’re seeing higher pricing for H100s than what we saw when they were first deployed in 2023.” As we read it, that is a different measure from the spot data above: Lochmiller compares today’s rental prices with earlier years, while our series compares Nebius spot hours with its list price, so both can hold at once. CNBC does not say whether Crusoe’s prices are spot or contract, and we have not seen Crusoe’s figures.

Connecting the Coverage

Our Nebius price piece of 8 October reported that Nebius raised its on-demand list prices from 1 October, with the B300 going from $7.85 to $9.50 per GPU-hour and the H100 from $3.85 to $4.50. Our Pichai piece of 21 October reported that Google is “capacity constrained” serving its newest Gemini model. Our SpaceX call piece of 20 October carried the CFO’s per-watt pricing remark quoted above.

Our DeepInfer piece of 20 October quoted DeepSeek’s revenue per GPU-hour, and our neocloud shakeout piece of 20 October covered The Information’s report that executives expect only seven to 10 of hundreds of neoclouds to survive. What is new here is a measured price series: 13 days of Nebius’s own spot data, chip by chip and region by region.

What It Means for Buyers of Compute

As we read it, the cheap end of the GPU market is now the older chips. A team that can run on H100s found spot hours at under a fifth of list in Nebius’s data, while B300 spot hours averaged about half of list or more.

As we read it, the region matters as much as the chip. The same H200 hour was priced about five times higher in one Nebius region than in another at the same moment, so a workload that can move between regions can buy the same compute for less.

The Structural Read

As we read it, the spot ceiling Nebius set, one cent under list, works as a test of scarcity: in the hourly data, the B300 reached it for between 12 and 31 hours in each of its three regions.

As we read it, Baker’s distinction is the one the table measures. Token prices and GPU-hour prices are two different prices, and Nebius’s numbers are the GPU-hour side only.

As we read it, the regional spread is part of the price. The same H200 hour cost about five times more in one Nebius region than in another at 09:29 UTC on 21 October, on our arithmetic.

Three Implications

FOR BUYERS OF COMPUTE In Nebius’s data, H100 spot hours averaged $0.80 against a $4.50 list price, 18% of list on our arithmetic.

FOR NEOCLOUD INVESTORS Volozh said on 21 October that “the prices are still there, demand is still there,” as Fireside Alpha transcribed it.

FOR TEAMS THAT CAN MOVE At 09:29 UTC on 21 October, an H200 hour on Nebius cost $4.49 in eu-west1 and $0.886 in us-central1, per its spot-price API.

The Business Engineer Lens

This story maps onto the Business Engineer framework The CFO’s Guide to the Token Economy.

The framework puts its thesis in one line: “The cost of intelligence is collapsing while the cost of deploying intelligence is exploding.”

As we read it, the Nebius data shows that line at the level of a single cloud’s price list: the GPU-hours on its newest chips sold closest to list as token prices fell, according to Baker. The framework was published in June 2026 and does not discuss Nebius.

What Is Not Established

Our figures come from one cloud’s public spot market. Nebius does not say how much of its revenue comes from spot capacity, and its price page lists its largest systems, GB200 NVL72 and GB300 NVL72, as “Contact us”, so their prices are not public. Averages of hourly prices are not a measure of what customers actually paid, because the data does not show how many GPU-hours sold at each price.

The Volozh and Baker quotes are from Fireside Alpha’s transcripts of video clips; we did not find a full recording of the session.

The Bottom Line

Nebius’s own price data shows its B300 spot hours averaging 49% to 63% of the $9.50 list price between 8 and 21 October 2026, while H100 hours averaged 18% of list. As we read it, that is the GPU-hour side of the token price story Baker described, measured in one cloud’s numbers.