A Nebius customer notice reshared on Reddit and X on 16 September lifts on-demand GPU-hour rates from 1 October: H100 $3.85 to $4.50, H200 $4.50 to $5.40, B200 $7.15 to $8.50, B300 $7.85 to $9.50. B300 is the largest step, $9.50 ÷ $7.85 = +21.0%; H100 the smallest, +16.9%. Old rates match Nebius's price page, still unchanged; Nebius has confirmed nothing. Shares closed that day at $209.37 and traded at $227.91 premarket on 17 September, up 8.9%, beside CoreWeave and IREN up 5% to 6%.
A fifth on the list need not be a fifth on revenue, and Nebius has never said how much revenue the list reaches. Its Q2 letter reports $574.9 million of AI cloud revenue and describes three contract types, the shortest three to six months, priced in annual contract value per megawatt: above $20 million on Q2's core deals, 70% with prepayments. On-demand never appears.
On-demand is a way of billing, not a new business line. Nebius reports three segments, Nebius AI cloud, Avride and TripleTen, and its Q2 operating review puts both pricing models inside the first: "We offer both on-demand 'pay-as-you-go' pricing and fixed 'reserved capacity' contracts." No filing splits revenue between the two. The list cannot reach the $37.5 billion of remaining performance obligations either: those cover prepayments and unbilled non-cancelable contracts, excluding anything of a year or less, and a pay-as-you-go hour commits nobody to anything.
At most $48 million a quarter, at Q2's usage
The Q2 financial statements do bound the revenue exposed.
- Three customers took 24%, 21% and 14% of Q2's $582.3 million of revenue: 59%, or $343.6 million.
- Nebius does not say how those three are billed. A customer paying a fifth of a quarter's revenue at list price would be unusual, so assume all three are reserved. That is our assumption, not a disclosure.
- The other 41% is $238.7 million. Removing TripleTen's $10.0 million and Avride's $1.0 million leaves $227.7 million, the most on-demand could have billed in Q2. The percentages are rounded, so the ceiling sits between $219 million and $237 million, and it still holds storage, inference and every smaller reserved contract.
- At unchanged usage, a 16.9% to 21.0% rise on $227.7 million is $38.5 million to $47.8 million a quarter, 6.7% to 8.3% of AI cloud's $574.9 million.
The softest step is how much of that $227.7 million is really on-demand, so here it is three ways:
| On-demand share of $227.7M | On-demand revenue | Uplift a quarter | Share of AI cloud revenue |
|---|---|---|---|
| 25% | $56.9M | $9.6M–$12.0M | 1.7%–2.1% |
| 50% | $113.9M | $19.2M–$23.9M | 3.3%–4.2% |
| 100% | $227.7M | $38.5M–$47.8M | 6.7%–8.3% |
Unchanged usage is the second assumption. It holds if on-demand is capacity Nebius has kept back and still sells out; the letter says it retains "some capacity for immediate customer needs" rather than selling its whole 2027 book. The rates start on 1 October, so 2026 gets one quarter of it, against the $2.0 billion to $2.4 billion the full-year guide needs from the second half.
Commitment pricing is the counter-case. Quoted at "up to 35% less than on-demand", new reservations could rise too, though the notice names on-demand only. Customers able to commit can also move to that discount, which would shrink the on-demand base the rise applies to. A disclosed on-demand share, or Q3 per-megawatt pricing moving with the list, would settle it.
Nebius updating prices starting October 1 — 16 September 2026
New rates and date are the customer notice's as reshared on Reddit. Current rates and the commitment discount are Nebius's price page, read 17 September; revenue, deal terms and the full-year guide its 12 August Q2 letter and call. Segments, customer concentration and remaining performance obligations are the Q2 interim financial statements, and the two pricing models its Q2 operating review, both filed 12 August. Share prices: 16 September close, 17 September premarket at 05:16 New York time. The on-demand ceiling, the share scenarios and every percentage are ours.
