Nebius buys Inferize to cut the idle GPU tax — utilisation is the cash print
Nebius today acquired Inferize into Token Factory. Terms were not disclosed. The cash print is elastic inference on racks already in the hall — not another megawatt.
WIRE · NEOCLOUDS · EU · Data as at 1 Oct 2026
Nebius today acquired Inferize and folded the team into Token Factory. Terms were not disclosed. The cash print is elastic inference on racks already in the hall — not another megawatt. On the neoclouds layer the stuck step is utilisation: who can sell the next token without holding spare GPUs warm to hit the SLA.
What happened
On 1 October 2026 Nebius said it has acquired Inferize and put the team into Token Factory. The named problem is cold starts — load time before the first request — which leave assigned GPUs idle at launch, when demand spikes, when new instances spin up, and when weights update mid-run (including reinforcement learning). That idle time, Nebius says, forces platforms to hold spare capacity to hit service-level targets. Inferize “cuts this idle GPU tax,” so capacity can scale closer to usage (Nebius newsroom, 1 Oct 2026 — company; same text on the EQS issuer wire, 13:00 CET). Founded in January 2026, Inferize had a working prototype within three months. Terms were not disclosed.
The same calendar day, the public on-demand GPU card stepped up from 1 October for B300, B200, H200 and H100 — the hike this desk covered on 21 September, now live (Nebius Compute pricing — company). Token Factory’s Q3 notes added Batch Inference 2.0, in private beta, which “scales to zero between jobs” (Nebius, What was new in Q3, 1 Oct 2026 — company).
| Figure | Date | |
|---|---|---|
| Event | Nebius acquires Inferize into Token Factory | Company, 1 Oct 2026 |
| Terms | Not disclosed | Same |
| Founded / prototype | January 2026 / within three months | Same |
| Named problem | Cold starts → “idle GPU tax” | Same |
| Eigen AI | ~$643m; signed 1 May, closed 10 Jun 2026 | Company |
| On-demand B300 | $7.85 → $9.50 / GPU-hour | Price list, from 1 Oct 2026 |
| On-demand B200 / H200 / H100 | $7.15→$8.50 / $4.50→$5.40 / $3.85→$4.50 | Same |
| Preemptible B300–H100 | Unchanged vs the pre-1 Oct on-demand column | Same |
On-demand percentages are our arithmetic: (new − old) / old. B300 is 1.65 / 7.85 = 21.0%. B200 is 1.35 / 7.15 = 18.9%. H200 is 0.90 / 4.50 = 20.0%. H100 is 0.65 / 3.85 = 16.9%. That 17–21% band is not Inferize’s price.
Why cash cares
On the neoclouds layer the usual earn is billed accelerator time: utilisation times realised price, after hardware carry, power, and hall cost. See What Is GPU as a Service?. Inferize does not add a hall. It tries to make more of the existing fleet billable. Who gets paid is the operator that can turn a spike into tokens without spare GPUs sitting warm. Who pays, on the Token Factory meter, is still the customer buying tokens. The second-order move is on Nebius’s cost line: every idle GPU-hour at today’s card is list carry with no token against it.
Our arithmetic, from the 1 October price list, not from a run:
- One spare B300 is $228 a day of list. $9.50 × 24 = $228 if the card sits assigned and serves zero tokens. One hundred spare B300s is $22,800 a day — $8.3m a year if that headroom never billed (100 × 9.50 × 24 × 365 = $8,322,000). That is list, not realised, and not a company saving. It is the size of the tax Inferize is hired to cut.
- The hike and the tax sit on the same day. On-demand is already 17–21% higher. Preemptible rows did not move — B300 preemptible is still $4.30 against $9.50 on-demand.
- Eigen was a $643m inference buy. Signed 1 May 2026, closed 10 June. Clarifai, 12 May, had no disclosed terms. Do not back-solve Inferize from Eigen.
The 21 September Wire on the October PAYG hike was the rate-card print. Today is software that tries to raise utilisation on the same card.
What idle GPU tax is
Idle GPU tax is not take-or-pay. Take-or-pay bills the customer for reserved hours after delivery, used or not. Idle GPU tax, as Nebius defined it today, is the operator holding extra GPUs so Token Factory can still hit latency when a model is loading, a replica is spinning, or weights swap mid-run.
Cold start. Weights have to land in GPU memory before the first token; until they do, assigned cards are on the power bill and off the token meter. Headroom. Token Factory advertised a 99.9% SLA at launch on 5 November 2025 — spare capacity sitting ready (Nebius, 5 Nov 2025 — company). Guy Bortnikov, Inferize’s co-founder and CEO: “Keeping spare GPUs running is the price of being ready for demand.” Mid-run updates. Reinforcement learning that hot-swaps weights is the same tax. The Q3 notes already advertise Batch Inference 2.0 that scales to zero between jobs — the batch twin of Inferize’s online claim.
Eigen made tokens cheaper per card. Clarifai was the system layer. Inferize is supposed to make the fleet thinner for the same SLA. Software does not connect a dark hall. It can change how many billed tokens a live hall produces. The Stack still has connecting contracted megawatts as the delivery bind.
The second-order read
| Who | Seat | What this means |
|---|---|---|
| Nebius / Token Factory | Operator | Paid on tokens; idle headroom is unbilled carry at the new card |
| Inferize team | Seller of elasticity | Joins Token Factory; cash terms not disclosed |
| Token Factory customers | Who pays | Still buy tokens — no automatic discount because Nebius cut its idle |
| Eigen + Clarifai | Already inside | Model/kernel ($643m) and system layer (terms n.d.) |
| On-demand GPU buyers | Same-day card | 17–21% higher from 1 Oct; preemptible unchanged |
| Nvidia / HBM / CoWoS / power | Upstream | No new tray or watt; more tokens from attached cards |
Who gets paid. Nebius, if spare-GPU hours become serving hours at a card that just went up. Inferize’s founders get paid on an undisclosed cheque. Who pays. Token Factory customers still pay the meter. Upstream. GPUs, HBM, CoWoS and power still bind the next hall. Downstream. Interactive agents feel cold-start first; batch can already scale to zero.
By clock. Rate card: 1 October on-demand step, live today. Software: Inferize acquired today, integration “starting.” Funding: Eigen closed 10 June at ~$643m; Inferize unknown. Delivery: 5 GW contracted versus 0.8–1 GW connected (Q2 guide). Our read: the stuck step is utilisation of live inference seats. Ready is not paid. Spare is a cost.
Two cautions
Terms are not disclosed, and no millisecond, GPU-count, or utilisation percentage printed. Do not treat “cuts this idle GPU tax” as a measured saving. Do not import Eigen’s $643m, or Token Factory’s November 2025 “up to 70%” fine-tuning claim, as Inferize results. Integration has started; it has not printed.
Do not mix books. Take-or-pay idle is a customer invoice after delivery. On-demand idle is unbilled operator carry. Preemptible / spot is a third clock — dynamic PVM pricing takes effect 8 October, announced 22 September, and is interruptible capacity, not Token Factory headroom (Nebius, 22 Sep 2026 — company).
What would change the view
- A dated Token Factory utilisation or cold-start print after Inferize is in production.
- Disclosed consideration, or a filing that puts Inferize next to Eigen’s $643m.
- On-demand rows reversing the 1 October step while installed GPUs keep rising.
- Connected power at year-end 2026 well below the 800 MW–1 GW Q2 guide.
What to watch
- Integration: next Token Factory note that names Inferize in production, with a cold-start or idle-GPU figure.
- Price list: whether the 1 October on-demand column holds through October, and whether preemptible rows stay flat.
- 8 October: PVM spot pricing goes live — a different idle pool.
- Q3 2026 results: connected megawatts versus the 0.8–1 GW guide, and any Token Factory utilisation language.
- Batch Inference 2.0: from private beta to a dated scale-to-zero print.
Read: What Does Nebius Do? The Neocloud Prepay Model · What Is GPU as a Service?
Get the Wire. A free brief every day, here and in your inbox. Subscribe free →
Sources: Nebius newsroom, Inferize acquisition, 1 Oct 2026 — company; EQS issuer wire, 1 Oct 2026 — company; Nebius Compute pricing — company (1 Oct on-demand rows); Nebius, What was new in Q3, 1 Oct 2026 — company; Nebius, Eigen AI agreement, 1 May 2026 — company; Nebius, Eigen AI close, 16 Jun 2026 — company; Nebius, Clarifai team and licence, 12 May 2026 — company; Nebius, Token Factory launch, 5 Nov 2025 — company; Nebius, PVM spot pricing, 22 Sep 2026 — company. Not investment advice.
