Affiliate disclosure: Some links in this article earn us a commission at no extra cost to you. We only recommend hardware we'd run ourselves. Read our full disclosure.

Cloud GPU compute just got more expensive again. Effective October 1, 2026, Nebius (NASDAQ: NBIS) raised its on-demand rates across all four of its current Nvidia platforms by roughly 12–21% — the second major increase inside a year. The H100 NVLink jumped from $3.85 to $4.50 per GPU-hour, the H200 from $4.50 to $5.40, and the flagship B300 now rents at $9.50 an hour. On its face it's one neocloud's price list. In reality it's the cleanest price signal yet that on-demand AI compute is being repriced as a scarce utility — and that shifts the break-even math toward hardware you own.

This is the third chapter in a story we've been tracking all year: the memory price crisis that tripled DRAM, the Rubin ramp squeezing day-0 supply, and now the rental side of the market repricing to match. The pattern is consistent — the demand for AI compute keeps outrunning wafer, HBM, and power supply, and every layer of the stack passes the shortage along.

The October 2026 rate card, line by line

Nebius publishes its list pricing openly, which makes the change easy to audit straight from the company's official price page. The on-demand, pay-as-you-go numbers that took effect October 1:

GPU platform Before After (Oct 1, 2026) Change
H100 NVLink$3.85 / GPU-hr$4.50 / GPU-hr~+17%
H200 NVLink$4.50 / GPU-hr$5.40 / GPU-hr+20%
B300~$7.85 / GPU-hr$9.50 / GPU-hr~+21%

Read together with independent EU cloud price trackers — which put the median H200 rental up roughly 25% year-over-year — the direction is unambiguous. And note the shape of the increase: it's the spread between committed and on-demand pricing that's widening, not demand cooling. On-demand is becoming the premium product; the discount goes to whoever signs a contract.

Why neocloud prices keep climbing

Three forces are compounding. First, HBM is the binding constraint: Rubin Ultra's memory demands and the broader AI buildout have absorbed the wafer and HBM supply that would normally flow into next-gen accelerators — the same dynamic inflating the DDR5 prices on our memory crisis tracker. Second, hyperscalers are signing away future capacity in multi-year blocks, which shrinks the pool available to spot buyers. Third, grid power has become a real queue: new data-center capacity waits on substations, not chips, so even fully-funded builds land late.

For neoclouds, the choice between reprice and ration isn't really a choice. When allocation is scarce, you raise the rate for the customers who didn't commit. That's mechanically what happened on October 1 — and with enterprise inference demand still compounding quarter over quarter, the pressure is structural, not cyclical.

The break-even math: renting vs owning

Here's the arithmetic that changed this month. An always-on H100 at $4.50/hr is about $3,240 per month, or ~$39,000 per year — before the March 2026 hike it was $27,750. That's the number small teams quietly absorbed when they "just left the instance running."

Now compare it to owning. A dual RTX 3090 workstation — still the value king for 70B-class quantized inference on vLLM or llama.cpp — runs $2,500–$3,500 on the used market, draws 700–900W under load, and costs roughly $60–$100/month in electricity. Against a continuous cloud workload it pays for itself in weeks; even at 50% utilization it breaks even inside a few months and then runs free. A single RTX 4090 build covers the same territory with better per-token efficiency for models up to ~34B at Q4/Q5, and a modern 96GB-class pro card like the RTX PRO 6000 holds entire 120B-parameter models locally — the same tier that European trackers now peg at ~€1,200/month to rent.

The honest caveats: local rigs don't do frontier-scale training, idle hardware still costs capital, and the memory crisis means GPU and VRAM prices themselves are inflated right now. The play isn't "local always" — it's that the threshold where owning wins dropped dramatically this month.

Four practical plays before your next bill

  • Lock committed-use pricing now. The widening on-demand spread means the discount for 1–3 year commitments just got materially more valuable. If your workload is steady, sign before the next reprice.
  • Move the baseline to hardware you own. Serve your steady-state traffic from a local rig and reserve the cloud for burst. Our on-device AI trend report shows the sub-34B tier is fully servable on consumer silicon in 2026.
  • Use spot for batch, never for latency-critical paths. Spot pricing still exists, but preemption risk means it belongs on batch jobs only.
  • Audit your idle GPU-hours. At the new rates, an autopilot'd H200 burning $5.40/hr over a weekend of zero traffic is a $260 mistake. Autoscale or consolidate — this is the cheapest "price cut" available.

Sizing a local rig before the next hike lands? 24GB-class cards are the value sweet spot:

Shop RTX 4090 inference cards on Amazon →

The bottom line: October's hike didn't change what AI costs — it changed who pays the shortage tax. Committed buyers and hardware owners are largely insulated; on-demand renters absorb the increase in full. As Rubin and MI450 supply land through 2027 most analysts expect rates to flatten, not roll back. Until then, every month of continuous on-demand inference is a month of equity you didn't build. If cloud is still your default, this is the quarter to price the alternative honestly.

FAQ

How much did Nebius raise GPU prices in October 2026? Effective October 1, 2026, on-demand rates rose roughly 17–21% across the four current Nvidia platforms: H100 NVLink $3.85→$4.50/GPU-hr, H200 $4.50→$5.40, B300 to $9.50. It's the second major increase inside a year.

Is this just a Nebius problem? No. Independent EU price trackers put the median H200 rental up ~25% year-over-year, and hyperscalers are locking capacity years ahead. On-demand compute is being repriced as the premium tier across the market.

When does a local inference rig beat renting? For continuous duty: an always-on H100 is ~$3,240/month at the new rates, while a dual RTX 3090 or 4090 rig costs $2,500–$3,500 to build and $60–$100/month in power — breaking even in weeks to a few months. Spiky training bursts still favor the cloud.

Will cloud GPU prices come back down in 2027? Supply improves as Rubin and MI450 ship, but HBM allocation and grid power remain the constraints. Expect flattening, not a reversal — which is exactly why owned capacity is the rational hedge.

Disclosure: This post contains affiliate links. We may earn a commission on purchases at no extra cost to you. See our affiliate disclosure for details.