Nebius just told its customers that AI compute is not getting cheaper. On September 17, 2026, the Amsterdam-based “neocloud” provider notified users it will raise pay-as-you-go prices for Nvidia GPU instances, AMD CPU capacity, and memory starting October 1, 2026. It’s the company’s second price hike in three months, and this time the increases stretch well beyond GPUs into the CPU and memory tiers that most cloud budgets treat as an afterthought.
The numbers are specific. Nvidia H100 instances rise roughly 17%, H200 climbs 20%, B200 goes up about 19%, and the newest B300 chips take the biggest hit at 21%, according to Yahoo Finance’s report on the Nebius announcement. AMD EPYC Genoa CPU rates jump 25%, and memory pricing rises about 41%, the steepest increase across the board. Investors read it as a sign of tight capacity, not weak demand: Nebius shares (Nasdaq: NBIS) jumped as much as 11.1% in pre-market trading the same day.
This is a Nebius cloud price increase story, but it’s really a story about what happens when AI compute demand keeps outrunning the physical supply of chips, memory, and power. Below is a full breakdown of what changed, why it changed now, how Nebius stacks up against AWS, Microsoft Azure, Google Cloud, and specialist rivals like CoreWeave and Lambda, and what it signals for anyone running AI workloads on rented infrastructure through 2027.
What Changed: GPU, CPU, and Memory Rates Broken Down
Nebius sells on-demand access to Nvidia GPUs, AMD CPUs, and memory-backed instances to AI labs and enterprises that don’t want to build their own data centers. The October 1 price sheet touches nearly every tier of that stack. Rather than a single across-the-board percentage, Nebius applied different increases to different chip generations, which is itself telling: the newest, most capacity-constrained hardware got hit hardest.
| Instance type | Price before Oct 1, 2026 | Price from Oct 1, 2026 | Increase |
|---|---|---|---|
| Nvidia H100 (per GPU-hour) | $3.85 | $4.50 | ~17% |
| Nvidia H200 (per GPU-hour) | $4.50 | $5.40 | ~20% |
| Nvidia B200 (per GPU-hour) | $7.15 | $8.50 | ~19% |
| Nvidia B300 (per GPU-hour) | $7.85 | $9.50 | ~21% |
| AMD EPYC Genoa CPU (per vCPU-hour) | Unpublished baseline | $0.015 | ~25% |
| Memory-backed instances | Unpublished baseline | Unpublished | ~41% |
Source: Yahoo Finance and Global Banking & Finance Review reporting on Nebius’s customer notice. For a team running a sustained H100 workload, that 17% translates to an extra $0.65 per GPU-hour, which compounds fast across a multi-GPU training cluster running around the clock.
What stands out is the memory line. A 41% jump on memory-backed capacity is the largest single increase Nebius applied, bigger than any of the GPU tiers. Memory rarely makes cloud pricing headlines because it’s usually bundled into instance pricing rather than billed separately, but Nebius broke it out this time, which suggests memory availability, not just GPU silicon, is now a binding constraint on how much compute the company can actually deliver.
Why October 1: Inside Nebius’s Second Hike in Three Months
Nebius didn’t invent this pattern in September. Around May 2026, the company already raised on-demand capacity prices by an average of 29% and spot capacity by 51%, according to reporting cited by Yahoo Finance’s coverage of the September announcement. That means Nebius has now raised prices twice in under half a year, with the spot-market increase in the spring actually exceeding the September GPU hikes in percentage terms.
Reuters framed the September move plainly: Nebius is raising prices because demand for computing power keeps rising faster than supply. That’s a different explanation than “costs went up, so we passed them on.” It’s closer to basic capacity rationing, the same mechanism that pushed cloud GPU rental rates up during the original 2023 H100 shortage, except this time it’s spreading into CPU and memory tiers that used to be treated as commodity add-ons.
Two price increases within three months also tells customers something about how Nebius is managing its own capacity commitments. Neoclouds like Nebius typically lock in multi-year supply deals with Nvidia and lease data center space and power years in advance, then sell that capacity on shorter-term contracts to customers. When retail-facing prices move twice in a single quarter, it usually means the wholesale side of that equation, chip allocation, power contracts, or memory supply agreements, is moving even faster underneath.
The Memory Squeeze Behind the 41% Jump
Every modern AI accelerator needs enormous amounts of high-bandwidth memory sitting next to the compute die, and every data center server around it needs conventional DRAM for the host system. Both markets have been tight through 2026 as AI infrastructure buildouts compete directly with PC makers, phone manufacturers, and enterprise server vendors for the same fabrication capacity at Samsung, SK Hynix, and Micron.
Nebius doesn’t manufacture memory, so a 41% increase on memory-backed instances is a direct pass-through of what it’s paying its own suppliers, or a signal that available memory capacity for new instances has simply shrunk. Either way, it’s the clearest read yet of how a hardware supply constraint outside AI chips specifically is bleeding into cloud bills. Newer GPU generations like the B200 and B300 also carry proportionally more HBM per chip than the H100 did, which is part of why B300 saw the steepest GPU-tier increase at 21%: it’s the most memory-hungry chip in Nebius’s current lineup.
For customers, this means the old mental model of “GPU cost is the whole story” is breaking down. A workload that’s light on compute but heavy on memory bandwidth, like serving large context-window inference, may now see cost increases closer to 41% than to the 17-21% headline GPU numbers, depending on which instance types it actually runs on.
How Nebius Stock Reacted: NBIS Jumps as Investors Cheer Higher Prices
Markets treated the price hike as good news, not bad. Nebius shares rose 11.1% in pre-market trading on the day of the announcement according to one Yahoo Finance report, while other outlets pegged the pre-market move closer to 9%. That gap between reports is normal for fast-moving pre-market trading, but the direction is consistent across every source: investors interpreted the hike as proof Nebius has pricing power, not as a sign the company is struggling with costs.
That reaction fits a broader pattern in how markets have treated AI infrastructure names throughout 2026. When a cloud provider raises prices because it genuinely cannot supply enough capacity to meet demand, investors read it as confirmation of a seller’s market. The alternative read, that rising input costs are squeezing margins and forcing defensive price hikes, would typically depress a stock rather than lift it. The market’s bet here is that Nebius’s customers have nowhere cheaper to go for the specific GPU generations they need, at least not without switching providers entirely.
Nebius vs AWS vs Azure vs Google Cloud vs CoreWeave: Who’s Cheapest Now
Cloud GPU pricing is notoriously hard to compare directly because providers mix spot, on-demand, reserved, and multi-year contract rates on the same public price list. But looking at H100 rental rates across the market as of September 2026 shows just how wide the spread is, and where Nebius now sits after its increase.
| Provider | H100 rate (approx., per GPU-hour) | Billing type |
|---|---|---|
| RunPod | $1.99 – $2.99 | On-demand / marketplace |
| Lambda | $3.29 – $3.99 | On-demand |
| Nebius | $4.50 (from Oct 1) | On-demand |
| CoreWeave | $4.25 – $6.16 | On-demand |
| AWS (P5, list price) | ~$6.88 (per-GPU basis) | On-demand |
| Google Cloud | Spot from ~$2.25; on-demand higher | Spot / on-demand |
Data compiled from CloudZero’s 2026 cloud GPU pricing comparison and IntuitionLabs’ Data Center GPU Pricing 2026 index. AWS and Google Cloud figures are derived from multi-GPU instance list prices divided down to a per-GPU basis, so they aren’t perfectly apples-to-apples with single-GPU neocloud rates, but they establish the rough band the market is trading in.
Even after the October 1 increase, Nebius’s $4.50 H100 rate lands in the middle of the pack, more expensive than RunPod and Lambda, cheaper than AWS’s list price, and roughly comparable to CoreWeave’s range. That positioning matters: Nebius isn’t pricing itself out of the market, it’s raising prices while staying competitive with the hyperscalers, which is exactly the kind of move a provider makes when it believes customers value reliability and available capacity more than shaving a dollar off the hourly rate.
FinOps Fallout: What the Hike Means for AI Startups’ Budgets
For a FinOps team, a 17-21% GPU increase and a 41% memory increase landing in the same billing cycle is not a rounding error. Startups that sized their runway around a fixed hourly rate now have to rebuild that model with a meaningfully higher baseline, and they have less than two weeks of notice between the September 17 announcement and the October 1 effective date.
A simple way to see the exposure is to model a mixed workload against the published percentage changes rather than trying to guess the company’s internal cost structure:
# Rough monthly cost delta after the Oct 1, 2026 Nebius hike
gpu_hours = 720 # e.g. one H100 running 24/7 for a 30-day month
old_gpu_rate = 3.85
new_gpu_rate = 4.50
cpu_vcpu_hours = 1440 # supporting CPU instance
new_cpu_rate = 0.015
memory_increase_pct = 0.41
memory_baseline_cost = 900 # example prior monthly memory spend
old_total = gpu_hours * old_gpu_rate
new_gpu_total = gpu_hours * new_gpu_rate
new_cpu_total = cpu_vcpu_hours * new_cpu_rate
new_memory_total = memory_baseline_cost * (1 + memory_increase_pct)
print("GPU delta:", new_gpu_total - old_total)
print("New CPU spend:", new_cpu_total)
print("New memory spend:", new_memory_total)Run that math across a fleet rather than a single GPU, and the memory line often ends up the biggest dollar swing, even though the GPU percentages get the headlines. Teams running memory-heavy inference workloads, long-context LLM serving, vector databases, retrieval-augmented generation pipelines, should audit that line item first, not last.
The practical response most FinOps teams are reaching for is contract renegotiation: locking in reserved or multi-year rates before the next hike, diversifying across two or three providers to preserve negotiating leverage, and re-testing whether older GPU generations (H100 instead of B200) can still hit performance targets at a lower blended cost.
Historical Context: How GPU Cloud Pricing Got Here (2023-2026)
The current cycle isn’t the first time GPU rental prices have moved sharply. When Nvidia’s H100 launched into the generative-AI boom in 2023, capacity was so scarce that on-demand rates carried steep scarcity premiums almost everywhere, with wide gaps between providers depending on who had secured allocation from Nvidia early.
Through 2024 and into 2025, as H100 supply caught up and H200 and early Blackwell systems started shipping, per-GPU rates broadly stabilized and even softened at some specialist providers as competition among neoclouds intensified. That’s part of why the current 2026 landscape shows genuine bargains at providers like RunPod and Lambda sitting well below $4 per H100-hour.
2026 broke that pattern. Instead of newer chip generations pushing older ones down in price as supply normalized, the market saw the opposite: broad-based increases across H100, H200, B200, and B300 simultaneously, plus the CPU and memory tiers that had previously been treated as stable. Nebius’s back-to-back hikes in May and September are the clearest documented example, but the direction, not the magnitude, is what makes this cycle different from 2023’s isolated H100 crunch: this time the scarcity is showing up across the entire hardware stack at once.
The Neocloud Business Model Under Pressure
Nebius belongs to a category the industry has started calling “neoclouds”: companies that don’t sell general-purpose enterprise cloud services the way AWS or Azure do, but instead specialize almost entirely in renting out Nvidia GPU capacity at scale. CoreWeave, Lambda, and Crusoe Energy occupy the same category. Their business model depends on locking in large, long-term GPU supply commitments from Nvidia, then financing the data center buildout against contracted customer revenue.
That model works well when demand is predictable and rising steadily. It gets stress-tested when demand spikes faster than the buildout schedule, which appears to be exactly what’s happening in 2026. A neocloud that raises prices twice in three months isn’t necessarily struggling, the stock reaction argues the opposite, but it is signaling that its existing capacity is fully booked and new capacity isn’t coming online fast enough to keep prices flat.
The risk for neoclouds broadly is reputational as much as financial: customers who feel repeatedly re-priced with two weeks’ notice have an incentive to diversify away, even if no single alternative is meaningfully cheaper today. Whether Nebius’s pricing power holds through 2027 depends heavily on how fast Blackwell-generation supply, and the memory that goes with it, actually normalizes.
Competitors’ Moves: Hyperscalers and the Broader Pricing Trend
Nebius isn’t the only cloud provider adjusting prices this fall. OVHcloud has scheduled its own pricing restructuring effective October 1, 2026, though its approach is different: rather than raising headline hourly rates, OVHcloud is unbundling services, like local storage and public IPv4 addresses, that were previously included in its B3, C3, and R3 instance plans, according to reporting from RocketDevs’ analysis of 2026 cloud cost trends. The advertised hourly instance price stays the same, but the effective bill for a fully configured server still goes up.
The available reporting doesn’t establish a confirmed 2026 price hike or cut from AWS, Microsoft Azure, or Google Cloud specifically tied to this memory and GPU capacity crunch, and none of the three hyperscalers has issued a comparable public price-increase notice to Nebius’s September 17 announcement. That asymmetry is itself notable: specialist neoclouds appear to be repricing faster and more visibly than the diversified hyperscalers, likely because GPU rental is a much larger share of their total revenue, leaving them less able to absorb rising input costs across a broader business.
That said, list prices from AWS, Google Cloud, and Azure already sit well above most neocloud rates for comparable hardware, so the practical effect for a customer weighing options is that the pricing gap between hyperscalers and specialists is narrowing, not widening, even without a formal hyperscaler price increase.
What This Means for AI Startups Choosing a Cloud Provider
Locking in rates versus staying flexible
The core tradeoff for any team picking a GPU cloud provider right now is between price stability and flexibility. Reserved and multi-year contracts, the kind CoreWeave and Nebius both offer, can lock in rates well below the on-demand price, sometimes reported as low as $2.25 per GPU-hour on 36-month terms according to IntuitionLabs’ pricing index. But those commitments remove the option to walk away if a competitor undercuts current rates next quarter.
Multi-cloud as a hedge, not just a resilience strategy
Running training and inference workloads across two or three providers used to be framed mainly as a reliability hedge against outages. In this pricing environment, it’s increasingly a cost hedge too: a team that can shift workloads between Nebius, RunPod, and CoreWeave has real leverage the next time any one of them raises prices with two weeks’ notice.
Predictions: Where AI Cloud Pricing Heads From Here
- Expect at least one more Nebius price adjustment before mid-2027 if the current pace, two hikes in three months, holds, especially on memory-backed instances where the 41% increase suggests the tightest constraint.
- Watch for other neoclouds, particularly CoreWeave and Crusoe Energy, to follow with their own increases within the next one to two quarters rather than absorbing rising input costs to undercut Nebius.
- Hyperscaler list prices are likely to stay nominally flat through the rest of 2026, but expect tighter reserved-capacity terms and less spot availability rather than outright rate hikes, since AWS, Azure, and Google Cloud have more room to cross-subsidize from other business lines.
- Memory pricing, not GPU silicon, becomes the pricing variable to watch most closely into 2027, since it’s the input showing the steepest percentage move in this cycle and it touches every instance type, not just the newest GPUs.
- Multi-year reserved contracts will likely become the default negotiating ask for any AI startup running sustained training workloads, replacing the on-demand-first approach that was common when GPU supply was looser in 2024 and early 2025.
Risks and Unknowns: Could Prices Reverse?
None of this is guaranteed to continue in a straight line. Cloud GPU pricing has moved down before, most notably during 2024 and early 2025 as H100 supply caught up with the initial post-ChatGPT demand spike. If Blackwell-generation production scales faster than expected, or if a slowdown in AI infrastructure spending materializes, the same neoclouds raising prices today could find themselves cutting rates to keep utilization high on capacity they’ve already built.
The bigger open question is memory. GPU supply constraints have a relatively well-understood fix: more fabs, more allocation, more time. Memory supply is entangled with consumer electronics demand (phones, laptops, PCs) in a way GPU silicon isn’t, which makes it harder to predict when the 41% pressure Nebius is passing through actually eases. Until that resolves, expect cloud bills tied to memory-heavy workloads to stay the least predictable line item in any AI infrastructure budget.
Frequently Asked Questions
Why did Nebius raise its AI cloud prices in September 2026?
Nebius said the increase reflects continued strong demand for AI computing power outstripping available supply. It’s the company’s second price increase in roughly three months, following an earlier round in May 2026 that raised on-demand capacity by an average of 29% and spot capacity by 51%.
How much more will Nvidia H100, H200, B200, and B300 instances cost on Nebius?
Starting October 1, 2026, H100 rises from $3.85 to $4.50 per GPU-hour (about 17%), H200 rises from $4.50 to $5.40 (about 20%), B200 rises from $7.15 to $8.50 (about 19%), and B300 rises from $7.85 to $9.50 (about 21%), according to Nebius’s customer notice reported by Yahoo Finance.
Why did memory pricing go up 41%, more than any GPU tier?
Memory-backed instances saw the steepest increase in this round because DRAM and high-bandwidth memory supply has stayed tight through 2026, with AI infrastructure buildouts competing against consumer electronics manufacturers for the same fabrication capacity. Newer GPU generations like the B300 also carry more onboard HBM, which compounds the effect.
Did Nebius stock go up or down after the price hike announcement?
NBIS shares rose, with reports ranging from roughly 9% to 11.1% in pre-market trading on September 17, 2026. Investors interpreted the price increase as evidence of strong, sustained demand rather than a cost-driven defensive move.
Is Nebius still cheaper than AWS, Azure, or Google Cloud after this increase?
For H100 instances, Nebius’s new $4.50 per GPU-hour rate remains below AWS’s list price of roughly $6.88 on a per-GPU basis, though it now sits above budget-focused specialists like RunPod ($1.99-$2.99) and Lambda ($3.29-$3.99). Direct comparisons are complicated because providers mix spot, on-demand, and reserved pricing on the same public sheet.
Are other cloud providers raising GPU prices too?
OVHcloud has a separate pricing restructuring effective October 1, 2026, unbundling services like local storage and public IPv4 that were previously included in some plans. No confirmed 2026 price hike from AWS, Microsoft Azure, or Google Cloud tied to this specific GPU and memory capacity crunch has been reported publicly as of this writing.
What should AI startups do to manage rising cloud compute costs?
FinOps teams are generally moving toward locking in reserved or multi-year GPU contracts before further hikes, spreading workloads across two or three providers to preserve negotiating leverage, and re-testing whether older GPU generations can still meet performance targets at a lower blended cost, particularly for memory-light workloads.
Will Nebius raise prices again in 2027?
There’s no official confirmation of further increases, but the pace of two hikes within three months suggests the company will keep adjusting prices if memory and GPU capacity constraints persist. Memory pricing in particular is the input to watch, given it posted the largest percentage increase in the September round.
