Neoclouds vs hyperscalers for GPU capacity: what the price gap buys and what it costs you
In brief
Some neocloud H100s list at under half Google Cloud's on-demand rate, but the gap mostly closes once you commit. The prices, the filings and a worked example.
Steve Elliott · · 6 min read
On this page
On published list prices as of November 2026, an NVIDIA H100 hour costs $4.50 at Nebius against $11.06 on demand at Google Cloud, about 2.5 times more. Once you commit or schedule capacity, most of that gap closes: Google's three-year commitment works out to $4.86 per GPU-hour and an AWS Capacity Block to $5.97.
So the headline gap is mainly an on-demand gap. In our view, the real choice between a neocloud and a hyperscaler is not a rate card comparison. It is a trade between a lower price for flexible capacity and the things a hyperscaler bundles: your data, your existing commitments and a balance sheet that doesn't depend on customer prepayments to build capacity.
How big is the H100 price gap, really?
The table below uses each provider's own pricing page, as of November 2026, converted to a per-GPU-hour rate. We use the H100, which every provider here lists in comparable form.
| Provider and offer | Per H100 GPU-hour | How it's sold |
|---|---|---|
| Google Cloud A3 High, on demand | $11.06 | $88.49 per 8-GPU VM-hour |
| Google Cloud A3 High, 1-year commitment | $7.67 | Compute resource commitment |
| Google Cloud A3 High, 3-year commitment | $4.86 | Compute resource commitment |
| Google Cloud A3 High, Dynamic Workload Scheduler | $4.79 Flex-start, $5.20 Calendar mode | Variable price, updated infrequently |
| AWS p5.48xlarge Capacity Block | $5.97 | Paid up front; price set by supply and demand |
| CoreWeave HGX H100, on demand | $6.16 | $49.24 per 8-GPU instance-hour |
| Lambda HGX H100, on demand | $4.59 | Self-serve instances |
| Nebius H100, on demand | $4.50 | Per GPU-hour, excluding tax |
| Crusoe H100, on demand | $3.90 through November 30, 2026; $4.50 from December 1, 2026 | Per GPU-hour |
Three cautions. First, bundles differ: Google's VM includes 208 vCPUs, 1,871 GB of memory and 6,000 GiB of local SSD; AWS's p5.48xlarge has 192 vCPUs, 2 TiB and 8 × 3.84 TB of NVMe; CoreWeave's instance has 128 vCPUs and 2,048 GB. Second, AWS and Microsoft publish on-demand GPU rates through their calculators and price-list APIs, which readers should check; we use AWS's published Capacity Block price and leave Azure out. Third, prices move and most rows name no region. AWS's Capacity Block prices are next due to change in January 2027, and Crusoe's H100 rate rises about 15% on December 1, 2026.
Why does the gap shrink once you commit?
Google's on-demand GPU prices carry a large premium over its own committed and scheduled rates. Its A3 High three-year rate is 56% below its on-demand rate, and its Dynamic Workload Scheduler Flex-start price is cheaper than its spot price. These machine types get no sustained use or flexible committed use discounts.
Neoclouds publish on-demand rates and mostly keep committed prices private. Lambda is a partial exception: it lists B200 cluster prices for terms of two weeks to a year, but not H100. CoreWeave offers "up to 60% discounts over our On-Demand prices for committed usage", which would put an H100 near $2.46 at the limit. Lambda, Nebius and Crusoe point reserved buyers to sales. The flip side: Google's on-demand buyers pay the full gap and, like any on-demand buyer, get no promise of capacity.
The comparison that matters, negotiated commitment against negotiated commitment, needs quotes in hand. For how hyperscaler commitments are negotiated, see negotiating cloud commitments for capacity.
What does the hyperscaler premium buy?
- Your existing commitments. Azure spend draws down a Microsoft Azure Consumption Commitment. Spend with a third party counts only if bought through Microsoft Marketplace as an "Azure benefit eligible" offer, through the Azure portal. Microsoft's page says 100% of the pre-tax amount of such purchases contributes. If you have an unspent commitment, a cheaper neocloud hour can still cost more in total.
- Proximity to your data. If training data sits in a hyperscaler's object storage, moving it out costs money. Google's Premium Tier charges $0.12 per GiB for internet egress, falling above 10 TiB a month to $0.08 to North America and $0.085 to Europe. CoreWeave lists egress as free, and Crusoe says it doesn't charge for ingress or egress "at this time."
- Published scheduled capacity. On-demand anywhere promises a price, not capacity. AWS Capacity Blocks and Google's Calendar mode sell scheduled windows at published prices. Among neoclouds, Lambda publishes B200 cluster terms and the others quote.
- Compliance scope you may already have. If your security team has already approved a hyperscaler account structure, identity setup and audit evidence, a neocloud needs its own review. CoreWeave's trust center lists SOC 2 Type II and HIPAA reports for services it labels "CKS and Inference" and "BMaaS", plus an ISO 27001 and 27701 certificate, among others; check that the attestations cover the specific service you'd use.
What do you take on with a neocloud?
Counterparty risk, mostly. The latest filings from the two neoclouds here that report publicly show fast growth funded by debt and customer money.
| Q3 2026 | CoreWeave (10-Q and results) | Nebius (6-K exhibits) |
|---|---|---|
| Revenue | $3.66 billion, up from $1.37 billion | $980.8 million, up 571% |
| Net result | $709 million net loss | $37.5 million loss from continuing operations |
| Debt on balance sheet | About $46.9 billion, recourse and non-recourse (our sum of line items) | About $14.2 billion (our sum of current and non-current) |
| Contracted backlog | $132.3 billion RPO, 36% expected within 24 months | $38.7 billion RPO |
| Concentration | Two customers were 30% and 24% of quarterly revenue | No percentage given; the letter names Microsoft and Meta agreements |
| Delivery | Refers to data center delivery delays announced in November 2025 | Some colocation delivery dates moved into the first half of 2027 |
| Deferred revenue | $10.0 billion | About $9.4 billion (our sum of current and non-current) |
We read three practical risks from this for a buyer. These are our interpretations, not findings in the filings.
- Delivery dates can slip. Both companies disclose delays, and a reserved cluster that arrives a quarter late moves your training plan with it.
- You compete with anchor tenants. When two customers account for over half of CoreWeave's quarterly revenue, a mid-sized enterprise has less bargaining power on scheduling and terms.
- A prepayment is a credit exposure. Nebius says over 90% of Q3 contract value carried a prepayment, and that prepayments covered more than 100% of the associated capex. If you prepay, you are partly financing the build. Ask what happens to that money if delivery fails.
None of this says either company is unsound: revenue at both is growing fast, and CoreWeave raised its full-year guidance on November 10, 2026. The risk sits in a different place than with a hyperscaler, and it should be priced.
A worked example: one 28-day training run
This example is illustrative. A team needs 64 H100s for a 28-day fine-tuning run: 64 × 24 × 28 = 43,008 GPU-hours. We use 28 days because AWS sells Capacity Blocks in 1-day steps up to 14 days and 7-day steps up to 182 days, and they can't be cancelled; Google's Calendar mode runs up to 90 days. The 200 TiB training set sits in Google Cloud Storage.
| Option | Compute | Data egress | Total |
|---|---|---|---|
| Google A3 High, on demand | $475,700 | $0 | $475,700 |
| Google A3 High, Calendar mode | $223,600 | $0 | $223,600 |
| AWS Capacity Block | $256,800 | $16,700 from Google | $273,500 |
| CoreWeave, on demand | $264,700 | $16,700 from Google | $281,400 |
| Nebius, on demand | $193,500 | $17,700 from Google | $211,200 |
Egress is our calculation from Google's Premium Tier table: after the first free GiB, 1,023 GiB at $0.12, 9,216 GiB at $0.11, and 194,560 GiB at $0.08 to North America or $0.085 to Europe. Nebius lists its H100s only in its eu-north1 region in Finland, which also matters for data residency and latency. One-year and three-year commitments don't fit a 28-day job unless other work fills the term.
On demand, Nebius saves about $264,500 against Google, roughly 56%. Against Calendar mode the saving is about $12,400, or 5.6%, and it depends on 64 Nebius GPUs being free for four weeks, which on-demand doesn't guarantee. We think that $12,400 is the real price gap for this job, small enough that capacity certainty, security review time and whether the spend retires an existing commitment can decide it.
For a steady inference fleet over three years, the arithmetic changes again: the comparison becomes Google's $4.86 commitment against whatever discount a neocloud quotes off its on-demand rate. Break-even logic for reserved AI capacity is set out in AI capacity reservations vs cloud commitments.
How should you decide between a neocloud and a hyperscaler?
We suggest six steps.
- Compare like with like. Set committed against committed and scheduled against scheduled, and normalize the CPU, memory and storage bundled with each GPU.
- Price the data path. Add egress from wherever the data lives today, and back again for checkpoints and outputs.
- Net off commitments. Count what an unspent hyperscaler commitment would otherwise cost you, and check whether the neocloud is sold through that hyperscaler's marketplace as an eligible offer.
- Read the filings. Look at debt, concentration, delivery delays and prepayment terms, and ask how your contract treats late delivery.
- Get capacity in writing. For on-demand capacity from any provider, a published price is not a promise of availability at scale.
- Model utilization. Track GPU utilization: a committed GPU you use 60% of the time costs 1.67 times its rate per useful hour, wherever you buy it.
A GPU contract is a funding decision with a delivery date, a counterparty and an owner, which is why, in our view, it belongs in the same operating model as the AI product it serves and the AI total cost of ownership behind it.
Sources
- Google Cloud, Accelerator-optimized VM pricing (A3 High on demand, spot, Dynamic Workload Scheduler and compute resource commitments), accessed November 22, 2026
- Google Cloud, Network pricing (internet data transfer, Premium and Standard tiers), accessed November 22, 2026
- AWS, Amazon EC2 Capacity Blocks for ML pricing, accessed November 22, 2026 (next price update scheduled for January 2027)
- AWS, Amazon EC2 P5 instances, accessed November 22, 2026
- AWS, Find and purchase Capacity Blocks, accessed November 22, 2026
- Google Cloud, Calendar mode future reservations overview, accessed November 22, 2026
- CoreWeave, GPU cloud pricing, accessed November 22, 2026
- CoreWeave, Trust Center: certification status, accessed November 22, 2026
- Lambda, GPU cloud pricing, accessed November 22, 2026
- Nebius, AI Cloud pricing and Compute pricing documentation, accessed November 22, 2026
- Crusoe, Crusoe Cloud pricing, accessed November 22, 2026
- Microsoft Learn, Azure Consumption Commitment benefit, last updated September 25, 2025
- CoreWeave, Form 10-Q for the quarter ended September 30, 2026
- CoreWeave, CoreWeave Reports Strong Third Quarter 2026 Results, November 10, 2026
- Nebius Group, Q3 2026 results press release, Form 6-K Exhibit 99.1, November 12, 2026
- Nebius Group, Q3 2026 shareholder letter, Form 6-K Exhibit 99.2, November 12, 2026
Frequently asked questions
Are neocloud GPUs cheaper than AWS or Google?
On published on-demand prices, yes, by a wide margin: as of November 2026, an H100 lists at $4.50 an hour at Nebius against $11.06 on demand at Google Cloud. Against hyperscaler commitments and scheduled capacity, the gap is much smaller. Google's three-year A3 High rate is $4.86 per GPU-hour, and an AWS Capacity Block for an H100 is $5.97.
Do neoclouds charge for data egress?
CoreWeave lists internet egress as free, and Crusoe says it does not charge for ingress or egress at this time. Nebius doesn't charge for compute networking egress, though one object storage tier does. The bigger cost is often the other direction: moving data out of a hyperscaler, where Google, for example, charges $0.08 to $0.12 per GiB on its Premium Tier.
Can neocloud spend count toward a hyperscaler commitment?
Only in limited cases. Microsoft says purchases of Marketplace offers it marks as eligible for the benefit, made through the Azure portal under your agreement, count 100% of the pre-tax amount toward an Azure consumption commitment. A neocloud contract signed directly doesn't count. Check whether your provider has an eligible listing before assuming a lower hourly rate means a lower total cost.
What counterparty risks should we check before signing a neocloud contract?
Start with the public filings. CoreWeave reported a $709 million Q3 2026 net loss, about $46.9 billion of debt and two customers making up 30% and 24% of quarterly revenue. Nebius reported colocation delays into 2027 and that over 90% of Q3 contract value carried a prepayment. Ask how your contract handles late delivery and protects any prepayment.
Steve Elliott
Published November 22, 2026
More on AI Economics
Tokens, inference cost, cost per task.
Keep reading
Hyperscalers' AI capex: what it means for the prices enterprises pay
Q3 2026 filings show about $237 billion of quarterly capex across five companies. How depreciation and utilization shape what enterprises pay for AI.
Negotiating cloud commitments when capacity is scarce
AWS and Google say demand outruns supply, and all three report backlogs over $550B. What that does to your bargaining power and which terms still move.
The Product Communication Gap That Kills User Adoption
How poor product communication causes missed features and outdated teams, and how audience-aware, multi-channel messaging drives real adoption.
Product Aligned vs Capacity Aligned: A Comprehensive Comparison
Product-aligned teams own a product area long term, while capacity-aligned teams are assigned work as demand shifts. Compare the trade-offs of each structure.