Blog · Global · · 5 min read

Neocloud or hyperscaler? What GPU capacity really costs.

Neocloud vs hyperscaler GPU rates for the H100, H200 and B200, read on 3 November 2026, with published commitment discounts and what else changes the bill.

Neocloud vs hyperscaler GPU pricing looks like a simple contest. The specialists charge less and the big clouds charge more. The rate cards say that is mostly true. They also leave out most of what you will pay and what you will get. This note sets out who sits on each side, what each publishes for the H100, H200 and B200, and what else moves the bill.

Who is on each side

A neocloud is a cloud built around GPUs. Newcomer, writing in May 2026, names Lambda and Crusoe as neoclouds that own their chips and sell the compute. CoreWeave and Nebius are two more. Regional ones sit in other markets: E2E Networks in India and Sakura Internet in Japan both publish GPU rates and commitment terms.

The other side is the hyperscalers: Amazon Web Services, Microsoft Azure, Google Cloud and Oracle. They sell GPUs as one product inside a much larger cloud.

What the rate cards say

All rates below are US dollars per GPU-hour, taken from each provider's own page on 3 November 2026. They are list rates. They are not what large buyers sign.

ProviderH100H200B200What the rate is
Lambda$3.99Not listed$6.69On-demand, 8-GPU instance
E2E Networks$3.99$5.29$7.99On-demand
Nebius$4.50$5.40$8.50On-demand, HGX
CoreWeave$6.16$6.31$8.60On-demand, North America
AWS$5.97$6.87$14.21Capacity Blocks, a short reservation
Google Cloud$9.80$10.71Not readVertex AI, managed service
Oracle$10.00$10.00$14.00Pay as you go
Microsoft Azure$12.29Not readNot readND H100 v5, East US

Three of these need a note. CoreWeave and Azure price a whole 8-GPU machine by the hour: $49.24 and $98.32 for the H100. The per-GPU figures above are our own arithmetic. AWS publishes a plain on-demand rate only in its calculator, so the AWS row is its Capacity Blocks price, and for the H200 it is the p5e machine (the p5en costs $7.90). Google's row is the Vertex AI managed service, which is not the same product as a bare virtual machine.

Even so, the spread is wide. The H100 runs from $3.99 to $12.29, about three times. The B200 runs from $6.69 to $14.21, about twice. Rate cards also move. In our read on 8 October, Nebius showed $3.85 for an H100. On 3 November it showed $4.50.

What commitments change

Every provider here says a commitment lowers the rate. Few publish by how much.

  • CoreWeave says it offers up to 60% off its on-demand prices for committed use. That is its own claim. The page lists no committed rate.
  • Azure publishes reservation prices through its public price API. For the H100 machine in East US, the one-year price works out 36% below pay as you go and the three-year price 56% below. We divided each term price by its hours.
  • Google Cloud lists Flexible Savings Plan rates on the Vertex AI page. For the H200 they are $9.64 for one year and $8.57 for three, which is 10% and 20% below $10.71.
  • E2E Networks says three-month, six-month and annual commitments can earn 10% to 30%, depending on length and volume.
  • Sakura Internet announced on 1 July 2026 a commitment plan for one H100 plan: 5% off for six months and 10% off for twelve.
  • AWS cut P5 (H100) prices by 44% for on-demand and 45% for the three-year Instance Savings Plan from 1 June 2025. Those are cuts from its earlier prices, not discounts off today's rate.
  • Nebius and Lambda name no percentage. Both point large buyers to sales.

The pattern is clear. The big clouds publish a formula. The neoclouds publish "up to" and a phone number. How GPU deals actually close covers what moves a private quote.

What else differs besides the rate

  • Network. Microsoft says each H100 in an ND H100 v5 machine has its own 400 Gbps InfiniBand link, 3.2 Tbps per machine. AWS says P5 machines offer up to 3,200 Gbps of EFA networking. The Lambda and CoreWeave pages we read give no per-GPU network speed. Ask for it in the quote.
  • Data out. AWS gives 100 GB a month free and meters the rest. Nebius lists traffic in and out as free on its price page. Lambda says it charges nothing for egress, its own claim.
  • Support. AWS Enterprise Support costs the greater of $5,000 a month or a share of your monthly AWS charges, at 10% of the first $150,000. We found no neocloud support terms on the price pages we read.
  • Contract length. Lambda's clusters run from two weeks to one year, and longer terms go through sales. AWS Capacity Blocks last days or weeks, start up to eight weeks ahead and cannot be cancelled. Azure reservations run one, three or five years.
  • Availability. AWS lists B200 Capacity Blocks in seven commercial regions. Every GPU cloud also needs a powered hall. CBRE put North American primary-market vacancy at a record 1.4% in the first half of 2026, with 80.4% of capacity under construction already preleased.
  • Credit. A contract is only as safe as the company behind it. SiliconANGLE reported on 11 August 2026 that CoreWeave ended the second quarter with $35 billion of debt, a $626 million net loss and a $104 billion backlog. The hyperscalers are far larger companies. A neocloud can still be the right choice. Ask for its accounts, and ask what happens to prepaid money if it fails. How to buy GPU capacity has the questions.

How to compare two quotes

  • Put both in one unit: dollars per GPU-hour, all in.
  • Match the chip, the network, the term and the start date.
  • Add or remove what differs: support, data out and storage.
  • Set the result against the rate cards above and a market index for the same chip.
  • Then ask what buyers of your size signed. That step needs executed prices, which no public page shows.

Casitel OS reads rate cards like these every day. A rate card is where a price starts. The gap between it and the signed price is what a benchmark has to close.

Sources

Comparing providers? Test the quote.

Put a neocloud quote next to a hyperscaler quote. See where each sits against the market, and what to ask for.

Benchmark a quote

Read next. More from the compute desk.

Ask an assistant about Casitel.

ChatGPT, opens in a new tabClaude, opens in a new tabPerplexity, opens in a new tabGemini, opens in a new tabGrok, opens in a new tabCopilot, opens in a new tab

The question: What is Casitel (https://casitel.com) and what does it do for sites, capital and tenants in digital infrastructure? Use casitel.com as the source.