Neocloud: what it is and how the GPU cloud market works

Definition

A neocloud is a cloud provider whose core product is GPU compute for AI training and inference. The term covers companies such as CoreWeave, Nebius, Lambda and Crusoe, plus dozens of smaller GPU clouds and marketplaces. Several started in other businesses. Crusoe sold its bitcoin mining business to NYDIG in 2025 to focus on AI infrastructure (Crusoe). IREN still mines bitcoin, but its AI cloud brought in $200 million in the September 2026 quarter against $28 million from mining (IREN 8-K).

What most neoclouds have in common:

• The latest NVIDIA parts first (and sometimes AMD Instinct), sold per GPU-hour.
• 8-GPU servers linked by InfiniBand or RoCE so jobs can span many nodes.
• Kubernetes and Slurm as the main ways to schedule work.
• A short product list: GPU instances or bare metal, storage, networking, and sometimes inference APIs.
• Revenue concentrated in a few very large AI customers.

Terms like SXM, NVL72 and RoCE are defined in our GPU cloud glossary.

Neoclouds vs hyperscalers

AWS, Azure, Google Cloud and Oracle also rent GPUs. The difference is focus. A hyperscaler wraps GPUs in a large platform: identity, managed databases, data warehouses, private networking to everything else you run. A neocloud gives you the GPUs and the fabric and expects you to bring the rest.

Hyperscalers sit in the same third-party ratings. SemiAnalysis's ClusterMAX 3.0 placed Oracle and Google Cloud at Gold, Azure at Silver and AWS at Bronze, while two neoclouds, CoreWeave and Nebius, were the only Platinum entries among neoclouds (SemiAnalysis). Those are SemiAnalysis's tiers, not ours. We did not collect hyperscaler GPU prices, so this guide makes no price claims about them.

How neoclouds sell compute

Four business models cover most of the market. Many providers run more than one.

On-demand instances

You pay per GPU-hour, often billed per second or per minute, and stop when you want. On 2026-11-09, published on-demand H100 SXM list prices ran from $2.45 at Denvr Dataworks to $6.155 at CoreWeave, and B200 from $5.43 at Massed Compute to $9.138 at Verda. The full table is on our GPU cloud pricing page, and the hardware trade-offs are in H100 vs H200 vs B200.

Reserved clusters

You commit to a block of GPUs for weeks to years in return for a lower rate and guaranteed capacity. Terms vary a lot. Lambda's 1-Click Clusters have a two-week minimum, Together AI publishes reserved rates from 3 days to 180 days, and DigitalOcean's reserved pricing needs a 12-month contract (Lambda, Together AI, DigitalOcean). Our GPU cluster rental guide covers terms and quotes.

Bare metal

You get whole servers with no hypervisor, which gives full control over drivers, kernels and the network stack. Latitude.sh, for example, publishes hourly prices for bare-metal GPU servers, with 8x B300 servers starting at $64 per hour on 2026-11-09 (Latitude.sh).

Marketplaces and aggregators

A marketplace lists capacity from many hosts or clouds. Vast.ai's prices are set by hosts and change hourly. Shadeform aggregates about 18 clouds and says you pay the same as going direct. SF Compute runs an order book for blocks of time (Vast.ai, Shadeform, SF Compute).

The neocloud market in 2025 and 2026

Equity rounds got very large

Crusoe raised a Series F of $3.9 billion or more, reported at about a $30 billion valuation, in September 2026 (Bloomberg Government). Firmus announced a fully subscribed US$2 billion equity investment in August 2026 (Firmus). Together AI closed an $800 million Series C at $8.3 billion in July (Latham & Watkins), and Fluidstack raised $830 million at $7.5 billion in January (Fluidstack). TechCrunch reported in October that Lambda was raising up to $4 billion at a $14.5 billion pre-money valuation and had pushed its IPO to 2027 (TechCrunch).

Debt is the stress point

GPUs are bought up front and paid off from contracts over years. Most of that gap is funded with debt.

• CoreWeave had $35.551 billion of debt principal at 2026-06-30. In the June quarter it reported $2.575 billion of revenue, a $626 million net loss and $640 million of net interest expense, up from $267 million a year earlier (CoreWeave 10-Q). On 2026-10-14 its 2032 notes reportedly yielded about 15% and the stock fell about 15% in a day. Nebius fell 11% the same day (The Wealth Advisor, citing Bloomberg and FT).
• Nebius closed $5.75 billion of convertible notes in August 2026 and a $3.5 billion, 364-day facility from Citi in October. The Citi release cites more than $40 billion of long-term contracts (W.Media, Nebius).
• IREN took $3.65 billion of GPU financing backed by its Microsoft contract and $3 billion of convertibles (The Block). Its Q1 FY27 report says about $6 billion of the remaining $15 billion of FY27 capex is still to be financed (IREN 8-K).
• Nscale filed an S-1 for a NYSE listing. Coverage of the filing reports a $1.02 billion net loss in the first half of 2026 on $140.6 million of revenue (Pulse 2).

Customers are concentrated

A few AI labs and hyperscalers buy most of the capacity. CoreWeave's 10-Q shows one customer at 36% of June-quarter revenue, a second at 26% and a third at 10% (CoreWeave 10-Q). Lambda's backlog went from $15 billion in June to $50 billion in September 2026, largely on a $35 billion Anthropic commitment (TechCrunch). IREN reports $20.3 billion of contracted backlog (IREN 8-K). At Applied Digital, one customer was 56% of revenue in the quarter to 2026-08-31 (Applied Digital 10-Q).

For a buyer, concentration cuts two ways. A large anchor contract funds the build-out you rent from. It also means smaller customers compete with that anchor for new capacity and attention.

Consolidation and renames

Several providers changed names or owners. If a vendor list or old blog post names one of these, this is where it went.

Some providers also moved away from public cloud. Fluidstack's site now focuses on building data centers for frontier labs, including a $50 billion Anthropic deal, and its old pricing pages return 404. CUDO Compute now quotes clusters only, starting at 128 GPUs (Fluidstack, CUDO).

Prices are rising, not falling

The common expectation was that GPU-hour prices would keep falling as supply grew. Published list prices in late 2026 point the other way.

• Crusoe's pricing page announces on-demand increases from 2026-12-01: H100 from $3.90 to $4.50, H200 from $4.29 to $5.40, B300 from $8.96 to $9.90, and a new B200 rate of $8.50 per GPU-hour (Crusoe).
• IREN's Q1 FY27 report says H100s were re-contracted "at double the previous pricing" after the quarter (IREN 8-K).
• DigitalOcean introduced new GPU pricing effective 2026-08-01, and Lambda publishes a new schedule effective 2026-11-16 (DigitalOcean, Lambda). Lambda shows current prices only after login, so we cannot say which way that change goes.
• SemiAnalysis writes that Nebius "commands premium pricing" with prepayments on one-year deals (SemiAnalysis).

Prices are also getting more visible. Nebius now publishes live spot prices and syndicates them to Shadeform, Silicon Data and GPU Finder (Nebius). We do not have a sourced 2025 price index, so we do not publish a percentage trend.

What this means if you rent GPUs

Our editorial view, based on the facts above:

• Lock in what you need, not what you might need. With list prices rising, a reserved term protects your rate. It also ties you to one provider's balance sheet. Size the term to your roadmap.
• Check the counterparty. For any reservation longer than a few months, read the provider's latest filing or funding news. Large debt, a single dominant customer or unfunded capex are reasons to ask harder questions, not automatic deal-breakers.
• Keep your exit cheap. Container images, checkpoints in object storage you control, and infrastructure as code make a move between providers a matter of days.
• Re-check prices often. We saw three providers change or announce list prices within four months. Our pricing table shows the date each price was seen.

Ready to compare? Start with how to choose a GPU cloud, our 11-step checklist.

Large neoclouds

These providers run their own multi-site GPU fleets and sell clusters of 1,000 GPUs or more.