Chapter 2.1010 of 12 in this part

The neocloud landscape

Two of these companies file with the SEC, and one discloses that a single customer is 67% of its revenue. The other five disclose nothing, because they don't have to. That is the tier boundary — not price.

10 min read·revised 2026-08-08

Every neocloud comparison you will read ranks them by price. Price is the least durable thing about them. A vendor that is cheap today and gone in eighteen months has not saved you money — it has handed you a migration.

Two of these companies are publicly listed and file audited disclosures. That makes a different kind of comparison possible, and it is more useful than the rate card.

What CoreWeave discloses

From CoreWeave's FY2025 Form 10-K, the customer table:

2025 2024 2023
Customer A 67% 62% 35%
Customer B 15% 17%
Customer C 21%

(— indicates the customer did not represent 10% or more of revenue.)

One customer is two-thirds of the business, and its share has roughly doubled in two years. The filing does not treat this as transitional:

We expect that our customer concentration with a limited number of top customers is likely to continue in future years because of the long-term nature of contracts with those customers.

The other side of that coin is the backlog. As of 31 December 2025 CoreWeave reported $60.7 billion of unsatisfied remaining performance obligations, an increase of 302%, with a disclosed recognition schedule: 43% over the initial 24 months, 38% between months 25 and 48, and the remainder between months 49 and 84.

Read those two facts together, because they pull in opposite directions. Sixty billion dollars of contracted future revenue is a real answer to "will this vendor exist in three years." A single counterparty at 67% is a real answer to "what happens if one relationship changes." Both are true, both are disclosed, and neither appears on a pricing page.

What Nebius discloses

Nebius Group files a 20-F. Its concentration profile is a different shape:

2025 2024
Customer A 25% 27%
Customer B 15%
Customer C 11%

Top-customer concentration of 25% against CoreWeave's 67% — materially more diversified on revenue.

But the filing discloses a second concentration that revenue tables hide. On credit exposure through accounts receivable: $597.0 million, 83% attributable to Customer D in 2025 — against $6.6 million, 59% to Customer A the prior year.

So the two companies carry concentration in different places. CoreWeave's is in revenue; Nebius's largest single-name exposure is in money already owed but not yet collected. A customer that stops paying hurts these two businesses through different line items, and only the filings tell you which.

The five that disclose nothing

RunPod, Lambda, Crusoe, Modal, and Vast.ai are not SEC registrants. Searching EDGAR's company index for them returns nothing, which is not a criticism — private companies are not required to file, and most good companies are private.

It is a statement about what evidence is available to you. For those five there is no audited revenue, no customer-concentration table, no going-concern opinion, no disclosed contract duration. You are evaluating them on their marketing pages, their uptime, and whatever your account manager says.

That is the real tier boundary in this market, and it does not correlate with price or quality. It correlates with whether an auditor and a regulator have looked at the numbers you would want before putting production on the platform.

The rates, per GPU-hour

Normalising everything to one H100 GPU-hour makes the comparison legible. AWS's p5.48xlarge carries eight H100 SXM, so its rates divide by eight:

Provider / product $/GPU-hour
AWS on-demand (p5.48xlarge ÷ 8) $6.88
Lambda 1-Click Cluster, 16 GPUs $6.16
Lambda 1-Click Cluster, 256 GPUs $5.54
Lambda instance, 2× H100 SXM $4.19
Lambda instance, 8× H100 SXM $3.99
Modal serverless H100 $3.95
RunPod Pod, H100 SXM $2.99
RunPod Pod, H100 PCIe $2.89
AWS 3-year committed (÷ 8) $2.59

The headline reading is the obvious one: the cheapest neocloud on-demand rate is 2.38× cheaper than AWS on-demand. That is the number the whole category is sold on.

The second reading is the one that matters. RunPod's on-demand $2.89 — the lowest neocloud rate in the table — is 11.7% more than AWS's three-year committed $2.59, which undercuts it by 10.5%. The neocloud discount is real against a hyperscaler's list price and disappears entirely against a hyperscaler's committed price.

That reframes the decision. If you have three years of conviction, the hyperscaler is already competitive on rate, and you are choosing neoclouds for availability, simplicity, or terms — not for money. If you do not have three years of conviction, commitment maths says you should not sign a commitment anyway, and the neoclouds win by a wide margin.

Lambda's pricing has a shape worth noticing

Two patterns in Lambda's own table, both counterintuitive.

Within instances, the per-GPU price falls as the node grows: 2× at $4.19, 4× at $4.09, 8× at $3.99. Standard volume behaviour.

Across products, the cluster costs more per GPU than the instance: an 8× H100 instance is $3.99/GPU-hour, while a 16-GPU 1-Click Cluster is $6.16 — 1.54× more. Buying twice as many GPUs more than halves your efficiency per dollar.

That is not a pricing error. The cluster product is a different good: guaranteed co-location and an interconnect fabric that makes 16 GPUs behave like one machine, which parallelism strategies explains is the difference between a training run that scales and one that doesn't. You are paying for the network, not the silicon. But it means "we need more GPUs" and "we need a cluster" are separate purchases at very different prices, and conflating them will surprise you by half.

Even at 256 GPUs the cluster rate ($5.54) never falls below the 8-GPU instance rate ($3.99). Scale does not buy you back into instance pricing.

Vast.ai is a different kind of company

Vast.ai describes itself plainly: prices are "set by supply and demand across 40+ data centers," and it runs a hosting programme through which third parties list their own hardware. It offers three tiers — On-Demand ("Guaranteed uptime. Best for production"), Interruptible ("50%+ cheaper… Preemptible — may be reclaimed"), and Reserved (1, 3, or 6 month terms).

This is a marketplace, not a cloud, and the distinction is the whole risk profile. On the others you have one counterparty who owns the machines. Here your counterparty is a platform, and the hardware belongs to hosts you did not select and cannot audit. Supply-and-demand pricing means your rate is not a rate — it is a quote that moves.

None of that makes it a bad choice. For interruptible batch work, a competitive marketplace is exactly the right structure and will beat any fixed rate card. It makes it a bad choice for anything where you need to know next month's price or who is physically holding your data.

The tiers, honestly

  1. Audited and disclosed — CoreWeave, Nebius. You can read the concentration, the backlog, and the contract durations before you commit. Both carry concentration; the filings tell you where.
  2. Established but private — RunPod, Lambda, Crusoe, Modal. Real hardware, real operations, published rates, no audited numbers. Judgement call on evidence you cannot verify.
  3. Marketplace — Vast.ai. Your counterparty is a platform; the machines are other people's; the price floats.

Tier is not quality. Several tier-2 vendors are excellent and cheaper than tier 1. Tier is how much you can know before deciding, and that should scale with what you are putting on the platform.

The diagnostic

  1. Does this vendor file with a regulator? If yes, read the concentration table before the rate card. If no, accept that you are deciding on unverifiable evidence and size the commitment accordingly.
  2. Have you compared against the hyperscaler's committed rate, not its on-demand rate? At $2.59 versus $2.89, AWS committed beats the cheapest neocloud on-demand price in this table.
  3. Do you need GPUs or a cluster? Lambda charges 1.54× more per GPU for the cluster product. That gap is interconnect, and you should only pay it if your job actually spans nodes.
  4. What is your counterparty? On a marketplace it is not the machine's owner.
  5. How long is your data on their platform? Concentration risk and egress cost compound: the vendor most likely to change is the one you can least afford to leave.
  6. Is the price a rate or a quote? Supply-and-demand pricing is a feature for batch and a liability for budgeting.

What this chapter is not saying

It is not saying use the listed companies. CoreWeave's 67% single-customer concentration is a disclosed risk that a private competitor might not have; disclosure reveals risk, it does not create it, and the vendors that reveal nothing may be carrying worse.

It is saying that "which neocloud is cheapest" is a question with a stale answer and "which neocloud has disclosed its concentration" is a question with a durable one. Two of these companies will tell you, in an audited document, how exposed they are. The rest will tell you their price. Those are not the same quality of information, and production workloads deserve the first kind.

Sources & methodcaptured 2026-08-08

Sources, captured 2026-08-08. CoreWeave: the customer-concentration table (Customer A at 67% / 62% / 35% for 2025 / 2024 / 2023, Customer B at 15% / 17%, Customer C at 21%), the statement that concentration "is likely to continue in future years because of the long-term nature of contracts with those customers," the $60.7 billion of unsatisfied remaining performance obligations, the 302% increase, and the 43% / 38% / remainder recognition schedule across months 1–24, 25–48 and 49–84 are quoted from CoreWeave, Inc.'s Form 10-K for the year ended 31 December 2025, filed 2026-03-02, read directly from SEC EDGAR. Nebius: the customer table (Customer A 25% in 2025 and 27% in 2024, Customer B 15% in 2025, Customer C 11% in 2024) and the accounts-receivable credit-concentration figures ($597.0 million / 83% attributable to Customer D in 2025; $6.6 million / 59% to Customer A in 2024) are from Nebius Group N.V.'s Form 20-F for the year ended 31 December 2025, filed 2026-04-30. Note that the two concentration figures measure different things — CoreWeave's 67% and Nebius's 25% are both shares of revenue and are directly comparable; the 83% is a share of gross accounts-receivable credit exposure and is not comparable to either, which is why the chapter treats it as a separate kind of concentration rather than a larger version of the same one. Customer identities are not disclosed in either filing and are not speculated about here. Rates: Lambda's instance prices ($3.99, $4.09 and $4.19 per GPU-hour for 8×, 4× and 2× H100 SXM) and 1-Click Cluster prices ($6.16, $5.85 and $5.54 per GPU-hour at 16, 64 and 256 GPUs) are from Lambda's public pricing page and are stated there as excluding applicable sales tax, VAT or GST. Vast.ai's "set by supply and demand across 40+ data centers" description and its three tier definitions are from Vast.ai's pricing page. RunPod's H100 PCIe $2.89 and H100 SXM $2.99 Pod rates and Modal's $0.001097/second H100 rate were captured in earlier chapters (hyperscaler versus neocloud, serverless GPU versus dedicated) and are reused rather than re-fetched. AWS's $55.04/hour on-demand and $20.69504/hour three-year all-upfront rates for p5.48xlarge were re-resolved from AWS's pricing feeds for this chapter. Computed by me and verified in a separate pass: the per-GPU-hour conversions (AWS $6.88 and $2.59 by dividing by the eight H100 in a p5.48xlarge; Modal $3.95 from its per-second rate), the 2.38× on-demand ratio, and the 1.54× Lambda cluster-versus-instance premium. The RunPod-versus-AWS-committed gap is stated from both bases deliberately — RunPod's $2.89 exceeds AWS's $2.59 by 11.7% measured off the AWS rate, and AWS undercuts RunPod by 10.5% measured off the RunPod rate. These are the same gap; an earlier draft attached the 11.7% figure to the "cheaper than" phrasing, which silently switches the denominator, and the verification pass caught it. Vast.ai's live marketplace rates are not quoted — the page renders them dynamically from current supply and demand, and a snapshot would misrepresent a floating price as a rate; the chapter's point about Vast is structural rather than numeric. Crusoe publishes no per-hour rate I could resolve and is described here only by tier, not by price. The absence of Lambda, Crusoe, RunPod, Modal and Vast.ai from SEC EDGAR was verified by searching the Commission's company-tickers index; absence there means they are not registrants, which is the only claim made about them. All per-GPU comparisons pair H100-class silicon but do not normalise for interconnect, storage, egress, or support, which differ substantially between these products — the Lambda cluster-versus-instance gap in the chapter is precisely an illustration of that. The tier framework and the diagnostic are my framing.

Want this done on your account rather than by you?

The handbook is the method, written out in full so you can run it yourself — that is the point of publishing it. If you would rather someone else did the first pass, the teardown is free and you keep the findings either way.