How the neocloud ranking works

Six public signals, fixed weights, one formula. The rank is computed from the sourced data on every build and nobody sets a position by hand. Method dated 11 Nov 2026.

Signals and weights

51 neoclouds

Signal	Weight	How points are given (0 to 100)	Disclosed by
ClusterMAX tier	30%	SemiAnalysis ClusterMAX 3.0 tier converted to points: Platinum 100, Gold 85, Silver 70, Bronze 55, Participation Ribbon 35, Not Recommended - Underperforming 15, Not Recommended - Unavailable or not listed 0.	47 of 51
Contracted power	20%	The larger of contracted and active MW, on a log scale across every neocloud that states one: the smallest stated figure scores 20, the largest 100.	11 of 51
Contracted backlog	15%	Revenue backlog, RPO or total contracted value in USD, on the same 20-to-100 log scale.	6 of 51
Capital raised	15%	Equity and debt raised as described in the capital note, on the same 20-to-100 log scale.	19 of 51
GPUs	10%	GPUs deployed, owned or under contract as stated, on the same 20-to-100 log scale.	9 of 51
Named customers	10%	10 points per publicly named customer, capped at 100 (10 customers).	21 of 51
Total	100%	Score = sum of (points x weight) / 100. Rank = position by score; equal scores share a rank.

Why these signals

ClusterMAX tier (30%) is a public, hands-on test of how well a GPU cloud actually runs clusters: SemiAnalysis rents capacity and grades reliability, networking, storage, orchestration and support. It gets the largest weight because it measures the service a buyer gets. Points per tier: Platinum 100, Gold 85, Silver 70, Bronze 55, Participation Ribbon 35, Underperforming 15, Unavailable 0. Providers SemiAnalysis lists as Unavailable score 0, the same as providers it does not list.

Contracted power (20%) and backlog (15%) measure scale that is already committed: megawatts a company has under contract and revenue customers have signed for. Both come from filings and investor releases where they exist. We use the larger of contracted and active MW, so a company that reports only live capacity still scores.

Capital raised (15%) measures the ability to keep buying GPUs. Neoclouds finance mostly with debt against contracts, so equity and debt both count. Each record’s capital note lists exactly which rounds, notes and facilities the figure covers.

GPUs (10%) and named customers (10%) round out scale and demand. Few large neoclouds publish a GPU total, so the weight stays low. Customers count only when the company, a filing or reputable press names them.

Scaling

Power, backlog, capital and GPUs span four orders of magnitude, so they are log-scaled: among the companies that publish a figure, the smallest scores 20 and the largest 100, with everything between placed on a logarithmic line. A company that publishes nothing for a signal scores 0 on it. That rewards disclosure: 1 of 51 neoclouds publish all six signals, and publishing a sourced figure is the fastest way up the table.

What counts as a source

Filings (SEC, ASX), investor releases and shareholder letters first, then the company’s own site and press releases, then named trade press (Reuters, The Energy Mag, W.Media, Pulse 2.0) for figures from filings we could not open directly. Every non-empty field lists its source and the date we read it. Targets, plans and speculative pipeline stay in the notes and never enter the score.

Scope and reporting cutoff

The dataset covers 51 GPU clouds and infrastructure providers. The September 2026 ClusterMAX appendix is a discovery list, not a claim that every reviewed cloud is included. AWS, Azure, Google Cloud, Oracle and other general-purpose hyperscalers are outside this neocloud list. Hot Aisle and TensorDock have no verified ClusterMAX 3.0 tier and receive 0 for that input.

At the 11 November 2026 cutoff, Nebius Q3 results were scheduled for 12 November. The latest reported quarter in this snapshot is Q2 ended 30 June, released 12 August 2026, with the reporting period and release date shown in the record: a year-end contracted-power target is not current MW, and capex is context rather than a scoring input. Nscale’s contract and power figures use its public S-1 filed 18 September, while its later GPU deployment release supplies the GPU count.

Counts and monetary measures have different scopes. A disclosed GPU cluster is a lower bound on known capacity, not a complete company fleet. Partner inventory, network-connected GPUs and maximum supported cluster sizes stay in capacity notes. Group-wide power can include general cloud and colocation loads. Contract value, customer commitments and ASC 606 RPO are labelled separately in backlog notes; future lease payments are not added to RPO. Capital may be a cumulative total, one round or one facility, exactly as each note states.

GPU models, operating footprint, orchestration, interconnect and capex enrich the records but do not change the six scoring inputs. Every newly added record has at least three substantive sourced facts beyond its name, URL and summary. An empty field means no verified value in this dataset, not proof that the company has never disclosed it. SemiAnalysis groups “Underperforming” and “Unavailable” under “Not Recommended”. Underperforming providers were tested hands-on; Unavailable providers could not be tested (no spare capacity, or they declined) or were judged too small to rate, so Unavailable is a testing status rather than a graded result.

Adding companies changes the minimum and maximum used by the log scales and can move existing scores without any change to their inputs. The weights and tier points remain the constants shown above.

Independence

Ranks are computed from the published inputs and weights; paid placement is separate. The Featured slot under the table is paid, labelled, and outside the ranking. The weights on this page change only with a dated note here.

Corrections and reuse

Send a correction with a public source to hello@neocloudrankings.com; it lands in the next build with a new verified date. The dataset is free under CC BY 4.0: CSV, JSON.

Field definitions

27 fields

Rank: Position by composite score, computed at build time from the signals below with the weights on /methodology/. Ties share the higher position. Never set by hand and never for sale.
Score /100: Composite score out of 100: ClusterMAX tier 30%, contracted power 20%, contracted backlog 15%, capital raised 15%, GPUs 10%, named customers 10%. A signal the company does not disclose scores 0.
ClusterMAX: Tier in SemiAnalysis ClusterMAX 3.0 (23 Sep 2026), the public hands-on rating of GPU clouds. Medal tiers run Platinum, Gold, Silver, Bronze; Participation Ribbon sits between Bronze and Underperforming. Underperforming and Unavailable are the two groups SemiAnalysis lists under "Not Recommended": Unavailable covers providers it could not test or judged too small to rate. Empty when the provider is not listed.
Contracted power MW: Power the company states as contracted, secured or committed in its portfolio, in MW. Speculative pipeline and planned campuses are excluded; the power note quotes the company's own word for it.
Active power MW: Power the company states as active, live, operational or managed today, in MW.
Power note: What the power figures cover, the as-of date, and targets or planned campuses that are not counted.
Backlog $M: Contracted revenue backlog, remaining performance obligations or total contracted value as the company states it, in USD millions.
Backlog note: Which backlog measure the company uses and its as-of date.
Quarterly revenue $M: Latest reported quarterly revenue, in USD millions, from the company's results release or filing. Company-level unless the note says otherwise.
Stated ARR $M: Annualized run-rate revenue exactly as the company states it, in USD millions. We never annualize a quarter ourselves.
Revenue note: Period, basis and any revenue targets the company has stated.
GPUs: GPUs deployed, owned or under contract as the company states it.
GPU note: What the GPU figure counts and when it was stated.
Capital raised $M: Equity and debt raised, in USD millions: the total when the company states one, otherwise the rounds and facilities named in the capital note.
Capital note: Exactly which rounds, notes or facilities the capital figure covers, with dates.
Named customers: Count of the customers listed in the customers field, computed at build time.
Customers: Customers named publicly by the company, in its filings or in reputable press.
Ownership: Publicly listed, privately held, or a subsidiary of another company.
Ticker: Exchange and ticker symbol for listed companies.
Parent: Owner or majority shareholder when the neocloud is part of another group.
Quarterly capex $M: Latest disclosed quarterly capital expenditures, in USD millions. Context only; excluded from the composite score.
Capex note: Period and scope of capital expenditures, including whether figures are actuals or guidance.
GPU models: Hardware advertised or tested, with reservations and future availability explicitly noted. Context only; not a fleet count or scoring input.
Operating footprint: Company or service geography as sourced; a general cloud footprint does not guarantee GPU stock in each region.
Deployment and orchestration: Available provisioning and cluster management services; future features are labelled as such.
Cluster interconnect: Published cluster network technology or bandwidth; not a measured service-quality claim.
Capacity scope: Partner supply, network coverage and other non-fleet capacity signals excluded from GPU scoring.