Most AI company numbers reach readers as a press release or a funding headline, written by the company. An S-1 is different in kind: it is the registration statement a company files with the SEC before selling shares to the public, and it has required sections. This guide uses CoreWeave's S-1, filed on 3 March 2025 by a company that rents GPU capacity to AI developers,[1] to show five checks that apply to any AI infrastructure filing. If you want the funding-round equivalent for private companies, see How to Check an AI Startup Funding Announcement Against SEC Form D, and for the claims side see AI Washing: What Four SEC Cases Show.
What an S-1 must contain, and why that helps you
The SEC's rule for risk factors, Item 105 of Regulation S-K, asks for "a discussion of the material factors that make an investment in the registrant or offering speculative or risky." It says risk factors should be organised under headings, that generic risks are "discouraged" and belong at the end under "General Risk Factors," and that a discussion longer than 15 pages needs a summary of no more than two pages.[2] That is why CoreWeave's filing carries a short bulleted risk summary ahead of the long section: read the summary for the company's own list of what could go wrong, then read the full risk factor for any bullet that touches the numbers.[1]
A second rule matters for AI companies that report adjusted figures. Item 10(e) says that when a filing includes a non-GAAP measure it must present the most comparable GAAP measure with equal or greater prominence, give a quantitative reconciliation, and state why management thinks the measure is useful.[3] The reconciliation table is therefore always in the document, and it is the table to read.
Check 1: who pays the revenue
CoreWeave's filing states that its largest customer was 16% of revenue in 2022, and that for 2023 and 2024 the largest customer was Microsoft, at 35% and 62% of revenue. It also says approximately 77% of 2024 revenue came from its top two customers and that no other customer reached 10%.[1] Revenue itself grew from $15.8 million in 2022 to $228.9 million in 2023 and $1,915.4 million in 2024.[1]
Two things to read from that. First, the growth is real in the filing's audited statements, but a single contract drove most of it: the filing says revenue recognised under the Microsoft master services agreement was $81 million in 2023 and $1.2 billion in 2024.[1] Second, the filing says negative changes in Microsoft's demand, ability or willingness to perform, or in the broader relationship, "would adversely affect" the business.[1] When you read any AI infrastructure filing, find the customer-concentration sentence first and ask what happens to every later figure if that customer slows.
Check 2: what "backlog" is defined to mean
The filing reports $15.1 billion of remaining performance obligations (RPO) at 31 December 2024, up 53% from $9.9 billion a year earlier, and defines the term as committed contract obligations yet to be delivered.[1] It says the vast majority of revenue comes from multi-year committed contracts on a "take-or-pay" basis, defined as a customer committing to pay for a minimum amount of services regardless of whether it uses them.[1]
The financial statement notes add the timing, which headlines usually omit: of the $15.1 billion, 54% is expected to be recognised within the first 24 months, 42% between months 25 and 48, and the remainder between months 49 and 72.[1] Our arithmetic from those percentages is roughly $8.2 billion in the first two years and $6.3 billion in the next two. So "$15.1 billion" describes delivery over up to six years, and it is delivery the company must still build capacity to perform. The revenue is contracted; it is not yet earned.
Check 3: how adjusted EBITDA is built
The filing defines adjusted EBITDA as net loss excluding fair-value losses, depreciation and amortisation, net interest expense, stock-based compensation, other income and income taxes.[1] Its reconciliation for 2024 runs as follows (US$ thousands):
| 2024 reconciliation line | Amount |
|---|---|
| Net loss | (863,448) |
| Loss on fair value adjustments | 755,929 |
| Depreciation and amortisation | 863,413 |
| Interest expense, net | 360,824 |
| Stock-based compensation | 31,487 |
| Other income, net | (48,194) |
| Provision for income taxes | 119,247 |
| Adjusted EBITDA | 1,219,258 |
Source: the filing's non-GAAP reconciliation.[1] The filing states a 64% adjusted EBITDA margin and a (45)% net loss margin for 2024.[1]
Reading this table: the two largest add-backs are the fair-value loss on financial instruments and depreciation. Depreciation of $863 million is about 45% of the year's $1,915 million revenue (our arithmetic), which is the cost of the GPUs and data centre equipment that the business exists to rent out. Adjusted EBITDA is a legitimate way to see operating performance before financing and asset costs, but in a business where hardware is the product, the depreciation and interest lines are not noise. Our interpretation is to read adjusted EBITDA next to operating income, which the filing gives as $324.4 million for 2024,[1] and to treat the gap between them as the price of the equipment.
Check 4: what the debt takes
The risk factor on indebtedness states total indebtedness of $8.0 billion at 31 December 2024, plus $4.4 billion of undrawn availability under its Revolving Credit Facility and DDTL 2.0 Facility.[1] The filing also separately cites $12.9 billion of debt commitments raised through the same date,[1] which is a different measure (commitments, not outstanding borrowing), so do not mix the two. We did not reconcile the $8.0 billion total to the balance sheet's $5.5 billion non-current debt line, and the filing's own tables are where to do that.
The cash-flow link is the more useful number. The filing reports $941 million of cash dedicated to debt service in 2024 ($588 million principal and $353 million interest, of which $169 million was capitalised interest), and says that was approximately 32% of operating cash flow before interest.[1] In an AI compute business funded by borrowing against contracts and GPUs, that percentage is a direct measure of how much slack the growth plan has.
Check 5: the estimates behind the profit line
Notes to the financial statements are where accounting choices live. CoreWeave's note says that effective 1 January 2023 it changed the estimated useful life of computing equipment in its data centres from five to six years, and lists technology equipment at six years.[1] A longer life lowers each year's depreciation for the same hardware. The filing gives its reason as hardware, software and data centre improvements, and its risk factors admit the company must estimate useful lives and "cannot guarantee" the estimates will be accurate.[1] That is not a criticism; it is an assumption to carry into your reading, because GPU generations turn over quickly and the filing's profit depends on that estimate holding.
The same risk section says customers have contractually specified NVIDIA GPUs, and that changing suppliers could affect the company's ability to meet its obligations.[1] Supplier dependence is a second input to read alongside depreciation.
What this guide does not tell you
It reads one preliminary filing and does not say whether CoreWeave is a good business or a good investment. The S-1 is dated 3 March 2025; we did not open the final prospectus or any later quarterly or annual reports in this run, so figures here are the March 2025 ones and may have been updated. We read the filing's text, not its exhibits, and we did not open the full risk-factor discussion of every bullet. Nothing here is investment advice.
A reusable checklist
- Find the customer-concentration sentence and note every named customer and percentage.
- Find how backlog is defined and the table that times its recognition.
- Take the adjusted measure's reconciliation and set each add-back beside revenue.
- Compare debt service with operating cash flow, and keep commitments, outstanding debt and availability apart.
- Read the useful-life and change-in-estimate notes, then the supplier-dependence risk factors.
Sources and method
The CoreWeave figures come from the preliminary S-1 on SEC EDGAR, which we downloaded and searched as text on 2026-10-28. The rule texts are the current Electronic Code of Federal Regulations versions of 17 CFR 229.105 and 229.10, read on the same day. Derived values (the dollar split of backlog by period and the depreciation share of revenue) are our arithmetic and are labelled as such.
