Anthropic's vendor file just got a lot thicker, and nobody at the bank had to ask for it. Its Class A shares are now listed on the Nasdaq Global Select Market under the symbol ANT, and with the listing come audited financials, a risk factor section the company's lawyers had to sign off on, and a duty to keep reporting every quarter. For a third-party risk team that has spent years working from a sales deck, a SOC 2 report and whatever the account executive was willing to say, that's a different kind of evidence.
I'm not offering a view on the stock, and nothing here is investment advice. My interest is narrower: what a public filing gives the people who buy AI for a bank, credit union or fintech, what I'd pull from it into the vendor file, and which parts of the relationship now answer to a different set of owners.
Where the offering stands
Unless I say otherwise, everything here comes from the company's filings on SEC EDGAR (opens in new tab). The company submitted a confidential draft in June and filed its public Form S-1 (opens in new tab) in late October. Three amendments followed in November. The SEC declared the registration effective on November 18, and the final prospectus, filed under Rule 424(b)(4) the next day, sets the terms:
- Size and price. 980,000,000 Class A shares at $83.00 each. The company sold 956,351,354 of them, and selling stockholders sold the remaining 23,648,646.
- Strategic buyers alongside. Under the section titled "Concurrent Private Placement," Google, NVIDIA, AMD, Broadcom, Micron, Microsoft and TSMC agreed to buy an aggregate of $26.0 billion of nonvoting Class N stock at the offering price, which the prospectus puts at 313,253,007 shares.
- Lockups. Most existing holders are restricted for up to 180 days. Co-founders and certain holders of more than 0.5% of fully diluted shares are restricted for up to 365 days. The prospectus also says 3.7 billion shares under the 180-day restriction "may become eligible for sale prior to the expiration of that period."
CNBC's pre-pricing guide (opens in new tab) put the marketed range at $70 to $78 on 825 million shares and said the valuation at that range was $1.405 trillion. I haven't found a valuation at the final price in the filing itself, so I won't quote one.
The part a buyer actually uses hasn't arrived yet: the periodic reports. Until the first one lands, the prospectus is the whole public record.
What the prospectus puts in the vendor file
A private AI lab answers a due diligence questionnaire however it likes, and you can't check much of it. A registrant answers to securities law for what it writes in a prospectus, and the risk factors in particular are written by people trying to make sure nobody can later say they weren't warned. They're the vendor's own list of the ways the relationship could go wrong.
Here's what I'd copy into the file, with the section it came from so the next reviewer can check it.
Revenue scale and trajectory. Under "Risk Factors," the company reports revenue of $41 million in 2023, $386 million in 2024 and $4.6 billion in 2025, then $33.2 billion for the first nine months of 2026. The same paragraph says "We do not expect our revenue will continue to grow at comparable rates." When a vendor grows that fast, support, account coverage and capacity planning get stretched, and you'll want to know that before the renewal.
Profitability, stated two ways. The summary reports an operating margin of (3)% and an operating loss of $567 million for the September quarter, against an Adjusted Operating Margin of 15% for the same quarter. I'd record both, label the second one as the company's non-GAAP measure, and track the GAAP number. For a vendor risk file, the GAAP line is the one that says whether the company can keep serving you through a bad year.
Customer concentration. The prospectus says no single customer represented more than 10% of revenue for the first nine months of 2026, and that about 79% of revenue in the latest quarter came from customers outside the top five, up from about 64% in the March quarter. The risk factors still warn that "a significant reduction in spending by, or the loss of, one or more of our largest customers" could hurt results. A bank isn't going to be one of those top five. That's the point: your leverage in a dispute is what it is, and the filing tells you roughly how much weight sits elsewhere.
Compute commitments. This is the number I'd put at the top of the file. The prospectus says that as of and after the end of the September quarter, the company has entered into purchase commitments and lease agreements with an aggregate committed spend of approximately $900.9 billion, "substantially all of which is related to compute and infrastructure." The section on access to capital adds that these commitments "generally must be satisfied regardless of our revenue or utilization." Obligations that large and that fixed explain why a vendor works hard to fill capacity. They're also why the filing ties its future to continued access to capital.
Partner dependence. The same section says part of the compute build-out "has been funded or supported by capital or credit support provided or facilitated by certain of our CSPs and chip partners," meaning the cloud service providers and chipmakers. A risk factor goes further: some arrangements depend on credit support "provided by a third party that is also a competitor." The filing doesn't name that party. The competition section names Google and OpenAI as its primary competitors. Google is also a cloud channel for Claude and a buyer in the private placement. One company, three seats in the relationship.
I argued earlier that most of the AI on a bank's roadmap traces back to three labs sitting on the same chips and the same pools of capital. Until now, the capital part was something I'd inferred. It's now in a filing, with a dollar figure on it.
Who the vendor answers to now
Governance usually gets skipped in vendor reviews because a private company doesn't have to explain its own. This one does, at length, and it matters to a customer for a plain business reason: it tells you who can change the company's direction and who can't.
The cover page lays it out. Anthropic is a Delaware public benefit corporation, required "to balance the pecuniary interests of our stockholders, the best interests of those materially affected by our conduct, and the specific public benefit purpose" in its charter. Through a special Class T share, the Long-Term Benefit Trust "generally has the authority to elect and remove the Class T directors, who together will constitute a majority" of the board. A Class F share held by an entity of the co-founders carries 50.1% of the voting power on matters submitted to stockholders generally, including electing the other directors. The Class A shares sold to the public carry one vote each. The company is a "controlled company" under Nasdaq rules and says it relies on some of the exemptions that come with that status.
I'd read that structure as a stability fact. Public holders of Class A shares can't easily force a change of strategy, a sale or a board turnover. That cuts both ways for a customer. A sudden activist push to raise prices or exit a product line is harder. A governance failure is also harder for outside shareholders to correct. Either way, it belongs in the file next to the ownership chart, with the section titles "Our Governance Structure and Its Goals" and "Description of Capital Stock" cited.
What does change is the rhythm of the business. Quarterly results now get read by analysts who model revenue against those compute commitments, and a vendor with a public earnings call has a reason to care about the quarter in a way a private one only partly did. I won't predict what that does to any customer's price. I'd just expect commercial terms, discount approval and end-of-quarter timing to start looking like every other public software vendor's, because the same pressures now apply.
Monitoring that runs on a filing calendar
The bigger change is that the evidence keeps coming. A public vendor has to file a material change, not just mention it when convenient. That turns part of ongoing monitoring from a questionnaire into a reading assignment, and the reading assignment has deadlines.
What I'd add to the monitoring plan for any AI vendor that files with the SEC:
- Subscribe to the EDGAR feed for the company. Every new filing becomes a ticket for the vendor owner, triaged within a few business days.
- Read every 8-K for four things. A change in a major cloud or chip agreement, a new financing that changes the commitment picture, a cybersecurity incident disclosure, and a change in leadership or auditor. Each one changes a row in the risk assessment, so it shouldn't wait for the annual review.
- Diff the risk factors each quarter. New risk factors and reworded old ones are the vendor telling you what its lawyers now worry about. A text diff takes minutes.
- Track three numbers from each 10-Q and 10-K. GAAP operating margin, the remaining committed spend, and the customer concentration language. Together they show whether the capacity you depend on is still funded.
- Put the lockup expirations on the calendar. Both restrictions are "up to" periods with early-release triggers tied to earnings, and some shares can trade from day one. Calendar the releases, not just days 180 and 365. Insider sales leave the service alone but can shift who holds the shares.
None of this replaces the questionnaire, the SOC report or the contract. A securities filing describes the company. It doesn't describe the controls around your data. What it adds is a second source with legal liability attached, and a schedule you don't have to negotiate.
It also changes what a vendor can reasonably say no to. If a fact is in the prospectus, there's no reason to treat it as confidential in a diligence call. I'd cite the section and spend the call on what only the contract can settle: data handling, exit rights and notice periods.
That's the same discipline vendor risk teams apply to us when we're the vendor. Serving 1,500+ financial institutions means sitting on the other side of a lot of these reviews, and the reviews that go well are the ones where the evidence already exists before anyone asks. A public AI vendor now has to produce much of that evidence on a fixed schedule, and the buyer's job is to read it and notice when it changes.
A filing tells you about a company's finances, its obligations and its governance. It tells you very little about a model's behavior, how a given release will change your workloads, or how a channel handles your data. Those still come from your own testing and your own logs. When we built AgentOS we put a single governed egress in front of the model calls, and the same logic applies here: the record of what went out and what came back has to be yours, whoever the vendor answers to. A securities filing doesn't change that, and neither does a single vendor relationship that covers every workload.