TheChek News Flow

The AI startup’s confidential IPO filing is giving financial regulators a chance to scrutinize its accounting and complex web of business relationships.

Few companies have gone public with books as unusual as OpenAI’s.

At first glance, the company portrayed by the AI startup’s financial statements, reviewed by The Information, resembles a lean, low-debt software business. Its balance sheet, as at March 31, had zero debt and less than $750 million of lease liabilities.

Its cash-flow statement, moreover, showed OpenAI—one of the most hardware-centric tech businesses—spending just $46 million on capital expenditures in the quarter, less than even Salesforce, which sells business software.

The reality is messier. The fine print of the financial statements show OpenAI has $665 billion of purchase commitments stretching out over the coming years for chips, energy and data centers. That’s because the company mostly rents computing capacity in other companies’ data centers, meaning it spends billions of dollars in leasing payments that flow through its income statement.

Financial regulators have had an opportunity to look at these and similar disclosures over the past two weeks, since OpenAI filed confidential paperwork for an initial public offering. Those documents likely lay out the company’s web of business relationships with investors like Microsoft, Amazon and Nvidia, and justify its accounting decisions around related-party transactions, purchase commitments and any nontraditional metrics it decides to use.

Another item that might catch the eye of auditors is a noncash accounting charge for warrants, the company’s largest expense, which has swelled due to OpenAI’s rising valuation. That liability, related to equity issued to investors and the creation of the OpenAI foundation, can be difficult to measure, accountants said.

To be sure, it isn’t likely that OpenAI will face any insurmountable hurdles as it responds to questions from the Securities and Exchange Commission about its filing. The agency is typically most concerned with making sure companies have adequate disclosures about their business. And under SEC Chair Paul Atkins, the agency has taken a deregulatory approach, with a mantra of “Make IPOs great again.”

“If I were to guess, the SEC will ask some questions about risk factor disclosures, business risks, demand and supply risks,” said Olga Usvyatsky, an accounting researcher who publishes a newsletter on regulatory developments.

A wider group of investors could get a chance soon to examine OpenAI’s business more closely. While the company has said it hasn’t decided when to go public, it could make its filing public as soon as a month from now and go public in late August or early September if it chooses.

Behind the scenes, OpenAI executives and advisers are managing an increasingly complex array of data center deals. Those deals include data center rental agreements from companies like CoreWeave and Cerebras, in which OpenAI holds lucrative stakes, as well as the stakes that hyperscalers Amazon and Microsoft hold in OpenAI. Then there’s OpenAI’s co-ownership of the Stargate data center expansion with Oracle and SoftBank, respectively a vendor and an investor in the project.

All these deals involve multibillion-dollar agreements that investors will need to make sense of. “We have all seen the press releases. But now, for the SEC filings, the company has to demonstrate they are 100% accurate with no exaggerations,” said Lise Buyer, a longtime IPO consultant for Class V Group. “The SEC takes its time being sure that the company has dotted all the I’s and crossed the T’s.”

It’s not unusual for tech companies like Amazon to have large off-balance sheet purchase commitments, said Usvyatsky. They are disclosed not on the balance sheet but in other sections of the IPO prospectus that go by the names of “commitments and contingencies” or “liquidity” under accounting rules.

Regulators could ask questions ensuring OpenAI makes adequate disclosures of these deals. But they highlight a more basic business question investors will have to understand, beyond just the numbers. “What happens if the demand doesn’t materialize? What happens if the demand falters?” Usvyatsky said. “Even after we see the S-1, many of those questions remain.”

Sarah Friar, OpenAI’s chief financial officer, told The Information’s Jessica Lessin at Davos in January that the company was “utilizing our [cloud] partners because it’s a way to keep a lighter weight on the balance sheet.” Lighter balance sheets tend to appeal to public markets investors, at least superficially, said Marius Skrondal, a tech investor at public equities fund Symbit Capital.

“It definitely helps the investor view,” he said. “Keeping it off balance sheet feels like it’s a higher return of capital than otherwise, even though it’s mathematically equivalent.”

OpenAI’s bottom line is sure to be an issue with investors. The heavy flow of rental payments for computing capacity led to a roughly $8.5 billion net loss in the first quarter, excluding an accounting charge that’s grown significant due to its rising valuation. Its cost of revenue alone, which is essentially the expenses required to run its AI models, was $3.5 billion in the quarter, roughly 75 times its capital spending.

Another dynamic to explain to regulators and investors is who sits on the other side of OpenAI’s spending. Nearly half of its total expenses last quarter—45%—went to related parties, an accounting term for financial arrangements between companies and individuals or entities with whom there’s a potential conflict of interest. In OpenAI’s case, those related parties include investors who are also its suppliers. OpenAI paid roughly 72% of its cost of revenue—mostly the money it spends to run its AI models—to related parties, likely Microsoft. OpenAI pays the companies that fund it for the chips and servers it runs on.

Revenue comes into OpenAI through related parties, too. OpenAI booked about $758 million in revenue from related parties in the quarter—12 times the figure a year earlier—meaning some of its investors are also its customers. And it pays for some of its computing in its own stock: $488 million of equity went to a related party for compute last quarter, a cost that never touches its cash.

There are other signs of complex dealmaking. On its income statement, OpenAI assigned nearly $5 billion of its loss to outside partners in a data center venture it controls and consolidates—likely tied to the Stargate project it runs alongside SoftBank and Oracle.

Staff at the SEC, who review IPO prospectuses and provide guidance before their public filing, have grilled companies about related party arrangements in recent years.

For example, when CoreWeave, a provider of AI computing services in the cloud, went public last year, the SEC made it name its largest customers and file as a public exhibit its contract with Nvidia—a supplier that is also an investor.

And when SoftBank-controlled Arm filed for an IPO in 2023, regulators forced it to add a risk factor about financing arrangements with its parent and to disclose specifics it had tried to keep private.

Those exchanges become public about 20 business days after a company lists, so investors will see exactly what the SEC pressed OpenAI on within weeks of its debut. Just because regulators demand better disclosures from a company doesn’t mean its IPO is in jeopardy. Both CoreWeave and Arm drew dozens of comments from regulators and still went public quickly.

OpenAI isn’t the only AI company with relationships that might get attention from regulators. Anthropic, too, is starting to take on more complex data center deals, expanding its data center footprint as it starts formal efforts to go public. The company has historically leaned largely on its investors, Google and Amazon, for cloud services, simplifying its books. More recently, it has struck dozens of data center leasing deals, The Information previously reported.

A significant portion of Anthropic’s data center expansion will flow through Fluidstack, an Alphabet-backed data center provider. Fluidstack told investors in a fundraising deck recently that Anthropic would pay it $4.5 billion in the coming years for a colocation contract. It is also leasing chips from Google and Broadcom through a special purpose vehicle, with Broadcom serving as the backstop for the $35 billion chip order, The Information previously reported.

“The deals have been getting more complicated, not less,” said Tanuj Thapliyal, CEO of Kos.ai, which builds software to review invoices and contracts for data center companies. “You’re getting new ecosystem players that are the transacting counterparties in these deals. The complexity has been going in one direction only, and that’s up.”