The financial position of a company at the time of its initial public offering is rarely as distinctive as that of OpenAI.
A review of the AI startup's financial statements reveals a surface-level appearance akin to a lean, low-debt software firm. As of March 31st, its balance sheet carried no debt, with lease liabilities under $750 million.
Furthermore, its cash flow statement indicated that for a technology company heavily reliant on hardware, its capital expenditures for the quarter were a mere $46 million, even lower than those of enterprise software giant Salesforce.
However, the reality is far more complex than the numbers suggest. Footnotes in the financials disclose that OpenAI has committed to a staggering $665 billion in future purchases for chips, power, and data centers over the coming years. This stems from the company's primary strategy of leasing computing power from third-party data centers, with the substantial costs continuously flowing through its income statement.
In the past two weeks, OpenAI has confidentially submitted its IPO registration materials, allowing financial regulators to examine these disclosures. The filing is expected to provide a comprehensive overview of its intricate commercial ties with investors like Microsoft, Amazon, and NVIDIA, while also explaining the accounting rationale for related-party transactions, purchase commitments, and its use of custom, non-standard financial metrics.
Another item likely to draw auditor scrutiny is the non-cash accounting expense from stock warrants, which represents OpenAI's largest expenditure. This expense continues to grow as the company's valuation rises. Accountants note this liability arises from issuing equity to investors and the establishment of the OpenAI Foundation, making it exceptionally difficult to value.
Nevertheless, it is unlikely that OpenAI will face insurmountable hurdles from the U.S. Securities and Exchange Commission's inquiries. The SEC's primary focus is on ensuring companies provide complete and adequate disclosure of operational information. Under current leadership, the regulatory body has adopted a deregulatory approach, with a stated goal of revitalizing the IPO market.
An accounting researcher specializing in regulatory trends stated, "In my view, the SEC will likely pose a series of questions regarding risk factor disclosures, operational risks, and supply and demand risks."
Soon, more retail investors will have the opportunity to scrutinize OpenAI's operations. While a final IPO timeline is not set, the company could publicly file its prospectus in as little as a month. If the process proceeds smoothly, a listing could occur between late August and early September.
Behind the scenes, OpenAI executives and financial advisors are managing an increasingly complex set of data center agreements. These include equity stakes in and leasing contracts with compute providers like CoreWeave and Cerebras, as well as shares held in OpenAI by cloud giants Amazon and Microsoft. Additionally, OpenAI is involved in the "Stargate" data center expansion project through joint ventures with Oracle (as an equipment supplier) and SoftBank (as an investor).
All these partnerships involve multi-billion dollar agreements, and investors will need to untangle the complex web of rights and obligations. A long-time IPO consultant noted, "We've all seen the press releases from various parties, but when filing with the SEC, a company must ensure all information is 100% accurate with no exaggeration. The SEC will scrutinize everything to ensure disclosures are watertight."
The researcher added that large off-balance-sheet purchase commitments are not uncommon for tech giants like Amazon. Under accounting rules, such obligations are not recorded on the balance sheet but are disclosed separately in the prospectus under sections like "Commitments and Contingencies" or "Liquidity."
While regulators may demand full disclosure of these partnerships, the agreements also highlight core operational questions investors must understand, raising deeper issues beyond the numbers. "What happens if market demand falls short of expectations or declines sharply? Even with a full S-1 filing, such critical questions remain unanswered," she said.
At the World Economic Forum in Davos this January, OpenAI's CFO stated that partnering with cloud service providers was a strategy to keep its balance sheet lean. A tech investor commented that a lean balance sheet is generally more appealing to secondary market investors, at least on the surface.
"It does improve investor perception. Even though the math on returns is the same, keeping these expenses off the balance sheet creates an illusion of higher capital returns," the investor explained.
The scale of losses on the income statement will undoubtedly be a key focus for investors. High compute leasing costs led to a net loss of approximately $8.5 billion last quarter, a figure that excludes the significantly larger warrant accounting expense driven by valuation increases. Operating costs alone, the fundamental expense of running AI models, reached $3.5 billion, roughly 75 times the capital expenditure.
The company must also explain to regulators and investors the composition of its counterparties for these expenditures. Last quarter, 45% of total expenses went to related parties—an accounting term for transactions with individuals or entities where potential conflicts of interest exist. For OpenAI, related parties are investors who are also suppliers. About 72% of its operating costs, primarily for running AI models, were paid to related parties, most likely Microsoft. OpenAI purchases chips and server compute from the very companies funding it.
OpenAI's revenue is also heavily sourced from related parties. Related-party revenue last quarter was about $758 million, an eleven-fold increase year-over-year, indicating that some investors are also customers. Furthermore, OpenAI settles some compute costs with its own equity: last quarter, it paid a related party $488 million worth of shares for compute, a transaction that used no cash.
Other details underscore the complexity of the partnership structure. On its income statement, OpenAI allocated nearly $5 billion in losses to external data center joint ventures it controls and consolidates, likely corresponding to the Stargate project operated with SoftBank and Oracle.
In recent years, SEC staff reviewing IPO prospectuses and providing pre-filing guidance have consistently probed deeply into related-party arrangements.
For example, when AI cloud compute provider CoreWeave went public last year, the SEC required full disclosure of its major customers and mandated the public filing of its contract with NVIDIA, which is both a supplier and an investor.
During SoftBank-owned Arm's IPO filing in 2023, regulators forced the addition of risk factors related to parent company financing and the disclosure of partnership details the company initially sought to keep confidential.
All correspondence between the SEC and a company becomes public approximately 20 business days after the company lists. Therefore, within weeks of OpenAI's IPO, investors will see the full scope of issues regulators focused on. Regulatory demands for better disclosure do not necessarily impede the IPO process. Both CoreWeave and Arm received dozens of SEC inquiries yet proceeded to list quickly.
OpenAI is not the only company with a partnership structure drawing regulatory attention. Anthropic is also negotiating increasingly complex data center deals as it prepares for a potential public listing while expanding its compute infrastructure. While its structure was previously simpler, relying on investors Google and Amazon for compute, recent reports indicate it has now signed dozens of data center lease agreements.
Much of Anthropic's compute expansion will occur through Alphabet-owned compute provider Fluidstack. Fluidstack previously informed investors that Anthropic is committed to paying $4.5 billion for colocation contracts over several years. Other reports suggest Anthropic is leasing chips from Google and Broadcom through special purpose vehicles, with Broadcom guaranteeing a $35 billion chip order.
The CEO of a data center contract review software service stated, "The complexity of these partnership agreements is only increasing. New parties are constantly entering the ecosystem, and the overall structure is becoming more convoluted."
Comments