Few companies go public with as special a financial situation as OpenAI. The media reviewed the financial report of this AI start-up,From the surface data, it looks like a software company with light assets and low debt.As of March 31, its balance sheet had no debt and less than $750 million in lease liabilities.
In addition, the cash flow statement shows that as a technology company that relies heavily on hardware, OpenAI’s capital expenditures in this quarter were only US$46 million, even lower than Salesforce, which mainly specializes in enterprise software.
But the reality is far more complicated than the paper numbers. The notes to the financial report disclosed that OpenAI has purchase commitments totaling US$665 billion in the fields of chips, power and data centers in the next few years.The reason is that the company mainly rents the computing power of third-party data centers, with huge rental fees.will continue to be reflected in the income statement.
In the past two weeks, OpenAI has submitted confidential IPO filing materials, allowing financial regulators to review the above disclosure information. This filing document is likely to completely sort out its intricate business cooperation with investors such as Microsoft, Amazon, and NVIDIA, and at the same time explain its accounting basis for related party transactions, purchase commitments, and custom non-traditional financial indicators.
Another item that is likely to attract the attention of audit institutions is the non-cash accounting expenses generated by warrants.This is also OpenAI’s largest expenditure; as the company’s valuation continues to rise, the scale of this expense continues to expand. Accountants said that this liability stems from the issuance of equity to investors and the establishment of the OpenAI Foundation, and is extremely difficult to measure.
Of course, in the face of inquiries from the U.S. Securities and Exchange Commission (SEC) regarding filing materials, OpenAI will most likely not encounter insurmountable obstacles. The focus of SEC supervision is usually to ensure that companies completely and fully disclose business-related information. Under current SEC Chairman Paul Atkins, regulators have pursued deregulation ideas under the slogan "Revitalize the IPO Market."
Olga Usvyatsky, an accounting researcher who focuses on regulatory developments and publishes industry newsletters, said: "In my judgment, the SEC will raise a series of questions on risk warning disclosures, operating risks, supply and demand risks, etc."
Soon more ordinary investors will have the opportunity to take a closer look at OpenAI's operations. Although the company has not yet finalized the listing timetable, the prospectus documents can be made public in as soon as one month; if progress goes smoothly, it is expected to be listed from the end of August to the beginning of September.
Under the table, OpenAI executives and financial advisors are dealing with a number of increasingly complex data center cooperation agreements: including equity holdings in two computing companies, CoreWeave and Cerebras, and data center lease contracts signed with them, as well as OpenAI shares held by Amazon and Microsoft, the two major cloud vendors. In addition, OpenAI also jointly operates the "Stargate" data center expansion project with Oracle and SoftBank respectively - Oracle is the project equipment supplier and SoftBank is the investor.
All the above-mentioned cooperations involve tens of billions of dollars in agreements, and investors need to sort out the intricate relationships between rights and responsibilities. Liz Beyer, who has been engaged in IPO consulting for a long time and is a senior consultant at Class V Group, said: "We have all read the press releases issued by all parties, but when submitting SEC filings, companies must ensure that all information is 100% accurate and without exaggeration. The SEC will check item by item to ensure that the information disclosed by the company is leak-proof."
Uswiatsky said it is not unusual for technology giants like Amazon to have large off-balance sheet purchase commitments. According to accounting standards, such obligations will not be reflected in the balance sheet, but will be disclosed separately in the "Commitments and Contingencies" or "Liquidity" section of the prospectus.
Regulators may require OpenAI to fully disclose all cooperation details, but these agreements also expose core operating issues that investors must understand. There are more deep questions beyond the numbers. She said: "What should we do if market demand is less than expected and demand drops sharply? Even if the complete S-1 prospectus is made public, such key questions remain unresolved."
At the Davos Forum in January this year, OpenAI Chief Financial Officer Sarah Fryer said in an interview with reporters that the company chose to rely on cloud service providers in order to maintain a lightweight balance sheet. Marius Sklondahl, a technology investor at public equity fund Symbit Capital, said that lightweight statements are, at least on the surface, more favored by secondary market investors.
"This does improve investor perception. Although there is no difference in earnings from a mathematical perspective, taking the related expenses off the balance sheet can give the illusion of a higher return on capital."
The scale of income statement losses must be a core issue that investors focus on. The high cost of leasing computing power resulted in the company's net loss of approximately US$8.5 billion in the first quarter (not counting the substantial increase in warrant accounting expenses due to the increase in valuation). Operating costs alone (i.e., the basic expenses incurred to run the AI model) amounted to $3.5 billion, approximately 75 times the capital expenditure.
Companies also need to explain to regulators and investors the composition of counterparties for capital expenditures: 45% of the total expenses last quarter went to related parties - an accounting term that refers to financial transactions between companies and individuals and entities with potential conflicts of interest. For OpenAI, related parties are investors who are also suppliers. About 72% of the company's operating costs (mainly used for running AI models) are paid to related parties, most likely Microsoft. OpenAI purchases chips and server computing power from companies that invest in supporting itself.
OpenAI's revenue also comes largely from related parties. Related parties contributed revenue of approximately US$758 million in the quarter, an increase of 11 times year-on-year, which means that some investors are also customers of the company. In addition, OpenAI will settle part of its computing power costs with its own equity: last quarter, it paid US$488 million worth of shares to a related party to purchase computing power. This expenditure did not occupy any cash.
Other details also illustrate the complexity of the cooperative structure. In the income statement, OpenAI allocates nearly US$5 billion in losses to external data center joint ventures it controls and consolidates statements. This loss basically corresponds to the Stargate project it jointly operates with SoftBank and Oracle.
In recent years, SEC staff who review IPO prospectuses and provide regulatory guidance before companies make public applications have continued to conduct in-depth inquiries with reporting companies regarding related party cooperation arrangements.
For example, when CoreWeave, an AI cloud computing service provider, went public last year, the SEC required it to fully disclose its top customers and archive its cooperation contract with NVIDIA (both a supplier and an investor) as a public attachment.
When Arm, a subsidiary of SoftBank, submitted its IPO application in 2023, regulators forced it to add risk warnings related to parent company financing, and at the same time disclosed details of the cooperation that the company had originally tried to keep confidential.
Approximately 20 working days after the company is listed on the market, all answers to inquiries from the SEC and the company will be made public. Therefore, within a few weeks of OpenAI's listing, investors will be able to fully see all the issues that the regulatory agency has focused on verification. Regulation requires companies to improve information disclosure, which does not mean that the IPO process is blocked. Both CoreWeave and Anmou received dozens of inquiries from the SEC that year, and finally completed their listings quickly.
It’s not just OpenAI’s collaborative structure that will attract regulatory attention. Anthropic is also negotiating data center cooperation with increasing complexity, and is constantly expanding its computing power infrastructure while preparing for its official listing. Previously, the company's computing power mainly relied on investors Google and Amazon, and its accounting structure was relatively simple; media previously reported that it had recently signed dozens of data center lease agreements.
Most of Anthropic's computing power expansion projects will be implemented through Alphabet's computing power service provider Fluidstack. Fluidstack previously disclosed to investors in financing promotion materials that Anthropic will pay it $4.5 billion for computer room hosting contracts in the next few years. It is also reported that Anthropic leases chips from Google and Broadcom through special purpose vehicles, and Broadcom fulfills chip orders totaling US$35 billion.
Tanuki Tapriyar, CEO of Kos.ai, a data center contract review software service provider, said: "The complexity of various cooperation agreements is only increasing. New transaction participants are constantly emerging in the industry chain, and the overall structure will only become more and more cumbersome."