Why This Course Follows Documents, Not Names

On January 21, 2025, at the White House, OpenAI, SoftBank, Oracle and the investment firm MGX announced a venture called Stargate to build artificial intelligence infrastructure in the United States. Two years earlier, a list of the companies building the largest American data centers would not have included OpenAI as a builder at all. It was a customer, renting capacity chiefly from Microsoft. The Stargate announcement is a useful reminder that the cast of this subject changes faster than the physics of it.

The same point can be made with Oracle. For much of the last decade Oracle was treated as a distant fourth among cloud providers. Then, with its results for the first quarter of its fiscal 2026, reported in September 2025, Oracle disclosed remaining performance obligations of roughly 455 billion dollars as of August 31, 2025. Remaining performance obligations are contracted revenue that a company has not yet recognized, and Oracle attributed much of the increase to large cloud contracts with artificial intelligence companies. A single disclosure moved Oracle's standing in this conversation more than years of product announcements had.

A third example sits outside the five entirely. Companies now called neoclouds, such as CoreWeave and Nebius, rent graphics processing units, or GPUs, to artificial intelligence developers, and some of their largest customers are the hyperscalers themselves. Microsoft signed a five-year capacity agreement with Nebius in September 2025, and Meta signed a capacity contract with CoreWeave the same month. A course organized around five company names would struggle to explain why the largest cloud providers were buying capacity from firms that barely existed in their current form a few years earlier.

So this course is organized around a method rather than a roster. The five companies appear throughout, but as dated examples of things that recur: a capital spending disclosure, a campus announcement, a lease, a partnership. The names will change. The documents they file, and the questions worth asking of those documents, change much more slowly.

The first element of the method is dating everything. A statement such as Microsoft is building its largest campus in Wisconsin is not wrong, but it is incomplete. A better version is: in a company blog post on September 18, 2025, Microsoft said it expected to bring its Fairwater data center in Mount Pleasant, Wisconsin, online in early 2026. The dated version tells you what was claimed, by whom and when, and it tells you what to check next. By the time you read this, the claim has either come true or it has not.

The second element is knowing each company's calendar. Amazon, Alphabet and Meta report on calendar years. Microsoft's fiscal year ends on June 30, so its fiscal 2026 ran from July 2025 to June 2026. Oracle's ends on May 31. When a headline compares the companies' spending for 2025, it may be mixing periods that overlap by only half a year. The error is common and it is easy to avoid if you check the period printed at the top of each table.

The third element is naming the source. Company statements come in several strengths. A number in an annual report on Form 10-K, filed with the Securities and Exchange Commission, or SEC, is audited and carries legal liability. A number in a quarterly report on Form 10-Q is reviewed but not audited. Guidance on an earnings call is a forecast. A press release about a campus is a statement of intent. A report in the trade press, citing people familiar with a matter, is useful but unconfirmed. Each has its place, and each should be labeled for what it is.

The fourth element is humility about categories. The boundary between a cloud provider, a platform company, a model developer and a neocloud has shifted several times in the last few years, and it will shift again. Faraday's read is that the category a company belongs to matters less than the question of who is paying for the building and who is promising to fill it. Those two questions have answers in public documents, and those answers can be checked.

None of this requires special access. Every document described in this course is public, and most can be found on the companies' investor relations sites or in the SEC's Electronic Data Gathering, Analysis and Retrieval system, known as EDGAR. What it requires is patience, and a willingness to write the date next to every fact before trusting it.