The railroad boom changed America. It also left a lot of carnage behind. In the years before the Panic of 1873, railroads expanded, investors poured money into new lines, and construction ran ahead of demand. When financing dried up, railroad companies failed, banks went down with them, and the country entered a long economic contraction. People lost work, workers faced wage cuts, and the pain reached well beyond the people who had made the bets.
A little more than a century later, we did a version of it again. Venture capital poured into internet companies while telecom companies built fiber networks for the traffic everyone expected to come. When the dot-com boom collapsed, companies failed, funding tightened, and people lost their jobs. In Santa Clara County, unemployment rose from 3% in January 2001 to 7% by the end of that year, as the U.S. entered a recession. Global Crossing, one of the companies that built a fiber network, filed for bankruptcy in 2002; a business connected to it reported defaults and doubtful receivables.
AI is changing work. Subscribe to The Human Part to explore what that asks of us as leaders.
The railroads were real. The internet was real. And the fallout was real for the people who worked at those companies and depended on them. That history is why I look at today’s AI infrastructure boom and see winners and losers emerging quickly. I can’t tell you exactly which companies will survive or when the reckoning will come, but the money and the relationships connecting them deserve a much closer look.
I keep coming back to a question about the AI boom: what happens when one of these companies fails? Not because I know which company that will be or when it will happen, but because the more I follow the money, the harder it is to see any of the major players as completely separate. They’re investing in each other, buying services from each other and building infrastructure for each other. At some point, one of the shells is going to crack, and I want to know what happens to everything connected to it.
This week, Jon Stewart spoke with Ed Zitron about the AI boom and its financing. The house-of-cards conversation stuck with me because it’s much bigger than Oracle, OpenAI and NVIDIA. Add Anthropic, Cursor, SpaceX, Microsoft, Amazon, cloud providers and the firms financing all those data centers, and the relationships start to look less like a simple circle and more like a web.
Here are some of the connections I found. These are disclosed relationships, not a complete map, and an announcement or a future commitment does not mean every dollar has already changed hands:
NVIDIA and OpenAI: OpenAI announced $30 billion from NVIDIA as part of a February funding round, alongside plans for more NVIDIA computing capacity. OpenAI needs the computing power; NVIDIA supplies systems used across the ecosystem and was also named as an investor in that round.
NVIDIA, CoreWeave and Anthropic: NVIDIA said it invested $2 billion in CoreWeave, whose cloud is built on NVIDIA infrastructure. CoreWeave later announced an agreement to provide computing services to Anthropic, the company behind Claude.
NVIDIA, Nebius and Nscale: NVIDIA announced a $2 billion Nebius investment while Nebius was already deploying NVIDIA infrastructure. NVIDIA also participated in Nscale funding rounds; Nscale announced a Microsoft deal involving approximately 200,000 NVIDIA GPUs. The Nscale figure is a contract, not a count of chips already delivered.
Anthropic, Microsoft and NVIDIA: Anthropic committed to Azure computing capacity and announced plans for NVIDIA systems, while Microsoft and NVIDIA committed to invest up to $5 billion and $10 billion respectively in Anthropic. Those are commitments, not proof that the full amounts have been paid.
SpaceX, Anthropic and NVIDIA: SpaceX’s June prospectus described agreements under which Anthropic would access computing capacity including approximately 325,000 NVIDIA GPUs. SpaceX also announced an approximately $40 billion financing lease for more NVIDIA equipment this week; that lease remains subject to closing conditions and future deliveries.
SpaceX, Cursor and OpenAI: Cursor says SpaceX acquired it in August. Cursor had previously named OpenAI as an investor. After the acquisition, OpenAI said it intended to wind down its contract supplying models to Cursor, with a proposed November 12 shutoff. A relationship can be both financially connected and strained at the same time.
Oracle, OpenAI and Anthropic: Oracle and OpenAI agreed to develop data-center capacity, and OpenAI has described Oracle delivering NVIDIA racks at their Abilene site. Oracle also says its own code scanners have called models from both OpenAI and Anthropic. That doesn’t establish how much Oracle pays either company or that Oracle owns OpenAI.
Amazon and the labs: Amazon has announced an investment and cloud partnership with OpenAI; Anthropic has separately announced Amazon investment and computing agreements. Both labs are buying or planning capacity from companies that also have a financial interest in them.
There are more connections than I can verify in one list. Some investments are private, some allocations aren’t disclosed, and access to NVIDIA chips can come through a cloud provider rather than a direct purchase. I can’t trace a particular dollar through this entire web. But the same names and commitments keep appearing in different places, and to me it’s starting to feel like a Silicon Valley circle jerk and a shell game happening in real time.
Here’s why that matters. A company building a data center has bills for land, power, construction and equipment. It may borrow against the expectation that a major customer will pay for computing capacity over time. If that customer can’t meet its obligations or cuts back, the provider may lose expected revenue. Lenders and investors then have to reassess the project, and a chip supplier that is also an investor may face more than one kind of exposure. Workers, customers and communities depending on the project can feel the effects too. A contract may have guarantees, termination rights or another customer waiting; the outcome depends on the actual agreement, not just the arrows on a chart.
The scale of the buildout makes the question harder to ignore. Companies are committing billions to infrastructure before they know exactly how much paying demand will be there years from now. OpenAI has said the vast majority of its regular users accessed its technology for free, in a February submission. In March, it reported more than 900 million weekly active ChatGPT users and more than 50 million subscribers. I wanted to know the exact percentage paying, but those are different measures and don’t establish the 20% figure I originally guessed. OpenAI also has business and API revenue, so consumer subscriptions alone don’t tell us whether its finances work.
That’s the part I don’t want us to miss when the next giant partnership is announced. Who actually owes whom? Which commitments are firm, which ones can be cancelled, and who is left holding the debt if the expected customers don’t materialize? When one shell cracks, those answers will matter more than another headline about how much money is moving around the table.
Prepared with AI assistance from my interview answers.
Dr. James Kelley
drjameskelley.com
Get The Crucible’s Gift on Amazon
Check your AI fluency
If this gave you something to think about, I’d love to have you back. Subscribe for more on AI, leadership, and the human part of work.
Sources
The Weekly Show with Jon Stewart: Bursting the AI Bubble with Ed Zitron, October 28, 2026; episode description.
OpenAI funding and NVIDIA partnership, February 27, 2026; NVIDIA and CoreWeave, January 26, 2026; CoreWeave and Anthropic, April 10, 2026.
NVIDIA and Nebius, March 11, 2026; Nscale funding, September 25, 2025, and Microsoft GPU agreement, October 15, 2025.
Anthropic, Microsoft and NVIDIA, November 18, 2025.
SpaceX prospectus: Anthropic compute agreements, June 2026; SpaceX GPU financing, October 2026; commitment subject to conditions.
Cursor acquisition, August 14, 2026; Cursor Series A, August 22, 2024; OpenAI on Cursor, August 28, 2026.
OpenAI and Oracle infrastructure, July 22, 2025; Oracle’s model access, October 2026; OpenAI and Amazon, February 27, 2026; Anthropic and Amazon, April 2026.
OpenAI submission on free use, February 2026; OpenAI user and subscriber figures, March 31, 2026; different measures, not a conversion rate.
Global Crossing annual filing, 2004; IMPSAT quarterly filing, 2003; Federal Reserve Bank of New York on railway mania, 2015.
Federal Reserve History: banking panics and railroad finance; Library of Congress: the Great Railroad Strike of 1877; Bureau of Labor Statistics: Silicon Valley after the dot-com bubble.


