Sam Altman once said, in effect: we do not know how we will make money. When the general system exists, we will ask it how to return capital to investors.
That line used to land as a joke. In 2026 it reads like a confession. The AI trade has grown so large that a leaked US Treasury draft reportedly warns the sector is more embedded in the economy than the dotcom firms were. Stock markets, private credit, data-center lenders, cloud providers, chip makers, and utilities all sit on the same assumption: demand for frontier models will eventually pay for the iron.
Vanessa Wingårdh’s OpenAI: A Bubble Bigger Than Dotcom walks the public case. Below is the structure in plain numbers and incentives, not hype.
The filing, the leak, the delay
OpenAI confidentially filed toward a public company in June 2026, a week after Anthropic’s similar move. Early talk pointed at a timeline as aggressive as September. Then independent reporting (Ed Zitron and others) put hard numbers on the burn.
The rough picture from that coverage:
| Claim | Scale |
|---|---|
| Last year’s burn | on the order of ~$38B (including restructuring costs as they shifted for-profit) |
| Spend vs revenue (cited path) | roughly $21B spent to make about $13B revenue |
| Target IPO story | up toward a $1T valuation while still loss-making |
| Capex-style commitments | on the order of $1.4T in multi-year infrastructure promises that cash flow does not fund today |
A few days after the numbers spread, reports said OpenAI was delaying the public path, with SpaceX’s volatile trading used as a soft excuse. The simpler reading: once unit economics and cash burn are public, a trillion-dollar story is harder to sell to ordinary investors.
When the CFO says no and the CEO still sells the dream
Reporting around CFO Sarah Friar’s pushback listed practical reasons to wait: Sora underwhelming as a product bet, executive churn, legal noise with Elon Musk, and commitments the company could not pay from operations.
That is a normal CFO job. The darker detail in the same reporting cycle: Altman allegedly kept her out of investor meetings about spending, which is exactly the domain she owns. Private caution, public confidence. That pattern is not unique to OpenAI. It is how bubbles talk.
Who is actually paying?
Consumers
OpenAI still leads usage share (the video cites roughly 54% market share, Gemini ~28%, Claude ~9%, Grok ~2%). Only about 5% of ChatGPT users pay. Billions in sales and marketing have not turned free users into a $20/month habit at the scale the cost base requires. Creators and power users openly switch to Claude or Gemini when quality or style feels better.
Even if every free user upgraded tomorrow, $20/month is not the price that funds this infrastructure stack. The product people love is not priced like the factory that trains it.
Enterprises
IBM’s CEO survey (cited in the video): only ~25% of AI initiatives delivered expected ROI; only ~16% scaled company-wide. Gains cluster in coding help, support assistance, and drafting. Everything else is still a pilot with a slide deck.
That matches what LLMs actually are: language tools. They are good at text. They are being sold as general labor replacement. Gap between pitch and payback is the bubble’s air.
Accidental $500M months
Usage-based enterprise billing made cost real. One company reportedly blew through on the order of $500M in a month on Claude after telling developers to “use AI for everything.” The industry swung from “AI everywhere” to “meter everything.” The next swing is cheaper supply: Chinese open and low-cost models (GLM, DeepSeek, Kimi).
OpenRouter token mix (as cited): Chinese models from ~1% of usage in late 2024 to over 60% by early 2026, ahead of US models. Coinbase cut AI spend ~50% by switching. Cursor built coding work on Kimi. Microsoft tested DeepSeek inside Copilot economics. Price is winning where quality is “good enough.”
Partners peeling off
If the product were locked-in and irreplaceable, partners would deepen ties. Instead:
- Apple’s Siri path moved toward Gemini, not ChatGPT as the long-term default story.
- Microsoft is no longer an exclusive cloud host and no longer pays a revenue share to OpenAI in the old structure. Joint “everything is fine” statements arrive on a regular schedule.
- OpenAI paid about $6.4B for io (Jony Ive’s hardware bet). Apple is now suing OpenAI over recruiting and alleged pressure on candidates to bring confidential device work into interviews.
- Key leaders keep leaving. That is not a vibe. That is operating risk.
A company can survive drama. It cannot survive drama and a cost structure that only works if the whole world pays premium prices forever while cheaper models close the gap.
Meta and xAI: from “we need more GPUs” to “please rent our GPUs”
Last year Meta said internal demand for compute was hard to meet even with 2025 capacity. Now Meta is selling data-center capacity outward. xAI is leasing capacity to Anthropic and Google. Leaders who said supply was the bottleneck are advertising spare capacity.
That does not prove AI is fake. It proves capacity was built ahead of durable paid demand. Overbuild is how every infrastructure boom looks right before the spreadsheet turns.
Oracle borrowed on the order of $43B for data centers, with huge contracted future revenue (hundreds of billions cited) and a large OpenAI slice. Then came reports of Stargate-related Texas expansion plans being cut, and Oracle job cuts tied to cash pressure from the AI build. At the same time, ~$64B of data-center projects have been blocked or delayed by local opposition. Power, land, water, and politics are choke points the pitch decks skipped.
Treasury draft vs the podium
The most important part of Wingårdh’s piece is not gossip about one CEO. It is the leaked Treasury draft for senior financial officials: AI firms sit deeper in markets and credit than the 2000 internet cohort. A sharp AI reset would hit equities, private credit, data-center finance, clouds, chips, and utilities together.
That is “systemic” language. While analysts draft that risk, public remarks from senior officials still celebrate hyperscaler spend on the order of $750B this year and ask everyone to “keep an open mind.” Private risk memo, public growth speech. Retirements, power bills, and tax-supported infrastructure sit in the middle.
The draft’s three danger conditions are already easy to check off in the video’s framing:
- Missed productivity goals (enterprise ROI mostly missing)
- Choke points (power, local bans, financing concentration, partner pullback)
- Tighter money (OpenAI reportedly shopping preferred equity with ~17.5% minimum return to private equity, a sign of expensive last-resort capital)
Preferred equity at that coupon is not a victory lap. It is the cost of capital when ordinary confidence is thin.
“Federal backstop” and the 5% idea
Friar’s earlier “federal backstop” talk (later walked back) and reports that OpenAI floated a 5% stake for the US government fit one pattern: privatize upside if the boom continues, socialize downside if it does not. Direct government ownership stakes across industry (the video cites tens of billions across dozens of deals since early 2025) make “too big to fail” less theoretical.
Whether Congress ever rubber-stamps a sovereign stake is politics. The ask itself is the signal: the private market alone may not want to carry the full bag.
What this is not saying
AI is not worthless. Coding assistants, support triage, and drafting already move real work. Open models will keep improving. Energy and chip supply chains will still matter.
The bubble claim is narrower:
- Prices and valuations assume AGI-scale payoffs on LLM-scale products.
- Cash burn and infrastructure promises outrun paid demand.
- Cheap substitutes are already eating token share.
- The loss path is no longer contained inside a few venture portfolios. It sits under pensions, power grids, and public balance sheets.
Dotcom had pets.com and also Amazon. The mistake is treating every GPU lease as Amazon in 1999.
Practical takeaways
- Separate “AI is useful” from “this equity is priced for miracles.” Useful tools can still destroy capital if cost of compute stays above willingness to pay.
- Watch free-to-paid conversion, enterprise ROI studies, and open-model share, not demo day videos.
- Treat circular infrastructure finance as credit risk, not pure demand proof. See also the Nvidia circular-deal story on this blog.
- If your company mandate is “use AI everywhere,” put a meter on it before you get a $500M invoice.
- Regulators muttering “systemic” in private while cheering capex in public is late-cycle behavior. Plan household and portfolio risk as if the soft landing is optional.
Sam’s original plan was to ask the model how to pay investors back. The better question for everyone else is simpler: when the music slows, who is still holding the GPUs, the debt, and the power contracts?