AI's Bill Came Due in Four Places Thursday

A gas pipeline in New Mexico, a toll booth inside a chatbot, a bid that died, and a sector that rose while tech fell. Four stories, one balance sheet.

HeyTheo Research — September 25, 2026

Quick Read

Oracle told its landlord it may not pay on time because a gas pipeline is late. Meta explained how its AI agent will actually make money, and the answer is a cut of your shopping. Cybersecurity rose while the rest of tech fell. And the man who tried to buy $18 billion of casinos specifically because AI cannot disrupt them walked away. Thursday was not four stories. It was one question asked four ways: what does the AI trade actually cost, and who collects?

The S&P 500 finished Thursday down 0.07%. On paper, nothing happened.

Underneath, four things happened that will matter more in six months than the index print. A 2.45 gigawatt data centre campus ran into a permit problem. A company with 3.5 billion users described a business model that charges per transaction instead of per impression. A basket of security software outperformed the Nasdaq by a wide margin for the third session in a row. And an $18 billion take-private died.

Each one is being covered as a separate item. They are not separate. They are the same story at four different points on the same invoice.

The day in four numbers

Source: TheStreet, Yahoo Finance, Bloomberg, Axios

Backdrop: the 10-year Treasury yield closed at 5.192%, a fresh 52-week high. The 30-year hit 5.476%, the highest since 2004. Brent settled at $104.70.

The power bill

Oracle sent a force majeure notice to Blue Owl Capital, the developer of Project Jupiter, its flagship Stargate campus in New Mexico.

Read the mechanics, because the headline undersells what happened. Force majeure here does not mean Oracle is walking. It means Oracle has reserved the right to delay payments if the site misses its 2028 target. Oracle says the project remains on schedule. Blue Owl says the financial commitments are unchanged. Both can be true and the notice still tells you something.

What it tells you is the reason. The campus is designed for 2.45 gigawatts, powered by gas fuel cells rather than the grid. The gas has to arrive through a pipeline. That pipeline has been pushed back roughly six months, to February 1, 2027, after permit denials. A separate air-quality permit for the fuel cell system faces a state deadline of November 23.

So the constraint on the largest AI infrastructure project in the country right now is not chips, not capital, not demand. It is a pipeline route and an air permit.

That reframes a lot. For two years the AI capex debate has been about whether the compute gets built and whether anyone will rent it. Thursday added a third question that is harder to model: can the electrons physically get there on time? Gigawatt-scale campuses need power that does not exist yet at the location where it is needed. Building generation, transmission or pipeline takes years and requires permission from people who are not thinking about your earnings model.

Now put that next to Oracle's own numbers. In its most recent quarter the company reported revenue of $19.3 billion, up 30%, with cloud infrastructure up 121% to $7.4 billion. Remaining performance obligations reached $664 billion. Those are extraordinary figures.

The other side of the ledger is why the stock fell 6.44%.

Source: Oracle investor relations, company filings

Three things in that table deserve more attention than they get.

Free cash flow is negative by $5.4 billion, against negative $362 million a year earlier. Gross margin fell nearly eleven percentage points in twelve months, because renting out GPUs is a fundamentally lower-margin business than selling database licences. And interest expense is growing 55% a year against a $125 billion debt load, at a moment when the 30-year is at a 22-year high.

A company funding a $90 billion capex programme with debt, in a market where the long bond yields 5.48%, has a narrow path. It works if the backlog converts on schedule. Management expects about half of the $664 billion to become revenue over 36 months. Thursday was the market being reminded that "on schedule" now depends partly on a pipeline permit in New Mexico.

The toll booth

The same week, Meta answered the question Oracle cannot yet answer: where does the revenue come from?

Muse, Meta's personal AI agent, launched on September 8 with three tiers. Free, a $20 monthly tier, and a $100 monthly tier, differentiated by weekly token allowance rather than features. Then at Connect the roadmap landed: an on-screen avatar for video chat, wake-word control through smart glasses, the ability to operate any app on a Mac while the user steps away, and an email address of its own so you can forward tasks to it.

The commercial detail is the one that matters. Zuckerberg said Meta expects to profit by taking a small fee from transactions the agent completes. Integrations were announced with Stripe, Shopify, PayPal, Best Buy, Gap, Sephora, Walmart, Wayfair, Expedia and Instacart.

That is not a subscription business dressed up as AI. It is a take rate.

Consider what changes if it works. Meta's existing model sells attention and charges per impression or click, with the actual purchase happening somewhere Meta cannot see. A fee on completed transactions moves the company from the top of the funnel to the bottom of it. Advertising revenue is cyclical and competitive. Payment take rates are neither, which is precisely why payment networks trade the way they do.

The obvious caution is that none of this is proven. Meta has not published per-task limits, only weekly token ceilings, so nobody outside the company can compute a cost per completed transaction yet. Agentic commerce has been announced far more often than it has been used. And the free tier requires a payment card on file before you can use it at all, which tells you Meta expects abuse and is screening for it.

But the direction is clear, and it is the first credible answer any large consumer platform has given to the question of how AI pays for itself at scale. The market gave Meta 4% for it on a day when Oracle lost 6.44% for the opposite reason.

The side effect nobody priced

Now put the two together, because this is the connection most coverage missed.

Meta just described an agent that will hold your payment credentials, operate your desktop unattended, receive email at its own address, and transact on your behalf with a dozen retailers. Multiply that by every company shipping something similar.

Every one of those agents is a new identity that has to be authenticated, a new set of credentials that can be stolen, and a new path into systems that were designed assuming a human was at the keyboard.

Which is roughly what the tape has been saying for two weeks. On Wednesday, CrowdStrike, Palo Alto Networks, Okta and Palantir each rose about 4% while the Nasdaq-tracking QQQ fell 0.83%. The cybersecurity ETF rose 1% on the same session. Earlier in the month the moves were larger, with CrowdStrike up 15%, SentinelOne up 14% and Palo Alto up 10% across a week.

There was no earnings surprise behind any of it. No company-specific disclosure has been verified for these names. It is a sector bid, and sector bids without catalysts usually reflect positioning rather than news.

The positioning logic holds up better than most rotations do. Security spend is close to non-discretionary, because systems already deployed still have to be defended whether or not the budget is growing. It is capital-light, so it does not need cheap debt to expand, which matters at a 5.48% long bond. And the spread of autonomous agents widens the attack surface rather than narrowing it, so the AI buildout increases the addressable market rather than threatening it.

That last point is the interesting one. Almost every other software category faces the question of whether AI replaces it. Identity and endpoint security faces the opposite question. If agents proliferate, there is more to secure, not less.

We covered this sector in our earlier note on where capital rotated across US industries. The difference now is that the bid has a specific mechanism behind it rather than a general theme.

The hedge that failed

The fourth item looks unrelated until you read why the bid existed.

Barry Diller's People Inc., which holds about 27% of MGM Resorts accumulated since 2020, offered more than $18 billion in June for the rest of the company. Diller sits on the MGM board. On Wednesday he withdrew, saying the mix was not coming together and that he was passing at this time. MGM fell 9.98%. The board said it remains excited to lead MGM as a standalone company. Diller said he stays open to future transactions.

The reason for the original bid is the part worth keeping. Diller framed casinos and live entertainment as a hedge against AI disrupting media. Physical experiences, the argument went, are insulated from a technology that is currently dismantling the economics of everything that can be delivered through a screen. Buy the thing the software cannot eat.

He was not wrong about the premise. He simply could not make the arithmetic work at 5.2% on the ten-year.

That is the quiet lesson. The AI-proof asset is a real category and the thesis is sound. But AI-proof assets tend to be capital-heavy, and capital-heavy assets get priced off the cost of debt. When money was cheap you could buy your way out of technological risk. At current rates that exit is expensive enough that one of the more experienced deal-makers in American media looked at it in June and passed in September.

Two caveats we will not paper over. The reporting attributes the withdrawal to deal structure rather than explicitly to financing costs, so the rate connection is our reading rather than a stated reason. And a 27% holder walking away from a control premium leaves an obvious question about what he saw in the diligence that the rest of us have not.

What the four have in common

Strip the specifics and the same sentence fits all four.

Oracle shows the cost side of AI arriving faster than the revenue, and running into physical limits that money cannot compress. Meta shows the first plausible revenue model, and it is a toll rather than an advertisement. Cybersecurity shows the second-order winner, the business that grows because the buildout creates work for it rather than despite it. And MGM shows that the escape hatch, buying assets AI cannot touch, has its own price now that debt is expensive.

For an investor the practical translation is that "AI exposure" has stopped being one thing. In 2024 it was a single factor and nearly everything with the label moved together. Thursday was the market pricing four different relationships to the same technology in four different directions on the same afternoon.

The rules HeyTheo tracks

Themes like this are only useful when they become something you can check.

Inside HeyTheo, the relevant baskets here are AI infrastructure, security software, and consumer platforms, each with rules that trigger on fundamental and technical conditions rather than on narrative. Where a name in this note sits inside one of those baskets, the trigger history is visible, and you can check the rule behind any trigger rather than taking the output on trust.

Ask Theo is the faster route for the specific question this piece raises. Useful ones to put to it: which names in the AI infrastructure basket carry negative free cash flow, how the security software basket has behaved on days when the Nasdaq fell more than 1%, and which consumer platforms have disclosed transaction-based revenue.

You trade through your own broker. HeyTheo helps you decide.

The Invoice

Three things go on the desk calendar from Thursday.

November 23 is the air-quality permit deadline for Project Jupiter's fuel cells. February 1, 2027, is the revised date for gas to reach the site. Between those two dates sits the credibility of a $664 billion backlog, and by extension a large part of the AI infrastructure trade.

The second item is the next hyperscaler capex guide. Oracle has now shown what the cost curve looks like when revenue is real, growth is 121%, and free cash flow is still negative. Whoever guides next will be read against that.

The third is whether security holds its bid on the next genuinely red day. A rotation that works when the Nasdaq falls 0.83% is a preference. A rotation that works when it falls 3% is a position.

None of this needs a view on whether AI is a bubble. It needs an understanding that the technology now sends four different kinds of bill, and that owning it is no longer a single decision.

FAQs

What does Oracle's force majeure notice actually mean?

It is a contractual protection, not a withdrawal. It preserves Oracle's ability to delay payments to the developer if the New Mexico campus misses its 2028 operational target. Oracle has said the project remains on schedule and the developer has said financial commitments are unchanged. The notice matters mainly for what it reveals about the cause, which is delayed energy infrastructure rather than construction or demand.

Why is Oracle's free cash flow negative if revenue is growing 30%?

Because capital expenditure of $28.5 billion in the quarter exceeded operating cash flow of $23.0 billion. The company is building data centre capacity ahead of the revenue that will eventually use it. Guidance implies $90 to $95 billion of capex for the full fiscal year. The gap is a timing question rather than a demand question, but it is being funded against $125 billion of debt at a moment when long-term rates are at 22-year highs.

How would Meta make money from Muse?

Through a fee on transactions the agent completes on a user's behalf, alongside subscription tiers at $20 and $100 a month that differ by usage allowance rather than features. Meta has announced retail and payment integrations to support this. The model is unproven and the company has not published the per-transaction economics.

Why did cybersecurity stocks rise while the rest of tech fell?

No verified company-specific news explains it, so the move reflects positioning. The argument being expressed is that security spending is difficult to cut, requires little capital to grow, and expands as autonomous agents multiply the number of identities and endpoints that need defending. That makes the sector a beneficiary of the AI buildout rather than a target of it.

Does the MGM bid withdrawal say anything about the broader market?

Possibly. The bid was framed as a hedge against AI disruption, on the logic that physical entertainment is insulated from software. Withdrawing it while the 30-year Treasury yields 5.476% is consistent with expensive debt making large take-privates harder, though the stated reason was deal structure rather than financing. Treat the rate connection as interpretation rather than fact.

Sources: Bloomberg, Reuters, TechCrunch, SiliconANGLE, Axios, Skift, CNBC, TheStreet, Yahoo Finance, Oracle investor relations, 24/7 Wall St.

Disclaimer: HeyTheo is a research and education platform, not an investment adviser or broker-dealer. Nothing here is advice to buy, sell, or hold any security. You trade through your own broker; HeyTheo helps you decide. Backtested results are hypothetical and do not guarantee future returns. References to governments, officials, or policies are for market context only and are not political endorsements. All investing involves risk, including loss of principal. Data is as of the dates noted.