The Investor

The legal substance is narrow. The Information reports the notice asserts Oracle's right to withhold payments if delays persist, and Oracle posted the same morning that Jupiter "remains on our planned schedule." A company can send a withholding notice and still call the schedule intact; that is what negotiating instruments are for. The delay itself is a power problem. New Mexico denied a natural-gas pipeline permit in July and the project switched to Bloom Energy fuel cells. Analysts put that setup at roughly $8B, likely billions over the original plan. No one has disclosed the figure; it is an analyst estimate.

Two OpenAI megaprojects, two risk structures

Nvidia is spec-setter, supplier and creditor to the same tenant. It hired Chris Malone, OpenAI's former head of data centers, to run DSX, the unit that designs customer sites to Nvidia's specifications. A supplier with pricing power lends to keep its order book full, and vendor financing extends the cycle, or rather it extends the cycle and concentrates the unwind on one balance sheet if OpenAI has trouble paying.

The tenant with legal leverage gets a free option on delay. Developers and their capital providers sit in the middle with pricing power over neither side, and every quarter the site sits unpowered is a quarter that capital is not funding a permitted one. Blue Owl's lenders should be the first to mark this.

Where the reporting diverges

The Information has Oracle and Blue Owl each down more than 5% in early trading. Pivot 5 puts Oracle's move at about 4%, probably a different hour's print. Both companies say their commitments are unchanged. Bloomberg notes the site already faced community opposition alongside the regulatory setbacks, so the binding constraint is local consent, not one permit.

A pattern, not one site

• Other slips. AI Breakfast reports Nscale's Loughton site in the UK moved from 2027 to the early-to-mid 2030s, a decade rather than a quarter.
• Financing migration. A Brookings estimate by Van Nieuwerburgh puts the US AI buildout at $10.3T from 2025 to 2032, and warns that financing is moving off corporate balance sheets into JVs, private credit and SPVs.
• Credit stress. The Information ran a headline saying Jane Street-linked data center debt is souring. No details were reported, so confirm before relying on it.
• Bench depth. OpenAI's head of data centers left in August amid a wave of executive departures while the company targets a 2027 IPO. Those lease commitments are likely to surface as S-1 risk factors.

The next public test is Firmus, a ~$5B ASX data center IPO, with a bookbuild opening October 6 against a $77M first-half pro forma loss reported by AI Breakfast. It covers at book and nothing here got priced. It covers at a discount and public markets have finally put a number on unsecured power. It fails and the category stays with private credit on private credit's terms.

A signed AI capacity lease is only as firm as its force majeure clause, and this notice showed who gets stuck paying for the site while everyone waits.

Underwriting used to ask whether a tenant signed. It now asks which clause lets the tenant stop paying, and who carries the site meanwhile. This is probably wrong if power queues clear faster than anyone expects, but secured, permitted power should earn a premium from here, and on-site generation vendors gain leverage every time gas permitting fails. Bloom's substitution already cost billions more than the pipeline would have.

The price only makes sense once you know how the deal started. Latent.Space reports it grew out of joint fraud work between the two companies. OpenRouter blocked 10x more fraudulent dollar volume in one month than in the month before. The fraud came from stolen cards, resale that broke terms of service, hacked accounts, companies compromised without knowing it, and runaway agents. Anjney Midha invested early and appears to sit on OpenRouter's board. He calls the deal "a security story for the frontier AI ecosystem."

That framing fits Stripe's history. Its edge was never payment code. It was absorbing fraud losses and turning that data into Radar, its fraud-scoring product. OpenRouter sees traffic across dozens of labs, so Stripe now has arguably the broadest cross-model view of token fraud available anywhere.

What the multiple might be

Neither revenue nor take rate was disclosed. Latent.Space ran an illustrative underwrite that assumes a ~5% take rate. On that basis the price is roughly 38x net revenue if routed tokens average $1 per million, and ~19x at $2. These are estimates, not reported figures. Velocity matters more than either number. At ~9% week-over-week growth, token volume doubles about every eight weeks, so any multiple halves quickly while growth holds.

Treat the inputs cautiously. The $7B figure comes from a podcast teaser, not a formal disclosure. The developer count moved from a 4M guess to "over 10 million" pending de-duplication. The 9% growth rate was described as "sounds about right."

The router itself is getting cheaper

Three developments show routing logic losing its scarcity even as the routing position gains value:

• Open-source decision models. NVIDIA and Stanford released CLM-8B with code. It treats tool choice and request routing as retrieval against cached embeddings. Daily Dose of Data Science reports it matches TypeSafe's Jev on computer-use and gaming at up to 9x lower latency, though it trails slightly on tool calling.
• Bundled routing. Techpresso reports that Microsoft's Copilot puts OpenAI, Anthropic and its own MAI models behind one switcher. An Auto mode picks a model for each request, and agent work is billed by usage.
• Orchestration spreads. The Batch reports that Cognition's Devin Fusion pairs a frontier lead model with its SWE-2 sidekick. The combination tied Claude Code's score of 62 on Artificial Analysis' coding agent index at $7.90 per task, versus $12.40.

Taken together, the spread between frontier and cheap tokens is real and a business can capture it. But the classifier that picks the route is becoming a feature. What survives is ownership of the request path and the data generated on it.

Who gets squeezed

Latent.Space says general-inference resellers are squeezed from two sides. Fraudsters blocked at OpenRouter move on to them. Meanwhile labs are pushing customers into committed contracts and bring-your-own-inference arrangements, which erodes markup. Independent gateways now compete with a Stripe-backed incumbent that has a fraud moat. Adyen and other payments incumbents have no AI token flow of their own, which makes them plausible acquirers of token-fraud startups.

In a multi-model market, the router's logic is heading toward free; the flow it sits on, and the fraud data that flow produces, is what commanded a strategic price.

The smart move is to separate the companies in your portfolio and pipeline that own flow from those that only own logic, and to price the second group as features.

The usage data under the repricing is thin. Azeem Azhar cites early, low-confidence figures of about 3M Muse downloads and roughly 500K users. That implies one active user for every six downloads, and more than 95% of users were already on Facebook. Pivot 5 reports Meta told employees it added a human agent layer so Muse's calling feature could complete calls. Newcomer reports that Instinct's early users are drowning in CAPTCHAs. None of this is a churn or GMV figure, and no such figure sits behind the stock moves either.

Access is becoming a negotiated asset

Insurify co-CEO Giorgos Zacharia made the merchant case plainly: "A quote without its context is not a fair comparison. It is a number." The pattern is that supply owners set the terms. Incumbents are happy to run their own agents but refuse to let outside agents shop them. Google has historically served long-tail merchants poorly, which makes them the swing supply.

The loyalty problem in the business model

Zuckerberg told developers Meta will profit by taking a small fee on transactions made through Muse. Azhar's counterpoint is that the best agent jobs so far work against the merchant. An early Muse user recovered $250 of Delta credit in five minutes. Azhar's own agent won him a refund and filed a complaint with a hotel. A butler funded by partner fees, Expedia among them, is conflicted on exactly those jobs. As Azhar puts it, "The butler's loyalty can't be in two places. It will follow the coin." That favors user-paid or success-fee agents. It also points to a liability, insurance and anti-steering audit layer that does not exist yet.

What a $10B Instinct round would buy

Newcomer reports that Instinct is seeking money at $10B right after a $2.5B round. Its credible bull case is neutrality: a user-aligned agent that works where Muse is blocked. But if the Shopify integration isn't exclusive, integrations are table stakes, and Newcomer expects many more entrants. Paying winner multiples before holiday data arrives prices that neutrality edge as already won.

Sorting the selloff

Newcomer's framework is the right one: separate friction retention, where customers stay because leaving is annoying, from value retention, where leaving is costly. The first group is exposed to agents that switch and cancel on command. The second is being marked down alongside it. Aggregators look most exposed. Expedia fell even after signing its own Muse integration, and agents may go straight to the Amadeus and Sabre booking systems instead.

In agentic commerce, the gate belongs to whoever owns the supply, and the durable agent will be the one users pay for, not the one merchants pay for.

The verdict data arrives with holiday 2026: Muse Charm sell-through, whether Amazon's block holds, and Q4 churn commentary from the names that sold off.