The Investor

a16z's diagnosis, which I mostly share, is that the queue is mostly noise. ERCOT's study rules were built for roughly 50 large loads at a time, and 2025 brought 225 new requests by mid-November. Developers file across multiple sites, and many have no customer. Planners won't make ratepayers fund upgrades for projects that may never show up, so everyone waits. Queue positions are cheap. What's scarce is a funded project that can actually be energized.

The new deposit rules filter lightly. Public Utility Commission of Texas rules, effective October 8, add a $100,000 study fee plus $50,000 per MW, or $50M for a 1 GW campus. A project that slips two years forfeits only 20%, which is why a16z calls it a "thin filter." My guess is the capital-light flippers leave and the funded speculators stay, which shrinks the queue more than it cleans it.

Flexibility is the price of grid access

The approved projects share one trait. Each promised to back off the grid when called.

• xAI won grid power in Memphis by showing its batteries could carry its full load for four hours.
• The 760 MW Freestone campus was approved only on condition it can shed load within 30 minutes.
• Duke's Tyler Norris estimates ERCOT could add about 10 GW if new load accepted 0.5% curtailment.

Vantage markets its 1.4 GW Shackelford campus as off-grid, and it has a switchyard beside a high-voltage line. Off-grid, in practice, is a bridge to the grid. That makes site batteries and aggregators durable demand, with grid-enhancing technology in the same column. Bridge gas is a cyclical trade, and GE Power's $22B write-down after the last turbine boom is the precedent.

Several jurisdictions, different brakes

The jurisdictions below are slowing data center capacity for different reasons. Texas and the federal and Virginia measures concern grid load and who pays for it. Finland's halt rests on environmental grounds, a separate permitting risk. I would not read it as the Texas queue constraint reaching Europe.

Contracted capacity is the mirror image. Morning Brew reports Constellation Energy rose 12.25% to $300.40 in one session on a 20-year Google supply deal, which Bloomberg ties to 890 MW of new reactor capacity. The public market already pays a premium for signed offtake. Private companies with the same profile haven't been repriced yet, which is the opportunity, unless the public market is simply overpaying.

Source bias: a16z backs Heron Power and Base Power, both in the flexibility stack it favors, and the Piq study it cites uses Base Power's fleet data. I'd discount the flexibility-capacity math until someone outside the Base Power fleet reproduces it.

The structure of the deal matters more than its size. In TLDR Hardware's reading of the Financial Times report, Nvidia disclosed a stake in SpaceX in August, alongside a $500B Wall Street financing push. So the buyer is borrowing from private credit to pay a supplier that owns part of the buyer. In vendor-financed cycles, order books tend to look strongest just before buyer leverage gets tested. The same analysis names the uncomfortable precedent: late-1990s telecom equipment.

The collateral is a customer

The Information sets the numbers side by side. SpaceX burned $25B in the first half of 2026. The new raise is about 1.6x that burn and roughly 47% of the up-to-$84.5B Anthropic has agreed to pay SpaceX through 2029. Lenders are therefore underwriting Anthropic's ability to keep paying. Anthropic, for its part, needs SpaceX's financing to get the capacity it has contracted for. Each side's credit is the other's collateral.

That loop matters because Anthropic's own risks are mounting. a16z's policy brief notes that its Claude exclusion from Department of War procurement was upheld in court, and that it faces an FTC probe and an antitrust suit. None of that threatens the commitment directly. It does mean the anchor customer carries legal tails the bond buyers are implicitly pricing.

The 2027 collision

The financing is expected to close in 2027. Lisa Su says that is the year AMD will "substantially increase" chip supply, after growing it through 2026. If both land together, lenders take peak exposure just as scarcity eases. Whether or not the debt is formally secured by hardware, its economics depend on compute staying valuable. The variables to watch are GPU rental rates and secondary hardware values. There is another tell from The Information: SpaceX's AI unit will use competitors' models to power Grok Bot. Even the borrower is treating models as procurement. The chips' value therefore depends on utilization, not on any in-house model advantage.

Cost of capital now splits the compute field

Techpresso points out that with about 75% of the package in investment-grade paper, bond buyers, not growth equity, now price GPU capex for the largest buyers. That widens the gap with equity-funded neoclouds. Lambda shows the other side of the market. It is reportedly raising up to $4B at a $14.5B pre-money, led by Coatue and Blackstone, ahead of a 2027 IPO, and that listing has to clear about $18.5B for the round to work. Both financings are timed to the same 2027 window.

What is not yet established

Reuters could not verify the FT report, and SpaceX, Apollo, Nvidia and Pimco have not commented. The Information flags as low confidence two claims that rest only on URL slugs: that SpaceX stock hit an all-time low, and that a prior Apollo deal was about $3B tied to xAI. Don't cite the $40B in an IC memo as settled fact. The broader direction holds without it: compute financing is migrating to credit, and credit spreads, not venture rounds, increasingly set compute economics.

Reflection AI, the Nvidia-backed lab, is aiming its first open-weight model, Beam, at coding and agentic tasks, according to Morning Brew. That is arguably the most commercially proven thing the closed labs sell, and it is the large, steady-state enterprise and government spend an IPO story leans on, so the question worth asking is which revenue gets hit. Western buyers already avoid Chinese open models on security grounds. A trusted-origin open model that can be fine-tuned and budgeted predictably wins in procurement, where a spreadsheet beats a leaderboard most weeks.

Nvidia's role is a familiar one, a supplier funding its own demand, since every self-hosted open-weight deployment is GPU demand outside the closed-lab and hyperscaler channel. It is one of the older moves in hardware, and it looks clever for as long as the customers keep growing. AI Breakfast has a separate warning for vertical AI. HuatuoGPT-3, a 27B open medical model, beat GPT-6 Astra on HealthBench. When a model that runs on a single box beats a flagship API in a vertical, the moat sits in the data and the workflow.

Why the price gap won't hit closed-lab revenue one-for-one

• Self-hosting favors scale. Mistral Large 4 activates 49B parameters per token and still has to hold all 1T in memory, as Simplifying AI points out, which is a lovely efficiency number until someone has to buy the memory. The savings go mostly to large buyers with steady utilization. Mid-market buyers will use hosted inference, a serving-layer opportunity of its own.
• Closed labs can cut price. Morning Brew's base case has hyperscalers with long-dated power contracts lowering token prices to keep share. Share holds, revenue holds up, and gross margin compresses during the IPO window. The multiple follows the margin, so it is a multiple problem all the same.
• The claims are unproven. Mistral's 93% Cybench score and its visual-grounding win over Astra are self-reported. Beam's technical details arrive later this month.

Mistral's sovereignty pitch comes with sovereign risk

Mistral calls its model the "best outside China," a phrase MIT Technology Review flags. This is probably an uncharitable reading, but the phrase concedes the open-weight frontier and stakes the moat on sovereignty and political backing, which Macron frames as Europe's "third way." The backer is looking shakier. Morning Brew reports France's 10-year yield above 5% for the first time since 2002, French stocks down 5% in a month, populists leading the election race, and a government struggling to keep its deficit under 5% of GDP. Three fives in one sentence is a rough backdrop for a national champion. A populist government might favor a national champion, but it will have less money to spend on one. Government AI procurement in France is politically attractive and fiscally constrained.

The late-October release is an underwriting event. The useful evidence lands between release and consensus, and press claims in either direction make a poor basis for repricing. I'd rather wait for third-party Cybench runs than reprice off a self-reported score.