The AI Buildout Is Burning Cash and Losing Permission at the Same Time

Alphabet's latest quarter put the AI buildout in a posture it had never struck before. The company reported its first-ever negative free cash flow, minus $5.9 billion, in the same window where it raised capital-spending guidance toward $205 billion for the year ^[1]. The cushion shrank while the commitment grew. Hold that shape, because the whole industry is starting to take it. Meta's quarterly free cash flow dropped 91%, to $784 million ^[1]. OpenAI's quarterly operating loss widened to $12.3 billion on $6.7 billion of revenue, with margins falling even as revenue grew 18% ^[2]. Z.AI, a Chinese lab whose annualized revenue has grown sixfold in five months, went to Hong Kong for its second fundraise of the year and pulled in $5 billion — a share placement priced at a 10% discount to the stock, plus $3 billion in convertibles. The market marked the stock down by more than a tenth on the news, leaving it more than 60% below its June high ^[3]. A raise that immediately costs the raiser a tenth of its value is not relief arriving; it is a symptom being priced. This is the pattern that has the analyst Joe Albano calling the boom's financing a "house of cards," and that verdict stays his ^[4]. The ledgers are uncomfortable enough without it. The refusal from the balance sheet is one half. The second came from the opposite direction, from the public, and it arrived with a bill attached. The labs want AI treated as essential public infrastructure, the next power grid ^[5]. The public has obliged, by treating the buildout as a tenant that raised the rent: household electricity rates rose nearly 10% in a single year, an increase attributed in significant part to data-center hunger ^[6]. In New Jersey, where household electricity costs jumped 22%, Governor Mikie Sherrill signed a law requiring large data centers to pay for their own grid upgrades, and more than 25 municipalities have banned the facilities outright. New York wrote the country's first statewide moratorium on new large-scale centers ^[7]. Texas produced the summer's neatest bureaucratic object: a state that blocked new data centers from its grid pending a water audit that will not finish until December, when its own official water plan contains no projections for data-center water use at all, so the audit must first find out what the plan never thought to ask ^[8]. By midsummer, refusals had spread across at least eight states: Missoula County, Montana, froze development for a year; Florida counties recommended one- and two-year pauses; Indianapolis proposed a countywide freeze; Virginia's Frederick County struck data centers from its zoning entirely ^[9]. By early this year, communities had blocked or delayed more than $130 billion of projects, a $1 billion Google proposal and a $3.6 billion Amazon campus among them ^[10]. Held apart, these are two bad stories. Held together, each one makes the other harder to survive. OpenAI's plan to grow from $25 billion of revenue this year to $200 billion by 2030 rests on capacity the company does not own: Oracle's $300 billion compute contract, which starts in 2027, and 10 gigawatts of custom chips from Broadcom ^[11]. Every gigawatt a state now refuses to connect is a subtraction from the projection that justifies the spending. The worry has already reached the middle of the chain: Oracle's shares have nearly halved on concerns about OpenAI's ability to pay, despite a $638 billion backlog ^[12]. The pressure runs the other way too. The one demonstrated route through a local refusal was bought: in Michigan, the Stargate developers, Oracle and OpenAI among them, sued Saline Township over a rezoning denial and settled on a consent judgment that included $14 million in community benefits ^[10]. Consent has a cash price. Cash is what the first refusal is draining. The case that none of this bites is genuinely strong, and it deserves its full form. Jefferies tells clients the buildout's binding constraint is the supply chain, not politics: in 2025 only 8.9 gigawatts of data-center capacity came online against 21.1 gigawatts of demand, with roughly $2 trillion in cloud backlogs waiting ^[13]. Piper Sandler began covering five AI chip stocks this weekend with buy ratings on all five, projecting a fivefold compute market by 2030 and order books full through 2028 ^[14]. Nvidia just reported $89 billion of data-center revenue in one quarter, up 117% ^[15]. Anthropic posted its first operating profit, $559 million on $11.6 billion of revenue ^[2]. And the analysts reading the New York and Texas halts call them temporary and expect demand to outlast them ^[16]. All of it can be true, and the collision still happens on the calendar. A halt needs no permanence to matter; it needs only to stand past 2027, when Oracle's contract starts billing, and Texas has already said its audit will not answer until December. Anthropic's profit does not pay OpenAI's leases. The demand is enormous. The test of the next two years is cruder: whether permitted concrete and patient cash can be had at the same time. On September 13 the argument about pace stopped being abstract. The labs set one pole: Dario Amodei published a manifesto asking to slow both AI development and the construction of the data centers themselves, backed by Sam Altman, Elon Musk and Demis Hassabis; President Trump rejected it, and skeptics read the gesture as the strongest labs asking regulators to freeze the field with them on top ^[17]. Washington set the opposite pole, with Treasury Secretary Scott Bessent demanding faster construction and refusing any pause ^[18]. The states pulled the brake, New York and Texas moving to halt new data centers ^[16]. Slow down, speed up, stop: three levels of authority now demand three opposite paces for the same construction, and none of the three can set one. Asked why the United States could not pause, Bessent answered with a single sentence.

If they were to pull ahead of us on AI, then nothing else matters. — Scott Bessent