Who Gets to Build AI Is Now a Balance-Sheet Question

In March, the Joliet city council voted 8–1 to approve a $20 billion, 1.8-gigawatt technology campus. It was the friendly vote, and it still came with $2.1 billion in local tax revenue earmarked over thirty years, written into the annexation agreement, with no tech tenant signed ^[1]. Permission, it turns out, has a sticker price.

This annexation agreement sets clear expectations for infrastructure, services, and community benefits while helping ensure the project moves forward in a responsible way. — Beth Beatty

Seven months on, that price has gone general. In one October week, all four hyperscalers put roughly a billion dollars each on the table for host communities: Amazon and Meta in incentives, Google in community funds plus full property taxes, Microsoft by forgoing tax breaks ^[2]. Indiana, where thirty counties restricted data centers in July, counted at least fifty by late October ^[3]. Around twenty gubernatorial candidates are now running on moratoriums ^[4]. An industry taskforce that includes Nvidia, Microsoft and Cisco has now carried the same logic abroad, asking the UK government to pre-purchase consent before the opposition to data centers even exists ^[5].

Britain has a window – we think of roughly 12 months – to build a compact that earns public consent before that pattern takes hold. — Sovereign AI Leadership taskforce

Those pledges landed in the same days that lenders signaled they were increasingly unwilling to finance the buildout regardless of cost ^[6]. A neocloud's IPO was pulled, OpenAI's listing was pushed out, and CoreWeave's bonds were trading near 15% yield ^[7][8][9]. The companies' stated purpose for the payouts is to ease local opposition, not to answer bond yields ^[2]. The two rejections run in parallel. What fuses them is the shared denomination: cash. And cash is the one thing the industry's two kinds of builders hold in unequal amounts. The neocloud tier, the firms renting GPU capacity between the chipmakers and the companies buying AI, is splitting on capital structure rather than on demand. Iren, an Italian group, funds 45% to 55% of its GPU spending through customer prepayments and holds a current ratio of 3.55; Nebius self-funds on $9 billion of prepayments ^[9][10]. CoreWeave grew revenue 112% and carries a $104 billion backlog, and funds it at 8.94 times debt to equity, with negative free cash flow of $5.74 billion and a current ratio of 0.46 ^[10][9]. Same sold-out demand, opposite books. Nvidia's chief executive has named the bottleneck himself.

What is missing in the middle is the working capital to fuel that growth… because there's just so much demand across so many sectors of AI, there's a scarcity of capital — Jensen Huang

Nvidia is CoreWeave's second-largest shareholder, and even Huang's public defense of its finances could not stop an 11.9% share plunge ^[11]. The market is already pricing the two books differently ^[9]. None of this is a spending collapse. Hyperscalers are still expected to put up roughly $720 billion in capex this year, and chipmakers are posting record revenue on sold-out capacity ^[12]. What decelerates is the growth rate, not the level: Morgan Stanley projects capex growth slowing from nearly 100% in 2026 to 14% by 2028 ^[13]. Calling this a funding crisis mislabels a sort as a collapse. Huang's word for the squeezed tier is "the middle," and the ledger's answer is that the middle is what the sort removes. Gartner expects more than 80% of neocloud providers to cease operating as independent businesses after 2030 ^[14]. Bain counts $68 billion of projects delayed or blocked by local opposition and power shortages in a single quarter ^[15]. That is a sorting fee, not a stop sign.