Cheap Intelligence Is Making the AI Buildout Bigger

One Amazon project running on Anthropic's Claude Sonnet came in 860 percent over budget, a $1.8 million bill. Uber spent its entire annual AI budget in four months and answered by capping what employees can spend on AI coding tools at $1,500 per employee per month ^[1]. These are the bills of a cheap technology. OpenAI cut the fees on its Luna model by 80 percent this summer to hold customers against cheaper Chinese rivals, Anthropic released a high-performance model at half the cost of its top system, and by September the industry had largely finished converting itself into a metered utility, charging for intelligence by the unit rather than by subscription ^[2][3]. Cheaper than ever, and the bills keep climbing. Wayne Liu, an industry analyst, calls it the token paradox: as the cost of a single AI answer falls, the number of answers bought rises, and the spread of AI agents, software that makes its own requests instead of waiting for a person to ask, multiplies the effect ^[4].

Cheap intelligence creates value only when it improves judgment. — Liu Ruilin

Customers have started acting on the arithmetic. PNC's chief executive has explained why the bank will own its GPU compute: whatever productivity AI lends a bank, the cost of tokens, the metered units of AI use, can take back ^[5].

Any impact that AI can have on the productivity of a bank, that productivity can be taken away by the cost of tokens. — Bill Demchak

The efficiency push is real, though, and the industry is candid about where it came from. The expensive part of AI stopped being teaching the models and became answering with them: inference now accounts for 70 to 85 percent of what enterprises spend once their systems are live, which makes power and silicon, not model size, the binding constraint on the business. Market share is migrating to whoever prices efficiency best: OpenAI's fell from half to about a third while Anthropic's doubled ^[6]. The industry has a name for the bind, and a name for the way out. It calls the point where scaling models stopped returning enough to justify the power and capital the "Utility Wall," and its new headline metric is the Unit Cost of Intelligence ^[7]. The way out was pitched, from the first, as adaptation to the wall rather than escape from it. In the spring, two suppliers, SiTime and Confluent, proposed precision-timing and data-streaming changes to cut AI's wasted energy, in the same weeks that Goldman Sachs projected global data-center power demand would rise 165 percent by 2030 and as some U.S. regulators began limiting large data center developments ^[8].

The transition from batch processing to data in motion isn't just an upgrade - it's the only way for AI's future. — Simon Laskaj

Power has become a critical constraint even at the level of the chip, and the engineering bends around it: chipmakers now optimize the movement of data, because the moving eats more energy than the math ^[9]. The cleanest test of whether the plan works is OpenAI, because OpenAI ran the experiment on itself. In March it behaved like a company that believed its own doctrine: it canceled Sora, its experimental video model, and killed a $1 billion Disney licensing deal, shut down Instant Checkout, ran a ChatGPT advertising trial annualizing at $100 million, and cut its infrastructure-spending target for 2030 to $600 billion, from an earlier plan of $1.4 trillion ^[10]. By September the target had crept back. Expected infrastructure spending from 2026 through 2030 now stands at $856 billion, against plans to spend $278 billion more cash than the business generates across those years. OpenAI is seeking a valuation of $1.2 trillion to $1.5 trillion to bridge the gap, on revenue that annualized above $40 billion in July, expects to run through the $122 billion it raised in March by 2028, and is co-designing an inference-optimized chip with Broadcom, code-named Jalapeño. The doctrine has reached the silicon ^[11]. OpenAI is the pattern, not the exception. The hyperscalers, the big cloud providers, keep raising their 2026 and 2027 capital-spending plans even as the cash on hand to cover them thins, a gap Goldman's own data compares to the late-1990s dot-com build, at stock-market concentrations comparable to Japan at its late-1980s peak ^[12]. Morningstar's Philip Straehl put the mechanism into words more than a year ago ^[13].

the AI party ... started at 9pm. It's now 10pm. That party goes to 4am. — Dan Ives

The direction of the price war is the tell. The U.S. labs keep discounting to hold share, while DeepSeek, the Chinese lab whose cheap models forced the repricing in the first place, is raising prices from traction. OpenAI's operating loss, for its part, widened to $12.3 billion in the second quarter ^[2][14]. Meanwhile, the physical layer keeps raising its numbers. The world's data centers drew 787.8 terawatt-hours of electricity in 2025, up 92 percent since 2020, and S&P Global Energy expects 1,550 terawatt-hours by 2030, the point at which it judges energy availability may become the primary constraint on deploying AI ^[15]. In the United States, of 3,969 announced data-center facilities, 802 are actually under construction. The rest wait on shortages of labor, materials and chips ^[16]. The refusals have moved into the price of credit. Opposition groups have more than doubled, to 833 across 49 states ^[17], and JPMorgan, Morgan Stanley and Bank of America now formally weigh community sentiment and permitting readiness in data-center lending, after at least 75 projects worth about $130 billion ran into local resistance in the first quarter ^[18]. Loudoun County, Virginia, which collects $1.2 billion a year in data-center taxes, about 40 percent of its budget, voted a 12-month pause on new applications anyway, after residents reported respiratory problems and noise from gas turbines and diesel generators ^[19]. The debt is repricing in real time. Oracle's $18 billion of loans against Project Jupiter, an AI campus built to serve OpenAI whose 2.2 gigawatts of turbines are waiting on a natural-gas pipeline, have traded at 89 to 91 cents on the dollar since a state land office blocked that pipeline amid local air and water objections; S&P has since cut Oracle to one notch above junk, and the banks Santander and Jefferies are holding more of the debt than they planned to ^[20]. One state land office, in other words, repriced $18 billion of syndicated debt. Apollo's chief economist, Torsten Slok, reads the same signal across the sector: the price of insuring the big cloud companies against default has risen roughly 60 basis points more than the banks' own since October ^[21].

What the market is repricing is hyperscaler credit fundamentals, namely a debt-financed AI capex cycle with rising leverage, negative free cash flow and uncertain payback on depreciating assets. — Torsten Slok

The search for a way around has been taken literally: data centers in orbit, from Google's Project Suncatcher to Aetherflux's Galactic Brain to ventures by Eric Schmidt, SpaceX and Blue Origin, at a prohibitive $1,500 per kilogram of launch cost ^[22].

The elephant in the room is that our current energy plans simply won’t get us there fast enough. — Baiju Bhatt

The industry is squeezed from both ends. The price of what it sells keeps falling, while Nvidia and Samsung have raised server and chipmaking prices by up to 15 percent because supply keeps falling short, and Broadcom negotiates more than $60 billion in debt with the cost of insuring it rising ^[23]. There is a serious case that none of this hardening matters yet. Treasury Secretary Scott Bessent has said AI companies issue debt whatever borrowing costs are doing ^[24], and the money has not stopped: Amazon and Alphabet are funding $725 billion of capital spending with record bond sales in euros, yen and sterling ^[5]. Spending on information-processing hardware reached $752 billion in the second quarter, past its early-2021 peak and now larger than all U.S. residential investment ^[24]. The banks writing community sentiment into their credit models are the same banks still committed to the sector on Goldman Sachs' forecast of more than $6 trillion in big-tech AI spending through 2030 ^[18]. Anthropic, the revenue leader, has booked its first operating profit, $559 million, and projects revenue climbing from a $47 billion run-rate to $200 billion by 2028 ^[14]. And Glen Kacher of Light Street Capital argues the buildout is a 15-to-20-year infrastructure cycle rather than a bubble, in which the energy demands and community opposition are, in his words, "education challenges rather than fundamental investment risks" ^[25]. The industry's two ledgers do not reconcile yet. BCA Research calculates that AI needs roughly $10 trillion in annual revenue to justify current capital spending, and that depreciation charges at the hyperscalers alone will more than double, from $255 billion in 2026 to $581 billion by 2029 ^[26]. Against that, a Maximand review of three years of earnings calls at the 60 largest U.S.-listed financial firms found one able to cite a realized dollar return from AI. One ^[23]. The industry can now measure, to the fraction of a cent, the one number that keeps falling: the cost of a unit of intelligence. What would justify all of it, the number a unit of intelligence earns back, remains mostly a forecast.