AI Is Getting Cheaper and More Expensive at the Same Time

In December, the company that had spent two years telling the world its models were everything was being priced on exactly that — and losing. GPT-5 had landed to tepid reviews, Google's Gemini 3 was winning downloads, and investors in OpenAI's last round were bracing for a down round, a deal that values a company below what it was worth the time before ^[1]. Sam Altman told his staff the mood would not improve quickly.

need to stay focused through short-term competitive pressure . . . expect the vibes out there to be rough for a bit. — Sam Altman

By March, the same company was worth $852 billion ^[2]. The new number arrived bolted to physical commitments: Nvidia agreed to put in as much as $100 billion and to stand up at least 10 gigawatts of computing — a gigawatt is roughly the power appetite of one very large data center — against a $638 billion backlog of contracted orders ^[2]. It was not a one-off. When Anthropic closed its $65 billion round in May at $965 billion, briefly worth more than OpenAI, the deal came bundled with five gigawatts of Amazon capacity, five gigawatts of Google-built chips, and access to the giant Colossus data center fleet ^[3]. One venture investor now says the logic out loud. Hans Tung of Notable Capital argues that dominance now runs through physical compute rather than model quality — the United States holds at least 47 gigawatts of data-center computing against China's 24 ^[4].

If you don’t have enough compute to meet demand, it doesn’t matter if you have a better model. — Hans Tung

Strip away the round sizes and a simpler fact remains: one industry, two prices, moving in opposite directions. The intelligence — the models, the tokens, the software that sells access to them — keeps getting cheaper. The capacity — the chips, the power, the buildings — keeps getting dearer. On the intelligence side, the price war is no longer a rumor. Usage-based token billing gave corporate buyers sticker shock — one ride-hailing company blew through a year's token budget in months — and firms now route routine work to cheaper open models while OpenAI is reported to be weighing price cuts of its own ^[5]. The software that sells intelligence has been repriced launch by launch: Anthropic's April announcements, a $30 billion revenue run rate and a batch of "agent" products, sent ServiceNow and Salesforce down within the week, and its security model knocked Qualys and Zscaler ^[6]. Salesforce, Adobe, and ServiceNow together shed roughly $160 billion of market value in 2025 ^[7]. The incumbents are behaving like companies in a shrinking business — Salesforce borrowed $25 billion in March to buy back its own stock even as it licenses the very models disrupting it, and Airtable's $1.285 billion sale was read as a new, lower floor for the whole category ^[8]. Jensen Huang's version is that AI is now eating the software that once ate the world ^[7]. One fund that owns Salesforce, Artisan Partners, argues the drop reflects a general economic slowdown rather than AI disruption, and that any real threat is years away ^[9] — a fair doubt, but a doubt about pace, not direction. Meanwhile the machines get dearer, and here the numbers turn strange. Compute has grown its own spot market — short-term rentals, like a hotel's nightly rate against a long lease — where a gigawatt of capacity now rents for $40 to $61 billion a year, against a long-term floor near $20 billion ^[10]. That premium is about four times the $11.6 billion per gigawatt that Goldman Sachs reckons the giant cloud companies, the hyperscalers, need to earn to clear a 15 percent return ^[10]. The same bank warns the spot price is not built to last, propped up by venture-funded startups training models rather than earning money, on contracts cancellable with 90 days' notice ^[10]. The shortage reaches down into unglamorous parts. Intel's server-chip orders stretch past a six-month backlog, and it has raised prices 10 to 35 percent with another 10 percent coming ^[11] — though Intel's stock has still fallen 19 percent since July, a reminder that even the physical layer's shares have not tracked its demand ^[11]. The grid has caught the same fever: capacity prices in the mid-Atlantic power market, anchored by northern Virginia's "Data Center Alley," jumped from under $50 per megawatt-day to roughly $333, a six-fold rise ^[12], and AI demand has lifted electric bills for some 67 million Americans across 13 states ^[13]. For the first time since the early 1980s, personal computers and their parts are getting more expensive, up more than 3 percent a month on a memory-chip shortage expected to last through 2027 ^[14]. The two blades are now crossing, and AI's own customers are walking from the cheap side to the expensive one. Microsoft's newest laptop runs AI tasks on the device itself rather than in the cloud, which the company frames as managing its energy and token budgets at once ^[15]. Then there is the bank. PNC is building its own GPU compute because its chief executive worries the cost of tokens could erase whatever productivity AI promises a bank.

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

A bank has decided that the way to keep the intelligence affordable is to own the scarce stuff underneath it. That is the quiet inversion of the past year: the company the market judged on model reviews last December is now priced on what it controls of the scarce stuff — the chips, the power, the buildings — and its customers are reaching the same conclusion from the other direction. The cheapest intelligence in the world still has to run somewhere, and the somewhere keeps getting dearer.