The artificial-intelligence boom is running into a very old problem: infrastructure has limits.
Alphabet, Meta, Microsoft, Amazon, and Oracle have disclosed more than $3 trillion in combined commitments for leases, buildings, electricity, and computing equipment. Google alone accounts for more than $1 trillion—nearly eight times its figure from last year. But the companies financing this expansion are becoming less willing to treat AI growth as inevitable. Banks are demanding stricter terms, concerned about construction delays, uncertain returns, and a flood of AI-related stock and bond offerings.
Even OpenAI appears to be feeling the pressure. Reports suggest that it reduced a planned funding round from $30 billion at a $1.4 trillion valuation to roughly $20 billion at $1.25 trillion. Investors must weigh that opportunity against a technology sell-off and Anthropic’s expected initial public offering, reportedly at a valuation of $1.5 trillion or more. The message is clear: AI may be software, but its expansion depends on some of the most capital-intensive infrastructure ever built.
The constraint is increasingly local. In Arizona, opposition to data centers is growing across party lines as the state faces severe water stress and has lost nearly 30 percent of its Colorado River allocation. Yet the debate is complicated by competing industrial demands. TSMC’s Phoenix semiconductor plant, praised for creating more than 3,500 jobs, reportedly uses in nine days more water than the proposed Project Blue data center would consume in a year. Agriculture still accounts for roughly 70 percent of the region’s water use.
That comparison does not make data centers environmentally harmless. It illustrates the political challenge: communities must decide which industries deserve scarce power and water, and at what price. The AI build-out is also following energy availability. Envision plans to complete a 3-gigawatt data-center campus in Inner Mongolia by 2028 and is considering another 10 gigawatts overseas. Boom Supersonic, meanwhile, is reportedly shifting attention from its delayed Overture airliner to 42-megawatt gas turbines for data centers. The company’s supersonic ambitions may be on hold, but the demand for electricity is not.
Nuclear power is attracting new supporters as well. The Green Party of England and Wales recently abandoned its longstanding opposition to nuclear energy, reflecting a broader reassessment of how to provide reliable, low-carbon electricity for an increasingly electrified economy. The tradeoff is familiar: nuclear plants offer dependable power but require enormous capital, long construction timelines, and difficult decisions about safety and waste.
Electricity is only one part of the supply chain. AI systems also consume vast quantities of memory and networking equipment. Flash-memory manufacturers such as Kioxia, once pressured by customers including Apple to accept steep discounts, are gaining leverage as hyperscalers compete for limited supply. Sam Altman has argued that memory in personal computers is badly underused—often idle for most of the day—and could instead be pooled in the cloud. For computer enthusiasts running AI models locally, that proposal sounds less like efficiency than a plan to turn privately owned hardware into another resource controlled by cloud providers.
Bandwidth is becoming a target for policymakers, too. French lawmakers voted 11–10 for a levy of €0.055 per gigabyte on large content companies such as Google and Netflix. Supporters say it could raise as much as €10 billion annually. Germany is considering a different approach: Family Minister Karin Prien has backed a tax on AI-generated revenue to help offset job losses. These proposals reflect a growing political question: if AI systems rely on public infrastructure, workers, and networks, how much of their revenue should return to society?
At the same time, institutions are struggling to distinguish human work from machine-generated output. AI detectors such as Pangram are increasingly used in publishing disputes, prompting writers to remove stylistic features that algorithms may interpret as evidence of automated prose—including em dashes, lists of three, and words such as “delve.” Linguists are beginning to describe a hybrid dialect shaped by both human habits and language models. Detection systems may be useful, but their uncertainty creates a new risk: writers could be penalized not for using AI, but for sounding like it.
The incentives to automate are not limited to writing. The US Justice Department charged five people in what it described as its first broad crackdown on “ghost students”—people who used stolen identities and AI tools to submit financial-aid applications, complete coursework, and attend online classes. The alleged scheme was linked to $32 million in losses. In San Francisco, some startup founders are applying the language of optimization to dating, treating relationships as a sales funnel and paying coaches as much as $5,000 a month to improve their “conversion rate.” The same metrics that govern software growth are being applied to education and intimacy.
Medicine is also being reshaped by computation and biotechnology. Eli Lilly reported third-quarter revenue of $24.4 billion, up 39 percent, as its weight-loss and diabetes portfolio expanded. The company’s retatrutide trial produced weight loss of about 28 percent, intensifying expectations that obesity drugs could generate enough revenue to finance work in gene editing, cell therapy, and cancer treatment.
The effects are already reaching industrial logistics. Hundreds of offshore workers in the United Kingdom are using weight-loss drugs to meet a new 124-kilogram helicopter limit, determined by the lifting capacity of rescue winches. The example is unusual but revealing: a medical technology developed for metabolic disease is changing who can safely perform a job and how companies manage physical risk.
Other safety questions are less visible. Regulatory reports obtained through freedom-of-information requests record 67 drone incidents near 28 US nuclear sites, including Virginia’s Surry plant near Langley. All were classified as “closed unresolved.” The designation does not establish malicious intent, but it highlights how difficult it can be to identify, investigate, and respond to small autonomous aircraft around critical infrastructure.
Far beyond Earth, NASA is confronting a different kind of resource constraint. Voyager 1 is expected to reach one light-day from Earth on November 18, after engineers replaced three heaters with lower-power alternatives to conserve its dwindling plutonium supply. The probe has even survived a 48-hour communications blackout. Hubble faces a more immediate logistical problem: NASA is asking private companies how they might raise the telescope into a safer orbit, service it, or deorbit it by 2032. No mission has yet been funded.
Meanwhile, AI systems are beginning to acquire physical bodies. Meta reportedly built a five-foot humanoid prototype called Max with China’s Fourier Intelligence in roughly three months, intending it to serve as the physical embodiment of its Muse agent. The project comes as the Federal Communications Commission considers tighter restrictions on Chinese humanoid imports, underscoring how robotics is becoming entangled with trade and national-security policy.
Software agents are also improving at operating inside simulated worlds. Ryan Madden reported that GPT-5.6 Sol completed a peaceful game of Factorio in 13.5 hours, outperforming GPT-5.5, which failed to master the game’s conveyor-belt systems. Yet Sol reached the final rocket only by manually carrying supplies between production lines it had not successfully connected. The result captures the current state of AI: impressive competence, brittle planning, and a persistent dependence on human workarounds.
Across data centers, factories, laboratories, classrooms, and spacecraft, the same pattern is emerging. AI can accelerate systems, but it cannot repeal physical limits, institutional rules, or economic tradeoffs. The next phase of the technology race will be decided not only by who builds the largest model, but by who can secure water, power, memory, labor, trust, and public consent.