The AI industry is beginning to look less like a software business than a new industrial economy—one that consumes vast amounts of capital, electricity, hardware, and political attention. Anthropic’s IPO filing offers a revealing snapshot. The company says Claude now writes more than 90 percent of its new code, while revenue has grown 86-fold and headcount fivefold. Its use of Claude, measured at retail prices, has reportedly exceeded its monthly cash payroll since April after rising 250-fold this year.
That productivity comes with an extraordinary bill. Anthropic has committed $895.4 billion through 2045 for computing capacity, including $121.3 billion with Google and $120.5 billion with SpaceX. The figures illustrate both the promise and the paradox of advanced AI: models may reduce the labor required to produce software, but they demand enormous physical infrastructure to run.
Investors are already treating that infrastructure as a new strategic asset. Spark Capital’s backers, including pension funds, could share in a windfall of more than $9 billion from the firm’s $777 million Anthropic stake if the company lists at a valuation of $1.5 trillion. Yet the industry’s financial momentum is colliding with uncertainty about what today’s systems can actually do.
Different tests produce sharply different pictures of progress. SemiAnalysis’s GDPx benchmark places AI agents in real-world jobs; the best-performing model scored only 12 percent. Claude Opus 5.5 led overall, GPT-6 Astra performed best on recursive self-improvement tasks, and almost every model scored zero in finance and law. By contrast, OpenAI’s LifeVal benchmark—covering caregiving, health, and administrative work—showed average rewards rising from 49.3 percent for GPT-5.2 to 80.6 percent for GPT-6 Astra.
The gap suggests that “general intelligence” is not a single capability. Tasks involving routine coordination or personal assistance may be easier to automate than work requiring accountability, specialized judgment, or complex institutional knowledge. Humans are responding accordingly. In an MIT survey of 268 young mathematicians, 75.5 percent said AI had reduced their likelihood of remaining in academia, largely because they feared being overtaken or “scooped” by machine-generated research.
The pressure is spreading beyond software. Stephen Gallagher’s test of 13 models found that Claude Opus 5.5 could produce a permit-quality house blueprint in about 12 hours for a few hundred dollars. He argues that architects may face a disruption similar to the one already reshaping programming. Cisco executive Jeetu Patel has described giving a 27-year-old employee, found several management layers below him, unlimited AI tokens to refactor 98 percent of the company’s legacy code over two years.
But generating code is easier than maintaining a shared system. The Linux 7.4 merge window set a record with 19,671 commits, overwhelming maintainers with AI-generated patches. PostgreSQL 19 may be delayed as back-patches rise by more than 60 percent. The lesson is becoming clear: abundance can create its own bottleneck. When production accelerates, review, testing, security, and governance become the scarce resources.
Even AI companies are discovering that not every token is economically rational. Elon Musk acknowledged that SpaceX’s Grok Bot costs remain “unsustainably high relative to revenue,” despite the company’s claim of 800,000 paid weekly users. Markets are beginning to price the risks. After the Nasdaq’s worst month in more than a year, CoreWeave’s trading volatility reached a record, credit-default swaps for major technology companies rose to their highest levels since 2021, and Jump Trading opened a desk to hedge falling GPU rents.
The infrastructure race is also reshaping the semiconductor industry. Marvell plans to acquire Lumentum for $115 billion, betting that optical lasers will carry more of the data moving through AI systems. The deal values Lumentum at roughly 33 times projected operating profit and was described by Marvell CEO Matt Murphy as “a double down on supporting Nvidia.” Marvell shares nevertheless fell more than 10 percent, a sign that investors are increasingly demanding evidence that AI spending will produce durable returns.
The political backlash is arriving at the same time. In 21 of the 36 states electing governors tomorrow, at least one major-party candidate supports a data-center moratorium. Local bans have increased from roughly 40 to more than 500. Communities are objecting to electricity demand, water use, noise, land consumption, and the possibility that they will bear the environmental costs while companies capture the profits.
OpenAI is preparing for that political environment with a post-midterm campaign for limits on frontier-AI development and a rebranding of data centers as “AI compute campuses.” Its Grid Protectors Program promises facilities that pay their own way and reduce consumption when the electricity system is strained, beginning with 1,000 megawatts in Georgia. Google’s energy chief argues that permitting, rather than power generation, is now the main constraint. Sending computation into orbit, he says, can cost up to 100 times more—though Google’s Suncatcher project is exploring exactly that possibility.
Compute is becoming a strategic reserve as well as a commercial resource. After Ukrainian drones struck Russian data centers, Vladimir Putin ordered private operators to reserve 10 percent of their capacity for the state. Meanwhile, suspected cyberattacks on Danfoss refrigeration systems disrupted supermarkets in five countries; one Swedish freezer reportedly reached 50 degrees Celsius. The incidents show how digital systems can turn ordinary infrastructure—cold storage, energy networks, logistics—into geopolitical targets.
Algorithmic decision-making is moving into more sensitive settings too. Michigan plans to pilot Recidiviz software that scores inmates’ violence risk in real time. Such systems may help officials identify people who need support, but they also raise familiar questions about biased data, due process, transparency, and whether a statistical forecast should influence punishment or confinement.
Physical technology remains slower to mature than software. Apple’s first home robot, known internally as J595, reportedly had finished hardware two years ago while its software lagged, pushing the launch to 2027 or later. Carmakers are taking a more incremental path: Ford, Rivian, and General Motors are expanding 3D printing for tooling and production parts, and Rivian says some printers repay their cost within a year. Toyota’s battery-electric sales rose 127 percent even as its China sales fell nearly 19 percent, evidence that the auto industry’s transition is advancing unevenly across markets.
The same pattern appears in communications and space. After being rebuffed by the largest US carriers, SpaceX has reportedly approached EchoStar about Boost Mobile, offering a route into a $235 billion wireless market. US intelligence has released the first clear image of China’s secretive spaceplane, sharpening competition beyond Earth. At the other end of the scale, new simulations suggest comets rarely delivered the hydrogen cyanide early Earth needed for life, although Mars may have received much more. The chemistry of life may have been largely local rather than imported.
That possibility has an intriguing parallel in aging research. Mice whose leg bones were mechanically compressed twice a week released anti-aging proteins, developing stronger bones, improved cognitive performance, and nearly three additional months of life. The result is preliminary, but it points to a broader theme: biology, like computing, may be influenced by signals that researchers are only beginning to understand.
The year’s developments share a common message. AI is advancing rapidly, but its consequences are determined not only by model capability. They depend on energy systems, labor markets, institutional trust, physical supply chains, and political consent. The central question is no longer simply how intelligent machines can become. It is whether society can build the rules and infrastructure needed to use them without allowing their scale to outrun our ability to govern them.