The artificial-intelligence economy is beginning to acquire the infrastructure of a mature industry: enormous financing, standardized units of exchange, specialized hardware, enterprise data rules, and political backlash. The scale is already difficult to absorb. Anthropic reportedly expects a potential public offering to rival SpaceX’s record $75 billion share sale, supported by a projected annualized revenue run rate of $65 billion and second-quarter revenue above $11.5 billion—up from $787 million a year earlier.
That growth depends on a new industrial base. Broadcom is reportedly discussing more than $60 billion in financing for custom AI chips, potentially reaching $100 billion with a junior tranche. The arrangement resembles sovereign-scale borrowing, reflecting how much capital is required to build the computing systems behind frontier models. Anthropic is also moving toward enterprise-style data governance: business customers may be allowed to retain the 30 days of model-interaction logs they are required to keep on their own cloud infrastructure rather than Anthropic’s.
The industry’s basic economic unit is the token—the fragments of text and other data that models process and generate. Stripe’s reported acquisition of OpenRouter, valued at more than $7 billion, would give the payments company control of a 90-person service that routes requests among more than 400 models from 80 providers. Stripe CEO Patrick Collison has described tokens as “the central currency for companies building with AI.” The purchase is also a bet that the market will remain pluralistic: companies may use different models for different tasks rather than choosing a single dominant system.
Demand is already immense. Meta has reportedly become one of Microsoft’s largest AI customers, consuming trillions of tokens each week through Azure, much of them to help write software. Collaboration tools are adapting accordingly. Slack Code gives Claude, Devin, Copilot, ChatGPT, and Vercel agents dedicated channels where they can write, review, and deploy code transparently, with conversations archived when work is complete. The shift is subtle but important: AI is moving from a private assistant window into the shared operational record of organizations.
Open models are expanding the same ecosystem. Google says its Gemma family has passed one billion downloads and spawned more than 100,000 community variants. Those models are being adapted for settings ranging from satellites to an Indian health app with more than 100 million downloads. Google’s new “Gemmaverse” directory is an attempt to catalog that proliferation, which increasingly resembles an open-source software movement rather than a market dominated solely by a few laboratories.
Even prompting is changing. Generalist’s GEN-1.5 model can learn a physical task from a three- to 12-second demonstration inserted into its context window—a “physical prompt.” It reportedly achieved 59 percent performance with no gradient updates and 83 percent after five minutes of additional data. The result suggests that some robots may acquire useful behaviors through brief examples rather than lengthy retraining. More intriguingly, the system began using tools in improvised ways, hinting that capabilities can emerge from a model’s pretrained representations rather than being explicitly programmed.
That possibility matters as AI leaves the screen. Nevada has approved as many as 7,000 robotaxis for the Las Vegas area from Tesla, Waymo, and Uber partners. Waymo is also developing a custom 1,000-teraflop-per-second chip fabricated on TSMC’s five-nanometer process, seeking faster responses and less dependence on Nvidia. Nvidia, meanwhile, is reportedly considering a Groq-licensed inference-chip variant for China by the end of the year, although the company denies having such a roadmap. The episode illustrates a broader constraint: as demand for AI inference grows, access to specialized processors is becoming a strategic and geopolitical bottleneck.
Autonomous delivery is expanding too. Amazon Prime Air plans to reach nearly 500 cities, a sixfold increase, promising deliveries of items weighing less than five pounds in as little as 30 minutes. But the system’s failures remain revealing. One Texas customer’s first delivery reportedly ended in a swimming pool. Automation may solve parts of the last mile while creating new problems in the final few meters—where weather, architecture, safety, and human behavior are hardest to model.
China is deploying physical AI at scale. SUPCON’s humanoid robots have reportedly issued 170,000 traffic warnings in Hangzhou. Chinese manufacturers are selling humanoids to state training centers that in turn generate teleoperation data for improving the machines. That feedback loop—robots deployed in public settings, human operators correcting them, and the resulting data sold back into development—could accelerate progress. Unitree’s reported $50 billion debut would symbolize the financial ambitions surrounding the sector, while its robotic dogs embody a different history: highly capable designs descended from DARPA-funded research that the United States published but did not mass-produce.
The buildout is provoking a democratic response. Residents in roughly a dozen US towns are seeking to recall officials over data-center agreements, with early efforts beginning September 1 in Independence, Missouri. Opposition is crossing party lines, appearing in an NRSC “sleeper issue” memo as well as satirical campaigns offering to mail urine samples to selected data centers—a crude expression of concern about water use and environmental impact.
Anthropic CEO Dario Amodei has linked this distrust to a broader crisis of confidence in institutions, arguing that every industrial revolution creates social disruption that must be addressed through transparency. Some proposed remedies are already appearing. Japan has approved a “comply or explain” code encouraging AI companies to disclose information about their models and training data. Apple Music plans to label songs made with AI, even as fully synthetic music reportedly accounts for about a third of submissions but less than 0.5 percent of listening. A Berkeley professor was also caught using AI to edit an opinion article calling for more standardized testing—apparently without realizing that she was effectively being tested by an AI detector. Former Spirit Airlines flight attendants are contesting Google’s reported $10 million bid to use their data for model training, raising familiar questions about consent, ownership, and who benefits from the conversion of human experience into training material.
Evidence about AI’s effects on people remains less dramatic than the surrounding investment figures. An eight-year Finnish study found that greater childhood screen time predicted better teenage cognitive performance, complicating simple claims that digital exposure is inherently harmful. In finance, the US Treasury is teaching newborns to invest through 530A Accounts restricted to broad, low-fee equity funds—an example of how automated, long-term systems may be designed to reduce speculation rather than amplify it.
The same period is producing more speculative ambitions. Representative Eric Burlison is helping people who worked on unidentified anomalous phenomena escape nondisclosure agreements, while Elon Musk says Starship could attempt its first reflight within months, including a tower catch. Musk has called the moment a fork in the road for consciousness reaching the stars.
That rhetoric is grander than most of the technology now being deployed. The more immediate transformation is quieter: tokens becoming a corporate cost center, custom chips becoming strategic infrastructure, robots entering public services, and data rights becoming a battleground. The central question is not whether AI will scale. It is whether institutions can make that scale accountable—before the systems built to process intelligence become too deeply embedded to govern.