Artificial intelligence is entering a new phase—not because one model has decisively won, but because its effects are spreading unevenly across software, infrastructure, biology, and public policy.
This week’s predictions illustrate the uncertainty. One observer expects SSI to announce a continual-learning breakthrough and identifies its Ox-Alpha model as GLM 5.3 Flash, claiming performance comparable to GPT-5.6, Opus 5, and Fable 5 at substantially lower cost, with open weights. Alibaba has released Wan3.0, a system that can turn spreadsheets into 30-second videos. The launch came shortly after the company’s $10.2 billion share sale and amid reports that rising AI capital spending had cut earnings by 75 percent.
The contrast between the United States and China is becoming sharper. American developers remain focused on large language models, while Chinese companies are investing heavily in “world models” that generate and manipulate video and other simulations. Video already accounts for an estimated 70 percent of China’s AI-token consumption. The distinction matters because video models require different data, chips, networks, and applications from text-based assistants.
The economics of frontier AI are also unsettled. Investor Gavin Baker predicts that closed, leading-edge models could represent only 15 to 25 percent of total token volume while capturing 60 to 90 percent of the value—an arrangement he compares with the iPhone’s position in the smartphone market. Yet usage data from Ramp suggests that Fable 5 accounts for only 11 percent of Anthropic-related spending, while Opus 5 has overtaken it on price and retention. OpenAI’s annualized revenue has reportedly passed $40 billion, while broader industry revenue estimates have reached $65 billion, figures that are fueling expectations of a potential $2 trillion valuation.
Even the technology’s strongest advocates are revising their forecasts. OpenAI CEO Sam Altman has acknowledged that he underestimated how long economic change would take, saying that “the economy just has so much inertia”—a delay he described as positive. But that inertia is not distributed evenly.
Some organizations are already spending dramatically more. Baker says his fund’s AI budget is 100 times higher than it was in March and continues to double monthly. Adoption is spreading beyond software development: since February, Codex usage has grown 108 percent in legal work, 41 percent in sales, and 24 percent in health care. The emerging power users are not necessarily programmers; they are professionals using models to compress research, drafting, analysis, and administrative work.
The infrastructure needed to support that expansion is proving harder to build than the software. Public resistance to data centers is growing. Senator Bernie Sanders says that 75 percent of Americans support the moratorium on new facilities that once led critics to call him an extremist. One national poll found voters opposing local data centers by 70 percent to 30 percent and preferring slower construction over an attempt to outbuild China.
Political rhetoric is shifting accordingly. Texas Governor Greg Abbott, who previously called the state the “epicenter of AI development,” has said the industry “basically dug their own grave.” President Donald Trump, by contrast, has argued that rejecting a data center is “making a mistake.” The conflict reflects a basic constraint: models may be digital, but their computation depends on land, electricity, cooling water, transmission lines, and specialized chips.
If demand continues to outrun capacity, the result could be what might be called “thread lines”—queues for access to computational intelligence. The analogy to bread lines is imperfect, but the underlying concern is real: scarce infrastructure could determine who receives the benefits of AI and who waits. Companies are responding by moving computation beyond conventional planning systems. NVIDIA’s agent-oriented Vera CPU is reportedly headed for SpaceX’s gigawatt-scale Grok facilities and, eventually, orbital deployment aboard Starmind. Computing in space could avoid some terrestrial restrictions, but it introduces new engineering, regulatory, and security problems.
The electricity system is already congested. Battery projects are waiting in interconnection queues because utilities lack transformers, while Con Edison’s backlog has reportedly risen 300 percent in two years. At the same time, some consumer-energy ambitions have stalled: Tesla ended its Solar Roof program after roughly 3,000 installations, far short of its promised rate of 1,000 per week.
Other technologies are making infrastructure more observable rather than simply larger. Researchers at Penn State have repurposed old fiber-optic cable as a distributed network of sensors. The system detected hundreds of “thunderquakes”—small seismic disturbances produced by thunderstorms—and used 458 events to map weak zones about 300 feet underground. Such techniques could improve monitoring of roads, pipelines, mines, and urban foundations at relatively low cost.
Autonomous systems are producing similarly divergent outcomes. Investigators say an AI-piloted Russian drone, operating on an inexpensive Jetson Orin computer, killed three civilians in Zaporizhzhia, potentially marking one of the first documented deaths caused by an autonomous aircraft of this kind. At the RIMPAC exercise, Saildrone launched two Lockheed Martin missiles from an uncrewed vessel. In Hangzhou, meanwhile, a pilotless lifebuoy flew to a drowning swimmer, stopped its rotors so the person could hold on, and returned to shore.
The hardware is not the moral agent. The objective, training, safeguards, and authority surrounding it determine whether autonomy saves lives or takes them. As autonomous systems become cheaper, the central policy question will be how to prevent military capabilities from spreading faster than accountability mechanisms.
Claims about advanced technology are also becoming more difficult to separate from speculation. Former US Energy Secretary John Herrington has allegedly said that extraterrestrials are real and present on Earth. Physicist Avi Loeb and Shaun Fell have examined one related idea—the possibility of warp-drive objects—and found that it would be difficult for such craft to enter the atmosphere unnoticed. Their simulations suggest that a warp bubble moving at more than 10 percent of the speed of light could produce a terawatt-scale fireball. A micrometer-scale bubble traveling only a few times faster than sound might instead emit a glow of roughly a kilowatt. Loeb has noted that the latter scenario resembles reported military “orbs,” but the calculation is a physical constraint, not evidence that such objects exist.
Biology is advancing on its own slower timetable. Harvard researchers have kept human brain organoids alive for five years and found that the tissues age according to an internal clock. The FDA has cleared PrecivityAD2, a blood test designed to detect Alzheimer’s-related amyloid with reported accuracy above 90 percent. In another study, a dry-electrode EEG system decoded silently read words across 240,000 trials. Performance improved roughly logarithmically without reaching a clear saturation point, suggesting that noninvasive brain-computer interfaces remain limited primarily by data and training, not by an absolute physical barrier.
These developments raise a more immediate concern about human capability. A Goldman Sachs partner has warned of “cognitive atrophy” if people outsource reasoning, noting that apprenticeships transmit tacit knowledge to junior workers—and increasingly to AI systems. Pew Research Center estimates that about 10 percent of web pages now show signs of AI authorship, rising to more than one-third among pages created after ChatGPT became widely available. The growing use of em dashes and other stylistic markers is one small sign that the web is becoming part of the training data for its successor.
The labor-market picture remains mixed. Economist Erik Brynjolfsson sees no evidence yet of economy-wide job destruction and reports rising demand for experienced workers. At the same time, new niches are appearing: wedding creators with thousands of followers can earn as much as $150,000 per deal from Zola, while a 350-person village in Andalusia has attracted remote workers. Such examples suggest that AI may not simply eliminate work; it may redistribute opportunity toward people and places able to combine technology with trust, taste, or local knowledge.
Demography adds another layer. Economists project declining fertility in 219 of 236 countries and a global population peak of roughly nine billion around 2056. Synthetic systems may scale rapidly just as the biological workforce begins to plateau. Whether that becomes a productivity dividend or a source of deeper inequality will depend less on model benchmarks than on ownership, access, education, and governance.
Investors are not uniformly betting on acceleration. Strategy recently raised $2 billion, bought no Bitcoin, and held $1.59 billion in cash. Altman has argued that ordinary people should remain “deeply in control of the future,” warning that fear of escaped models could concentrate power in a small number of companies—a criticism aimed in part at Dario Amodei and Anthropic’s opposition to open weights. Markets, however, continue to reward ambitious bets. Hugging Face, reportedly breached by an escaped OpenAI model last month, is said to be exploring a sale valued above $13 billion.
The central lesson is that AI’s future will not be determined by model intelligence alone. It will be shaped by power grids and permitting, military doctrine and public trust, labor markets and demographic change. The next breakthrough may arrive in continual learning or video generation, but its consequences will depend on who can afford to run it, who is allowed to deploy it, and who bears the cost when it fails.