The next phase of artificial intelligence is being shaped less by a single breakthrough than by a contest over infrastructure, access, and control. Models are becoming persistent agents, scientific instruments are being connected directly to software, and the computers needed to run them are pushing against the limits of electricity, chips, and public trust.
That tension was visible in a dispute involving OpenAI, SpaceX, and Cursor, the AI coding company. OpenAI told Cursor it would wind down the company’s access to its models by November 12, citing a record of contract violations by companies associated with Elon Musk. Cursor co-founder Michael Truell said OpenAI accounted for only about 5 percent of its traffic, but argued that the company had relied on OpenAI as “neutral infrastructure.” Anthropic quickly offered a counterpoint, describing Cursor as a trusted partner since Claude Sonnet 3.5 and promising additional computing capacity.
The episode illustrates a growing strategic reality: access to frontier models is becoming a business dependency, while the companies that control models, chips, data centers, and networks increasingly compete across the same markets. Alliances that once looked temporary can become decisive. Some observers see the possibility of a Dario Amodei–Elon Musk alignment leaving OpenAI chief Sam Altman confronting two powerful rivals. Others argue that Musk’s interest in space-based computing reflects a broader shift: whoever controls abundant energy and computing capacity may control the next stage of AI.
Inside OpenAI, Altman has reportedly suggested that artificial general intelligence could arrive as soon as this year, while the company previews a new model, reportedly called Astra. Its Codex system is also evolving from an interactive programming assistant into a persistent agent that can continue working until it is “put to sleep.” Such systems promise productivity gains, but they also raise a practical question: how much autonomy should software have when it can plan, execute, and revise work over long periods without continuous human supervision?
The model race is also widening geographically. Z.ai released the weights of GLM-5.3 after a two-week safety review, replacing a permissive MIT license with terms that require security screening for organizations operating more than $10 billion in computing infrastructure. The decision reflects a difficult balance between open research and misuse prevention. The model had recently topped the CyberGym benchmark after identifying 2,436 vulnerabilities, demonstrating both the defensive value of capable systems and the risks posed when similar abilities are used offensively.
Political assumptions are becoming another source of scrutiny. A new AI Political Compass evaluated 57 models and found that 54 clustered in the left-libertarian quadrant, with Grok a notable exception. The researchers also found that when models are asked questions built on premises shared across much of the political spectrum, they landed in that quadrant 19 times out of 20. The result does not prove that models possess a coherent political ideology. It does suggest that training data, safety policies, and the wording of questions can produce systematic patterns that users may mistake for neutral judgment.
Science is beginning to operate as an agentic loop: a system proposes an experiment, runs or directs it, interprets the result, and designs the next test. Google’s “Co-Scientist” reportedly designed a safe synthesis route for MXene nanomaterials and tested it on a real deposition reactor. It also predicted the behavior of swarming E. coli in a way that matched unpublished laboratory observations and proposed an architecture that outperformed six frontier models.
The significance lies not simply in faster literature searches. Scientific progress is often limited by the time required to translate a hypothesis into a physical experiment. Connecting AI to laboratory equipment could compress that cycle from weeks to hours. Anthropic’s proposed Model Hardware Standard aims to make that connection easier by allowing agents to control microscopes, robot arms, and other instruments through common interfaces. Standardization could reduce integration work dramatically, but it also makes mistakes easier to scale. A poorly specified instruction could damage equipment, contaminate samples, or generate misleading results at machine speed.
The need for safeguards is becoming clearer in government and security. A judge blocked the Pentagon’s effort to blacklist Anthropic, ruling that national security does not give the government “a blank check to punish and retaliate against government critics.” In Texas, officials paused funding for Flock’s automated license-plate camera network before a reported $30 million investigation into surveillance practices. Meanwhile, Immigration and Customs Enforcement has explored technologies including robot dogs and electrical shock gloves.
These developments expose a central problem in AI governance: systems built for safety or efficiency can also expand the reach of institutions that already possess coercive power. More than 150 companies, including former rivals, signed OpenAI’s call for governments to secure critical infrastructure during what it called the “defenders’ window”—the period in which defensive AI capabilities may outpace offensive attacks. Yet technical defenses cannot substitute for legal limits, oversight, and clear accountability.
The physical infrastructure required to run these systems may be an even more immediate constraint. Washington is considering new semiconductor tariffs and restrictions on China’s remote access to advanced chips. Architect Labs, meanwhile, has announced what it describes as the first AI chip designed end-to-end by AI. The claim points toward a future in which machine learning accelerates not only software development but also the design of the hardware on which machine learning depends.
Electricity, however, may prove scarcer than logic. Musk has warned that roughly 15 gigawatts of planned 2027 computing capacity cannot simply be switched on in 2027, because turbines, transmission lines, and other equipment take years to build. SpaceX is reportedly establishing its own turbine-blade factory. Germany has promised to quadruple computing capacity by 2030, while labor unions have defended data centers as sources of jobs. At the same time, a bot network of roughly 200,000 Chinese-linked accounts reportedly promoted the claim that data centers drive up household power bills—an example of how infrastructure debates are becoming targets for information warfare.
Governments are responding with both industrial policy and emergency powers. Washington has pursued a major oil agreement with Venezuela and issued Executive Order 14420, which bars risky foreign grid equipment amid concerns that AI-enhanced attacks could magnify vulnerabilities. The tradeoff is familiar: efforts to secure energy and communications systems can also centralize authority and weaken transparency.
Space is increasingly presented as the escape valve. Musk has recalled locating the future Starbase by scrolling through satellite imagery. Mach33 analyzed the propellant economics of SpaceX’s proposed $100 billion Louisiana campus, where producing fuel on site could be roughly ten times more efficient than trucking it in; Musk called the analysis “mostly correct.” The President has announced plans for a nuclear-powered Mars ship in 2028 and a United States Space Academy to train its crews. Such ambitions remain technically and politically uncertain, but they reflect a growing belief that AI, energy, and space development are becoming one industrial project rather than separate fields.
That project is attracting capital. Andreessen Horowitz has raised $1.1 billion for AI’s physical buildout. Meta is testing robots capable of replacing data-center cables, a development that has reportedly unsettled technicians whose work could eventually be automated. Hugging Face has opened $399 preorders for Microduck, an open-source bipedal robot with a grasping beak. Even consumer software is feeling the pressure: Google has imposed memory limits on Android as AI applications consume more DRAM, while one hacker connected Minimax H3 Max to Twitch to generate an effectively endless stream of video.
The cultural consequences are arriving alongside the machines. In China, as many as 95 percent of short dramas are reportedly AI-generated, and actors are increasingly being reduced to reusable digital tools. Australia removed AI-generated songs from its charts after a synthetic Madonna cover reached number one. Research from Pew found that ChatGPT rewrites human prose in recognizable ways, including roughly twice as many em dashes and three times as many constructions of the form “it’s not X, it’s Y.”
Meta’s reported plan to replace as much as 60 percent of its staff with agents triggered an employee backlash, even as the company remains one of Anthropic’s largest customers and prepares for a possible $2 trillion valuation. In Beijing, an “AGI Bubble” bar turns the technology’s financial mythology into a joke. Wall Street, meanwhile, has begun looking beyond GLP-1 drugs toward baldness treatments, with MANE shares reportedly rising 500 percent. The pattern is revealing: speculative enthusiasm does not disappear when one technology matures; it moves to the next promise.
The oldest systems on Earth may offer a different perspective on intelligence. The Earth Species Project’s BirdCODE initiative is attempting to decode vocalizations across 9,000 bird species. Researchers have also reconstructed a Jurassic soundscape from 165-million-year-old insect wings, including ultrasonic calls from a period long before bats evolved. These projects remind us that communication and coordination are not inventions of computers. AI is powerful partly because it can imitate and amplify capacities that biology has been refining for millions of years.
The question now is not whether AI will become more capable. It is whether societies can build the institutions, energy systems, legal constraints, and scientific norms needed to direct that capability. The future may be shaped as much by who supplies the power and sets the rules as by who trains the largest model.