The next bottleneck in artificial intelligence may not be algorithms, chips, or even money. It may be electricity.
As capital floods into AI infrastructure, every constraint becomes valuable. Data centers compete for grid connections, water, land, and specialized equipment. The result is an emerging economy in which the ability to deliver a megawatt of reliable power can matter as much as access to the latest model.
Early performance figures suggest why. Nvidia’s Vera Rubin NVL72 systems may generate as much as 10 times more tokens per megawatt than Blackwell systems, although independent comparisons remain limited. If that advantage holds, energy efficiency will increasingly determine which companies can afford to deploy models at scale. Buyers are already turning to unconventional sources: reports describe firms removing engines from private aircraft to supply generators, while Caterpillar, Cummins, GE Vernova, and Siemens Energy race to sell the equipment needed for rapidly expanding data centers.
The geography of wealth is shifting with the power. The two richest counties in the United States are also the country’s two largest data-center hubs. But the benefits come with familiar costs. SpaceX’s Memphis operation plans to recycle 10 million gallons of water a day, potentially ending its dependence on local aquifers, while Malaysia’s semiconductor and data-center boom helped push economic growth to 6 percent amid protests over energy and water use. Nvidia is reportedly considering a $3 billion investment in an Ohio campus operated by SB Energy for OpenAI—another sign that AI infrastructure is becoming a strategic industrial asset.
The global model ecosystem is changing just as quickly. Open-weight models—systems whose parameters can be downloaded and adapted—are increasingly being developed in China. Qwen reportedly surpassed 3 billion downloads in six months, compared with 418 million for Google’s models and 227 million for Meta’s. A recent survey of open-model activity found Chinese laboratories leading US releases in most months, while Qwen had accumulated 151,448 derivatives. US participation has increasingly come from hardware companies rather than independent model labs.
The shift matters because open models spread capability beyond the companies that train them. They allow startups, researchers, and governments to customize systems without paying an API provider for every query. But as model performance becomes more widely available, proprietary data becomes more valuable. Contractors are approaching startups to buy old Slack conversations, customer-support tickets, and other records that can be used to improve future systems. Data ownership and consent are becoming central questions in the next phase of AI competition.
Provenance is another unresolved problem. Future Claude models are expected to include an invisible watermark that changes the statistical choice among equally plausible words. The alteration would not make text visibly different, but could allow it to be identified later—an approach intended to help meet requirements under the European Union’s AI Act. Such systems may assist with accountability, though they will not solve the broader problem of determining who wrote, edited, or authorized a piece of generated content.
AI is also beginning to function as a scientific instrument. Faraday, a 27-billion-parameter “AI scientist,” was trained to reproduce figures from papers it had not seen. In reported tests, it outperformed Opus 4.8 and GPT-5.5 across every category, while using a more capable coding agent as a tool. In another experiment, 153 autonomous runs of Claude Fable 5 spent as long as eight days optimizing nanoGPT. The best result closed 81.7 percent of the gap to the human record, although none of the runs discovered a genuinely new method.
Other systems have done better. An automated research loop reportedly found a 232-fold speedup for a kernel used in QR decomposition. And after 10 open mathematical problems were solved with substantial help from language models, mathematician Timothy Gowers argued that these systems are particularly effective at search-heavy proof discovery. They can explore many possibilities cheaply, while humans remain better at deciding which deep lines of reasoning deserve sustained attention.
That division of labor may be more important than claims that AI will simply replace scientists. The near-term advantage is likely to come from systems that generate, test, and discard hypotheses at high speed, with people providing judgment, interpretation, and responsibility. The open question is whether automated search will produce new concepts or mainly accelerate paths that experts already understand.
Capability, meanwhile, is running ahead of trust. Anthropic chief executive Dario Amodei has rejected accusations that warnings about AI are merely doom-mongering. He has described public pessimism as part of a much older crisis of confidence and argued that the strongest evidence for technological progress will be practical achievements such as curing cancer, not promotional claims. On regulation, he has said the debate need not be a choice between industry capture and indiscriminate restrictions. Anthropic’s proposals, he argues, could slow frontier laboratories while exempting smaller challengers.