Three years into the generative-AI wave, business adoption of AI looks surprisingly flimsy. The Census Bureau asks American firms whether they have used AI "in producing goods and services" in the past two weeks; by late 2025 the employment-weighted share of Americans using AI at work sat at about 11%, and had recently fallen by a percentage point. Adoption has fallen sharply at the largest businesses, those employing over 250 people. Other surveys find higher levels—Stanford University's Jon Hartley and colleagues found 37% of Americans used generative AI at work in September 2025, down from 46% in June—but even unofficial trackers point to stagnating corporate uptake.
Nearly two-thirds of executives at S&P 500 companies mention AI in earnings calls, and 87% of executives say they use AI on the job, according to Dayforce, a software firm. Yet just 57% of managers and 27% of employees do. Middle managers may set up AI initiatives to satisfy superiors' demands only to wind them down quietly later.
Usage of AI is booming among American state lawmakers and their staff. A survey by the National Conference of State Legislatures found that 44% of state legislative staff used AI in their work in 2025, up from just 20% in 2024. Legislators say AI is especially valuable for part-time lawmakers with minimal staff: South Dakota, for example, has just 60 staffers supporting 70 representatives. Some legislators use AI to fact-check lobbyists during committee hearings.
Studies from Stanford, Harvard and King's College London have found that firms adopting generative AI in America and Britain tend to hire fewer junior white-collar workers. In November 2025 6.8% of 20- to 24-year-olds with a bachelor's degree in America were unemployed. More than half of university graduates are underemployed (working in jobs not requiring a four-year degree) a year after graduation, and 73% of those who start out underemployed remain so a decade later. Almost 60% of new chip-manufacturing and design jobs projected to be created in America between 2023 and 2030 are expected to remain unfilled because of a lack of skilled workers, according to a study by the Semiconductor Industry Association and Oxford Economics.
From 2025 to 2030, big tech firms are expected to spend $5trn on infrastructure to supply AI services. To justify those investments they will need on the order of $650bn a year in AI revenues, according to JPMorgan Chase, up from about $50bn a year in late 2025. People paying for AI in their personal lives will probably buy only a fraction of what is ultimately required; businesses must do the rest.
A survey of executives in America, Australia, Britain and Germany, conducted by researchers from the Federal Reserve Bank of Atlanta, Macquarie University, the Bank of England and the Bundesbank, found that almost three-quarters of businesses are using AI in some way. Yet 86% of bosses across these four countries report the technology has had no impact on labour productivity over the past three years. In December 2025 OpenAI reported that ChatGPT Enterprise was saving users an average of 40-60 minutes on each day they used it.
Despite fears of mass displacement, white-collar workers have continued to do well since the arrival of ChatGPT in November 2022. America has added roughly 3m white-collar jobs—including management, professional, sales and office roles—while blue-collar employment has remained flat. Some occupations regularly cast as AI's early victims are on a tear: America has 7% more software developers, 10% more radiologists and 21% more paralegals than three years ago. White-collar workers now earn a third more than blue-collar ones, controlling for education, age, gender, race and other characteristics—nearly triple the premium in the early 1980s.
Roles that combine technical expertise with oversight and co-ordination have enjoyed the biggest gains: employment among project managers and information-security experts has risen by around 30%. Only routine back-office work has shrunk: insurance-claims clerks have fallen by 13% and secretaries and admin assistants by 20%. The share of Americans in clerical and administrative work has fallen from 18% in the 1980s to 10%.
AI is already generating all-new jobs: data annotators, forward-deployed engineers to guide clients through AI implementation and, in the C-suite, chief AI officers. The fastest-growing white-collar occupations in recent years are those without settled names: "other mathematical-science occupations" have swelled by about 40% since late 2022; "business operations specialists, all other" have jumped by almost 60%.
Evidence from Anthropic, drawing on millions of anonymised interactions with its models, finds that only around 4% of occupations use AI across three-quarters or more of their tasks; hardly any roles can be automated in full. Benchmarks created by METR, a research group, suggest that AI can write software by itself for five hours straight, and that this figure has been doubling roughly every seven months.
The unpredictability of AI's strengths and weaknesses—what Ethan Mollick, a professor at the Wharton School at the University of Pennsylvania, and others have christened the "jagged frontier"—means that it takes time for workers to develop intuition for how to use the technology. An AI model can outperform the world's best mathematicians while still being stumped by the number of "r"s in "strawberry". Bret Taylor, the chairman of OpenAI and co-founder of Sierra, likens the current moment to "we're all accountants and Microsoft Excel was invented last weekend."
Vibe-coding—using natural-language prompts to get an AI to write a computer program—makes it much easier for novices to create apps and features. Coding tools like Claude Code and platforms like Lovable or Replit allow end users and product managers to demonstrate what they want rather than writing specification documents. The phrase "demo, don't memo" is circulating inside some tech firms. But vibe-coding shifts the bottleneck from writing code to reviewing it. Johnson & Johnson found that 85% of the value generated by its AI experiments was attributable to just 15% of its applications, leading it to switch from a let-a-thousand-flowers-bloom ethos to a more focused approach overseen by a central AI council and a data council.
Demand for AI computing is shifting from training models to getting them to answer real-world queries (inference). McKinsey estimates that by the end of the decade inference will account for three-fifths of demand in AI data centres. The two stages place different demands on hardware. Training relies on enormous numbers of calculations conducted in parallel. Inference unfolds in two phases: prefill (processing the prompt and converting it into tokens) and decode (generating the response token by token, drawing on the model's weights stored in memory).
Modern GPUs struggle with inference because it requires constant access to off-chip memory (DRAM), which can be ten times slower and far more energy-hungry than on-chip memory (SRAM). A study by Amir Gholami of the University of California, Berkeley, and colleagues finds that over the past two decades computing performance has roughly tripled every few years, whereas off-chip memory bandwidth has improved by a factor of only about 1.6. This "memory wall" has become the main bottleneck in increasing the speed of AI inference. A crop of startups—including Cerebras, MatX, d-Matrix and Etched—is building chips aimed at running AI models faster and more efficiently than Nvidia's GPUs. Researchers at the Chinese Academy of Sciences have proposed embedding model weights directly into hardware by physically encoding them in the layout of metal wires, removing the need to fetch parameters from memory entirely.
By early 2026 artificial intelligence faced a severe supply problem. Weekly token consumption (snippets of text counted from large language models) quadrupled between January and March 2026, partly owing to the growing use of coding tools. Rationing spread across the sector: Anthropic adjusted terms to deter heavy peak-hour use; Amazon cited "capacity constraints" limiting growth; OpenAI's finance chief said the company lacked processing power for every opportunity and scrapped its video-generation model.
Adding capacity quickly is difficult. Local opposition has slowed data-centre construction in America. Shortages of transformers, switchgear and gas turbines cause delays (some equipment takes two to five years to arrive). The tightest bottleneck is processors—Nvidia chips remain scarce, as do memory chips and CPUs.
When hardware is expensive, balance-sheet size matters enormously. Only five data-centre hyperscalers—Amazon, Alphabet, Meta, Microsoft and Oracle—have the muscle to lock up needed hardware, spending over $750bn on capital expenditure in 2026. OpenAI and Anthropic announced hundreds of billions in partnerships to secure capacity. Nvidia secured most of its 2026–27 memory needs in advance.
Greatest profits concentrate at choke points. Nvidia's gross margin reached 75% (up from 60% in 2019); TSMC's exceeds 60%, roughly double many other manufacturers'. Software makers are responding by designing custom chips costing about half as much as Nvidia's, but only Alphabet has succeeded in large volumes, having started over a decade ago. Displacing TSMC is harder still; Intel and Samsung have struggled to match it.
Though inference prices fell five- to tenfold in a year and AI firms offer cut-price subscriptions in places like India, this obscures massive cash burn. OpenAI and Anthropic are expected to lose billions in coming years. As demand grows and model-makers pass rising compute costs to users, prices will rise.
"Distillation" refers to the practice of feeding prompts to a rival's AI model and harvesting its responses to train one's own model more cheaply. American labs spend billions creating the training data in the first place: they pay human experts—mathematicians, say—to write step-by-step solutions to hard problems, creating worked examples for models to learn from. Unlike the diffuse knowledge gained from the web, the know-how needed to complete a specific task is extractable. American labs have committed $5trn of data-centre investment between now and 2030, according to JPMorgan Chase; that largesse is challenged by Chinese rivals that are nearly as good but much cheaper. In February 2026 Anthropic and OpenAI each disclosed evidence that leading Chinese AI labs, including DeepSeek, had illicitly used their models via distillation. It is technically difficult to detect and prevent, warn American labs, and doing so is made harder by the increasing sophistication of Chinese efforts, including routing online traffic circuitously to shield its origins.
China has taken the lead in AI talent and is continuing to extend it. In 2025, for the first time, more studies presented at the world's top AI conference had lead authors based in China than in either America or Europe. At the December 2025 Conference on Neural Information Processing Systems (NeurIPS), the world's largest AI gathering, 51% of presenting researchers began their careers in China, up from 29% in 2019. Over the same period, the share who started out in America fell from roughly 20% to 12%. Nine of the top ten institutions where NeurIPS 2025 authors earned their undergraduate degrees were in China; graduates of Tsinghua University alone accounted for 4% of researchers, compared with 1% from MIT.
Among authors affiliated with American institutions, roughly 35% have a Chinese undergraduate degree—as many as have an American one. When Meta announced the researchers staffing its "superintelligence lab" in June 2025, a leaked list revealed that half were described as being from China. Of 483 contributors to OpenAI's GPT-5, 15% had at least one degree from a Chinese institution.
According to Digital Science, a data firm, China now has more active AI researchers than America, Britain and Europe combined, though it still trails the West per head of population. China's cohort skews younger: 47% are students, compared with about 30% in the West. Around two-fifths of Chinese university students study STEM subjects, roughly double America's share.
More Chinese researchers are choosing to stay at home. In 2019 roughly a third of NeurIPS authors who completed undergraduate degrees in China remained there; by 2025 that had reached 68%. None of the core contributors to DeepSeek R1 held degrees from outside China. Initiatives to lure researchers back, such as the Qiming Plan, offer salaries exceeding 700,000 yuan ($100,000), generous research grants and help with housing. Meanwhile, America has become less attractive: funding cuts, visa uncertainty and suspicion of loyalties have unsettled applicants. In 2025 Purdue University rescinded offers to more than 100 graduate students, most of them Chinese, after lawmakers asked it to document researchers' ties to Chinese institutions.
In 2019 just 12% of Chinese NeurIPS researchers who had earned graduate degrees abroad had returned to China; by 2025 that share had more than doubled to 28%. Using NeurIPS authors as a metric, around 37% of the world's top AI researchers now work in Chinese organisations, compared with 32% in American ones. If the trend of the past decade continues, by 2028 top China-based researchers could outnumber America-based ones by two to one.
Large language models are being used to accelerate progress in pure mathematics by streamlining the formalisation and verification of proofs. Formalisation—the painstaking process of checking proofs symbol by symbol—is, according to Patrick Shafto of DARPA, "a core bottleneck in mathematics, which is trust". LLMs work through mathematical logic very differently from humans: they run as a "stream of consciousness", working out what they think should come next, rather than devising a plan, according to Terence Tao of UCLA.
Startups including Harmonic and Math, Inc. are developing bots that can translate human-written proofs into the Lean coding language, verify each step and fill in gaps. Google DeepMind's AlphaEvolve can generate proofs for optimisation problems using natural language. At DARPA, Shafto hopes to automate translation between natural language and formal languages such as Lean, and to decompose complex proofs into their constituent propositions—bringing what he describes as "a hot mess of papers, text books and human heads" into a unified body of work.
Timothy Gowers of Cambridge University notes that LLMs struggle to transfer what they learn solving one problem to another, and lack the "aesthetic sense" that prompts human mathematicians to seek neater proofs. Donald Knuth of Stanford was impressed by Claude Opus 4.6's work on a travelling-salesman variant, though the model eventually malfunctioned when pushed further. A model capable of reasoning about mathematics could also be taught to reason about other quantitative fields, from economics to physics.
Goldman Sachs produces an index of companies with the largest estimated potential earnings change from AI-driven productivity gains, including Ford, H&R Block and News Corp. By late 2025 the index was trailing the broader market: investors did not see AI adoption translating into improved profitability. A poll by Deloitte and the Centre for AI, Management and Organisation at Hong Kong University found 45% of executives reported returns from AI below expectations; only 10% said expectations were exceeded. McKinsey found that for most organisations AI had not yet significantly affected enterprise-wide profits.
Erik Brynjolfsson of Stanford University has described a "productivity J-curve": introducing AI may temporarily reduce productivity as IT systems and workflows are rewired, before efficiency eventually shoots up. A paper by Yvonne Chen of ShanghaiTech University and colleagues identifies "genAI's mediocrity trap"—with the tech's assistance, people produce work that is "good enough", which helps weaker workers but can harm the productivity of better ones, who decide to work less hard.
One nice thing about egotists: they don't talk about other people.