to America. Global bond markets also do not seem to be pricing in all that dramatic an uplift to growth. So far the only profession truly upended by AI is software engineering. Many coders now hardly touch code themselves; instead, they supervise armies of AI agents. America’s “information” sector, which includes software publishing, has the highest and fastest-growing AI penetration rate in the economy, according to figures from the US Census Bureau (see chart 3). Mercor, an AI-data company, claims to spend more on AI agents than on employee salaries. That is extreme, but not entirely unrepresentative. Tech companies’ AI spending looks nothing like the rest. The average VC-backed company spends around $100 per employee per month on AI, according to data from Ramp (see chart 4). Regular firms spend only $10 or so. Many, after a wave of “tokenmaxxing” AI enthusiasm, are becoming more cautious. In October Goldman Sachs, a bank, banned non-engineers from using AI coding tools, citing disappointing productivity gains and concerns about cyber-security risk. (Though JPMorgan, a rival, has since announced that it is doubling down on the technology.)

Where tremendous gains have been most visible is in the underlying capabilities of AI models. A report by METR, a non-profit outfit that embedded researchers at Anthropic, found the pace of improvement in Anthropic’s models may well have doubled in the past year. Recent months have brought a drumbeat of announcements of AI solutions to unsolved mathematics problems, or novel AI-enabled cyber-attacks. But the surging investment in training AI models has also meant that each leap in model intelligence has been more expensive than the last. That raises the commercial imperative still further to deploy this intelligence profitably. For now, selling access to AI has been an excellent business. Improvements have arrived fast enough, and the best models have been appealing enough, to let developers charge healthy margins. Anthropic’s gross margin in this year’s first half was 85%. At first glance, then, AI might remind investors of the previous generation of runaway software success stories. But in their fundamental economics, AI and conventional software are quite different businesses. Software products, from software-as-a-service tools like Salesforce to Google to the social-networking giants, had several appealing characteristics from an investor’s perspective. Switching to a competitor was costly, so their moats were wide. Users would have to relearn how to use a complex tool from scratch. Network effects benefited the biggest players and stymied upstarts: why join a social-media site with no other users, or an ad network with few prospective customers? And, once the software was built, it was cheap to distribute. So, the bigger a company’s customer base, the more users it could spread its fixed costs across, and the tougher any rival would find it to compete on price. Selling AI models has, so far, been something like the reverse. Switching is trivial. For developers, tools like OpenRouter make it as simple as changing a line of code. For consumers, one chatbot looks much like another. Many products, like Cursor, a coding agent, offer a drop-down menu to toggle between different model providers. Others like Instinct, a personal-assistant agent, make even that decision invisible by juggling among models on the back end, balancing performance and costs. And the cost of serving a new customer is not near-zero, as with regular software. Each new query must be run in an expensive data centre, on a scarce chip.

Still, Anthropic’s and OpenAI’s pricing power has held up thus far. Breakneck revenue growth relies on that position lasting. The firms can manage that either by continuing to make their current business model work, or by finding new ones. They are trying both. So far, though switching costs are low, users have been willing to pay a premium for the best models. But at the current pace of development, each gets only a narrow window, weeks or at most months, to recoup its training costs before a successor surpasses it. Will cheaper models become a threat? Once a certain level of model intelligence is achieved, it tends to be commodified and plummet in price (see chart 5). That reflects competition between American AI players as well as the threat of Chinese fast-followers. Today’s best Chinese offering on the Epoch Capabilities Index, which aggregates a wide range of AI benchmarks, is Kimi K3.1 Pro, which scores around the same as GPT 5.5 Pro from OpenAI, which was the strongest publicly available American model just over six months ago, in April. Over the past few years, Chinese models have been consistently about that far behind. Cheap models could be especially worrying if trained cleverly. Researchers at Bridgewater, a hedge fund, using tools from Thinking Machines, an AI startup, found that a small fine-tuned Chinese model beat the then-leading

models from Anthropic and OpenAI at processing financial documents. Such fine-tuning requires investments in data and computing that only big firms will deem worthwhile, but if the finding holds more broadly, it will be a tempting way to lower the AI bill. Still, some uses of AI will always demand the absolute best, especially those that are adversarial, like trading or cyber-security. There, success means staying ahead of rivals. And in other cases too it may be worth paying a premium. Large unsupervised swarms of AI agents, of the sort that OpenAI accidentally set loose to hack Hugging Face, an AI startup, earlier this year, still need top-flight intelligence. The big developers’ lead on capabilities also could grow. The computational power needed to train the latest models is becoming a moat, keeping upstarts out. Chinese developers are constrained by a (thus far) smaller data-centre buildout and American restrictions on chip sales to China. And Anthropic says that it is getting better at impeding distillation, the process where less capable models are trained on the output of better ones, which has helped Chinese firms stay in the race on the cheap. AI developers could find ways to shift their business to make spending stickier. In recent months Anthropic and OpenAI have announced a crop of specialist tools for law, science and finance. Anthropic has also set up a joint venture with Blackstone, a private-equity firm, Goldman Sachs and others to start selling bespoke AI services to businesses. Each is an effort to find revenue streams where the company can embed itself and dig a deeper moat. But AI products can only ever be so sticky. By making computers easier to use, and more responsive to human-language requests, AI itself has simplified the process of moving from provider to provider. Anthropic cannily capitalised on this fact earlier this year, when its spat with the Pentagon drove a wave of sympathetic switchers from ChatGPT. Upon setting up Claude, users were given a prompt that they could paste into ChatGPT, asking the rival model to summarise what it knew about them. Claude would then digest the output, and be able to offer personalised responses from the get-go.

Another option for the model companies is proprietary trading. This could be of the literal sort. Nearly every rumour imaginable has been attached at some point to Safe Superintelligence, the secretive AI-research company run by Ilya Sutskever, a former chief scientist of OpenAI. A popular, if barmy, one is that it is using its models for high-frequency trading. A more likely version of that business model for Anthropic and its peers is prop trading of the metaphorical variety: monetising AI by using it directly, potentially in a joint venture, in domains like drug development or robotics, where breakthroughs could be highly profitable. For sufficiently smart models, restricting access to a few ultra-high-paying customers, which would also choke off the threat of distillation, could be a shrewder move than a wide release. Politics could get in the way. America’s midterms were the first big elections where AI was on the ballot, and it fared poorly. Ahead of the results, The Economist found that Americans living near a data-centre site swung towards the Democrats by around five percentage points over the year, compared with voters with the same demographic characteristics elsewhere. Voters opted to restrict data centres in all but one of the 41 local referendums on the topic. Anti-data-centre candidates won stunning victories. Most consequential was the surprise election as governor of Texas of Gina Hinojosa, a Democrat who campaigned for a pause in data-centre construction. Texas is host to around a third of America’s data-centre buildout. The potential moratoriums would hit projects not yet under way, most of them due in 2028 or beyond. In states that elected governors who support a data-centre pause, there are 25.6 gigawatts’ worth of not-yet-started data centres due in 2028, finds Maya Barkin of SemiAnalysis, a research firm. That is equal to just under half America’s current data-centre capacity. Still, new data centres are comfortably profitable enough to fund tax cuts and payouts for the communities where they operate. Politicians and developers may well be able to negotiate win-win deals—though that may be easier in places with more Republican voters, who are a little less opposed to data centres. Ms Hinojosa has already toned down her rhetoric. After the election, she clarified that Texas is still pro-business. “It is not that

I hate data centres,” she said, though she did accuse the industry of “taking advantage of Texas”. A bigger risk is federal regulation. None of the proposals floating around Congress with any chance of passing so far would interfere much with the AI developers’ business models. But the new Democratic majority will be keener on regulating AI, and could pivot quickly if, say, a serious cyberattack makes AI risk a top-tier issue. None of that will stop Anthropic from raising gobsmacking sums. Nor has the buildout of data centres yet shown much sign of slowing. But the business models and political environment that brought AI into the hundreds of billions in revenue won’t carry it into the trillions. For that, the industry will need to adapt and evolve. A looming question is whether Wall Street will have the patience to let it. ■ This article was downloaded by zlibrary from https://www.economist.com/briefing/2026/11/12/can-the-ai-boom-make-enough-money-to- sustain-itself

· Asia

Inside America’s first new submarine base abroad for 53 years Indonesia’s vice-president resigns, escalating a destabilising feud The Japanese region where “eight-testicle women” outdo men Can anyone fix the China-Japan relationship? Why Indians should have higher standards