The AI investment tsunami that began with the public launch of ChatGPT in November 2022 is starting to crash against the rocks of reality. An early AI adopter was Uber. As far back as 2016 the ride-sharing company had used machine learning for dynamic pricing and fraud detection. In 2022 it used neural networks to improve the accuracy of drop-off and pickup time estimates. Further AI investments followed, one of which was the use of large language models to generate software code. However, in May 2026 its chief operating officer, Andrew McDonald, went on the record to say that he is not seeing a connection between spending on AI and improved outcomes. “That link is not there yet, right?” This came after Uber’s CEO Praveen Neppalli Naga said in April that Uber had already spent its annual Claude code budget in four months.
Uber is not alone in starting to question its AI spending. Duolingo’s CEO Luis von Ahn had announced in April 2025 that the company would be “AI-first” and that employees would be assessed on their AI use. He walked this back in April 2026, saying: “The reality is it’s not yet the case that AI is better at coding than humans. I think you still really need engineers, and you’re going to need them for a long time.” Microsoft has started cancelling Claude licenses, though that may be more to do with a desire to push employees towards its own Copilot than an issue of license costs.
A broader survey by Walkme in April 2026, involving 3,750 staff at large companies including 1,750 executives, found that 55% of employees trust AI only for simple tasks. 88% of executives believe AI tools are adequate, but only 21% of workers agree. According to the report: “Workers reported spending more time on AI-related friction: verifying outputs, reworking prompts and moving context between systems“.
The issue that Uber found on buying AI tokens from vendors such as Anthropic and OpenAI is particularly problematic since at this point AI tokens are essentially subsidised by the investors of these companies. Both companies are growing fast but are still loss-making. Open AI’s CFO Sarah Friar recently admitted that it had missed its own revenue growth targets. Anthropic seems to be faring better, with revenue of $4.8 billion in Q1 2026, and reckons that it will more than double that in the second calendar quarter of 2026. Nonetheless, as these companies get closer towards their planned initial public offerings, increased scrutiny will be placed on their path to profitability. It is likely that token pricing will rise, since that is one clear way to improve profits. If that happens, though, then enterprise projects justified on current token pricing will be harder to sustain.
These case studies give further evidence that AI, especially large language models (LLMs) like Claude and ChatGPT, are struggling to deliver perceived benefits at any scale in the enterprise. A National Bureau of Economic Research study in February 2026 of 6,000 executives found that: “Nearly 90% of firms said AI has had no impact on employment or productivity over the past three years.” “AI is everywhere except in the incoming macroeconomic data,” according to Apollo chief economist Torsten Slok.
AI, specifically LLMs, has proved wildly popular amongst consumers, with an estimated 3.8 billion people having used a large language model. However the vast majority of these users are not paying for it: OpenAI has a huge user base (845 million in May 2026) but only about 5% of those have paid licenses. Anthropic has a much smaller user base, estimated at between 18 and 30 million monthly users. However, it has focused more on enterprises, with perhaps 300,000 business customers. This focus has led it to overtake OpenAI in the enterprise in terms of market share. However, if those customers are to keep paying then at some point the CFOs of those enterprises are going to want to see a return on that investment. Of course, AI costs may still collapse over time. Open-source models, custom inference stacks, and hardware improvements could dramatically reduce pricing. But current enterprise behaviour is shaped by today’s economics, not hypothetical future cost reductions that may neve materialise.
An enterprise with budget concerns might consider lower cost LLMs rather than slashing licenses of the frontier models like Claude. Chinese open-source models like DeepSeek, Qwen and KimiK2 are considerably more cost-efficient. Although they may not quite have the capabilities of Claude and ChatGPT, they are not far behind on benchmarks. Indeed, a lot of US AI companies (maybe 80% of start-ups) have adopted models from Chinese vendors. The market is evolving quickly, and there is clearly a trade-off between cost and state-of-the-art models. We can expect to see this tension between capability and cost play out further if OpenAI and Anthropic raise prices higher. Enterprises need to be alert to this, and insulate themselves as far as they can from dependency on a specific model. In a fast-changing market, they may have to switch AI horses in mid-stream. For smart enterprises, loyalty to a specific LLM may turn out to be a token gesture.
