Copy article

Could Runaway API Token Costs Actually Bankrupt a Company?

ended 29. May 2026

Axios reports one company just got a £380 million AI bill for a single month.

Axios reports an unnamed firm spent half a billion dollars in one month after failing to cap employee usage of Claude licences. Microsoft has cancelled most of its internal Claude Code licences by June 30 over spiralling costs, forcing staff to use a cheaper internal model. Uber burned through its entire $3.4 billion AI budget in four months. 

Companies risk denting their balance sheets to fund AI that hasn't proven its commercial value out in the field. Processes requiring interpretation and judgement are messy for machines and need a lot of scaffolding to assess and verify the output.

UK businesses are currently planning a 40% increase in AI investment over the next two years, averaging £15.94 million per organisation.

The FT reports that Amazon has scrapped its AI leaderboard to stop workers chasing usage scores rather than efficiency gains.

We want your views:

  • If Uber's 5,000 engineers blew a £3 billion budget in four months, what governance should UK firms have in place before rolling out enterprise AI licences at scale?
  • Nvidia's Jensen Huang said engineers should be evaluated by how many AI tokens they consume. One company followed that logic and got a £340 million bill. Where does the accountability sit when the vendor culture actively encourages overconsumption?
  • What if layoffs are being used to offset AI costs, not because AI replaced the roles. Should UK employment regulators be scrutinising redundancy justifications that cite AI automation?
  • UK firms are forecasting a 40% increase in AI spend while some US enterprises are discovering unitended consequences of deployment. Is the UK about to repeat the same cycle, or is there a structural difference that protects against it?

2 responses from the Newspage community

Copy all

Star Quote
Copy

Colette Mason
AI Ethics Consultant at Clever Clogs AI
This is the enterprise equivalent of a parent discovering their child has spent £1,000 on in-game credits in the app store, except the child is a Fortune 500 company and the bill is £380 million in a single month.

Demoralised software engineers who read the model card prices pointed out the real cost of transformation months ago, but were labelled as naysayers.

Someone in this company didn't tick the budget cap box, and I suspect it wasn't someone in the IT team but someone in accounts who didn't realise that Claude Opus 4.6 costs can run up to £19 per million output tokens.

Just summarising a single 10-page PDF can burn through 20,000 tokens. Process just 50 PDFs and you've hit that £19/million, the same spend for a subsidised flat-rate subscription a back-office member might have been using all month. API costs are an entirely different beast.

Employees may be less glamorous than AI, but the costs are a lot more predictable. We need humans in the loop.
Copy

Premium
Debbie Porter
Managing Director at Destination Digital Marketing
I find this really amusing, especially given that AI has been marketed as the low-cost solution that will replace employees. It's so easy to just throw a document into Claude or ChatGPT and ask for a summary rather than just actually read it yourself. Once you then start to ask questions conversationally to develop the thought process as each response comes back, it can soon burn through those tokens. It encourages people to abandon their own critical thinking processes, and become reliant on AI to do what humans are perfectly capable of doing - without the surprise costs at the end of the month.