My latest opinion piece for Reuters is out today. In it, I focus on the rising costs of building 1GW of data centre capacity and compare it to the expected annual revenue from 1GW of compute. and let me tell you, the numbers don’t look good at all…
Data-centre reality check could slam brakes on AI earnings boom
LONDON, Aug 5 (Reuters) - U.S. corporate earnings are already running almost 60% above trend, with rapid growth still expected for years, mostly thanks to the AI boom. But the questionable economics of new AI data centres could pop this bubble, sending earnings tumbling back toward their long-term average – generating significant losses for investors along the way.
Read the full piece here (paywall): https://www.reuters.com/commentary/reuters-open-interest/data-centre-reality-check-could-slam-brakes-ai-earnings-boom-joachim-klement-2026-08-05/
According to consensus, hyperscaler free cash flow will drop from USD 300bn to under USD 50bn, while semiconductor free cash flow will climb by a similar amount. This matters because 97% of chipmaker revenue is circular, according to the Bank for International Settlements' Annual Economic Report 2026. It is the same money booked twice, as capex leaving the hyperscalers and as cash arriving at their suppliers.
This treats hyperscaler capex as fixed rather than as a decision. Consensus assumes they will keep spending without weighing the cost properly: a payback longer than the assets' useful life, funded anyway, quarter after quarter. If even one of them changes course, it breaks the semiconductor half of the chart.
The Reuters piece was really, really balanced. Very few articles focus on facts rather than pander to a side or an opinion. Thanks for the link!