Tencent used its 12 August second-quarter call to make an unusual disclosure: the roughly $53bn of AI hardware it has bought could be rented out profitably today, and the company is choosing not to.
The numbers executives put on it
Chief strategy officer James Mitchell said depreciation costs could be recovered "almost immediately" by renting capacity, at a profit margin above 30%. President Martin Lau agreed renting would "achieve a decent return in an immediate timeframe" — then said Tencent is "allocating a very substantial proportion of the new" hardware "to building our own models", converting compute into what he called superior intelligence and, eventually, token sales.
What it costs to say no
Capital expenditure hit RMB 52.8bn in the quarter, up 176% year on year. Free cash flow came in at negative RMB 13.8bn; excluding prepayments for compute procurement it would have been positive RMB 37.6bn. Revenue was RMB 204.8bn (up 11%), while IFRS profit attributable to shareholders rose just 0.7% to RMB 56.0bn — roughly a 10% miss against consensus. The shares fell about 3%.
What the compute is for
Tencent said its Hy3 production model ranks top-three globally by token consumption on OpenRouter, and confirmed a larger-parameter Hy4 is coming later in 2026 — with no date and no parameter count. Internally the compute feeds WorkBuddy, which recorded more than 20m PC visits in June, and CodeBuddy.
Read the margin claim carefully
The 30%-plus rental margin and the "almost immediately" recovery are management assertions on an earnings call, not audited figures or a modelled disclosure. Note also that RMB 51.8bn is operating capex and RMB 52.8bn the total — the 190% and 176% growth rates attach to different lines.
The bet underneath
Neoclouds exist because renting GPUs is a viable business. Tencent is saying it agrees, and is passing anyway. That is a wager that model quality compounds faster than rental yield — and it is being funded by a profit line that has stopped growing.