MiniMax published its first-half 2026 results this morning, roughly three hours after the Hong Kong bell and exactly on the date it pre-announced on 14 August. It is the clearest look anyone has had at what it costs a Chinese frontier lab to operate.
The headline growth is real
Revenue reached US$116.6m, up 283.1% year on year. Cash stands at roughly US$1.32bn. On top-line growth alone this is one of the faster-scaling AI businesses anywhere.
The cost side is the story
Research and development consumed US$296.9m — 2.55 times the company's entire revenue. Gross margin came in at 17.9%, which means about 82 cents of every revenue dollar is absorbed before a single fixed cost is paid. For a software company that number would be alarming; for a business whose cost of goods is GPU time serving tokens, it is the honest arithmetic of inference at open-weight prices.
What the common framing gets wrong
The release contains two loss figures moving in opposite directions, and both are correct. The IFRS loss narrowed, from US$402m to US$358m. The adjusted loss more than doubled, from US$139m to US$293m. Whichever a headline picks determines whether the company appears to be converging on profitability or diverging from it. The gap is definitional — IFRS carries non-operating items that adjusted figures strip out — so "MiniMax narrows losses" and "MiniMax losses double" can both be filed off the same document. Neither alone is a description of the business.
What it says about cheap Chinese models
The prevailing read on Chinese open-weight pricing is that it reflects structurally lower cost. This disclosure argues otherwise: at a 17.9% gross margin, the price leadership is being funded, not engineered. That is a solvable position with US$1.32bn of cash and 283% growth — but it is a subsidy, and subsidies have a horizon.