Chinese AI labs are also developing other sources of revenue. MiniMax, which makes roughly 60% of its sales from overseas, generates subscriptions from consumer apps. Moonshot, the owner of the Kimi models, is demanding revenue-sharing agreements with major users of its models - a practice that is set to become more common. So long as revenue grows faster than the costs of training new models, the payback period on R&D should shorten.
The other promising trend is in inference, where trained models process and respond to prompts. China's obsession with efficiency and scale is starting to pay off as serving APIs - the biggest drag on gross profit - is getting cheaper. Z.AI says per-unit inference costs are down 80% since the start of the year, lifting gross margins at its API division to 25% in the six months to June. That's a notable improvement from the negative 0.4% a year earlier. Separately, MiniMax credits "improving infrastructure efficiency" for a 5.8 percentage point increase in overall gross margins over the same period, to just under 18%. These levels of profitability are still puny, though. Anthropic's gross margin in the second quarter was 44%, PitchBook estimated. Even that falls far short of the 70%-plus level earned by many of the cloud software specialists that AI is meant to replace.
For Chinese companies, extracting greater performance from models while meeting growing demand increasingly rests on bulking up in-house computing capacity and expertise. Yet the sums remain tiny when compared with the U.S. infrastructure splurge. Alibaba racked up $10 billion in capital expenditure in the second quarter, up 75% from a year ago and dragging the e-commerce giant's cash flow into the red. But the company has a cloud computing business to cross-sell APIs, storage, higher margin IT services and even excess computing power to its already sizable customer base. That puts boss Eddie Wu in an enviable position: Alibaba says it expects a three-year payback period on its AI capex. Standalone AI labs with fewer revenue sources will probably struggle to match that.