AI money pours in, but few pay, and the math remains unsettled

This week's AI economics stories tell two things at once: a great deal of money is flowing into the industry, but it is still unclear who will ultimately foot the bill.

Money floods into China

DeepSeek is close to raising at least $12 billion, according to Bloomberg, with Tencent and CATL among the major investors. That is far more than the original plans. Moonshot AI and Manus have also raised, or come close to raising, large sums, and all three are considering or preparing an IPO.

But the stock market is no sure path to riches. MiniMax, which is already publicly listed, has lost more than 80 percent of its value since its peak in March and is now raising new money. Investors clearly believe strongly in some companies, while others are punished hard.

Few pay, and those few pay a lot

A new report shows that only 4.5 percent pay for ChatGPT, Gemini or Claude. The top 1 percent of users account for nearly a fifth of usage. That is a fragile business if it rests on such a narrow layer of users. a16z's list of consumer apps also looks, for the first time, at what people pay for. ChatGPT still leads, and Claude is number three.

At the same time, the picture for Claude is mixed. The app's US downloads have more than halved, from 7.42 million in March to 3.26 million in September, but it has gained a much larger share of daily users. Fewer new downloads do not necessarily mean fewer active users.

Cheaper AI, costlier computing power

When AI gets cheaper, companies may end up using more of it. That is the view of a16z and Citadel Securities, though they stress that the signs are preliminary. The price of renting the chips that power AI is not falling correspondingly. That could squeeze providers.

Pricing is also a theme at Anthropic, which will shut down its cheapest model, Haiku 4.5, on October 15, 2026 at the earliest. That could determine whether developers will have to pay double for the small model going forward.

Companies use AI where it pays off

At the other end of the market, large organizations are moving in. Fujitsu has gone from 7,500 to roughly 100,000 Copilot licenses in under a year, but the figures come from Microsoft's own customer story. Ecosia is dropping France's Mistral after just a few months and plans to use open models, including Chinese ones. The founder says Mistral is a year behind.

What can you do? If you buy AI for a business, expect prices to shift, and avoid tying everything to a single model. Watch whether the Chinese IPOs go through, and whether more than a small group of users ends up opening their wallets.