← Writing

China's AI Bubble Already Burst

China's main index of AI and chip stocks lost about $480 billion in under four months. Its data centers deflated a year before that. Its AI companies are still selling. What that could mean for the US, and a price check anyone paying for AI tools can run on their own contracts.

China’s STAR50, a mainland index of AI and chip companies, fell about 40 percent in less than four months and erased about $480 billion of market value, Reuters reported on October 26, 2026. Reuters put its conclusion in the headline: in China, the AI bubble has already burst.

I agree with the headline, and I’d add to it. China’s AI boom came apart in layers, and the order is the part an American executive should study.

  • 2025: the data centers deflated. Buildings full of AI chips sat mostly empty.
  • 2026: the stock prices fell. One index lost about $480 billion in under four months.
  • Still standing: the AI companies. They are still selling and still raising money.

If your company pays for AI tools, the most useful section is the price check near the end. Jump to it.

40%fall in China’s STAR50 index of AI and chip stocks, in less than four monthsReuters, October 2026
$480Bof market value erased by that fallReuters, October 2026

The short version

This is my read of what happened and what it means.

  • Yes, China’s AI stock bubble has burst. Prices fell hard, and they fell across the sector at once.
  • The data center bubble deflated a year earlier, in 2025, when buildings full of AI chips sat mostly empty.
  • China’s AI industry has not collapsed. Its best-known AI companies are still selling and still raising billions.
  • If the US has a correction, I expect a similar order. The first trouble would show up at the companies that borrowed heavily to rent out AI chips.
  • If you pay for AI tools, a stock crash doesn’t change whether they pay off. A price change does, so test your numbers at today’s price, at half and at double.

What happened in China, layer by layer

2025: the data centers

Starting in 2023, China built AI computing capacity, meaning buildings full of AI chips rented out by the hour, much faster than anyone could put it to work. MIT Technology Review laid out the scale in March 2025:

  • More than 500 new data center projects were announced across 2023 and 2024.
  • At least 150 of them were finished and running by the end of 2024.
  • Up to 80 percent of the newly built capacity sat unused, according to estimates in two Chinese trade outlets. Treat that as a rough size of the problem.

The rent on the chips fell by more than half. A server with eight Nvidia H100 chips, the standard hardware for training AI, rented for about 75,000 yuan a month by early 2025, down from about 180,000 yuan at the peak. Light Reading put those at about $10,000 and $25,000.

$25Ka month to rent an eight-chip Nvidia H100 server in China at the peakMIT Technology Review and Light Reading, 2025
$10Ka month for the same server by early 2025MIT Technology Review and Light Reading, 2025

The buildings were meant for companies training their own AI models, and most of those companies stopped. Of more than 144 Chinese companies registered to build ChatGPT-style AI models in 2024, only about 10 percent were still doing large-scale training by the end of that year, per the same MIT Technology Review piece.

The sharpest warning came from inside the chip business. Zhao Haijun, co-CEO of SMIC, China’s largest chipmaker, told analysts in February 2026 that “Companies would love to build 10 years’ worth of data center capacity within one or two years,” as reported by Bloomberg and carried by Tom’s Hardware. He said what those data centers will do has not been fully thought through. He was talking about the buildout around the world, China included.

2026: the stock prices

The stock market came apart about a year after the data centers. These figures come from the Reuters analysis:

  • The STAR50 fell about 40 percent in less than four months.
  • China’s index of its biggest companies fell about 20 percent from its June high.
  • AI-related stocks make up about 42 percent of the value of China’s stock markets and 57 percent of the trading, by Goldman Sachs’s count, so when they fell, the broader market fell with them.
  • Unitree, a robot maker, had 8,000 times more orders from small investors than it had shares to sell when it listed. It has lost half its value since it started trading in August.
  • CXMT, a memory chipmaker, lost more than a quarter of its value in the month before the Reuters report.

The two best-known Chinese makers of ChatGPT-style AI that trade publicly, Zhipu and MiniMax, are listed in Hong Kong. As of October 22, 2026, Chinese business outlet 36Kr reported:

  • Zhipu was down about 85 percent from its June peak.
  • MiniMax was down nearly 90 percent from its March peak, and it had closed below its HK$165 IPO price, about US$21, for the first time.
  • At their peaks, investors valued MiniMax at more than 600 times its last twelve months of sales and Zhipu at more than 1,000 times.
Price paid for each $1 of last year’s profit
China’s AI index still costs three times the Nasdaq.
DOLLARS PAID PER $1 OF PROFIT$250$100$30STAR50 AT PEAKSTAR50, OCT 2026NASDAQ 100
Source: Reuters, October 26, 2026, which gives the STAR50 at nearly 250 times trailing earnings at its peak, about 100 times in October, and the Nasdaq 100 at about 30 times. Bars drawn to scale.

Reuters found no single trigger. Traders pointed to three pressures:

  • High prices after the run-up.
  • Tax bills.
  • The threat of bans, as the US and other countries moved to keep Chinese equipment out of their data centers.

Wang Zhuo, a Shanghai fund manager, summed it up for Reuters: “All bubbles inevitably burst.”

Beijing has noticed. On October 9, 2026, with the slide well underway, China’s Communist Party and cabinet told officials to avoid asset bubbles, “blind investment,” and “swarm-like” rushes into high-tech industries, Reuters reported. Officials responsible for major losses from blind investment will be held accountable.

Still standing: the companies

While the share prices fell, the businesses kept booking sales and attracting investors.

  • MiniMax’s revenue for the first half of 2026 was $117 million, up 283 percent from a year earlier, per 36Kr. It also lost about $293 million in that half, so it is growing fast and still losing money.
  • DeepSeek, a private company, was raising at least 80 billion yuan, about $11 billion, with Tencent and battery maker CATL among the largest investors, Bloomberg reported via Euronews on October 6, 2026.
  • Moonshot AI, also private, raised money at a valuation of about $50 billion, up from $31.5 billion in a summer round, per the same report. It is aiming for a Hong Kong IPO in early 2027.

Two cautions:

  • Most of the revenue figures above predate the worst of the slide. The next two quarters will show whether customers keep paying.
  • Moonshot’s IPO could slip. Hong Hao of Lotus Asset Management told Reuters he expects the selloff to scare companies away from listing.

The takeaway: the bubble was in the prices people paid for AI, first for buildings and chips, then for shares. Judge an AI vendor on its sales and how long its cash lasts, and leave its share price to the traders.

Is it a burst or a correction?

The best argument against the word is that this is a sharp correction in a market that will recover. Goldman Sachs added two of the fallen names to a recommendation list in October, and William Bratton of BNP Paribas told Reuters he sees buying opportunities.

I still call it a burst, for two reasons:

  • It may not have found its floor. Even after the fall, the STAR50 costs about three times what the Nasdaq 100 costs for the same dollar of profit.
  • The losses are real for whoever paid peak prices. Anyone who bought MiniMax at its March peak was down nearly 90 percent by October 22, whatever happens next.

What it could mean for America

This section is my take. China’s market differs from ours in ways that could change the ending.

ChinaUnited States
Who built the AI computing capacityLocal governments and companies, with more than 500 data center projects announced in 2023 and 2024Mostly five companies (Alphabet, Amazon, Meta, Microsoft and Oracle), with more than $690 billion of capital spending in fiscal 2026
Stock prices per $1 of profit, October 2026STAR50 about $100, after a peak near $250Nasdaq 100 about $30
How far stocks have fallenSTAR50 down about 40 percent in under four monthsNasdaq down 6 percent from its early-October record

Sources: MIT Technology Review, FactSet, Reuters.

Where the US looks like China:

  • The US may be building ahead of demand too. SMIC’s warning covered every country building data centers, and the US is the biggest builder.
  • More of the US buildout runs on borrowed money each year, according to FactSet. The figures are below.
  • The US has a version of China’s rent-out-the-chips business. CoreWeave, the largest listed company that rents out AI chips, reported $35 billion of debt as of June 30, 2026, according to the trade outlet Capacity Media. Like China’s data centers, these companies need tenants to arrive on schedule.
9%of the five biggest US builders’ capital spending paid with new debt in fiscal 2024FactSet, 2026
32%paid with new debt in the twelve months before June 2026FactSet, 2026

Where the US looks different:

  • US tech stocks are priced much closer to their profits, as the chart above shows.
  • Most US capacity is being built by five companies with large, profitable businesses underneath.

My call: if the US goes through a version of this, it starts with the companies that borrowed to rent out AI chips. The five giants’ stock prices come later, if at all. Those borrowers’ results and debt costs are what I’d watch.

What it means if you pay for AI tools

You won’t be buying data centers. A crash in AI stocks would reach a company your size in three ways.

  • Through prices on software you already rent. Increases arrive as AI add-ons bundled into Microsoft 365, CRM and ERP renewals, as seat minimums, and as switches from flat fees to usage meters.
  • Through your vendors. MiniMax went from HK$1,330 a share, about US$170, in March to below its HK$165 IPO price by October 22. A vendor whose value collapses may cut staff, get bought or change its pricing, and you hear about it afterward.
  • Through commitments you aren’t using. Seats bought for 400 people when 60 use them, and prepaid credits sitting idle, are your version of China’s empty data centers.

A stock crash doesn’t change whether your AI tools pay off. A price change does.

Prices could move either way, so plan for both:

  • Down, if there’s a glut. China’s glut cut chip rents by more than half.
  • Up, if vendors need to repay debt. Vendors carrying debt have reason to push prices up, through the add-ons, seat minimums and usage meters above.

Either way, keep funding the AI work that pays for itself.

Run the price check on your own contracts

Here is a composite company, invented for illustration. A $120M regional distributor pays $180,000 a year for AI tools in order entry. The tools let it avoid two and a half hires, worth about $210,000 a year all in. At today’s price, the program comes out $30,000 a year ahead.

Now run the same program at other prices. Treat double as a stress test.

AI price vs. todayYearly AI costYearly savingNet per year
Half (a glut)$90,000$210,000+$120,000
Same as today$180,000$210,000+$30,000
Up 17 percent (break-even)$210,000$210,000$0
Double (vendors reprice)$360,000$210,000-$150,000

Here is what the table tells this company:

  • It makes money at half price.
  • It stops paying at a 17 percent increase. So the term it needs at renewal is a yearly price cap below that break-even.
  • A 5 percent cap keeps the program ahead through year four.
  • A 10 percent cap puts it underwater in year three.

The check takes a few hours per contract with your own numbers.

The takeaway: find the price increase that breaks each AI business case, and get a cap below it into the contract before renewal.

What to do this quarter

The Bubble Check

  1. List every AI cost. Include AI add-ons inside Microsoft 365, CRM and ERP renewals, and anything bought on company cards. Note each renewal date, price-increase term and exit term. IT with procurement, about two weeks.
  2. Test leaving each smaller vendor. For each AI vendor that isn’t a public company or a major platform, request a full export of your data, prompts, settings and workflow definitions, load it somewhere else, and confirm it works. At the next renewal, add a 30-day data return and deletion clause. IT with legal, about 30 days.
  3. Cap your prices. Find each AI business case’s break-even price increase, and negotiate a yearly cap below it. Finance, a few hours per contract.
  4. Stop paying for unused seats. Sign no seat or credit commitment above what a named workflow uses today, and negotiate quarterly true-ups so each quarter’s bill follows what people used. IT, starting at the next renewal.
  5. Keep funding what works. That means the AI workflows that pay for themselves in cash or avoided hires, checked each quarter. The business owner of each workflow.
  6. Brief your board. Write a one-paragraph answer: which AI vendors you depend on, what each would cost to replace, and what you’d keep running if prices doubled. The CIO, before the next board meeting.

China went first. The companies with paying customers kept selling, and the people who paid peak prices for buildings and shares took the losses. Be a paying customer of AI that works, on contracts you can leave, and a crash costs you a renegotiation while the program keeps running. If you’ve gotten a hard price cap into an AI renewal, I’d like to hear how.


How RLK Can Help

I help CIOs and CFOs test their AI business cases and contracts against price swings and vendor failure before a board asks. Start a conversation.


Sources

Ryan King

About the author

Ryan King

Fifteen years in technology strategy at McKinsey and Deloitte. Now running RLK Consulting: enterprise-caliber tech strategy, one strategist doing every hour of the work. Over $10B in documented value capture across 50+ engagements and 12 industries.