Nvidia China Share Falls to 34% as AI Cash Tightens

Nvidia held 34% of China’s data center AI chip market in the first half of 2026. Shipments into China still fell 44% from a year earlier, to 678,000 accelerators. This week memory names gave back $108 billion of market value. Broadcom’s chief warned that money is the next hard stop.

China volume is leaving Nvidia

IDC figures put Nvidia at 34% of China’s data center AI accelerator market from January through June, on 678,000 chips. That count was 44% lower than a year earlier.

US export controls limit which Nvidia parts can be sold into the second largest economy. Beijing adds a second gate. Buyers need approval even for Nvidia chips that still meet US rules. A supplier can win the technical comparison and still miss the order.

Huawei shipped 572,000 accelerators in the same six months, up from more than 300,000 a year earlier, for a 28% share. The gap is about six points of share and roughly 100,000 units.

Domestic chips are no longer a footnote

The next names add up. Alibaba’s chip unit held 7%. Cambricon held 5%. Baidu’s Kunlunxin held 4%. Hygon held 3%. Nine of the 11 Chinese chip makers IDC tracks shipped more than they did a year earlier.

Total accelerator shipments in China reached 2.008 million units in the first half. Graphics processors fell 26% to 991,000 units. Other accelerators rose nearly 80% to 1.016 million units and passed graphics processors by count.

IDC expects total accelerator shipments in China to grow 19% in 2026 and graphics processor shipments to decline 9%. Many of the extra units are local designs for running models under the export rules that remain. Those parts can take an order Nvidia is blocked from filling. The China unit line is the cash that moved.

Memory cash looks great until supply shows up

Cathie Wood is staying away from traditional memory chip makers. Memory runs in vicious boom and bust cycles. Prices fall fast once supply gets ahead of orders. She told investors in Hong Kong that extra Chinese capacity can make the next bust worse.

ChangXin Memory Technologies reported progress on a new memory cell and wants a server memory product by the end of 2026. Samsung is chasing a similar design, with mass production talked about closer to 2028. A working sample is a long way from a high factory yield. The headline was still enough to knock a combined $108 billion off Micron, SK Hynix, and SanDisk.

Samsung has estimated a record third quarter operating profit of 107.4 trillion won, about $80.2 billion. Micron’s latest fiscal quarter showed an 86.8% gross margin. Customers placed about $12.7 billion of contract deposits over the fiscal year. A margin at that level describes a shortage. Deposits show buyers afraid of running out.

Elon Musk made the same split. Standard phone and laptop memory looks easier for China to scale. High bandwidth memory, and stacking it on the processor without overheating, is harder. Extra supply in ordinary memory is enough to push prices down.

Capital is the constraint Tan is watching

Hock Tan at Broadcom named capital as the constraint that may outlast the others. Memory could look different within a year. The wafer squeeze should ease by 2028. Power is there for plans aimed at 2028 and 2029.

He put worldwide spending on running models at about $200 billion a year, against about $150 billion of revenue. Frontier labs, in his split, turn about $100 billion of spending into about $120 billion of revenue. Makers of open models spend a similar sum and get about $30 billion back. Most buyers of that output still show no return.

OpenAI was at about $50 billion of annualized revenue as of September, and still loses money on training. Anthropic, on Tan’s numbers, is near $65 billion of annualized revenue on less than 2.5 gigawatts, about $30 billion per gigawatt. He put a gigawatt at about $10 billion a year over five years, so about $20 billion stays with the model company and chip sellers fight over the rest.

Broadcom’s AI semiconductor revenue was $16.7 billion in the fiscal third quarter, up 221% from a year earlier, with $21.7 billion expected in the fourth quarter. The full year figure in use is $58 billion, with a sketched path toward $115 billion in fiscal 2027 and $230 billion in fiscal 2028. Almost all of that AI revenue comes from about six customers. Public reports put more than $50 billion of debt in discussion so OpenAI can buy custom chips, with Apollo and Blackstone among possible lenders. Banks are also syndicating a package near $60 billion tied to Anthropic, including about $42 billion of senior loans that use Broadcom’s credit rating. Tan still sees demand ahead of supply for another two years if people keep using the models.

What this means for income investors

The cash sits with companies selling chips, networking, and scarce memory. It is real this year. It is a weak base for copying the same growth rate into 2028.

• China units and China share are different numbers. A 34% share can sit on shipments that are down 44%. Later payout growth that depends on revenue will feel the missing units first.
• Memory margins above 80%, plus $12.7 billion of customer deposits, are peak cycle signals. Chinese capacity can hurt the price without taking the hardest product. A server part aimed at the end of 2026 is a date worth tracking.
• If the next leg of orders needs debt packages of $50 billion and $60 billion, the risk sits in the credit market as well as the chip factory. High interest works while frontier revenue holds. A slip in that revenue leaves the debt in place and thins the equity story.

Buyers are fewer, more of the spending is borrowed, and local parts are taking China volume that used to pad US chip profits. Judge the toll from orders and borrowing costs.