Huawei Moved the Ascend 960 Up Nine Months, Its System Shrank 73%

Huawei moved a date forward in Shanghai today.

Its new AI chip, the Ascend 960DT, will arrive in the first quarter of 2027 instead of the fourth. Nine months early.

The system unveiled on the same stage was about a quarter the size of what was promised a year ago for that same generation.

What Was on Stage

David Wang, Huawei's rotating chairman, made the announcement.

The Ascend 960DT lands in the first quarter of 2027 and its sibling, the Ascend 960PR, in the third — one quarter early. The company also placed the Ascend 970 in 2028 and the Ascend 980 in 2029.

A Huawei spokesperson put it briefly: "Ascend 960 chips are launching ahead of schedule, doubling performance."

The system is called the Ascend 960 SuperPoD. A SuperPoD is a cluster of chips wired to behave as one machine; Huawei calls these supernodes.

This one holds 4,096 cards, 8 EFLOPS of FP8 compute and up to 1 petabyte of HBM. It was described as the first supernode to use NPO, which moves the optical link next to the chip package.

What Was Promised a Year Ago

Huawei's own September 2025 keynote, still posted on its corporate site, describes something else.

There, the Ascend 960 chip sat in the fourth quarter of 2027. The node for that generation, the Atlas 960 SuperPoD, was to carry 15,488 chips, deliver 30 EFLOPS of FP8 compute and hold 4,460 TB of memory.

The node shown today has 4,096 cards and 8 EFLOPS.

What moved forward was the calendar.

What did not move forward was the chip count.

Same Total, Half the Silicon

The shrinkage is not a downgrade. The chip itself delivers what was promised.

Huawei set a target of 2 PFLOPS of FP8 per Ascend 960 chip last year. Multiply 4,096 cards by 2 PFLOPS and you get exactly 8 EFLOPS. What shrank is the node, not the chip.

The interesting part is where that total lands.

With the line fixed, the engineering gain goes into how many customers can be served, not into node size.

The Bottleneck Is Memory

Huawei described the constraint itself: output does not cover what Chinese customers alone are asking for. Serious overseas selling, the company says, is not realistic at this stage.

Price confirms it. Bloomberg reported on September 10 that the indicated price of the Ascend 950DT rose roughly 60% in three months to 250,000 yuan, about $37,300, citing tight supply of key components.

The key component is HBM: high-bandwidth memory stacked vertically beside the chip. Three companies make it, one of them Micron, and most of 2026 output is already sold.

Demand makes the gap plainer. According to secondary reports, DeepSeek wants at least 160,000 Ascend 950DT chips for a data center in Inner Mongolia, and Huawei cannot fill an order that size.

What the Market Did

Nvidia traded up about 2.5% on the day, near $219.

That move cannot be tied to the Huawei announcement. The Nasdaq 100 rose 1.7% and the S&P 500 1.1% the same day, as the market recovered from the Fed's rate increase a day earlier and oil fell.

The instructive data here is the absence of a move. China's largest chip designer said it was pulling its AI processor in by nine months, and Nvidia's price showed no break that reflects it.

What Is Still Open

Huawei did not say the 15,488-chip node was cancelled. What it did not do was repeat the figure.

The roadmap relayed by TrendForce still shows a liquid-cooled Atlas 960 SuperPoD in the third quarter of 2027, with no chip count attached. Analyst Rui Ma flagged the gap between today's 4,096 and the earlier 15,488.

The second open question is memory. Huawei is developing its own HBM; if that line works, the node can grow again.

The 2025 roadmap figures come from Huawei's own corporate keynote; the 2026 announcement from event reporting. The price increase rests on Bloomberg's reporting and the DeepSeek order size on secondary accounts.