• Kirin 9050 Pro is the first commercial chip built on Huawei’s LogicFolding architecture.
• Malaysia is said to be considering Ascend 910C silicon that predates the architecture entirely.

Huawei launched the Kirin 9050 Pro in Guangzhou on Monday, the first commercial chip to use the LogicFolding architecture the company has spent four months defending in research papers. Bloomberg reported the same day that Malaysia is seriously evaluating Huawei’s Ascend 910C accelerators for a RM2 billion (US$494 million) sovereign AI initiative.

The two developments are closer than they look, and the gap between them is the point. LogicFolding is the technology Huawei says will keep it competitive without access to extreme ultraviolet lithography. It has now been in a shipping product for one day. The Ascend 910C that Malaysia is reportedly considering predates it entirely.

The chip debuted in the Mate XT 2, a trifold handset starting at 19,999 yuan (US$2,980). Richard Yu Chengdong, chairman of Huawei’s consumer business group, said the Kirin 9050 Pro delivered a 42% overall performance improvement over the previous-generation Mate XTs and called it the most powerful Kirin the company has built. The device is also the first to run HarmonyOS 7, which began rolling out to existing Huawei phones on Monday.

https://x.com/Huawei/status/2096942699689726061

What LogicFolding actually does

Chipmakers with access to the newest lithography tools improve performance mainly by printing smaller transistors. Huawei cannot, so it is buying performance with geometry instead.

Conventional 3D integration stacks largely self-contained dies on top of one another, like floors in a building that each operate independently. LogicFolding splits core computing functions across two active layers designed to work as a single unit. 

Folding the circuit shortens the physical distance signals travel, cutting both delay and the energy lost in transit across the neural processing, graphics, central processing, and digital signal units that dominate a smartphone chip’s thermal budget.

That principle sits under what Huawei calls its Tau Scaling Law, introduced in a May paper by He Tingbo, chairwoman of the company’s Scientist Committee and president of its semiconductor business department. The paper claimed LogicFolding could reach transistor densities equivalent to a 1.4-nanometre process by 2031 without EUV machines, which remain blocked for sale to China under US-led export controls.

The Kirin 9050 Pro numbers need reading carefully

Huawei published a third paper on ChinaXiv on Friday, and according to South China Morning Post, the tech giant addressed the objection analysts had raised most consistently, which is that stacking active logic vertically would trap heat. He wrote that measurements on the new chip turned the objection on its head. The paper said the chip ran cooler than its predecessor while packing 55% more transistors per square millimetre and cutting power consumption by up to 66% on some tasks.

Two caveats belong with those figures. The paper has not been peer-reviewed. And the density claim is measured against the Kirin 9030 Pro, while the 42% performance figure compares device generations, so the two headline numbers are not describing the same comparison. No independent teardown of the Kirin 9050 Pro has been published.

The approach also carries a manufacturing cost. Bonding two active silicon layers requires precise wafer thinning and microscopic vertical wiring, steps that can weigh on factory yields, according to Leslie Wu, chief executive of semiconductor consultancy RHCC, in a research note. Wu described LogicFolding as a compensatory route and a choice born of external constraints. For companies that can simply print smaller transistors on one layer, he said, the economics are less compelling.

Malaysia’s decision sits on the other side of the roadmap

Venkat Somala, a researcher at Epoch AI, wrote on Friday that smartphone processors offered a far gentler starting point than AI chips. Smaller dies make production losses easier to absorb, lower power draw reduces the thermal problem, and high handset volumes give assembly lines the repetition they need to mature. Scaling the same architecture to Ascend accelerators, which are larger and run far hotter, would be considerably harder. Huawei has said it expects to rely on established 3D stacking for the rest of the decade before bringing LogicFolding to its flagship Ascend 990 around 2030.

That timeline is what regional buyers are actually purchasing against. Anyone evaluating Huawei AI silicon this year is evaluating pre-LogicFolding hardware, and Bloomberg’s reporting, quoting people familiar with the matter, suggests Malaysia is looking at the 910C rather than the newer Ascend 950 because the 950 remains in limited production. Supply, not architecture, is setting the terms.

The politics are unresolved as well. Malaysia’s deputy communications minister announced a national AI system running on 3,000 Ascend chips in May 2025, then retracted the statement within days. The investment ministry said at the time that Malaysia was committed to export control compliance while reaffirming its sovereign right to set policy in the national interest. 

Washington has warned that use of the 910C could breach US export rules. Bloomberg reported that Anwar Ibrahim’s administration is also discussing a separate tender that would let Nvidia compete for the same contract.

Huawei is carrying the cost of the workaround on its own balance sheet in the meantime. First-half research and development spending rose 25% year on year to 121.4 billion yuan (US$18 billion) while net profit fell 36% to 23.8 billion yuan. Revenue grew 9.6% to 467.8 billion yuan, helped by a rebound in smartphone shipments and demand for its AI processors, and the company held 22.6% of the Chinese smartphone market in the second quarter, on a 19.4% jump in shipments, according to IDC.