The reported evaluation raises questions about government data, accelerator availability, and what Malaysia's vendor-neutral policy means in practice.
Malaysia is evaluating Huawei's Ascend 910C accelerators for a RM2 billion sovereign AI initiative, Bloomberg reported on September 7, citing unnamed people familiar with the matter. The report identifies state-linked Telekom Malaysia as the operator selected through a public tender. It also describes US concerns about sensitive information and discussions of another tender that could give NVIDIA a more direct opportunity to compete. Chip quantities remain undisclosed. The procurement claims have not been independently confirmed for this article. Bloomberg, republished by Free Malaysia Today.
The significance is what a government would need to secure alongside the hardware: delivery commitments, a usable software environment and control over who can access its data. A chip choice alone cannot establish those conditions.
Malaysia's openness to competing chip suppliers predates this report. In a parliamentary reply reported in November 2025, then investment and trade minister Tengku Zafrul Abdul Aziz described a strategy supporting access to both US and Chinese AI technology while protecting national interests. Malay Mail/Bernama.
That policy provides context for a Huawei evaluation. It does not prove that a government previously selected or deployed Chinese accelerators. Nor does evaluating one supplier necessarily commit Malaysia to a single ecosystem across all its AI projects.
Malaysia also has a more specific precedent. In May 2025, Deputy Communications Minister Teo Nie Ching promoted a Huawei-linked AI infrastructure initiative — a project with Skyvast Corporation that was announced as deploying 3,000 Huawei Ascend chips by 2026. References to Huawei were subsequently removed from the original news report. SCMP.
The investment ministry then issued a statement identifying the venture as privately driven and denying government development, endorsement or coordination. It said government AI infrastructure would require appropriate due diligence. MITI's May 21 statement, as reported by Malay Mail.
The distinction matters when assessing the present report. A prospective system procured for government use raises questions about public accountability and operating control that a commercial partnership cannot answer. But it would still be premature to treat an evaluation as an installed national capability.
The reported interest in an older accelerator deserves scrutiny, but specifications alone cannot explain a procurement decision.
Huawei's original CloudMatrix384 engineering paper describes the 910C as a dual-die package with 128GB of memory and up to 3.2TB/s of aggregate memory bandwidth. Trade-press accounts of Huawei's Connect 2025 presentation put the newer 950PR at 128GB and 1.6TB/s; Huawei's own published keynote states neither figure, describing the 950PR's HiBL 1.0 memory only in cost terms. On those numbers, the 910C would have twice the package-level memory bandwidth — but the two figures do not sit at the same evidentiary level, and neither establishes twice the performance, or better training performance across workloads.
Huawei's product distinction is deliberate. The company describes the 950PR as designed for inference prefill and recommendation systems, while the 950DT targets inference decode and training. Its roadmap specifies different proprietary memory systems for those uses, with 144GB and 4TB/s for the 950DT.
For the Atlas 350 card built around the 950PR, reported specifications give 112GB of memory, 1.4TB/s of bandwidth, and maximum power consumption of 600W. These are card specifications, distinct from the processor's advertised maximum configuration — and they trace to the Chinese outlet Mydrivers reporting on Huawei's March 2026 partner conference, not to a published Huawei spec sheet.
The provenance pattern is itself relevant to a buyer. The Atlas 350 figures and the revised cloud date both reach English-language readers through the same outlet's reporting on Huawei events, and the 950PR memory figures come from comparable trade-press relays of a Huawei presentation. Huawei has published the 950DT's memory numbers and the workload split; it has not published the rest. A procurement evaluation runs on documents a vendor will stand behind contractually, and on that test most of the public 950-series record does not qualify.
The engineering question is therefore workload fit. Procurement evaluation would need to compare the relevant models, precision formats, interconnect, software, and sustained performance. Peak memory bandwidth is one input; it is not an acceptance test. For the software context, see SCN's analysis of DeepSeek V4-Pro running on the Ascend 950PR.
Availability also needs a precise definition. Huawei originally scheduled the 950DT for the fourth quarter of 2026. In June 2026, Chen Lin, Huawei vice-president and president of Huawei Cloud China, was reported to have brought that forward to an August deployment on Huawei Cloud. But a part reaching a Chinese cloud service is not the same event as a part a foreign government can buy. It does not establish when the same hardware could be exported, installed, and supported in Malaysia — which is the only date a procurement schedule can be built on.
Without a disclosed delivery commitment, it is not possible to say that Malaysia is choosing the 910C because it can arrive sooner. The same caution applies to claims that supply constraints make it unsuitable for a multi-year program. A buyer would need binding allocations, replacement parts, a support term and an upgrade path before judging that risk.
US export controls add a separate procurement problem. BIS's May 2025 guidance identifies several Ascend models, including the 910C, as presumptively subject to General Prohibition 10. Its list is non-exhaustive. The 950 series' absence therefore does not establish an exemption or a lower-risk choice.
The analysis depends on whether an item is subject to the Export Administration Regulations, the relevant transaction history, and knowledge of a violation. That calls for evidence about provenance and authorizations, rather than a conclusion based solely on the model number. Export controls can shape system design as much as system procurement: SCN's reporting on the CPU-only machine China built around them describes a program that removed the choke points rather than negotiate them.
Two other instruments should not be confused with that guidance. The AI Diffusion Rule and the administration's non-enforcement policy have their own procedural history: in May 2026, GAO concluded that the policy was a rule subject to Congressional Review Act submission requirements. That decision was not a judicial order resolving the legality of a Malaysian Ascend purchase.
The US–Malaysia trade agreement signed in October 2025 is another distinct instrument. Its text contains export-control commitments and provisions on ICT suppliers. Those provisions should not be characterized as an explicit ban on buying Huawei accelerators; nor should the absence of that product description be treated as proof that the provisions could never apply.
Malaysia's own advanced-chip directive addresses export, transit and transshipment, with a notification and assessment process. MITI Directive No. 1/2025, issued 14 July 2025, applies to high-performance AI chips of US origin. That is different from an acquisition for domestic operation, and it does not settle the application of US controls.
The reported budget cannot establish the system's capacity. Without a configuration, pricing and contract scope, an accelerator estimate would be guesswork. Buyers need to know what is included in the program and what performance the operator is obliged to deliver.
They also need to distinguish hardware ownership from data control. The CLOUD Act concerns relevant service providers and data within their possession, custody or control, including data held abroad. Buying a US-designed accelerator does not, by itself, establish that its manufacturer controls the data processed on it. Equally, choosing Chinese-designed silicon does not by itself demonstrate operational independence.
Meaningful procurement terms would address administrator access, encryption-key custody, software support, model portability, and the ability to continue operating if a supplier relationship changes. A second procurement process could clarify those requirements, but the mere presence of competing vendors would not answer them. Korea's sovereign AI program is the nearer precedent: its sovereignty conditions held where a ministry could enforce them by administrative decision, and dissolved at the infrastructure layer, where the state needed private bidders, capital and accelerators it could not itself supply.
Malaysia's next useful disclosure would be the system it intends to buy: the workloads, delivery obligations, operating arrangements and measurable service commitments. Those details would show whether vendor neutrality is translating into an AI capability the government can control.
The SCN Staff is a small AI editorial squad working under human direction. Each agent owns one job.
Scout does the research. It runs down primary sources and checks what's already been published, on SCN and everywhere else, before a story gets written. If a claim can't be traced back to a real document, Scout flags it.
Forge writes. It takes what Scout found and turns it into a draft, argument and sentences and all. Every SCN piece starts here, then gets sharpened.
Cipher handles search: the titles, descriptions, and keyphrase work that decides whether a good article ever gets found. Least glamorous job on the squad. Also one that matters more than it looks.
Pixel makes the visuals. Images, charts, the occasional diagram, all built to SCN's brand instead of pulled from a stock library. When something's easier to see than to read, it goes to Pixel.
Editorial judgment and the final call stay with the humans. So does the fact-checking.
