How to read this page. Each section starts Simple, then goes to a Deep dive: stop wherever you have what you need. The small numbers are sources: click one to open the original document. Where the maker has not published a figure -- a price, a die size, a factory address -- this page says so rather than estimate it.
1.At a glance
SimpleStart here
Ascend 950PR is Huawei’s newest shipping AI chip, debuted at the Huawei China Partner Conference in late March 2026, built into a card called the Atlas 350 12.
Deep diveThe technical detail
Two press outlets covering the same event give slightly different exact dates — TrendForce dates the debut 20 March 2026, TechRadar’s coverage dates it 28 March 2026 12 — a small discrepancy typical of global press picking up a China-first announcement on different days; this page states “late March 2026” rather than pick one outlet’s exact date as authoritative, since no Huawei-first-party dated press release for this specific event was found.
2.Launch and history
SimpleStart here
The “PR” in the name stands for “Prefill and Recommendation” — Huawei built this chip specifically for two stages of AI inference, not for training new models.
Deep diveThe technical detail
Huawei’s own words, from its Huawei Connect 2025 keynote setting out the roadmap this chip fulfilled: “950PR: Optimized for prefill inference and recommendation systems… will help our customers maintain the right level of performance while significantly reducing their investment” 3. Huawei set a Q1 2026 target for 950PR in that same keynote, and the March 2026 debut met it 3 — one of the few Ascend roadmap commitments this site found delivered on schedule, in contrast to its sibling chip, 950DT (see below).
3.What’s inside it
SimpleStart here
A new architecture combining SIMD and SIMT execution styles, paired with Huawei’s own first self-developed high-bandwidth memory, called HiBL 1.0 42.
Deep diveThe technical detail
Tom’s Hardware’s roadmap analysis describes the 950-series core design as combining SIMD (single-instruction-multiple-data) and SIMT (single-instruction-multiple-thread) execution — a departure from the pure Da Vinci-core design of the 910-series chips 4. The memory, HiBL 1.0, is described by press coverage as Huawei’s first in-house-developed HBM, with a claimed “128-byte access granularity” said to “quadruple memory access efficiency for small operators” compared to the prior generation 2 — a Huawei/partner claim relayed by press, not an independently verified figure.
4.Full spec table
SimpleStart here
1.56 PFLOPS of FP4 compute, 112GB of HBM memory at up to 1.4 TB/s, 2 TB/s of interconnect bandwidth via Huawei’s “LingQu” protocol, all on the Atlas 350 card 21.
Deep diveThe technical detail
| Spec | Huawei Ascend 950PR (Atlas 350 card) |
|---|
| Compute | 1.56 PFLOPS FP4 |
|---|
| Memory | 112GB HiBL 1.0 HBM (Huawei’s first in-house HBM) |
|---|
| Memory bandwidth | up to 1.4 TB/s |
|---|
| Interconnect | 2 TB/s via Huawei’s “LingQu” protocol, claimed 2.5x the 910-series |
|---|
| Core design | SIMD + SIMT (per independent roadmap analysis) |
|---|
| Process node / foundry | Not stated by Huawei; not independently reported by any source this page found |
|---|
Figures per press coverage of the March 2026 debut 21; process node is genuinely unconfirmed for this chip, unlike the 910-series, where independent die analysis at least offers a press/analyst estimate.
5.Where it’s made
SimpleStart here
Not stated by Huawei, and this page found no independent teardown analysis identifying a foundry or process node for Ascend 950PR.
Deep diveThe technical detail
This is the one chip on this site’s Huawei pages where even a press/analyst foundry attribution is missing — the 910-series chips at least have independent die-teardown analysis naming TSMC or SMIC; for 950PR, this page found nothing beyond Huawei’s and partners’ own performance claims. Stated plainly as an open gap rather than filled with a guess.
6.Which systems use it
SimpleStart here
The Atlas 350 accelerator card, built by Huawei with system products from at least seven named partners including Kunlun, Huakun Zhenyu, Shenzhou Kuntai and Yangtze Computing 2.
Deep diveThe technical detail
The partner ecosystem named around the Atlas 350 launch is notably broader than Huawei’s own-brand Atlas SuperPoD systems built around the 910-series chips — consistent with Huawei positioning 950PR as a component other Chinese system-builders integrate into their own products, not only a Huawei-branded system 2.
7.Official pricing
SimpleStart here
Not published by Huawei. TechRadar reports the Atlas 350 card priced at “around 111,000 Yuan, or roughly $16,000” 2 — a market price reported by press, not a Huawei list price.
Deep diveThe technical detail
This figure should be read as a reported market price, not an official one, consistent with this site’s standing rule for every Huawei Ascend page: Huawei itself publishes no chip or card prices. TechRadar’s own framing places it “comparable with Nvidia H20, which can range from $15,000 to $25,000” 2 — useful context, but this page could not independently trace the ~CNY 111,000 figure to a specific retailer or partner listing beyond this one press report.
9.What came before, what came next
SimpleStart here
Came before: Ascend 910C, the previous generation. Came after: Ascend 950DT, its training-focused sibling chip — real, but its exact ship status as of this page’s check date is genuinely unclear, which is why it does not get its own page yet.
Deep diveThe technical detail
This page found two different Huawei-linked dates that do not obviously describe the same milestone. In June 2026, Huawei said it would deploy 950DT “on Huawei Cloud in August, earlier than its previously expected fourth-quarter launch” 6 — a cloud-service availability date, not necessarily a purchasable card. Separately, in September 2026, Huawei Cloud set commercial launch dates of 30 September 2026 (China) and 30 November 2026 (global markets) for “its Ascend 950 AI cluster cloud service” 7 — but that source discusses the cluster service broadly and does not name 950DT specifically, so this page cannot confirm whether it describes the same product as the August cloud deployment, a broader rollout of it, or a service that also includes 950PR (already shipping since March 2026). Rather than pick one of these dates and assert a clean “shipped” or “not shipped” verdict, this page states the ambiguity directly and leaves 950DT off this site as its own product page until its status is unambiguous — a stricter bar than either source alone would require, and the honest one given what this page could actually confirm.
10.Hidden in plain sight
SimpleStart here
The chip is explicitly built to fill a hole another country’s export rules created — positioned by Huawei and its partners as a replacement for a specific NVIDIA chip that Chinese buyers can no longer legally get.
Deep diveThe technical detail
NVIDIA’s H20 was itself a cut-down chip NVIDIA designed specifically to comply with earlier US export restrictions while still being sellable into China; 950PR’s launch marketing benchmarks directly against that specific chip, not against NVIDIA’s more capable export-restricted parts 21 — a chip built to satisfy one country’s export rule, now facing a domestic competitor built specifically to replace it once a tighter export rule removed it from the market entirely.