Huawei's physical-AI report grades cross-layer compute at 2 out of 5
Huawei wants to sell the industry a coordination layer for physical AI. Its own report puts that layer in the least mature row of the table.
Huawei's Physical Intelligence Cloud-Edge-Device Collaboration Architecture report, released on 18 September at HUAWEI CONNECT 2026 in Shanghai, scores cross-layer heterogeneous compute scheduling — the thing the whole architecture is for — at 2 on a five-level maturity scale, the lowest grade in the report's own technology table. The document runs that table across cloud, edge, device and cross-layer capabilities, using a scale that climbs from concept validation to prototype validation to commercial deployment, scale replication and finally infrastructure. Device-side motion control and the safety loop scores 5, the only entry to reach the top: Huawei calls the PLC-and-real-time-controller stack the "certainty base" physical AI already runs on. Cloud data governance, model training and evaluation, model repositories and global operations all score 4. Unified scheduling across CPU, GPU and NPU and across cloud, edge and device scores 2, with the report noting that cross-vendor, cross-body resource awareness, task orchestration and dynamic scheduling remain in early exploration. Task offloading and live migration also score 2: state migration, latency prediction and safe continuation are unresolved.
Read that as an admission with a product attached. Huawei's executives used the same launch to say the company will bring out chips, accelerator cards and modules for device-side computing, and to pitch "imperceptible" scheduling that hides the underlying compute network from developers. Selling the coordination layer while grading coordination lowest is not contradiction for its own sake — it is a vendor arguing the hard part is also the unclaimed part, which is where margin lives. The report's reference architecture is called "three-domain collaboration, one network": application, data and compute coordination across cloud, edge and device, stitched by device-internal, device-to-device and device-to-cloud networks that lean on 5G-A, TSN, industrial Ethernet and Huawei's own NearLink short-range link.
The release is a standards play as much as a technical one. Huawei published it with the Chinese Academy of Sciences' Shenyang Institute of Automation, EY China and the Global Computing Consortium — a list the company's own release confirms — and the report's governance chapter points at existing standards bodies for testing and certification. Because the Chinese embodied-AI market has no shared answer for how a robot hands a task to an edge node and back, the vendor that writes the reference architecture writes the vocabulary everyone else evaluates against. Huawei's timeline gives that consensus three phases: fixed roles with edge deployed on demand through 2028, then regional autonomy with edge nodes as local intelligence hubs to 2030, then dissolved boundaries and dynamic coordination after that.
Two things are worth holding apart from the framing. The report contains no funding figures, percentages or unit-shipment numbers at all — the statistics circulating around it in Chinese coverage belong to IT Juzi, Morgan Stanley and CCID Consulting, not to Huawei, so treat them as context rather than evidence of anything Huawei measured. And the document is Chinese-only: there is no English edition, which makes it a domestic standards instrument before it is a global one. Huawei's manufacturing chief put the commercial clock at three to five years, comparing physical AI to autonomous driving in 2017 and 2018, with lighthouse projects in automotive, emergency response and logistics first. We have tracked Huawei's habit of announcing the silicon before the ecosystem is ready — Huawei names its 2027 Nvidia challengers: the 960DT and Ascend 960PR.
Google has started testing a Buy button inside Gemini and AI Mode in India that sends shoppers to a Flipkart-branded checkout without leaving the AI interface, according to TechCrunch — and Google's own marketing blog confirms Flipkart as an agentic-commerce partner but not the test itself. Google's India marketing team published a page this month stating that merchants "like Flipkart are already partnering with Google to offer this agentic experience to Indian shoppers," built on the Universal Commerce Protocol, the open standard Google introduced in January to let agents reach a merchant's inventory and complete checkout. The reported test covers a small selection of products — smartphones, electronics and accessories — for some users, with a broader rollout targeted at later in October, ahead of India's festive shopping season. Google invested roughly $350 million in Flipkart in 2024 and holds a minority stake; a Google spokesperson gave only a generic statement about testing new features, and Flipkart did not respond. Amazon listings surfaced in the same AI interfaces without a purchase option, per the same report. The honest framing is a two-tier story: the partnership is on the record at Google, the Buy button is not, and neither company will confirm the mechanics.
What to watch: whether Huawei's maturity table moves off level 2 by the 2028 checkpoint, and whether India's festive quarter becomes the first real test of agentic checkout at scale.
Would you trust an AI assistant to press Buy with your saved card — and whose fault is it when it picks wrong? Tell us in the comments.
Sources: Physical Intelligence Cloud-Edge-Device Collaboration Architecture report (Huawei, PDF) · ITHome · CCIDnet · Zhidx · TechCrunch · Think with Google
