|
|
|
|
Bottom Line Up Front Infinigence published its PDD architecture for serving one AI model from separate Chinese data centers linked over ordinary wide-area Ethernet, cutting 99th-percentile latency to a reply's first word by 51.5% and improving price-performance by 37.5%. Moonshot scaled Kimi K3 on efficiency, reaching close to 2.5 times the training efficiency under fixed compute on an estimated thousands of Nvidia Hopper-equivalent cards rather than large fleets of GB200 NVL72 racks, per a Wallstreetcn breakdown. An unverified screenshot points to a closed beta of DeepSeek Harness, which would be DeepSeek's first in-house agent product. |
| Infinigence publishes a method for pooling separate Chinese data centers to serve one AI model Infinigence AI, a Chinese AI infrastructure company, posted the full technical report for an architecture it calls PDD to a public code repository. PDD serves one large model from several geographically separate data centers linked over ordinary wide-area Ethernet, the commodity networking used between sites. It splits the serving pipeline into three stages, prefill, relay decode and main decode, so mismatched chips can each take the stage they run better. Against an eight-GPU flagship baseline, one unnamed vendor's accelerator reached about 81% of baseline on prefill, the compute-bound stage that reads the user's prompt. On decode, the memory-bound stage that writes the reply one word at a time, the same chip hit about 210%. Infinigence first showed the design at the World Artificial Intelligence Conference. Its own measurements put the result at 51.5% lower P99 latency before a reply's first word and a 37.5% better price-performance ratio. |
Moonshot carried Kimi K3 to 2.78 trillion parameters through efficiency changes, not added chipsKimi K3's training efficiency gain is close to 2.5 times under fixed compute, per a Wallstreetcn technical breakdown. CTD covered the same report's withheld training data on July 29th. K3 carries 2.78 trillion total parameters, the values a model learns during training, but activates only about 104.2 billion of them per token. It uses a mixture-of-experts design with 896 specialized sub-networks, roughly 16 of which run for any given token. A redesigned attention mechanism compresses long-context compute cost, and expert parallelism holds down communication load. Full attention across 93 layers and 1 million tokens of context would have been computationally out of reach, so K3 uses linear attention, whose cost grows closer to proportionally with input length. The estimated resource envelope is thousands of Nvidia Hopper-equivalent cards, without large fleets of Nvidia's tightly coupled GB200 NVL72 racks. Moonshot has not disclosed the accelerator count, chip model or GPU hours behind K3's main pretraining run. |
| Unverified screenshot points to a closed beta of DeepSeek's agent software A screenshot circulating in Chinese AI communities describes an internal recruitment notice for a closed beta of DeepSeek Harness. 36Kr, which reported the screenshot, said it could not verify its authenticity. A harness is the software layer around a model that turns it into a working agent, handling tool calls, session state and multi-step tasks. DeepSeek formed the Harness team in March 2026 under Cui Tianyi. The notice as described says selected users must submit personal details and sign a confidentiality undertaking, and that a leak would cost them future access to DeepSeek betas and partnerships. DeepSeek's job postings for the team reference long-context cache pruning and compression, chained tool use with automatic retry and task-planning-graph generation. Cui has said Harness will be released at the same time as the full version of DeepSeek's V4. |
|
| Chinese court penalizes a patent claimant for suing robot maker Unitree during its listing preparation A Chinese court ordered a patent claimant to pay Unitree 80,000 yuan ($11,000) in legal expenses, finding that its suit over the A2 four-legged robot was malicious litigation. The claimant, whose registered main business was general merchandise, acquired the patent on June 25, 2025 and filed the first of its suits against Unitree five days later, per state broadcaster CCTV. The Supreme People's Court, China's top court, affirmed on Feb. 3, 2026 that a parallel suit over Unitree's Go2 robot dog failed because the product lacked features the patent claimed. In that Go2 suit, CCTV said the claimant's damages demand could push Unitree's exposure above 70 million yuan ($9.9 million) while keeping the plaintiff's filing fees low. The A2 suit was filed during Unitree's listing tutoring period, the regulator-supervised phase before a Chinese initial public offering. The court wrote that the conduct amounted to "enforcing rights in name while infringing in substance." |
|
| III | AI & FOUNDATION MODELS | |
| ByteDance dissolves its standalone workplace app team into its AI units ByteDance said on July 30 that it is splitting the team behind Feishu, its workplace collaboration app known outside China as Lark, into two parts. One part joins the team behind Doubao, ByteDance's AI assistant. The other joins Volcano Engine, the company's cloud and model services arm. Feishu's product team merges into a new Doubao product team led by Doubao head Zhao Qi, and Feishu head Xie Xin now reports to him. ByteDance said its large language model business reached a $4 billion annualized revenue run rate measured by average July consumption, a projection of current usage-based revenue over a full year. The company said that figure exceeds the combined run rate of every other Chinese model developer, a comparison based on its own numbers. |
| Tencent open-sources an inference toolchain it reports speeds AI replies up to 2.4 times Tencent's Hunyuan team open-sourced AngelSpec, a framework for speculative decoding, on July 30. Speculative decoding runs a small, fast model ahead of the main one, drafting several tokens that the large model checks in a single pass and keeps where they match its own output. Across six benchmarks and concurrency levels from 4 to 64, Tencent reports average speedups of 1.98 to 2.40 times over a conventional one-token-at-a-time baseline. Average accepted length was 4.79 tokens, meaning close to five drafted tokens survived verification per round. Tencent also released the drafter weights and training code for its Hy3-A21B model, so the reported speedups can be tested outside the company. |
|
| AI memory demand quadruples Shenzhen DDR5 prices and freezes graphics card shipments Dealers in Shenzhen's Huaqiangbei electronics market said the factories of every major graphics card brand have frozen inventory and stopped bulk outbound shipments, per state broadcaster CCTV Finance. They tie the freeze to a new round of Nvidia price increase notices, and report increases of about 30% over the two days after the notices went out. Nvidia's RTX 5080 gaming card sold for about 8,000 yuan ($1,100) in the market on July 22 and now goes for close to 12,000 yuan ($1,700). Measured against September 2025, 16GB DDR5 memory modules rose from 450 yuan ($63) to 1,850 yuan ($260), up 311%. |
| Chip design startup says AI tools let 20 people build a new processor in two months Sunzhan and a partner brought a new CPU to market within two months, founder Zeng Yi said. The company, founded in 2022 with about 20 employees, sells customized processor cores built around a single workload rather than general purpose computing. Zeng credits the pace to using AI inside the company's design engine to generate candidate designs quickly. He said architectural decision rights have moved from chip companies to AI model developers, because different models need different custom instructions. On closed architectures such as Arm and Intel's x86, Zeng said, those custom instructions either do not exist or run inefficiently. |
| Chipmaker NOVOSENSE launches domestic angle sensing chips for robot joints and industrial motors NOVOSENSE launched its NSM350x series of magnetic encoder chips on July 30. A magnetic encoder reads the angle of a rotating shaft by sensing a magnet attached to it, which is how a robot joint or industrial motor tracks its position. The company said the series is China's first domestically developed encoder of its kind. It uses a dual-track vernier design built on the Hall effect, the voltage shift a conductor shows in a magnetic field. Two rings of magnetic poles are read together, letting the chip report an absolute angle within a single turn rather than counting steps from a reference point. NOVOSENSE said the NSM3052 series is open for sample requests. |
|
| V | ROBOTICS & AUTONOMOUS SYSTEMS | |
| State power group SPIC puts a driverless 100-ton electric mining truck into commercial service A 100-ton-class driverless electric mining dump truck entered service on July 30 at the South Open-Pit Mine in Holingol, Inner Mongolia, per state broadcaster CCTV News. The truck runs a closed loop of automatic loading, self-driving and dumping with no operator aboard. It pairs a combined charge and swap system with semi-solid-state batteries, which use a gel-like electrolyte rather than a fully liquid one. SPIC said that combination addresses the long recharge times that have limited electric heavy trucks. Yu Haili, director of SPIC's Inner Mongolia South Open-Pit Mine, said energy cost per unit of output runs 65% below a comparable diesel truck and continuous single-vehicle range is at least six hours. China has fielded these trucks in batches, including 100 driverless electric mining trucks XCMG delivered to the Huaneng Yimin open-pit coal mine in May 2025. |
|
|
|