AI×HPC Weekly Watch (2026-10-16): 10K-P Domestic AI Compute Goes Live, the Supernode Density Race, and AI4S Enters Its Verification Era
This is the AI-HPC expert team's weekly industry watch, based on public reports from October 10–16, 2026. Core judgment: this week the AI×HPC narrative shifted from "building machines" to "running systems" — on the compute side, the competition is about supernode density and the system-level economics of "compute–power coordination"; on the research side, AI for Science is entering a "large-scale verification" phase, from a domestic foundation model proving a world-class math conjecture to platform vendors opening their most powerful models to scientists.
The week of October 10–16 was event-dense: on October 12 the 26th China International Industry Fair (CIIF) opened in Shanghai, with domestic GPU vendors showcasing supernodes; on October 14 the first two compute satellites of the Pazhou Space Computing Constellation reached orbit from Jiuquan, and Google announced in Dublin that it would open its most powerful AI models to 200 European scientists; on October 15 NVIDIA's Vera Rubin MGX AI Factory ecosystem partner list surfaced; on October 16 Pengcheng Cloudbrain III declared 10,000 P of domestic AI compute in trial operation — the same day Chinese scientists announced the first world-class math conjecture cracked by a domestic foundation model. The timeline is below; the body follows four threads: compute infrastructure, hardware race, AI4S, and new compute scenarios.
| Time | Event |
|---|---|
| 2026-10-10 | Pujiang Innovation Forum wraps up: AI for Science framed as the "fifth research paradigm"; first "Pujiang Science Questions" released |
| 2026-10-12 | CIIF opens; domestic GPU supernode "density race" on display (64/128/144-GPU rack designs) |
| 2026-10-13 | Anew Labs, ByteDance's spun-out AI4S drug-discovery unit, closes a $290M Series A at a $1.5B valuation |
| 2026-10-14 | First two Pazhou space compute satellites reach orbit from Jiuquan (Zhuque-2 improved rocket) |
| 2026-10-14 | Google opens AlphaFold 3 and other flagship models to 200 European scientists via the EU RAISE program, plus $3M in cloud credits |
| 2026-10-15 | NVIDIA Vera Rubin MGX AI Factory ecosystem partner list surfaces (Foxconn, Midea, Luxshare, etc.) |
| 2026-10-16 | Pengcheng Cloudbrain III puts 10,000 P of domestic AI compute into trial operation; the 3D Ghys conjecture is cracked by a domestic foundation model and formally verified |
I. Compute Infrastructure: Pengcheng Cloudbrain III Goes Live at 10K P; "Compute–Power Coordination" Becomes a New Throughline
On October 16, the 2026 World Computing Network Conference (WCNC 2026) was held in Nanshan, Shenzhen, where the latest progress of Pengcheng Cloudbrain III was announced: 10,000 P of domestic AI compute formally entered trial operation, marking the system's move into a "build while operating" phase. Cloudbrain III uses a first-of-its-kind domestic AI supernode architecture built on Ascend's latest 384-supernode configuration; its first-phase equipment was powered on in March this year, and full build-out will form a 16,000 P (16 EFLOPS) cluster. For reference, Shenzhen's built and under-construction intelligent compute reached 54.3 EFLOPS by early this year, and the city's three-year action plan (issued in September) targets a core industry scale above RMB 300 billion and an agent adoption rate above 90% ["https://news.qq.com/rain/a/20261016A0CY2Y00"].
Two signals from the conference are worth noting. First, the "computing network" is being elevated into a national-infrastructure narrative: Pengcheng Laboratory's chief scientist Gao Wen said the China Computing Network (C²NET) pilot has run for four years and now targets the "ZPG goal" — Z-scale compute, P-scale communication, and GW-scale green power — while listing five remaining challenges: compute supply, network interconnection, task scheduling, data security, and compute–power coordination. Second, compute–power coordination is being elevated to white-paper status: the conference released the World Computing Network Development White Paper and the Compute–Power Coordination Technology Development White Paper, the latter arguing that compute, as a flexible new load type, can help absorb renewables — while still facing time-space supply/demand mismatches and insufficient load-regulation capability; the roadmap plans system building in 2026–2030 and coordinated global optimization in 2031–2035 ["https://news.qq.com/rain/a/20261016A0CY2Y00"]. Lu Yutong, director of the National Supercomputing Center in Shenzhen, made the same point in one direction: competition has moved from "single-system efficiency" to "large-computing-system efficiency," and a computing network matters less for wiring supercomputers and data together than for building the software ecosystem and service platform on top of it.
II. Hardware Race: Single-Rack Supernode Density Reaches 144 GPUs; Vera Rubin Ecosystem Surfaces
At this week's CIIF (October 12–16, Shanghai), domestic GPU vendors stopped racing on single-chip specs and instead pitched "a bigger computer": Moore Threads-style 64-GPU full-interconnect racks (MetaX S600), Enflame's ESL64-O, and — the headline — Tianshu Zhixin's Tianqian 300 IXL144, a single rack integrating 144 GPUs and 18 CPUs with ~10.1 TB of HBM, currently the industry's highest-density single-rack GPU supernode, with 100-nanosecond-scale single-hop latency and up to 15% effective-bandwidth gains in MoE All-to-All communication. The 128-GPU rack had been the representative high-density design among domestic vendors (Moore Threads, Biren, and Alibaba Cloud all have such products); the 144-GPU design pushes the density race one notch further ["https://36kr.com/p/4028460994777479"] ["http://m.toutiao.com/group/7696064150086910499/"].
The substance of the supernode race is "system-level economic efficiency." Tianshu VP Guo Wei argues that customers must clear three ledgers: migration cost (whether existing models and code need rewriting), facility cost (whether racks require upgraded power, load-bearing, and cooling or even new data centers), and operation cost (whether per-million-Token cost falls). IDC estimates that agent workloads push 2026 China intelligent compute to 2,576.5 EFLOPS, up 87.9% year on year, and supernodes are moving from technical validation to volume procurement ["http://m.toutiao.com/group/7696045064212202035/"].
On the overseas side, NVIDIA's Vera Rubin MGX AI Factory ecosystem list surfaced on October 15, with TSMC, Foxconn, Midea, Luxshare Precision, Sungrow, Lingyi iTech, and Megmeet among the partners spanning server OEM/ODM, liquid cooling, power, power distribution, and data-center infrastructure. Vera Rubin adopts a 100% liquid-cooled architecture, and the Vera Rubin NVL72 delivers up to 10× the AI-inference performance-per-watt of Blackwell, with mass production planned for autumn 2026; industry observers note that liquid cooling is becoming the standard rather than an option for high-end AI compute hardware ["http://m.toutiao.com/group/7696885566134157843/"].
III. AI4S: A Domestic Foundation Model Cracks a World-Class Math Conjecture; Platforms Accelerate "Opening the Strongest Models"
On October 16, Xu Disheng, a professor at the Great Bay University, announced that his team, using the Shanghai AI Laboratory's "Shusheng" (Intern-S2) scientific multimodal foundation model along a "general + specialized integration" path with deep human–machine collaboration, cracked the 3D Ghys conjecture, open for over 30 years — a recognized benchmark of mathematical reasoning in complex dynamics; professors Dong Bin and Liu Ruochuan's team at Peking University completed the formal verification confirming the result. This is the first time a domestic foundation model has cracked a world-class math problem: Intern-S2 ran seven batches of successive reasoning, expanding the problem nodes from 15 to 389, producing a thousand pages of rigorous derivation and dozens of novel proof strategies; the paper is on arXiv and the formal proof code is public on GitHub. The team plans to package the loop into a reusable math-verification module of the "Shusheng·Duanyan" science discovery platform ["http://m.toutiao.com/group/7697231169699578377/"].
The same week, platforms moved visibly toward "opening the strongest models to scientists." At an AI summit in Dublin on October 14, Google announced it would give 200 European scientists access, through the EU's RAISE program (Resource for AI Science in Europe), to DeepMind's flagship tools — AlphaFold 3, AlphaGenome, WeatherNext, AlphaEvolve, and AlphaEarth Foundations — plus $3 million in Google Cloud credits; awardees are selected by the European Commission from RAISE beneficiaries ["https://www.infobae.com/tecno/2026/10/14/google-da-acceso-a-sus-modelos-de-ia-mas-potentes-a-cientificos-europeos-para-impulsar-nuevas-investigaciones/"]. Combined with the Pujiang Innovation Forum (October 10) naming AI for Science the "fifth research paradigm" and releasing the inaugural ten "Pujiang Science Questions" ["http://www.xinhuanet.com/liangzi/20261010/be01c3243b374e2ebb0376af1cc60e8a/c.html"] — "open models + compute supply + verification systems" is crystallizing as the main battlefield of AI4S.
IV. New Compute Scenarios: Space Compute Satellites in Orbit; ByteDance's AI4S Unit Raises Independently
On October 14 at 19:40, the first two compute satellites of the Pazhou Space Computing Constellation were launched into orbit from Jiuquan aboard a Zhuque-2 improved carrier rocket, moving Haizhu's "compute in orbit, decide in orbit" (tiānshù tiānsuàn) route from a competition brief to orbital capability. The constellation plans 24 satellites to support stable 24/7 on-orbit operation, upgrading the "space-based world model" from short-term on-orbit validation to long-term on-orbit service; Haizhu plans to commercialize 24+ satellites within three years, building China's largest space-compute cluster and a ten-billion-yuan industry ["http://m.toutiao.com/group/7696537847066542632/"].
On the capital side, Anew Labs — the AI4S drug-discovery unit spun out of ByteDance — disclosed this week a $290M Series A at a $1.5B post-money valuation led by HSG, IDG Capital, and Hillhouse, the largest single-round AI-pharma financing in China this year; its core pipeline is a small-molecule inhibitor of the IL-17 family's AA/AF/FF dimers, built on Protenix-v2, an Apache-2.0 open-source biomolecular structure prediction model (464M parameters, matching or surpassing AlphaFold 3 on several benchmarks) released in April. TikTok CEO Shou Zi Chew said publicly in Singapore on October 12 that within two to three years, real scientific breakthroughs in physics, biology, chemistry, and materials will arrive thanks to AI ["https://36kr.com/p/4022968162195079"].
Takeaways
- Procurement decisions should read "the system ledger," not single-chip specs. Domestic supernodes pushed single-rack density from 64/128 to 144 GPUs this week, but vendors are already framing the discussion as three ledgers — migration, facility, and operation. Research groups and engineering teams should compare rack density, ecosystem compatibility, and per-million-Token cost on the same sheet.
- AI4S is moving from "demonstration" to "verification." A domestic model proving the Ghys conjecture with independent formal verification, and Google opening AlphaFold 3 to scientists, together suggest that "AI output + independent formal verification" is becoming the standard loop for scientific-AI results — research teams should build verification into their workflows early.
- The compute frontier is expanding: orbit and the grid both count. Space compute satellites and the compute–power coordination white paper landed in the same week, a reminder that next-generation compute planning must factor in space-borne capacity and power-grid coordination — consistent with this site's earlier "converged HPC×AI at scale" thesis.