AI×HPC Fusion Enters Harvest Season: 2026 Summer–Autumn Industry Roundup
This is the industry tracking roundup of the AI-HPC expert team, based on public reports from June to mid-October 2026. Core judgment: AI×HPC fusion has moved from "vision" to "delivery" — AI for Science is being elevated to national strategy by both China and the US, the compute foundation is shifting toward "converged HPC×AI at scale," and AI's role in science is changing from "assistant" to "the one who solves the problem."
Over the past four months, AI-HPC fusion has visibly accelerated: on June 23, China's supercomputer LingSheng topped the global TOP500 list with 2.19 EFlops sustained performance — the world's first supercomputer with sustained performance above 2 EFlops; in July, the all-domestic 100K-GPU AI supercluster Sugon 8000 (Dengfeng) was released and put into use; in late September, SemiAnalysis estimated that China's delivered AI datacenter capacity had surpassed 24 GW, second globally; in mid-October, NVIDIA's next-generation Vera Rubin NVL72 began global delivery, and Nature published two "no brute-force" AI research breakthroughs on the same day. The timeline of major events is below; the body follows three threads: policy, compute, and research.
| Time | Event |
|---|---|
| 2026-04-28 | CAS releases the "Panshi 100" science model system |
| 2026-06-23 | LingSheng tops TOP500 at 2.19 EFlops (67th list, ISC2026) |
| 2026-07-10 | Sugon 8000 (Dengfeng), the all-domestic 100K-GPU AI supercluster, goes live |
| 2026-07-22 | US DOE announces first Genesis Mission project selections |
| 2026-09-22 | Loongson releases the first software release of its accelerated computing platform for its in-house general-purpose GPU |
| 2026-09 end | SemiAnalysis: China's delivered AI datacenter capacity exceeds 24 GW, second globally |
| 2026-10-12 | Lenovo announces global delivery of NVIDIA Vera Rubin NVL72 at OCP 2026 |
| 2026-10-14 | Nature publishes two "no brute-force" AI research breakthroughs on the same day |
I. AI for Science Becomes a National Strategy
1.1 United States: Genesis Mission Enters the Project-Award Phase
China and the US turned "AI for science" into a national act within nearly the same window. On the US side, an executive order launched the Genesis Mission in November 2025, coordinated by the White House Office of Science and Technology Policy, aiming to build the "world's most powerful science platform" — integrating national-laboratory supercomputers, AI systems, experimental facilities and multi-domain datasets to double US scientific productivity and influence within a decade. On July 22, 2026, the US Department of Energy (DOE) announced the first project selections: 278 projects spanning energy, discovery science and national security, with 342 participating institutions including 16 national laboratories and 142 universities. The teams will develop and demonstrate AI-enabled scientific workflows on the Genesis Mission platform (agent frameworks, frontier models and national-laboratory HPC resources) ["https://www.energy.gov/articles/secretary-energy-chris-wright-announces-first-genesis-mission-projects-selected-accelerate"].
1.2 China: Science Foundation-Model Platformization and the WAIC Consensus
China's moves have been just as dense. On April 28, the Chinese Academy of Sciences (CAS) formally released the "Panshi 100" (ScienceOne 100) model system — built on scientific foundation models as the root and a backbone of domain-specific model clusters covering mathematics, physics, biology and more — seen as a milestone in research AI moving from fragmented exploration to platformized collaboration ["http://www.cas.cn/cm/202604/t20260429_5108338.shtml"]. During WAIC 2026 in July, "converged HPC×AI leading the research-paradigm transformation" became the theme of Shanghai's Science Intelligence segment, where Academician Zhang Tao presented the outlook for combining single-atom catalysis with AI ["http://sh.news.cn/20260719/8f4ae637575143289b892911a5752c2f/c.html"]. The "AI for Science Innovation Atlas 2026," released in March, shows global related paper output more than doubling over the past five years, with aerospace, quantum technology and materials science growing at over 30% annually ["http://www.cas.cn/cm/202608/t20260824_5118852.shtml"].
Judgment: AI for Science is no longer a concept confined to papers — both China and the US have moved it into national-level compute and research-resource allocation, and international competition around scientific foundation models, research data and compute platforms has entered a substantive phase.
II. The Compute Foundation: World Records, 100K-GPU Clusters, and New AI Factories
2.1 New No. 1 on the TOP500: LingSheng Tops the List at 2.19 EFlops
Supercomputer-ranking changes are the most direct signal. In the 67th TOP500 list released at ISC2026 in Hamburg, Germany, on June 23, LingSheng — an all-domestic supercomputer in Shenzhen's Guangming Science City — topped the list with 2.19 EFlops sustained double-precision performance, the first time a Chinese supercomputer has ranked first since Sunway TaihuLight in 2017. According to chief designer Lu Yutong, LingSheng uses an all-domestic software-hardware stack and an innovative large-scale converged HPC×AI, all-CPU mixed-precision architecture with an "Online Acceleration" structure, powered by the in-house LX2 CPU ["https://www.news.cn/tech/20260624/e4a7065291b64d2fa4f9861e9be4a70a/c.html"] ["https://www.news.cn/tech/20260627/1b36643871eb436a9988317d550f7e3d/c.html"]. The US El Capitan ranked second at 1.809 EFlops.
2.2 The 100K-GPU Era: Sugon 8000 (Dengfeng) Goes Live
Cluster scale and full domestic production are advancing in parallel. On July 10, at the Photon Ecology Organization 2026 Intelligent Computing Application Conference, Sugon announced that the Sugon 8000 (Dengfeng) — China's first all-domestic 100K-GPU AI supercluster — had officially gone live and been connected to the National Supercomputing Internet. Its in-house scaleFabric, an IB-like native RDMA high-speed network, supports highly reliable interconnect for 100K GPUs; ParaStor distributed storage ranked first in both the production full-node and 10-node performance categories of the 2026 global IO500 list. The system supports full-precision computing from FP64 to INT8, unifying supercomputing and AI-computing in a single system, and has completed optimization of more than 300 converged HPC×AI applications across more than 20 domains including large models, robotics and new materials ["https://www.sugon.com/cut?id=2927"] ["https://app.xinhuanet.com/news/article.html?articleId=2026071035cd0e84f4fc4ed28c5b916a6622c747"]. Sugon has also started development of a second all-domestic 100K-GPU system together with the Beijing Academy of Artificial Intelligence's Institute for Scientific Intelligence.
2.3 Global Compute Map Reorders: China at 24 GW, Second Globally
In late September, SemiAnalysis released its "China AI Infrastructure Boom" report. After interviewing 60+ companies and tracking more than 1,000 data centers, the team estimated that China's delivered AI datacenter capacity exceeds 24 GW, second globally (US: ~56 GW), with another ~20 GW under construction and ~30 GW planned — larger than Europe, the Middle East and Africa combined (~14 GW) and the rest of Asia (~15 GW), with ByteDance as the largest datacenter tenant ["https://www.icviews.cn/news/36271/7"] ["https://cj.sina.com.cn/articles/view/7879924057/1d5ae195902001nsl8"]. The report also highlighted a key shift: what constrains AI infrastructure is no longer chips but electricity.
2.4 Next-Gen Hardware Refresh: Vera Rubin NVL72 and Domestic Chips
At the OCP 2026 Global Summit on October 12, Lenovo announced it had begun global delivery of NVIDIA's next-generation Vera Rubin NVL72 rack-scale AI system (72 Rubin GPUs and 36 Vera CPUs with NVLink 6 interconnect), claiming up to 30× higher throughput per megawatt and up to 45× lower inference cost per million tokens versus the previous platform — delivering the equivalent AI compute with only a quarter of the GPUs ["https://lenovopress.lenovo.com/lp2549-lenovo-brings-the-full-nvidia-vera-rubin-platform-to-ai-cloud-providers"] ["https://www.nvidia.com/en-us/data-center/vera-rubin-nvl72/"]. Meanwhile, "compute-electricity synergy" became a keyword at the summit: NVIDIA, Google and Microsoft are driving 800 VDC within OCP as the open power-architecture standard for next-generation AI datacenters ["https://blogs.nvidia.com/blog/800-vdc-power-architecture-ai-factory/"], and vendors such as xFusion showcased 800V HVDC power systems and supernode components focused on power, cooling and interconnect root technologies ["https://www.xfusion.com/static/cn/news?list_per=30"]. Domestic chips advanced too: Hygon guided Q1–Q3 2026 revenue of RMB 14.28–15.58 billion, up 50%–64% year on year ["http://jjckb.xinhuanet.com/20261016/b50684ffc446421c846102637493bcf7/c.html"]; Loongson released in September the first software release of its accelerated computing platform supporting its in-house general-purpose GPU ["https://finance.sina.com.cn/tech/digi/2026-09-22/doc-inissitp2230220.shtml"].
Judgment: the compute race has moved from "individual chips" to a systemic contest of "cluster scale, power supply, and open ecosystem"; converged HPC×AI — unifying supercomputing's high-precision compute and AI-computing's large-scale training in one architecture — has become the technology of choice for leading players in both China and the US.
III. Scientific Breakthroughs: AI Starts Doing "Fine-Detail Work"
The October 14 issue of Nature published two AI studies that did not win by bigger models or more compute; instead, both applied AI to fine-grained steps long considered "untouchable" ["https://m.huxiu.com/article/4897690.html"].
3.1 Protein Design: The Baker Team's "Molecular Scissors"
The team of David Baker at the University of Washington Institute for Protein Design used an AI pipeline — "RFD2-MI/RFdiffusion2 generates backbones → ProteinMPNN designs sequences → AlphaFold 3 filters structures → wet-lab validation" — to design two classes of proteases from scratch. The metalloproteases selectively cleave three disease-related human proteins (TDP-43, amyloid-β, serum amyloid A), with catalytic rates up to 1 billion times faster than the non-catalytic background; 13 of 69 cysteine-protease designs were active (~19% success rate), six crystal structures deviated from design models by less than 1.2 Å, and the folds differ from any natural protease. The team also demonstrated "bio-orthogonal control": locking drug molecules inside caged cytokines and having a designed protease "snip" them open at a specified site.
3.2 Weather Forecasting: Microsoft Adds a "Correction Layer" to Subseasonal Forecasts
A team led by Microsoft Research and MIT proposed a probabilistic bias correction (PBC) framework: instead of rebuilding weather models, it learns and corrects the systematic biases of existing models on 2–6 week subseasonal timescales. Applied to ECMWF's AI subseasonal model, mean skill (RPSS) doubled; applied to physical models, 98% of grid points improved on precipitation forecasts; across 2,435 floods from 2016 to 2026, three-week-ahead forecast skill improved by 76%–92%. The whole system runs on CPUs, is retrained incrementally every day, is open-sourced under an MIT license, and took first place across all variables and all lead times in ECMWF's year-round live AI Weather Quest competition.
3.3 Other Protein Progress
Other protein advances accelerated in the same period: in July, the Duke University team published Raygun in Nature — an AI tool that, like a "translator," extracts patterns from massive data to extensively modify protein size and properties while preserving structure and function ["http://www.xinhuanet.com/liangzi/20260730/ce7e3b83349f41f7b9d4feb6c5c3f017/c.html"]; in late June, teams from the Technion and Tel Aviv University published BetaDescribe in PNAS, "translating" protein sequences into natural-language descriptions, opening a new path for drug development and new-material design ["http://www.cas.cn/kj/202607/t20260708_5114903.shtml"].
Judgment: the shared subtext of the two papers is that AI's sharpest use in science may not be "end-to-end replacement" but precisely targeting the most painful link in the system — Baker's team did not build a general protein foundation model; instead, they chained existing tools into a pipeline focused on peptide-bond hydrolysis. Microsoft's team openly conceded "you keep building the model; I'll just fix the bias." Understanding the problem is scarcer than stacking models.
IV. Three Consensuses Taking Shape
Put together, three consensuses are taking shape across the policy, compute and research threads.
First, from "stacking models" to "patching bottlenecks." Doubling subseasonal forecast skill came not from a bigger foundation model but from a lightweight correction layer that only learns "how the model errs"; decades-old bottlenecks in protease design were opened by an AI pipeline of "generate + design + verify." Methodology dividends are overtaking parameter dividends.
Second, the compute constraint is shifting from chips to electricity. AI factories are being built at megawatt and even gigawatt scale; SemiAnalysis says electricity — not compute — is the hard constraint on infrastructure, making "compute-electricity synergy," liquid cooling and 800V HVDC the keywords of OCP and other summits. Whoever solves the system-level problems of power and cluster engineering controls the initiative in the next round of scaling.
Third, converged HPC×AI and platformization are the shared technical route. LingSheng's Online Acceleration, Sugon 8000's native HPC×AI convergence, and the National Supercomputing Internet's connection into the national integrated compute network all point the same way: unifying supercomputing precision and AI-computing scale in one system. Supporting this is software-layer platformization — Genesis Mission, Panshi 100 and open-source toolchains are turning "continuously manufacturing reproducible software tools for real scientific problems" into a national-level act. This echoes the Scientific Software Factory vision that ai-hpc.org advocates: problem-driven, knowledge organized by problem domain, one problem one software — moving from initiative to industry mainstream.
V. Conclusions and Outlook
For the AI-HPC community, this summer–autumn news cycle offers several takeaways: reproducible, comparable software and toolchains are replacing "bare compute" as the new competitive focus; the boundary between supercomputing and AI-computing is rapidly disappearing, with single architectures and unified scheduling becoming the norm; and leaderboards, benchmarks and open communities (TOP500, IO500, OCP, AI Weather Quest) define "what real progress means" more than ever. As usual, the next TOP500 list will be released at SC26 in November — the next round of change on the converged HPC×AI track is worth following.
Sources: Xinhua/Xinhuanet (LingSheng tops TOP500, Sugon 8000, Raygun, WAIC 2026, Hygon) | US DOE (Genesis Mission first selections) | Chinese Academy of Sciences (Panshi 100, BetaDescribe, AI for Science Innovation Atlas 2026) | Sugon (Sugon 8000) | SemiAnalysis (via Sina Finance / ICViews) | Lenovo Press & NVIDIA (Vera Rubin NVL72) | xFusion (800V HVDC) | Sina Tech (Loongson) | Huxiu (Nature double feature) | Verified: 2026-10-16
Note: this roundup is compiled from public reporting; facts and figures follow the cited sources; opinions are this team's engineering-perspective analysis and do not constitute investment or technology-selection conclusions.