ASC2026 finals reveal: AI is lacking more than just compute power

Compute is booming. Who will drive it?

The ASC World Student Supercomputer Competition (ASC世界大学生超级计算机竞赛, ASC26) concluded at Wuxi College (无锡学院) last week, and the headline is stark: China’s raw compute capacity is racing ahead, but skilled engineers are not. Is compute alone enough to unlock science-grade AI and supercomputing? The finals — a seven‑problem gauntlet drawn from real Gordon Bell-winning applications, quantum toolkits and world‑model inference challenges — made the answer painfully clear: no.

Exponential capacity, acute human shortage

It has been reported that IDC data shows China’s intelligent computing capacity could reach roughly 1,460.3 EFLOPS by 2026, about double 2024 levels, and that China’s AI compute market is on a steep growth trajectory. Huatai Securities (华泰证券) reportedly calls 2026 “the inaugural year of domestic super‑nodes,” and projects large market upside through 2028. Yet the labour market tells a different story: spring recruitment figures showed supply‑to‑demand ratios for high‑performance computing engineers as low as 0.15 — one qualified candidate chased by seven employers — with similar scarcity for core AI algorithm roles. Reportedly, top AI roles at some firms command monthly pay in the tens of thousands of yuan.

The contest: a microcosm of human‑AI limits

ASC26’s problems deliberately blurred the line between traditional HPC and AI for science (AI4S) — from HPL/HPCG benchmarks to UnifoLM‑WMA‑0 inference, AMSS‑NCKU relativity simulation, QiboTN quantum simulation and a “mystery” LeWorldModel task reportedly from a Turing Award‑winning group. Teams reported heavy reliance on AI agents for routine code edits and IO scheduling — one Tsinghua University (清华大学) squad said its custom “MIA” agent handled roughly 80% of repetitive tasks — but the agents also introduced subtle precision errors that required hours of human debugging. Peking University (北京大学) contestants similarly found that an unforeseen change in random‑number logic broke AI‑led optimizations and only human algorithmic insight saved the run. The message: AI accelerates work, but cannot yet replace human judgment on correctness and scientific fidelity.

Talent, policy and the path forward

ASC26 exposed a structural bottleneck for China’s supercomputing strategy: hardware self‑reliance and cluster scale are necessary but not sufficient. It has been reported that China’s recent push for domestically controllable chips — partly driven by export controls and trade policy in the West — has accelerated infrastructure build‑out, but the sector now urgently needs cross‑disciplinary engineers who understand parallel computing, large‑model inference and agent frameworks, as well as domain science. The finals closed, but the real contest — scaling a workforce to match an industrial‑scale compute surge while keeping humans in the critical loop — has only just begun.