Scientists “Created” a New Universe — and Revealed the Real Bottleneck of the AI Era

Storage, not raw compute, is now the choke point

China’s landmark cosmological simulation Qianyan (千衍) did more than reproduce billions of years of cosmic evolution; it exposed a systemic fault line in modern high‑performance computing and AI: storage. Led by researcher Wang Qiao, the team used Sugon (中科曙光) infrastructure to track 4.2 trillion dark‑matter particles across a 12‑billion‑light‑year volume — it has been reported that the run produced roughly 13 PB of data. The surprising lesson? CPU and GPU flops are plentiful, but moving and holding data at the required scale is what actually breaks workflows.

A practical test of architecture and the rise of “compute‑storage” co‑design

Qianyan’s early attempts on other international supercomputers stalled not because of insufficient compute, but because of architecture mismatch: many systems favor many cores but offer limited memory per node. Wang says the team spent years debugging and rewriting code until they moved to a large‑memory architecture at the Chinese Academy of Sciences (中国科学院) network center’s Dongfang supercomputer and Sugon systems, which could sustain the continuous, high‑throughput writes and reads the project required. The experience mirrors a wider shift — when GPUs and accelerators race ahead, storage latency, bandwidth and local memory become the gating factors for scientific and AI workloads.

Industry implications and geopolitical context

This is no isolated anecdote. It has been reported that TrendForce data showed steep DRAM price rises in 2025 and that a growing share of global memory production is earmarked for data centers — trends driven by AI training and KV‑cache demands that can reach terabytes or petabytes per day. Vendors and hyperscalers are increasingly pursuing “storage‑as‑compute” and storage‑aware architectures that offload KV caches from GPU memory to ultra‑fast flash and that expose native KV semantics for efficiency. Geopolitics matters too: export controls and chip supply frictions have accelerated China’s push for domestic storage stack autonomy, making “domestication” and vertical integration strategic priorities for both research and industry.

From a scientific demo to a strategic front

If Qianyan is an extreme stress test, AI is the broad pressure test: both point to the same verdict — storage is no longer logistics, it is strategy. Sugon senior VP Li Bin (李斌) has argued that AI will reshape storage technology itself, not merely demand more capacity. Will future competition be decided by GPUs or by who can move and pre‑position data fastest? For the moment, the answer is clear: you can build the most powerful engine, but without a fuel system that can deliver at scale, it won’t go far.