Skip to content
SIGPULSE
China AI Watch Issue 4 4 min read raw .md ↗

Why Is China's Computing Power Collapsing Into Eight Hubs While Every Province Still Builds? The Take Inside China

Chinese original 2026-08-30 · 「AI机房遍地开花，算力为什么在向八个点塌缩」 · translated to English 2026-08-30

Two signals, opposite directions, same month: local governments keep signing AI-data-center deals (one toy factory crossed over into a ¥3.2B compute contract), while the National Data Administration’s numbers show built intelligent computing collapsing into eight national hub nodes. Our WeChat column’s diagnosis ran under “AI server rooms bloom everywhere — so why is computing power collapsing into eight points.” This entry translates it, with the official figures pinned to primary sources.

The numbers

  • Eight national hubs (ten clusters): 138.8万 PFlops built intelligent computing as of end-2025 — >80% of the national total (NDA chief Liu Liehong, official speech text)
  • National totals: 159万 PFlops end-2025 (People’s Daily: second globally) → 188万 PFLOPS FP16 end-March 2026 (Caijing) → 2,185 EFLOPS FP16 by end-June 2026, +177% YoY [unverified — July State Council briefing as relayed by financial media]
  • Utilization: national rack fill rate 71.4% [unverified — relayed]; regional split: East 55.9%, West 32.6%, Northeast 0.9% [unverified — same relay]
  • 70+ compute corridors built around hub nodes; inter-node network performance +10%; 80%+ green-power requirement for new hub facilities [as cited in the take]
  • Figure arbitration: the take presents 138.8万/>80% as a July disclosure; the primary source is Liu’s March 2026 forum speech, and the figure is an end-2025 snapshot. The direction (hubs >80%) persisted through at least Q1 2026 per Caijing. We keep the numbers, re-dated to their actual vintages.

The take inside China

The essay works as a systems diagnosis rather than a policy brief. Visible benefits first: a near-fourfold annual growth in intelligent compute that only concentrated deployment could have stacked that fast; a 71.4% rack-fill rate the take calls healthy for infrastructure; an emerging East-trains/West-supplies division of labor; and an energy dividend — green power gets built where compute clusters.

Then the second page of the ledger, three hidden costs. The regional divide gets redrawn: when computing becomes a factor of production like electricity, a 0.9% regional share shapes which provinces catch AI industries — with the honest boundary that this bites training, not inference (inference stays near users; local clusters and edge persist). Stranded-project risk: non-hub centers are easy to build and hard to feed; the average utilization masks the two ends, and trophy cross-over deals are, in glut industries, usually bad news — with the carve-out for hubs and stable-local-demand projects (government, financial disaster recovery). Grid reconstruction: single-point compute concentration means single-point power demand; concentration yields efficiency only if energy infrastructure keeps up — hence, says the take, the compute-electricity coordination push is a pre-emptive patch.

Its structural verdict: consolidation is not an administrative decree but scale economics’ physical law plus misaligned local incentives directing the same play — the former guarantees concentration happens; the latter determines how much waste precedes it. The official layered framework is read not as fighting consolidation but installing brakes and a steering wheel on it.

Closing re-aiming for three audiences: non-hub cities should re-benchmark from “we have a machine room” to “our industries can afford compute” (applications over infrastructure patronage); SMEs are net winners (compute retailed like utilities; surrender the “self-hosted = secure” fixation, except for actually classified data); and for individuals the effect is slowest but deepest — the compute map quietly rewrites city-level industry, employment and migration stories, and the Northeast’s 0.9% is the next engineer-drift chart being drawn now.

What the Chinese take left out

The utilization and regional-split figures ride on financial-media relays of the July briefing rather than a primary transcript we could check — tagged accordingly. Also absent: which non-hub projects have actually failed (the piece argues risk forward from the toy-factory anecdote rather than a failure count), and the electricity side’s own numbers (would the green-power buildout arrive faster than the demand?).

Why it matters outside

China is running the world’s largest deliberate experiment in compute geography — where training concentrates into eight sanctioned hubs while inference distributes — and the official figures (80%+ share, ~1.9 million PFLOPS-scale totals growing near 3x YoY) are the citable baseline for anyone analyzing Chinese AI capacity, energy planning, or the regional economics of data centers. The take’s two-force frame (physical law + incentive misalignment) is also a portable lens for the West’s own AI-data-center sprawl debates.

Sources

Provenance & disclosure. Originally published in Chinese on our WeChat channel on 2026-08-30 (“AI机房遍地开花，算力为什么在向八个点塌缩”); drafted with AI assistance under human editorial direction. Translated to English on 2026-08-30 (AI-assisted, human-reviewed). The hub-share and national-total figures were re-dated against the NDA’s official speech text and Caijing/People’s Daily coverage — the take’s July attribution was corrected to the figures’ actual vintages; relay-only numbers carry [unverified] tags. This is translated commentary — not a SigPulse measurement. Our first-party measurements live in the dispatches and the /data/ ledger.

Cross-checked sources (machine-readable in the raw markdown)

FAQ — Direct Answers

How concentrated is China's intelligent computing?
Per NDA chief Liu Liehong's China Development Forum 2026 speech (official text online): by end-2025 China's total intelligent computing reached 159万 PFlops (1.59 million), of which the eight national hub nodes (with their ten clusters) held 138.8万 — over 80%. By end-March 2026 the national total had reached 188万 PFLOPS (FP16) with the hubs' share still above 80%, per Caijing. The Chinese take also cites a July 2026 State Council briefing figure of 2,185 EFLOPS FP16 nationally, +177% YoY [unverified — relayed via financial media, not checked against a primary transcript].
Why do local AI data centers keep getting announced anyway?
The take's two-force diagnosis. Physics: large-model training punishes fragmentation — thousands of GPUs spread across ten small rooms cannot match one ten-thousand-GPU cluster, so training compute concentrates the way blast furnaces grew. Incentives: for a local government, an AI data center is visible GDP and reportable achievement, while utilization is a hidden bill that settles years later — eighty cities each betting they are the exception makes a shakeout arithmetic, not a possibility.
What are the hidden costs of consolidation?
Three, per the take: a redrawn regional divide (the Northeast holds 0.9% of intelligent computing — bearable for inference, which stays near users, but decisive for which regions catch AI industries); stranded non-hub projects (training orders flow to the eight hubs; trophy projects like a toy factory's ¥3.2B compute deal read as a glut signal, with the carve-out that hub cities and projects with stable local demand aren't in the risk zone); and grid pressure (concentrated compute means concentrated power demand, which the 'compute-electricity coordination' push pre-patches).
What is the official posture toward this?
Not resistance but guardrails. The take reads the official point-chain-network-plane framework as an admission that free-growing construction ran off course — the layered structure (hubs absorb training, edge serves inference, the network connects them) gives the spontaneous consolidation brakes and a steering wheel.