China's centralized AI facility airgapped
Tracker summary: China has aggressively hardened security by airgapping (closing external connections) and siloing internally.
At a glance
- Assessment: Not Yet Testable
- Confidence in assessment: 30%
- Predicted timing: Late 2026–Early 2027
- Primary source: February 2027: China Steals Agent-2, archived PDF p. 10; February 2027: China Steals Agent-2, archived PDF p. 11
Timing split: The mid-2026 creation of a centralized zone at Tianwan is an elapsed prerequisite and remains unverified. The 40% compute concentration and external-network disconnection are separate February 2027 components. Not Yet Testable (30%) reflects limited site and network visibility, not proof that the elapsed prerequisite failed.
What AI 2027 Predicted
Tracker summary: The scenario creates a Centralized Development Zone (CDZ) at the Tianwan Power Plant in mid-2026. By February 2027, the CDZ holds 40% of China’s AI-relevant compute, with external connections closed and internal siloing. The scenario describes security as “aggressively hardened” February 2027: China Steals Agent-2, archived PDF, p. 11. These are scenario events, not observed facts. Mid 2026: China Wakes Up, archived PDF, p. 7; February 2027: China Steals Agent-2, pp. 10-11.
Scenario source locators: February 2027: China Steals Agent-2: archived PDF, p. 10; February 2027: China Steals Agent-2: archived PDF, p. 11.
How We Track This
Confirm only with dated, credible evidence that a designated Chinese frontier-AI site has closed its external network connections. Qualifying evidence includes an attributable technical disclosure, documented network architecture, or independently corroborated reporting specifically establishing disconnection. Imagery can identify and date a facility, but buildings, fences and physical security alone do not show an airgap; combine imagery with evidence about external network connections. Record the site, effective date, coverage and any exceptions. If those facts remain unavailable, the claim remains Not Yet Testable rather than being disproved by silence.
We monitor:
- Intelligence community reports and think tank analyses on Chinese AI infrastructure
- Open-source satellite imagery of large-scale Chinese datacenter/AI facility construction
- Chinese government policy announcements regarding AI security and centralization
- Reports from CSET, RAND, and similar organizations on Chinese AI infrastructure security practices
- Cybersecurity industry analysis of Chinese network architecture
Current Evidence
Component assessment: November 8, 2026
Retained Not Yet Testable (30%). The archived scenario places the Tianwan zone’s creation in mid-2026 and the 40% compute concentration and airgapping in February 2027. The creation window has passed without a qualifying site-specific record in this bounded review; no later evidence cited here establishes late creation. Strider and SCSP’s May 2025 infrastructure report and CSET’s April 2026 law translation establish context, not that facility or its network architecture. The February 2027 concentration and disconnection tests remain ahead. Missing visibility is not proof of nonexistence or failure. See China’s research centralization for the related implementation comparison; these are not additional independently fulfilled predictions.
The cited evidence does not establish airgapping at a designated Chinese frontier-AI facility. The late-2026-early-2027 tracking window is open, while the scenario’s specific event is dated February 2027; public visibility into network security remains limited.
China AI Infrastructure Surge: Strider and SCSP’s May 29, 2025 report announcement describes a state-directed AI infrastructure buildout. It concerns scale and coordination, not evidence of a specific site’s external-network disconnection.
Chinese AI Governance: The Cybersecurity Law was amended on October 28, 2025. CSET’s April 14, 2026 translation incorporates those amendments into the original law and links the underlying Chinese texts. Its discussion of AI security establishes governance context, not an airgap at a designated facility.
- Cybersecurity Law of the People’s Republic of China: CSET translation, April 14, 2026, with original-text provenance and translation notes.
Centralization Precedent: China’s broader pattern of centralizing strategic technology programs (from nuclear weapons to space) suggests airgapping a critical AI facility would be consistent with historical practice. However, the cited evidence does not identify the scenario’s Tianwan CDZ.
Counterevidence & Limitations
- This prediction is fundamentally difficult to verify from open sources. An airgapped facility’s existence might not be publicly known.
- As of November 7, 2026, the tracker’s late-2026-early-2027 window is open; the scenario places airgapping in February 2027. The specific scenario date is still ahead, but the tracking window is no longer far off.
- China’s AI ecosystem is currently more distributed than centralized — companies like DeepSeek, Baidu, Alibaba, and Tencent operate independently. The centralization depicted in AI 2027 would require a significant policy shift.
- Airgapping a facility running frontier AI research would create significant practical challenges for data access, model training on web-scale data, and researcher productivity.
- The bounded source review did not establish the Tianwan CDZ’s creation; this is an evidence limitation, not an exhaustive claim about all public reporting.
What Would Change Our Assessment
- Upgrade to “emerging”: Credible reports of China constructing a dedicated, highly secured centralized AI research facility; evidence of mandatory information-sharing mechanisms between Chinese AI companies
- Upgrade to “on-track”: Satellite imagery or intelligence reports confirming a CDZ-like facility with military-grade physical security; policy mandates requiring AI companies to consolidate research
- Remain Not Yet Testable: If no qualifying evidence establishes or rules out a designated site’s external-network disconnection; a distributed ecosystem alone does not rule out a secured site.
Update History
| Date | Update |
|---|---|
| 2026-11-09 | Assessment reviewed. Status: Not Yet Testable. Confidence: 30%. Retained Not Yet Testable (30%). Split the elapsed mid-2026 Tianwan creation prerequisite from February 2027 compute concentration and external-network disconnection. The bounded review does not verify creation or late fulfillment; opacity is not proof of failure. Preserved the March 13 publication date and linked the related centralization comparison. |
| 2026-11-07 | Assessment reviewed. Status: Not Yet Testable; confidence: 30%. CSET’s translation of China’s amended Cybersecurity Law establishes general network-security obligations and AI-security policy, not a designated frontier-AI site’s external-network disconnection. Public visibility into such a hardened, air-gapped cluster remains insufficient to establish or rule out the claim; the low confidence reflects that thin evidence, not a measured probability of nonexistence. Specified a dated, site-specific external-network-disconnection test; construction imagery and general security law do not establish an airgap. The tracking window is open, while the scenario’s airgap event is February 2027. Clarified the Tianwan/CDZ chronology and linked the dated CSET translation. A distributed ecosystem alone cannot rule out an isolated site. |
| 2026-03 | Prediction timeframe not yet reached. No public evidence of a CDZ-like centralized, airgapped AI facility in China. Consistent with broader centralization trends but specific implementation remains speculative. |