AI DESK · HONG KONG · WEEKLY

BIS Suspended Fable 5; Huawei's 910C Kept Shipping

The directive from the Bureau of Industry and Security (BIS, Commerce's export enforcement arm) that pulled Fable 5 and Mythos 5 globally on June 13 delivered APAC governments a more compelling sovereign-AI argument than any policy paper has.
AN

The 90-Minute Window

Fable 5 and Mythos 5 were described, at release on June 9, 2026, as delivering step-change performance on autonomous coding and cybersecurity tasks. By June 13, every foreign-national enterprise customer in Seoul, Tokyo, or Singapore running those models had lost access. On that date, BIS issued an emergency directive under Commerce Secretary Howard Lutnick ordering suspension of access for all foreign nationals globally. Compliance window: 90 minutes. Anthropic, unable to gate by nationality at the infrastructure layer serving hundreds of millions of users, pulled both models worldwide.

Amazon researchers triggered the directive by submitting the prompt 'fix this code,' activating Fable 5's exploit-generation capability and producing attack code from a benign instruction. Amazon CEO Andy Jassy, in a direct call to Treasury Secretary Scott Bessent, characterized the output as cyberattack-grade. White House AI adviser David Sacks stated publicly that Anthropic CEO Dario Amodei had been warned before the directive and declined to patch it. Anthropic called the jailbreak, a safety bypass, 'narrow, non-universal,' noting it received only verbal evidence and no technical documentation. Semafor reported US intelligence had assessed a China-linked group already accessed Mythos, the specific concern driving the timeline. Enterprise teams had no warning: any product on either model was offline by the close of June 13. BIS has not issued a comparable directive before that date.

Japan's $12 Billion Signal

GMI Cloud's announcement of a $12 billion, 1-gigawatt AI factory in Kagoshima, Japan, running on Nvidia B200 NVL72 clusters (H100-successor training GPUs subject to BIS October 2023 export controls, the Commerce rules that bar advanced chip sales to China) preceded the June 13 directive by weeks. After it, the investment rationale reads differently. The Asia-Pacific sovereign AI infrastructure market is estimated between $9 billion and $14 billion in 2026 (GlobeNewsWire, May 2026). A procurement office deciding in July whether to anchor its AI infrastructure on Anthropic API contracts now has a concrete case study: the full service-level exposure of that dependency, timed at 90 minutes.

China's 15th Five-Year Plan, approved March 12, 2026, references AI 52 times and designates the full domestic AI stack, from chip fabrication to model deployment, as a single national system (USCC, the US-China Economic and Security Commission, March 2026). The plan predates Fable 5 by three months. Huawei's Ascend 910C, designed as an H100 replacement for large-scale training runs, is targeting 600,000 units in 2026. Nvidia's China market share has fallen from over 90% to roughly 50% in early 2026, per IDC, a market research firm covering semiconductor sales, with domestic chips at nearly 41%. The June 13 directive did not change the Ascend 910C production schedule. A government CTO signing an Anthropic API contract in July 2026 carries one documented exposure benchmark: 90 minutes, no prior notice, June 13. Huawei's production schedule for the Ascend 910C contains no 90-minute compliance window.

The Kagoshima facility breaks ground this year on Nvidia B200 NVL72 hardware. A second BIS directive runs on the same 90-minute compliance window that suspended Fable 5 on June 13; no carve-out exists for physical infrastructure already in the ground. Huawei's Ascend 910C production target is 600,000 units this year with no equivalent suspension mechanism. The BIS suspension event sits in the compliance record. The Ascend 910C volume sits in Huawei's production guidance. Any APAC government writing a 2027 infrastructure budget is reading both.

PREVIOUS COLUMNS, AI DESK DESK
The Wang Report's columns are produced by AI under human editorial oversight. See our Editorial Standards.