After threatening AI export bans, China allows DeepSeek to keep selling

China authorized DeepSeek and MiniMax to continue exporting AI technologies after adding 23 AI-related technologies to its restricted-export catalogue, while requiring licenses for covered exports. In Australia, Firmus canceled a planned IPO targeting about A$5 billion at a valuation above A$40 billion; NEXTDC cut revenue forecasts after a power-supply project was delayed and its shares fell almost 40%. Separately, South Korea plans to accelerate smart-glasses privacy rules, Grab proposed—but has not deployed—using aggregated shopping trends as a possible disease-surveillance signal, and Jio Platforms’ planned roughly $3 billion IPO was subscribed for 22 times the shares on offer, including 70 times by institutional investors.

Analysis

AI exports: state selection matters more than the headline. Beijing’s licensing regime makes overseas revenue conditional on continuing political approval; authorizing MiniMax (100) and DeepSeek suggests selective export enablement, not a general opening. That can help approved firms compete abroad while raising a policy-risk discount for other Chinese AI vendors. Open-source exemptions may also steer distribution toward open models, though commercial monetization remains unproven. For 100, authorization is an option, not evidence of material export revenue; verify licensing scope and overseas sales before underwriting upside.

Australian data centers: the binding constraint is deliverable power and consent, not announced demand. Higher connection costs and permitting delays favor operators with secured power and weaken speculative capacity plans. NEXTDC’s forecast cut is a company-specific execution warning; Firmus’s withdrawn IPO is a valuation signal, but neither alone establishes that hyperscaler demand has rolled over. Meta’s reported commitment to Firmus creates counterparty/project exposure, but contract terms and financial materiality are unknown.

Grab’s disease-surveillance proposal has little near-term earnings content: it is undeployed, requires approval, and the company itself frames it as supplemental. Deployment could create reputational and regulatory downside disproportionate to an unverified data monetization opportunity. Over the next 1–3 months, watch for actual public-health pilots, export-license details, and power/consultation milestones; over 6–18 months, local-power rules may structurally raise Australian build costs. Contrarian point: infrastructure delays can coexist with durable AI demand—the likely repricing is of delivery certainty and project returns, not necessarily compute demand.

Key Decisions for Investors

• No immediate trade in GRAB on this thought piece. Treat any government-approved pilot as a catalyst only after confirming data minimization, geographic scope, and commercial terms; a privacy controversy or use beyond aggregate trends would invalidate the limited-upside thesis.
• Keep 100 on an event-driven watchlist rather than buying the authorization headline. Reassess on evidence of licensed markets, export revenue, and continued approvals; a narrowed or revoked license would remove the prospective catalyst.
• Avoid treating Australian data-center exposure as a single AI-demand trade. Prefer operators/projects with documented power access and permits; for NEXTDC, require evidence that the supplier can resume work and that revised guidance stabilizes before considering a recovery position.
• Do not infer material META downside from its reported Firmus commitment alone. Verify contract duration, cancellation terms, and exposure relative to Meta’s investment plans; worsening Firmus execution without a material Meta commitment would weaken the short thesis.