Debt-Fuelled AI Arms Race Triggers BIS Warning as HBM Becomes Chokepoint

Three separate developments this week mark a qualitative transition in AI infrastructure finance. SpaceX's $25 billion investment-grade bond — exceeding its initial target — sets a market precedent that credit rating agencies and bond investors now treat AI compute as a creditworthy asset class. ByteDance's reported $20 billion offshore loan and Oracle's debt-financed data centre buildout, funded by 21,000 layoffs, complete the picture: across US, Chinese, and legacy enterprise players, the dominant funding instrument has shifted from equity to leverage.

This shift has structural consequences that go beyond cost of capital. Leveraged AI infrastructure players are now exposed to interest rate cycles and credit market sentiment in ways that equity-funded hyperscalers are not. If enterprise AI adoption plateaus or revenue-per-dollar-of-compute declines faster than expected — precisely the scenario the BIS is warning about — debt-service obligations could force capacity pullbacks or asset sales among the most leveraged builders. The BIS warning and the debt-market surge are not separate stories; they are two sides of the same risk.

South Korea's $1.3 trillion commitment, Micron's bullish forward guidance catalysing broad Asian equity gains, and Baidu's Kunlunxin IPO bid all converge on the same underlying reality: HBM is the component most directly constraining AI accelerator throughput, and control of its supply chain is now a matter of industrial sovereignty. Seoul's state-orchestrated investment — with the president physically present alongside Samsung and SK Hynix leadership — mirrors the CHIPS Act playbook: government as convener and signal-amplifier for private capital defending a structural advantage before competitors close the gap.

The competitive landscape is shifting from a de facto duopoly to a three-player supply chain as Micron scales HBM production, with China's CXMT a longer-horizon challenger. For infrastructure planners, this means HBM availability and pricing are the most sensitive leading indicators for AI accelerator costs — more immediately actionable than TSMC capacity signals. The Korea-Micron-China triangle will determine whether HBM scarcity eases by the 2027 accelerator generation or persists as a durable constraint.

Google capping Meta's Gemini access due to compute constraints is not a bilateral commercial dispute — it is evidence that frontier model capacity is genuinely rationed even among the largest operators. Combined with US export restrictions on advanced AI models prompting Austria to lobby the EU to host Anthropic, and Nvidia using infrastructure partnerships to lock in Southeast Asian GPU demand before competing architectures arrive, the picture is of a global AI landscape fragmenting faster than any single governance framework can contain.

The Baidu Kunlunxin IPO bid is the most consequential expression of this fragmentation: a $50 billion public listing specifically designed to fund a domestic Nvidia alternative, using Hong Kong's international capital markets to finance capabilities that US export controls are intended to retard. If successful, it would provide China's AI infrastructure with the capital base to sustain genuine competition in the domestic market, accelerating the bifurcation of global AI into parallel ecosystems.