AI Capital Flood Meets Chip Cold War at Beijing Summit

In a single week, Cerebras raised $5.55 billion in its IPO, Blackstone deployed over $4 billion across a REIT and a leveraged loan, Fractile closed a $220 million Series B, and Anthropic entered talks for a $30 billion raise. The aggregate capital mobilisation dwarfs anything the sector has previously seen in a comparable window. Yet the physical infrastructure these dollars are meant to fund faces grid interconnection queues of two to four years in the US, UK, and Europe, and hyperscale construction timelines of 18 to 36 months in best-case scenarios. Storage supply chains have already cracked — enterprise SSD prices in Japan have tripled — and optical interconnect components depend on III-V semiconductor fabs that are among the most supply-constrained in the industry.

Alternative asset managers have become the dominant capital formation channel, with Blackstone's dual public-and-private approach setting a template that PE peers will replicate. This shifts AI infrastructure financing from hyperscaler balance sheets to pension, insurance, and credit capital — sources with different return horizons and governance requirements. The risk is not a lack of money but a misalignment between financial commitments and physical deliverability, which historically resolves through either cost inflation as projects compete for constrained resources, or write-downs when timelines slip.

Two developments this week illuminate the same structural tension from different angles. Loongson's 3A6000 CPU — built on a proprietary instruction set with no Western IP — crossed one million units shipped, converting a proof-of-concept into a supply chain. Simultaneously, China's integrated circuit exports doubled year-on-year to $31 billion in April, demonstrating that global AI hardware demand is generating geopolitical leverage for Beijing even as controls restrict its access to leading-edge training chips. Jensen Huang's addition to the Trump delegation in Beijing — explicitly described as a bargaining chip on H200 export restrictions — illustrates how AI semiconductors have become a direct instrument of bilateral diplomacy rather than a downstream commercial matter.

The strategic calculus for Western policymakers is deteriorating. Controls are likely functioning as intended at the frontier: China still lacks leading-edge GPU equivalents for large-scale training. But the collateral effect is a parallel indigenous AI hardware and cloud stack — ByteDance's ArkClaw, Alibaba's cloud AI infrastructure, Loongson's CPU ecosystem — that will be commercially viable and exportable to Global South markets within a three-to-five year horizon. The question is no longer whether controls slowed China's frontier access; it is whether that delay was worth accelerating the creation of a structurally separate and exportable AI infrastructure.

Ramp's expense data showing Anthropic has overtaken OpenAI in business customer share is a concrete commercial signal that the model capability race is no longer the sole determinant of market position. Anthropic's simultaneous launch of a small business offering targeting 36 million US SMBs indicates a deliberate move to consolidate enterprise credibility before broadening — a classic crossing-the-chasm sequence. OpenAI's response is a TPG-backed consulting joint venture designed to accelerate enterprise deployment through services, while Microsoft's court-acknowledged dependency fears have triggered an active acquisition scouting posture to diversify its AI model exposure.

The bifurcation between enterprises that have crossed from pilot to production — legal tech, developer tooling, infrastructure — and the majority still in evaluation mode is where the next wave of commercial revenue will be unlocked. The constraint is not model capability but change management, integration cost, and ROI visibility. This explains why PE-backed services GTM models are attracting capital at scale, and why valuation multiples at the foundation model layer — Anthropic at a potential $900 billion on a few billion in ARR — are decoupling from conventional software benchmarks in ways that make them vulnerable to any demand deceleration signal.