Compute & Infrastructure

Japan approved ¥631.5 billion ($4 billion) in new subsidies for Rapidus Corp., bringing total government support above $16 billion for the chipmaking startup targeting 2nm AI accelerator production by 2027. Bloomberg reports the funding will accelerate Rapidus into direct competition with TSMC and Samsung in advanced logic manufacturing — a market where entrenched advantages in process yield, packaging integration, and customer trust create formidable barriers. The initiative represents Tokyo's most aggressive industrial policy play since the 1980s memory chip campaigns, now driven by concerns that dependence on Taiwan Semiconductor Manufacturing Co. creates unacceptable strategic vulnerability for Japan's automotive and robotics sectors.

The economics remain punishing: Rapidus faces capital expenditure exceeding $50 billion through 2030 to reach competitive scale, while TSMC and Samsung already operate at volumes that spread tooling costs across hundreds of customers. Japan's approach mirrors the EU Chips Act and US CHIPS Act in prioritising sovereign capacity over commercial viability, but the timing is particularly challenging — the company must achieve 2nm yields just as the industry transitions to gate-all-around transistor architectures that require entirely new manufacturing processes. Even with unlimited subsidies, recruiting the specialised process engineers and photolithography specialists needed to debug sub-3nm production at scale remains a global talent bottleneck.

PJM Interconnection issued an emergency procurement targeting 15 gigawatts of new generation capacity to prevent rolling shortages across the mid-Atlantic grid serving 65 million people, explicitly citing AI data centre expansion as the primary demand driver. Bloomberg reports the proposal would add generation equivalent to 15 large nuclear reactors or 25 natural gas combined-cycle plants — an unprecedented scale of emergency capacity addition driven by load forecasts that now assume data centres will triple their regional electricity consumption by 2029. The request follows PJM's January warning that reserve margins could fall below reliability thresholds during peak demand periods as early as summer 2027.

The grid operator's move exposes a fundamental mismatch between AI infrastructure buildout timelines and power system planning cycles. Data centre developers are signing leases and ordering servers on 18-24 month schedules, while natural gas plants require 36 months to commission and nuclear units need 7-10 years even with expedited permitting. PJM's emergency authority allows it to contract capacity outside normal market mechanisms, but the physics of generator construction remain unchanged — meaning some proposed data centres in Virginia, Maryland, and Pennsylvania will face connection delays or curtailment requirements regardless of capital availability. The situation is particularly acute because PJM's coal fleet retirements are accelerating faster than replacement capacity is coming online, shrinking the buffer that historically absorbed unexpected demand spikes.

CoreWeave secured multibillion-dollar capacity commitments from Anthropic and a separate $21 billion agreement with Meta within a three-day span, positioning the specialist cloud provider as a critical infrastructure layer between GPU supply and frontier AI labs. CEO Michael Intrator disclosed the Anthropic contract value during a Bloomberg interview, though neither company released specific terms or capacity volumes. DatacenterDynamics confirmed the arrangement represents a multi-year compute lease, likely spanning 2026-2029 based on typical AI infrastructure contract structures. The timing suggests Anthropic is diversifying beyond its primary AWS relationship as Claude adoption accelerates, while Meta's massive commitment indicates continued delays in bringing its own data centre capacity online at sufficient scale.

CoreWeave's rapid ascent as an infrastructure kingmaker reflects structural inefficiencies in how compute reaches AI developers — hyperscalers remain capacity-constrained for H100 and newer architectures, while CoreWeave's willingness to sign long-term GPU purchase commitments with NVIDIA provides allocation priority that cash-rich but capacity-poor AI labs cannot secure directly. The business model carries substantial execution risk: CoreWeave is essentially arbitraging its credit access and supply chain relationships against future utilisation rates, betting it can fill capacity between anchor tenant workloads. Any slowdown in AI training demand would leave the company holding expensive, rapidly-depreciating GPU clusters with insufficient revenue to cover debt service on the billions borrowed to acquire the hardware.

President Trump's initiative to expand AI chip exports to allied nations is breaking down under licensing backlogs, staffing attrition, and policy confusion at the Bureau of Industry and Security, according to Bloomberg reporting based on interviews with Commerce Department officials and semiconductor executives. The agency responsible for approving billions in sensitive technology exports has seen 30% turnover in key technical positions since January 2025, while the volume of license applications has tripled as companies seek clarity on which AI accelerators can ship to which countries under the administration's revised framework. Processing times for routine approvals have stretched from 45 days to over 120 days, effectively freezing sales to emerging markets the White House identified as strategic priorities.

The dysfunction creates enforcement vulnerabilities: delays incentivise gray-market distribution through third countries, while inconsistent approval standards make compliance unpredictable for manufacturers. Companies report receiving contradictory guidance on identical chip specifications depending on which BIS analyst reviews the application, while the agency has not published clear technical thresholds defining which performance levels trigger enhanced scrutiny. The breakdown is particularly consequential given simultaneous reports that DeepSeek is recruiting data centre engineers for Inner Mongolia facilities reportedly running banned NVIDIA Blackwell chips — suggesting export controls are failing to prevent advanced hardware from reaching Chinese AI labs despite rhetorical escalation.