Compute & Infrastructure

Apollo Global Management and Blackstone have closed a $35 billion debt financing package earmarked for Anthropic to expand its AI chip and infrastructure capacity, according to Bloomberg. The scale of this single transaction — structured as debt against expected compute utilization revenues — represents a maturation of private credit as a funding layer for frontier AI infrastructure, sitting between equity venture rounds and sovereign-backed national programmes. The deal is confirmed as finalized, not merely announced.

The strategic implication is that chip procurement is now being treated as a capital asset class, not an operating expense. By financing GPU purchases through structured debt, Anthropic can scale compute orders at a cadence that equity fundraising alone would not support, while Apollo and Blackstone take on collateralized exposure to AI demand. This model, if it holds, creates a new leverage point: private credit funds become de facto gatekeepers to frontier compute access for labs that lack hyperscaler balance sheets.

Google has contracted SpaceX for $920 million per month in computing power through mid-2029, a commitment totalling approximately $30 billion, per Bloomberg. This is Google's second such deal with an AI-adjacent competitor in recent weeks, indicating a deliberate strategy to source compute from non-traditional providers rather than exclusively expanding owned data centre capacity. The nature of SpaceX's compute offering — whether ground-based or tied to orbital ambitions — carries material implications for infrastructure risk.

SpaceX has publicly stated a goal of deploying 100 gigawatts of AI compute in orbit, and Starcloud CEO Philip Johnston has acknowledged the engineering challenges of building and maintaining orbital data centres at scale, per Bloomberg. Thermal dissipation, latency to ground users, and the absence of in-orbit servicing infrastructure remain unresolved constraints. The Google deal, structured through 2029, likely covers near-term ground-based capacity with orbital as a longer-horizon option — but the contract terms have not been disclosed. The distinction matters: orbital compute is speculative capacity; ground capacity from SpaceX's Stargate-adjacent operations is more credible near-term supply.

Nine U.S. trade associations have formally petitioned the Trump administration to intervene in an AI-driven DRAM and HBM shortage that they warn will constrain supply and inflate costs across automotive, medical device, consumer electronics, and broadband infrastructure sectors through at least 2027, per Tom's Hardware. The petition frames AI data centre memory consumption not as a market equilibrium issue but as a supply distortion requiring policy intervention.

The structural driver is HBM allocation. SK Hynix, Samsung, and Micron have redirected the bulk of leading-edge DRAM production capacity toward HBM3E for NVIDIA H100/H200 and Blackwell systems, leaving standard DDR5 supply constrained. The Semiconductor Engineering week-in-review also flagged HBM price hikes as a notable market signal this week, per Semiconductor Engineering. With Samsung still ramping HBM yield recovery and Micron the only U.S.-headquartered supplier, the geographic concentration of memory production — South Korea and, to a lesser extent, Taiwan — represents a supply chain vulnerability that extends well beyond the AI sector.

Meta is deploying temporary tent-based data centre structures across the U.S. that can be operational within three months, powered by jet turbine generators rather than grid interconnections, per Tom's Hardware. This approach compresses the traditional two-to-three-year data centre development cycle to a fraction of its length by circumventing the two primary bottlenecks: construction permitting timelines and utility grid interconnection queues, the latter of which currently runs three to seven years in many U.S. markets.

The operational model — on-site generation rather than grid draw — carries significant cost and emissions implications. Jet turbine generation is materially more expensive per MWh than grid power and produces direct carbon emissions at the site. This positions Meta's approach as a short-term capacity bridge, not a sustainable infrastructure model. The Seattle moratorium, discussed separately, suggests municipalities are beginning to develop regulatory responses to exactly this kind of rapid, non-grid-integrated deployment, which may close the permitting arbitrage window that makes the tent model viable.

Seattle's city council committees have passed a one-year moratorium on AI data centre construction alongside a resolution to study community impact, with full council approval widely expected, per Tom's Hardware. Seattle is not a primary data centre market — Northern Virginia, Phoenix, and Dallas dominate U.S. capacity — but its moratorium establishes a municipal policy template that other cities with denser populations and constrained grid capacity may replicate. The stated rationale covers water consumption, power grid strain, noise, and community displacement, framing data centre buildout as a zoning and environmental justice issue rather than purely an economic development question.

The timing is notable. As hyperscalers pursue modular, rapid-deployment strategies specifically to bypass traditional development timelines, municipalities are developing regulatory instruments that operate at the land-use layer, where federal preemption is limited. A wave of similar moratoriums — even temporary study periods — across secondary markets could meaningfully constrain the geographic diversification strategies that data centre operators are pursuing to spread power grid and climate risk.