Anthropic in talks to lease 1 gigawatt from Stream Data Centers- The Information

Anthropic is in preliminary talks to lease up to 1 gigawatt of data-center compute capacity from Stream Data Centers, which is majority owned by Apollo Global Management. The facilities could be equipped with Broadcom- and Alphabet-designed tensor processing units, alongside potential Nvidia GPUs, underscoring continued large-scale AI infrastructure demand. Google has also discussed potentially providing a credit guarantee, although the scope and terms remain unclear.

Analysis

The relevant read-through is not the headline capacity figure but a potential shift in AI infrastructure procurement: a direct lease backed by a hyperscaler credit enhancement would move demand from cloud capex onto private-data-center balance sheets. That structure is incrementally positive for APO because it can monetize development, financing and potentially asset-management fees while insulating the tenant from upfront construction capital; it also broadens the funding channel for AI load beyond the public hyperscalers' capex budgets. The key constraint becomes utility interconnection and power-delivery timing, making energization milestones—not signed leases—the value inflection.

AVGO has the cleanest upside asymmetry if custom TPUs are the primary deployment: a multi-year inference/training commitment would support visibility for its custom silicon and networking franchises simultaneously. NVDA is not necessarily displaced; heterogeneous clusters preserve GPU demand, but the relative risk is mix and multiple—evidence that frontier labs can scale through TPU-heavy deployments would reinforce the market's concern that custom ASICs capture the marginal inference dollar. GOOG's possible guarantee is strategically valuable if it secures TPU utilization and ecosystem lock-in, but economically ambiguous until the guarantee's recourse, duration and consolidated-capex treatment are known.

Near term, this is a sentiment catalyst rather than an earnings estimate change because no executed contract, power allocation, or equipment order has been disclosed. Over 1-3 months, confirmation of an executed lease plus a defined power-delivery schedule would favor AVGO/APO; over 6-18 months, the larger risk is that grid delays, financing spreads, or tenant concentration turn nominal gigawatts into stranded development inventory. A key falsifier is any indication that the project relies principally on NVDA GPUs or that Google declines credit support, which would weaken both the ASIC and private-financing interpretations.

Key Decisions for Investors

• Maintain/establish a 3-6 month long AVGO versus short NVDA relative-value position in equal dollar beta-adjusted sizing; add only on confirmation of TPU purchase commitments or a disclosed power schedule. Target 10-15% relative upside if custom-silicon backlog visibility improves; stop if disclosures indicate GPU-first deployment or AVGO AI guidance fails to improve at the next earnings update.
• Initiate a small 6-12 month long APO position or add to existing alternatives exposure, sized for a 1.5-2.0x upside/downside profile. The catalyst is lease execution accompanied by Apollo-managed financing or asset ownership; exit if credit support is absent and project financing requires materially wider spreads, as development economics would deteriorate quickly.
• Do not add directional GOOG exposure solely on this report. Set an alert for disclosures of guarantee amount, term, and whether the obligation is treated as incremental capex or off-balance-sheet support; a broad guarantee without commensurate TPU revenue would create a negative risk-transfer signal.
• Monitor utility interconnection filings, transformer procurement, and announced power-delivery dates over the next 90 days. Treat a delay beyond 2027 or a lower-than-1GW initial energization tranche as a reason to reduce AVGO/APO exposure, since the market is likely to capitalize announced AI power demand before revenue is realized.