Anthropic's Claude Takes Bigger Role in Building AI

Anthropic found that its Claude chatbot is leading 26% of its own AI research and development, underscoring accelerating adoption of AI tools in technical work. SoftBank is pursuing nearly $21 billion in potential new borrowings to expand AI financing capacity, while AI infrastructure company Crusoe has raised nearly $4 billion in fresh funding. The developments point to sustained capital investment in AI models, financing and compute infrastructure amid intensifying global scrutiny of the technology.

Analysis

The investable read-through is not broad software productivity yet; it is a deepening capital-intensity cycle in model development and inference. As frontier labs automate portions of their own engineering workflow, the near-term bottleneck shifts further toward scarce compute, networking, power delivery and data-center commissioning. This favors NVIDIA (NVDA), Broadcom (AVGO), Arista (ANET), Vertiv (VRT) and Eaton (ETN), but also raises the risk that customers pull forward infrastructure orders faster than end-demand monetization can validate them over the next 6-18 months.

Private AI financing increasingly substitutes leverage for public-equity discipline, creating a second-order benefit for hardware vendors with upfront payment terms while concentrating counterparty risk among GPU-cloud operators. The vulnerable link is not semiconductor demand in the next one to two quarters, but financing spreads and residual values for leased accelerators if model providers moderate training spend. SoftBank (SFTBY) is a liquid proxy for this reflexive cycle: additional financial capacity can support AI asset prices, but its equity should remain highly sensitive to credit conditions and any gap between AI valuations and realizable cash flows.

Consensus is likely underestimating the value of power and cooling relative to incremental GPU supply. Compute deployments delayed by interconnection queues or power constraints can defer revenue recognition for cloud tenants even when chips ship on schedule; that makes VRT, ETN and GE Vernova (GEV) more defensible 6-18 month beneficiaries than another high-beta bet on training compute. This thesis is falsified if hyperscaler capex guidance flattens materially, GPU lead times normalize alongside falling utilization, or high-yield spreads widen enough to impair private data-center financing.

Key Decisions for Investors

• Maintain a 6-12 month overweight in VRT and ETN versus NVDA: own the power/cooling bottleneck where incremental capacity is harder to substitute. Reassess if either company reports backlog conversion delays or if major hyperscalers cut 2027 infrastructure capex outlooks.
• Pair trade for the next 1-3 months: long ANET / short SFTBY. ANET has direct exposure to AI cluster networking spend and a cleaner balance sheet, while SFTBY embeds both AI upside and financing-duration risk; exit if credit spreads tighten materially and SFTBY outperforms ANET by more than 15% from entry.
• Use a basket rather than a single-name GPU-cloud exposure: long GEV, VRT and ETN in equal weights against a small short in the Global X Cloud Computing ETF (CLOU). The trade expresses physical-infrastructure scarcity versus potentially overcapitalized application/cloud capacity; size modestly because software multiples can expand if enterprise AI revenue inflects.
• Set a financing-stress alert rather than initiate a directional short: monitor high-yield spreads, private GPU-cloud funding terms and NVDA customer concentration disclosures. A sustained 75-100 bp widening in high-yield spreads or evidence of accelerator lease-rate compression would be a catalyst to reduce AI-infrastructure beta and consider SFTBY downside hedges.