Investors should maintain high conviction in NVIDIA (NVDA) as it remains the only hardware platform clients are willing to back with massive, 5-year "take-or-pay" contracts. To capitalize on the shift from AI training to execution, focus on companies specializing in Inference and Model Routing software, which optimizes costs by directing simpler tasks to cheaper, older chips. The primary bottleneck for AI growth has shifted from chip supply to Power Infrastructure, making data center providers with "powered shells" and secured energy access premium assets. Look for investment opportunities in the electrical supply chain, specifically firms providing transformers, battery backups, and specialized electrical labor. While CoreWeave remains private, its $10 billion backlog from financial giants like Jane Street signals a massive, untapped enterprise demand for dedicated GPU clusters.

CoreWeave (Private)

• CoreWeave is a "NeoCloud" provider specializing in GPU infrastructure for AI training and inference. • The company currently supports 9 of the top 10 AI labs globally (excluding China). • Customer Diversification: The client base is shifting from just hyperscalers and AI labs to include large enterprise and financial services firms. • Financial Services Growth: CoreWeave reports a backlog of nearly $10 billion from direct financial services clients (e.g., Jane Street), who are managing infrastructure directly rather than through an AI lab. • Contract Trends: Demand is shifting toward longer-term commitments. While 3-year contracts were standard, clients are now seeking 5-year "take-or-pay" contracts with fixed economics to ensure uninterrupted access to compute.

Takeaways

• Inference is the Growth Engine: Over 50% of infrastructure utilization on CoreWeave’s platform is now for inference (running models) rather than just training them. • Infrastructure Longevity: Contrary to fears of rapid obsolescence, older chips like the NVIDIA A100 and H100 are expected to have a functional life of 6–8 years due to "model routing" (using smaller, cheaper chips for simpler tasks). • Execution is the Moat: The primary bottleneck in the industry has shifted from GPU availability to "Powered Shells"—data centers that are fully energized, cooled, and staffed with specialized electricians.

NVIDIA (NVDA)

• NVIDIA remains the "de facto choice" for AI infrastructure according to CoreWeave. • Despite the rise of custom silicon (like Microsoft’s Maya or Google’s TPUs), CoreWeave sees almost no material demand from large-scale clients for anything other than NVIDIA chips. • The "CUDA" Moat: The software ecosystem and reliability of NVIDIA hardware make it the only platform clients are willing to sign multi-billion dollar, 5-year contracts for. • Product Pipeline: CoreWeave is already receiving testing racks for Vera Rubin, NVIDIA’s next architecture following the Blackwell (liquid-cooled) generation.

Takeaways

• Fungibility Issues: High-end compute is not yet a commodity. An H100 in one cloud may perform differently than in another based on the software stack and "good put" (efficiency), which favors established NVIDIA-centric operators. • Scaling Laws: AI labs believe that scaling (adding more compute) still yields better models, suggesting the "ceiling" for NVIDIA demand has not yet been reached.

Investment Themes & Sectors

AI Infrastructure & Data Centers

• The Power Bottleneck: The real constraint is the physical delivery of electrons to the rack. This involves long-lead items like transformers, battery backups, and a shortage of master electricians (a 5-year apprenticeship trade that cannot be easily scaled). • Financing Innovation: The sector is maturing financially. CoreWeave recently secured Investment Grade rated financing for GPU infrastructure, unlocking "insurance-grade" capital which lowers the overall cost of building these centers.

Model Routing & Efficiency

• The "CFO Reckoning": Companies are experiencing "sticker shock" from AI costs (e.g., Uber reportedly burning its 2026 AI budget in four months). • Opportunity: There is a massive upcoming investment theme in Model Routing. This involves software that automatically sends a simple query to a cheap, small model and only uses expensive "frontier" models (like GPT-4 or Claude 3.5) for complex tasks.

Compute as a Commodity

• Tradable Compute: While there is interest in "Compute Exchanges" (like CME-listed futures), the lack of fungibility (standardization) makes this difficult in the short term. • Risk Factor: Until different cloud providers can guarantee identical performance ("Good Put") for the same chip, a liquid trading market for compute remains a long-term prospect rather than a current reality.

Takeaways

• Bullish Sentiment: Demand for inference is "unrelenting" and "insatiable," with no signs of a pullback despite corporate budget concerns. • Risk Factors: The primary risks mentioned are supply chain constraints for data center components and the specialized labor required to build them, rather than a lack of demand for the AI itself.