Investors should prioritize companies that control physical power and grid connections, such as IREN, as energy capacity has replaced chip supply as the primary bottleneck for AI growth. High-performance hardware remains a stable investment, with NVIDIA GPUs like the A100 and H100 maintaining value through a long "inference" lifecycle and five-year enterprise contracts. For software exposure, look toward "orchestrators" like Perplexity AI that remain model-agnostic and capture high-margin recurring revenue from enterprise automation. Mistral AI offers a strategic play on "data sovereignty," as industries like finance and defense shift toward specialized, local models to protect proprietary data. Finally, consider the growing demand for high-end local workstations from Apple or Dell as businesses move compute away from the cloud to reduce long-term token costs and improve privacy.

CoreWeave

CoreWeave is a specialized cloud provider that pivoted from Ethereum mining to providing massive-scale GPU infrastructure for AI hyperscalers and model builders.

• Business Model: They operate as a "NeoCloud," sitting above the NVIDIA hardware layer but below the AI models. They focus exclusively on high-performance compute for AI training and inference.
• Asset Lifecycle: CEO Michael Intrator dismissed the idea of rapid GPU obsolescence as "nonsense."
  • CoreWeave uses a 6-year depreciation schedule.
  • Older chips (like the A100) are currently appreciating in value due to high demand for inference and smaller model experiments.
• Financing Innovation: The company uses a "Box" structure for financing. They secure a contract (e.g., with Microsoft), place the contract and the GPUs in a bankruptcy-remote vehicle, and use the cash flow to pay down debt.
  • They have raised $35 billion in 18 months using this method.
  • Principal and interest are typically paid off within 2.5 years of a 5-year deal.
• Supply Chain: They are the "tip of the spear" for NVIDIA, being among the first to deploy H100s, H200s, and GB200s at scale.

Takeaways

• Infrastructure Scarcity: Demand for compute remains "relentless" and overwhelms global capacity. The bottleneck is no longer just GPUs, but power, transformers, and memory.
• Investment Stability: Long-term contracts (average 5 years) with high-quality counterparties provide significant protection against market volatility.
• Secondary Market Strength: Investors should note that older GPU generations remain highly profitable for "inference" (running models) even after newer chips are released for "training."

Perplexity AI

Perplexity is an AI-native search engine and "answer engine" that is evolving into a "computer" capable of executing complex tasks.

• Product Evolution: They recently launched Perplexity Computer, which uses agents to perform multi-step tasks (e.g., researching LinkedIn, populating Google Sheets, or building a CRM).
• Switzerland Strategy: Unlike Google or OpenAI, Perplexity is model-agnostic. They orchestrate multiple models (GPT-4, Claude, Gemini, DeepSeek, Llama) to provide the best result for a specific prompt.
• Monetization: They have high-margin recurring revenue.
  • Enterprise Pro: $40/month.
  • Enterprise Max: $400/month (includes higher usage limits and advanced orchestration).
• Local vs. Cloud: They are moving toward a hybrid model where sensitive data stays on a local device (like a Mac Mini or Mac Studio) while heavy compute is handled in the cloud.

Takeaways

• The "Harness" Value: As AI models commoditize, the value shifts to the "orchestrator" or "harness" that knows which model to use for which task.
• B2B Growth: Their enterprise segment is growing faster than their consumer segment, indicating high corporate adoption for back-office automation.
• Hardware Trend: There is a growing trend toward high-end desktop "workstations" (e.g., Dell/NVIDIA collaborations) to run local models for privacy and cost-saving.

Mistral AI

Mistral is the leading European AI company, focusing on high-performance open-source and proprietary models.

• Open Source Advantage: Mistral argues that open-source models allow enterprises to keep their data on-premises, avoiding the security risks of sending proprietary IP to closed-source providers.
• Verticalization: They focus on "Mistral Forge," a product that helps companies in finance, engineering, and physics specialize models on their own private data.
• Human vs. Synthetic Data: While they use synthetic data to "warm up" models, they believe high-quality human expert feedback is essential for the final 10% of performance.

Takeaways

• Data Sovereignty: For sectors like defense or banking, Mistral’s "portable platform" (deploying models on the customer's own cloud) is a major competitive advantage over OpenAI.
• Specialization over Generalization: The future of AI investment may lie in "vertical" models trained on specific industry IP rather than one-size-fits-all global models.

IREN (formerly Iris Energy) (IREN)

IREN is a data center owner and operator that transitioned from Bitcoin mining to AI cloud services.

• Power as the Moat: The company owns 4.5 gigawatts of power capacity. In the current market, power is a scarcer resource than the chips themselves.
• Renewable Arbitrage: They locate data centers in West Texas and British Columbia directly at the source of excess wind, solar, and hydro power. This allows them to access low-cost energy that cannot easily be transmitted to cities.
• Jevons Paradox: CEO Daniel Roberts argues that as AI becomes more efficient and cheaper, demand will increase exponentially (induced demand), rather than decrease.

Takeaways

• Real Estate & Energy Play: IREN is essentially a "real-world" play on AI. Their value is in land, permits, and grid connections.
• The "Data Center is the Computer": Modern AI requires massive clusters where the entire building acts as a single machine. This requires specialized "trades" (electricians, HVAC) that are currently in high demand with rising salaries ($150k–$300k range).
• Nuclear Potential: While not currently using it, the company is tracking Small Modular Reactors (SMRs) as a future 24/7 clean energy source for data centers.

Investment Themes & Sector Insights

1. The Shift to Inference

While the "training" of models gets the headlines, the "inference" (using the models) is where the long-term monetization happens. This provides a long tail of utility for older GPU hardware.

2. Power is the New Oil

The primary constraint for AI growth is no longer chip production, but the physical infrastructure of the electrical grid. Companies with "behind-the-meter" power or direct access to renewable sources (like IREN) hold significant strategic advantages.

3. Agentic Workflows

The industry is moving from "chatbots" to "agents." Tools like Perplexity Computer and OpenClaw represent a shift where AI doesn't just talk, but executes work across different software applications.

4. Local AI (Edge Computing)

Due to privacy concerns and the high cost of cloud tokens, there is a burgeoning market for powerful local hardware (Mac Studios, specialized Dell servers) to run "local" versions of open-source models like Mistral.