Most important take away

AI infrastructure demand remains relentless and far outstrips global supply, with every CEO confirming there are no idle GPUs in the world. The companies winning this era are those that started early, secured power and land, built creative financing structures, and positioned themselves as orchestrators rather than trying to own every layer of the stack.

Chapter Summaries

CoreWeave: From Crypto to AI Infrastructure Giant (Michael Intrator)

CoreWeave originated as an algorithmic hedge fund that pivoted through crypto mining into GPU cloud computing. They donated early GPUs to the EleutherAI open-source project, which taught them how to run large-scale parallelized computing and organically generated their first customers. Their innovative “box” financing structure packages client contracts, GPU purchases, and data center leases into discrete cash-flow vehicles, allowing them to raise $35 billion in 18 months while dropping their cost of capital by 600 basis points. Intrator firmly rejects GPU depreciation concerns, noting average contracts are five years and A100 prices have actually appreciated, with six-year depreciation as the industry standard.

Perplexity: The AI Orchestra Conductor (Aravind Srinivas)

Perplexity has evolved from an AI-powered search engine to a multi-model orchestration platform. Their product progression went from model-agnostic search (Perplexity Ask) to agentic browsing (Comet) to full computer control (Perplexity Computer). Their key competitive advantage is being model-neutral, acting as an “orchestra conductor” that routes queries to the best-suited model. They are launching a personal computer product using Mac Mini as a local server for private data orchestration, with a hybrid local/cloud architecture. The enterprise business is growing faster than consumer, with enterprise max at $400/month and all revenue carrying positive gross margins.

Mistral: Enterprise AI with European Values (Arthur Mensch)

Mistral positions itself as the enterprise-focused AI company with a strong emphasis on data sovereignty and European regulatory compliance. They differentiate through forward-deployed engineers who work alongside enterprise domain experts to train specialized models. Mensch discussed the limitations of synthetic data (useful for model compression but ultimately requires human signal) and how enterprises need governance, observability, and deterministic control gates that consumer tools like OpenClaw cannot provide. Their approach involves strict data segregation and role-based access controls for mission-critical enterprise processes.

IREN: Powering AI from the Ground Up (Daniel Roberts)

IREN is transitioning from Bitcoin mining to AI data centers, leveraging eight years of land and power acquisition. They hold 4.5 gigawatts of capacity (comparable to the entire Bay Area’s annual power usage) and signed a $9.7 billion contract with Microsoft representing only 5% of capacity. Located in West Texas near excess renewable energy sources (wind and solar), they use 100% renewable energy by co-locating near generation sources. Roberts emphasized that the real constraint is now time-to-compute rather than power, and endorsed Jevons’ paradox: as compute becomes cheaper and faster, demand will compound rather than plateau.

Summary

Key Themes:

• Infrastructure demand is insatiable. All four CEOs confirmed that demand for AI compute far exceeds supply. CoreWeave has a waitlist comparable to 90s Knicks tickets, and IREN says there are zero idle GPUs in the world. The constraint has shifted from GPUs to memory, power, and construction labor.

• Creative financing unlocks scale. CoreWeave’s “box” structure for packaging long-term contracts into discrete financing vehicles allowed a relatively small company to raise $35 billion. This approach of asset-backed lending against five-year client contracts with creditworthy counterparties has driven their cost of capital down 600 basis points toward hyperscaler levels.

• Multi-model orchestration is an emerging moat. Perplexity’s model-neutral position lets them benefit regardless of which foundation model wins. As models specialize rather than commoditize, the value of an orchestration layer that can route to the best model for each task increases.

• Enterprise AI requires governance that consumer tools lack. Mistral and Perplexity both highlighted that enterprises need data segregation, role-based access controls, observability, and deterministic process gates. Simply connecting AI to enterprise data without these safeguards creates serious risks around compensation data leaks and compliance failures.

• Energy follows geography, and AI follows energy. IREN’s strategy of co-locating data centers near stranded renewable energy (West Texas wind and solar) rather than near population centers is proving highly effective. Fiber latency from remote Texas to Dallas is only 6 milliseconds, busting the myth that data centers must be near metro areas.

• GPU longevity is underestimated. Both CoreWeave and IREN pushed back on GPU depreciation fears. Older GPUs find new life in inference workloads, smaller model training, and emerging companies that could not access bleeding-edge hardware. The six-year depreciation standard reflects real commercial lifespan.

Actionable Insights:

• Investors and operators should think about AI infrastructure in terms of long-duration contracted cash flows, not speculative capex. The “box” model demonstrates how to de-risk massive buildouts.
• Companies building AI products should consider multi-model strategies rather than betting on a single provider, as model specialization is increasing.
• Enterprises adopting AI agents must invest in data governance and access control infrastructure before broadly deploying agents with access to sensitive internal data.
• The trades workforce (electricians, HVAC, construction) servicing data centers represents a significant and growing career opportunity, with salaries starting around $150K.
• Local/edge compute (Mac Mini, Dell workstations) paired with cloud orchestration is an emerging architecture that balances privacy with capability, worth evaluating for both consumers and enterprises.