AI summary by TwinMind
The Data Center is the New Unit of Compute: Crusoe CEO Chase Lochmiller
Sequoia Capital
18m 45s
Introduction and Context
- Welcome to Chase Lockhart Miller (Crusoe CEO) and partner David Khan
- Chase provides unique perspective on AI infrastructure focused on servers, steel, power - raw infrastructure beneath GPU level
- David Khan known for "$600 billion question" about AI investment vs application revenue gap
- Focus on AI industrial revolution, specifically steel and power components
Current AI Infrastructure Scale
- Race between most powerful companies to build largest GPU clusters
- Current scale: 100,000 GPU clusters expanding to 300,000 GPU clusters
- Big tech discussing million GPU clusters
- Crusoe currently has 5,000 people in Abilene, Texas building data center
- Working around the clock to accelerate AI infrastructure
AI Data Centers vs Traditional Data Centers
- Traditional data center serving web pages vs high performance AI data center comparison: "go-kart vs Range Rover"
- Northern Virginia (global data center hub): ~4.5 gigawatts total capacity built over decades
- Crusoe (7-year-old startup): 2 gigawatts currently under construction, 20 gigawatts in pipeline
- Massive transformation of both energy and data center landscapes
Supply Chain and Manufacturing Bottlenecks
- Bottlenecks constantly shifting: previously impossible to get chips, now eased with TSMC/Nvidia production increases
- Current major bottlenecks: power, steel, data center space
- Complex supply chains not easy to accelerate
- Crusoe's vertically integrated approach with domestic manufacturing
- Example: Low voltage switch gear quotes were 100 weeks, Crusoe built factory and reduced to 22 weeks
- Speed as critical competitive advantage
Big Tech Partnership Strategy
- Big tech companies structurally designed not to move quickly
- Microsoft example: 19 layers of approval for purchasing decisions
- Traditional big data center programs: 3-5 year timelines for 100+ megawatt facilities
- Crusoe's acceleration: First phase Abilene 300 days, second phase just over 200 days
- Massive time-to-market advantage as primary value proposition
Energy-First Approach and Geographic Strategy
- Energy-first approach: bringing computing demand to areas with low-cost, clean, abundant energy
- West Texas/Abilene selection: massive renewable overbuild, particularly wind
- Wind power generators getting negative pricing ~30% of time
- Abundance of curtailment (could produce power but don't due to negative pricing)
- Natural fit: AI needs lots of energy, area has energy surplus
Sovereign AI and Data Locality
- Geographic dispersion driven by data locality requirements
- Nation-state owned data that cannot leave country boundaries
- Catalyzing domestic AI strategies and geopolitical power considerations
- Stargate project involvement (rumored, not confirmed)
US Energy Landscape and Constraints
- US utilities historically "sleepy industry" with flat generation capacity
- Efficiencies previously outpaced demand growth
- AI creating step function change in demand
- "Sleeping giants being woken up" with gigawatt power requests
- Crusoe's first principles approach: if can't get utility power, build own power plant
Nuclear and Next-Generation Energy
- Three deals with small modular reactor companies
- AI as opportunity to accelerate next-generation energy technologies
- Vision: future with both intelligence and energy widely abundant
- First time in human history manufacturing intelligence (vs organic method of having children)
- Manufacturing intelligence with silicon
Energy Deal Negotiations
- Diverse team backgrounds including former "landman" (mineral rights negotiator)
- Significant inbound requests and opportunities
- More complex than just having power access
- Confluence required: power, networking, permitting, supply chain
- Abilene impact: 5,000 workers in town of 120,000 people
- Substantial skilled trades import required (electrical, mechanical, construction)
Data Center Design Evolution
- Traditional data center: 2-4 kilowatts per rack 20 years ago
- Current NVL72 GB200 racks: 120-130 kilowatts
- Future Vera Rubin Ultra: 600 kilowatt rack installation
- Multiple orders of magnitude power density increase
- Fundamental thermodynamics changes requiring new cooling architectures
Liquid Cooling Infrastructure
- Data centers resemble "water works" with massive pipes
- Each building: ~1 million gallons of water
- Direct-to-chip liquid cooling architecture
- 84 chillers per building for heat exhaust
- Network design at cluster layer, not chip/server layer
Data Center as Computer Concept
- "Data center is the new unit of compute"
- AI factory = data center scale computer
- Abilene: 8 separate buildings under construction, all interconnected
- 44-inch conduits connecting each network core
- One coherent cluster of GB200s - planned to be world's biggest
Company Culture and Mountaineering Philosophy
- Chase climbed 5 of 7 summits
- Company value: "think like a mountaineer"
- Founded Crusoe after returning from Mount Everest climb
- Mountaineering practices informing business practices
- Plan A with contingencies for things going wrong
- Resilience and robustness embedded in process and organization design
- Success through multiple chaotic periods in company history
