Episode

Key Takeaways

AI infrastructure investment faces potential overheating despite long-term demand.
Data center design requires flexibility for rapid GPU and power density advancements.
Coordination gaps among industry players significantly slow AI infrastructure deployment.
Addressing power grid limitations and community relations is critical for AI growth.

Discussion frames the AI bubble talk as a recurring pattern seen with new general-purpose technologies like the internet.
While AI itself may not be in a bubble, segments of the market show overheating due to unprecedented speed of capital expenditure.
A JP Morgan estimate suggests $5 trillion in AI investments by 2030 would require $650 billion in annual revenue for a 10% return.
Current AI services are generating tens of billions, creating a notable gap compared to the required revenue stream.
Infrastructure build-outs must consider flexibility, as the timeline for artificial general intelligence (AGI) remains uncertain.

Core markets, including North Virginia, Dallas, and London, report long queues and low data center vacancies.
Graphics Processing Units (GPUs) have a short 5-6 year shelf life, with NVIDIA's rapid advancements potentially leading to obsolete equipment.
Data center building shells have a 20-40 year lifespan, contrasting sharply with customer-provided server equipment's 3-6 year refresh cycle.
Evolving technology necessitates retooling existing data centers due to significantly increased power density per rack, requiring upgrades to power delivery and cooling systems.

Future-proofing data center designs for AI, particularly regarding liquid cooling and higher power densities, presents significant current complexity.
The cost for retooling data centers to accommodate hyperscalers' needs is typically covered by additional non-recurring charges within contracts.
A 15-year take-or-pay data center contract can involve negotiations for non-recurring charges and lease rate adjustments if new designs are introduced.
KKR has a track record of recontracting data center assets at potentially higher rates, driven by supply/demand imbalances and the difficulty of relocating established ecosystems.
Older contracts are often significantly below current market rates, creating negotiation points where hyperscalers may consider leaving or operators may upgrade capacity to increase lease rates.

KKR aims to re-engineer data center construction using a 'molecule to the rack' approach to address a 'coordination tax' within the industry.
This inefficiency stems from a lack of coordination among data center operators, power companies, and capital providers, slowing deployment cycles.
KKR integrates data center platforms with power generation, transmission, and utility relationships, including a partnership with ECP leveraging power generation assets.
A pilot project in Bosque, Texas, co-locates a data center with two hyperscalers adjacent to a power plant, requiring negotiations with regulators and communities.
The power plant can divert electricity from data centers back to the grid during strain, with data centers utilizing on-site generation, a complex engineering feat.

Hyperscale data center costs range from $12 million to $18 million per megawatt, with state-of-the-art liquid cooling adding $2 million to $5 million per megawatt.
The industry's current lack of standardization in liquid cooling technologies leads to varying costs for similar solutions.
Long-term success depends on building customer relationships and providing innovative solutions, rather than just maximizing short-term EBITDA.
The traditional view that hyperscalers require power 24/7 with no flexibility is evolving, with discussions around on-site batteries and generation.
There is optimism for operators innovating and hyperscalers embracing flexibility for efficient resource use, though full utilization of renewable resources via battery storage is still developing.

Significant obstacles to infrastructure projects include supply chain issues, community opposition, and rising electricity costs.
The U.S. faces a 'collision course' of ideas, balancing the need to win the AI race with protecting ratepayers and making long-term investments amid technological uncertainty.
Building approximately 70,000 miles of high-voltage power lines in the next decade in the U.S. would take 200 years at the current pace.
Optimism remains for solutions, including federal policies for grid efficiency, equitable cost-sharing, fast-tracking authority, and utilizing federal land for infrastructure.
The next few years will test operators' ability to prudently allocate capital and deliver mega facilities, potentially leading to a split between successful and struggling companies.