#180 - Data Centers + AI Capex - Boom or Bust?

“As DeepSeek has already demonstrated, there will be further savings with improved efficiencies in training large-language models and with the open-source release of better digital mousetraps. In another development, Prime Intellect, Inc. released an open-source training method that requires less than a tenth of the usual bandwidth, allowing it to use smaller, geographically dispersed training centers. Reduced training demands of the DeepSeek variety, plus more distributed training, could obviate the need for capital-intensive hyperscale data centers and their engineering headaches (e.g., cooling, networking and power systems). For eager coders, there will be every incentive to make the additional supply unnecessary by the time it arrives. “The hyper-capex AI model will start getting more (justified) questions,” tweeted our friend, the short seller Jim Chanos. We agree.”(Grant’s Interest Rate Observer)

“Big power, big needs: The $2 billion construction loan—one of the largest ever for a data center—will fund a 100-acre campus in West Jordan, Utah. The facility, developed by CIM Group and Novva Data Centers, will provide 175 megawatts of continuous power, enough for 175,000 homes. Notably, it is already fully leased.Record breaking: This marks the second $2 billion-plus data center loan this year, following a $2.3 billion deal for an Abilene, Texas facility. Traditionally, construction loans for data centers were below $1B, but loans are growing as AI drives demand for power-hungry infrastructure, pushing lenders to fund larger, pre-leased projects.Growing demand: North America’s data center supply surged in 2024, doubling to 6,350 MW under construction—a 12-fold jump from 2020—as AI and digital services fuel demand, CBRE reports. Despite a 34% YoY rise in completed projects, power delays and 36-month equipment wait times keep oversupply risks low.”(source: CRE Daily)

The technology industry is experiencing a historic surge in data center and semiconductor capital expenditure (capex), fueled by AI’s insatiable need for computing power. Microsoft, Google, and Amazon are investing billions into digital infrastructure, with Microsoft alone planning $80 billion in AI-driven data center spending for 2025.

But is this growth a permanent structural shift, or are we witnessing a speculative boom that could correct once AI adoption stabilizes? While AI’s demand for compute resources is undeniable, efficiency breakthroughs, power constraints, and rising capital costs could introduce limits.

AI: A Powerful Driver of Data Center Expansion

AI’s rapid development has led to a frenzy for GPUs and data center capacity. Large-scale AI models like GPT-4 require massive compute resources, and as AI applications expand—from robotics to automation and enterprise software—so does the need for high-performance data centers.

Some estimates suggest AI workloads could push future data centers to be 5–10 times larger than today’s, requiring over $1 trillion in total investment over the next decade. Major infrastructure investors like Blackstone and Brookfield see data centers as critical long-term assets, comparing them to utilities due to their role in the global digital economy.

Beyond cloud computing, autonomous vehicles, AI-powered logistics, and robotics also require continuous AI model inference and real-time data processing, further increasing the need for scalable AI infrastructure.

However, not everyone is convinced this expansion will continue indefinitely. The assumption that AI workloads will grow exponentially forever could be challenged if cost-effective compute methods emerge that reduce the need for hyperscale facilities.

Efficiencies: How Will New Models Effect the Capex Boom?

While AI currently requires enormous compute power, the long-term outlook for hyperscale data center expansion depends on efficiency breakthroughs.

New AI model optimization techniques, such as:

• Model pruning (removing unnecessary parameters),

• Transfer learning (reusing trained models), and

• Distillation (creating smaller, equally powerful models),

… have the potential to significantly reduce the computing needs of AI systems. For example, DeepSeek’s new AI training methods cut computing requirements by up to 90%, allowing for smaller, decentralized data centers instead of hyperscale facilities.

Furthermore, custom AI chips (e.g., Amazon Trainium, Google TPUs) and optimized software frameworks are increasing the efficiency of AI training and inference, meaning companies could achieve the same outcomes with less hardware.

Skeptics, like short-seller Jim Chanos, warn that today’s “hyper-capex AI model” is reminiscent of the fiber optic overbuild of the early 2000s, which led to billions in underutilized infrastructure for years. If AI efficiency outpaces compute demand, many of today’s AI-driven data centers may never reach full utilization.

Macro Constraints: Power & Capital

Even if AI demand remains strong, physical and financial constraints could slow the pace of data center expansion.

1. Power Availability

• AI data centers consume massive amounts of electricity, and many power grids are already at capacity.

• In Northern Virginia, Oregon, and parts of Europe, AI data center growth is being limited by energy constraints.

• AI mega-clusters require 100MW+ per site—enough to power a small city. Some projects are being delayed due to grid limitations.

2. Rising Interest Rates

• Building a $500M+ data center at 7% interest rates is far more capital-intensive than at 2%.

• Some AI firms are securing loans using GPUs as collateral, reminiscent of past tech bubbles driven by excessive leverage.

• Higher borrowing costs could force companies to prioritize profitability over expansion, slowing capex growth.

3. Supply Chain Disruptions

• Nvidia’s AI GPUs have multi-month backlogs, delaying new deployments.

• Transformers, cooling systems, and electrical components required for data centers have years-long supply chain delays.

• Even if companies want to build more data centers, they may not have the hardware or power infrastructure to do so quickly.

Taken together, these macro constraints could act as natural brakes on unchecked AI infrastructure expansion.

Other Factors Influencing Data Center Growth

While AI is the biggest driver of today’s capex surge, other factors also play a role:

1. Enterprise Cloud Migration

• Companies continue shifting IT workloads from on-premises to cloud, fueling steady demand for data centers.

• However, cloud adoption growth is stabilizing, suggesting hyperscale data center expansion could moderate in the future.

2. IoT & Edge Computing

• Smart devices, 5G, and industrial IoT are increasing local data processing needs.

• This could reduce reliance on massive hyperscale data centers, distributing workloads across smaller, regional facilities.

3. Sustainability & Regulation

• Governments are scrutinizing data centers’ environmental impact, particularly power and water consumption.

• Stricter regulations could limit expansion in certain areas, forcing a shift toward more sustainable, energy-efficient designs.

These trends suggest that AI won’t be the only factor shaping data center growth—and in some cases, alternative computing strategies could slow hyperscale expansion.

The AI-driven data center and chip capex boom is a mix of structural demand and speculative overbuilding. AI’s computing needs are massive and growing, but efficiency improvements, power constraints, and financial costs could temper hyperscale data center expansion over time.

Much like the fiber optic boom of the 2000s, today’s AI buildout could overshoot, leading to a temporary slowdown in capex growth and underutilized capacity in some regions. However, long-term digital infrastructure demand remains strong, suggesting a boom that persists, but with inevitable corrections along the way.

Further reading

1. Goldman Sachs Research – AI’s impact on data center energy demand

2. Dell’Oro Group – AI-driven data center capex forecast to hit $1 trillion by 2029

3. Bain & Company – AI efficiency breakthroughs reducing compute needs

4. The Wall Street Journal – AI’s impact on data center construction

5. Data Center Dynamics – Power constraints affecting AI data center growth

6. Meta’s AI CapEx Plans – Meta plans $60-65bn capex on AI data center boom

7. AWS AI CapEx Growth – AWS on track for $100bn revenue, increasing AI infrastructure spending

8. Bubble, Interupted - Grant’s Interest Rate Observer

9. Robots will be next which will propel AI capex spend to new heights - Coatue

10. Jonathan Gray from Blackstone ( BX 0.00%↑ ) and Bruce Flatt from Brookfield ( BAM 0.00%↑) are continuing to invest billions in data center development capacity and capex for large tech companies.

Investor’s Corner

Streamline 66 - Elliott Management’s presentation and position on optimizing Phillips 66 (Link)

Steel prices post tariffs - @SpecialSitsNews (Link)

Paul Singer | Podcast | In Good Company | Norges Bank Investment Management (Link)

Wedgewood Partners Fourth Quarter 2024 Client Letter (Link)

Ferrari 2024 Annual Report (Link) (Disc. I am not long, but may consider initiating a position) - A lot to like here. $EXOR owns 20% (John Elkann Chairman of Board), Piero Ferrari owns 10% of company (big inside ownership), incredible global brand, ROIC of 30%+, has repurchased ~4% of shares back in last 3 years, however valuation is pretty rich currently.

Howard Marks’ new memo (Link)

Horizon Kinetics’ Q4 Market Commentary (Link)

Blackstone’s Jonathan Gray on the Investment Boom around AI (Link)

Brookfield’s Bruce Flatt on Tariff Impact & Deal Making (Link)

Tech Corner

Robotics won’t have a ChatGPT Moment - Coatue (Link)

America’s Industrial Reboot: A Massive Tech Opportunity - Coatue (Link)

Grant’s Interest Rate Observer on the AI Capex Bubble (Link)

TD Cowen sounds the alarm on MSFT 0.00%↑ cancelling data center leases @Wallstengine (Link)