Hyperscaler AI Capex Surge Risks $600B Stranded Assets as Data Center Vacancy Hits Historic Low

Hyperscaler AI Capex Surge Risks $600B Stranded Assets as Data Center Vacancy Hits Historic Low

The Bottom Line

Hyperscaler AI infrastructure spending is accelerating despite early signs of overcapacity, creating a structural mismatch where $600B in annual capex risks becoming stranded assets while enterprise AI adoption lags behind buildout. By 2028, even a modest 20% utilization gap translates to $120B in annual stranded capital, forcing hyperscalers to shift from buildout to optimization and monetization of existing assets through AI-as-a-service offerings. Enterprises without multi-cloud strategies face vendor lock-in and potential price hikes as utilization falls.

What Happened

North American data center building boom showed signs of slowdown in H2 2025, with capacity under construction falling nearly 6% year over year despite record demand for AI and cloud services. Data center vacancy rate hit a historic low of 1.4% despite supply increasing 36% across markets. Four U.S. hyperscalers — Amazon, Google, Meta and Microsoft — increased data center capex by 76% year over year, with Amazon pledging to double its spending from $100B in 2025 to $200B in 2026, Google signaling more than double to $185B, and Microsoft investing $37.5B in Q4 2025 (annualizing to ~$150B).

The Financial Reality

At current trajectory, hyperscaler AI infrastructure spending could reach $600B annually by 2028. With AI revenue generation lagging infrastructure buildout, even a modest 20% utilization gap translates to $120B in annual stranded capital. For enterprises running $20M AI budgets, this creates a looming vendor lock-in trap as hyperscalers seek to monetize underutilized infrastructure through long-term AI service commitments, increasing switching costs and reducing negotiating power. The financial impact scales directly with the capex-revenue gap, creating a structural cost disadvantage for locked-in enterprises.

Under the Hood

Hyperscalers are building AI-optimized data centers at unprecedented scale, focusing on GPU-accelerated servers and specialized networking for AI workloads. Unlike traditional cloud infrastructure, AI data centers require higher power density, advanced cooling, and low-latency interconnects, driving up capex per server. The economics are driven by a race to capture AI workloads early: hyperscalers believe securing early AI demand justifies overbuilding, as they can later monetize through higher-margin AI services and long-term consumption contracts. However, this assumes AI adoption will keep pace with buildout — an assumption now showing cracks as vacancy rates rise and utilization metrics soften.

The Tension

Hyperscalers argue continued investment is necessary to meet exploding AI demand and maintain competitive edge, noting they are taking proactive measures to mitigate risks and optimize costs. They contend that temporary overcapacity is preferable to underinvestment in a strategic priority, and that AI workloads will rapidly absorb new capacity as model deployment accelerates. However, this ignores the structural lag in enterprise AI adoption: enterprises face organizational resistance, legacy system integration challenges, and talent shortages that slow AI deployment regardless of infrastructure availability. The break point occurs when capex continues to outpace demonstrable AI revenue growth, making overbuilding structurally unsustainable.

What Breaks Next

• Hyperscaler data center construction slows as vacancy rates rise and utilization metrics deteriorate
• Shift from net new construction to upgrade and replacement cycles begins within 12 months
• Enterprises face increasing pressure to renegotiate cloud contracts as hyperscalers seek to monetize underutilized capacity
• AI infrastructure spending transitions from growth to maintenance, forcing hyperscalers to develop consumption-based pricing models

Winners and Losers

Hyperscalers with scale and vertical integration (AWS, Azure, GCP) — can absorb overcapacity periods and leverage AI services to monetize infrastructure Chip manufacturers (NVIDIA, AMD) — benefit from sustained GPU demand regardless of utilization rates Enterprises that secure early-access AI capacity through long-term contracts — guarantee availability during peak demand periods Enterprises without multi-cloud strategies — face vendor lock-in and potential price hikes as utilization falls Smaller cloud providers and colocation operators — unable to match hyperscaler capex scale, losing market share Hyperscaler shareholders — face potential writedowns if infrastructure becomes stranded capital

What Nobody's Talking About

There is no enforcement layer to prevent hyperscalers from overbuilding; the market assumes demand will materialize organically, but if AI adoption lags due to enterprise readiness gaps or regulatory delays, capex becomes structurally stranded capital with no clawback mechanism or utilization guarantee. This creates a prisoner's dilemma where individual hyperscalers rationally overbuild to avoid missing the AI wave, but collective overbuilding destroys returns for all.

Where This Goes

Now (0-6 months): Hyperscalers pause or slow new data center construction as vacancy rates rise and utilization metrics soften, shifting focus from buildout to optimization of existing assets Next (6-24 months): AI infrastructure spending transitions from net new construction to upgrade and replacement cycles, forcing hyperscalers to monetize existing capacity through AI-as-a-service offerings and consumption-based pricing models By 2028: The majority of hyperscaler AI infrastructure spending shifts to maintenance and optimization, with new construction focused only on replacing obsolete gear

The Executive Playbook

1. Audit current cloud infrastructure contracts for exposure to utilization-based pricing and renewal terms — complete within 30 days
2. Measure actual AI workload utilization versus contracted capacity — establish baseline within 60 days
3. Create a multi-cloud workload placement strategy to avoid single-vendor lock-in — implement within 90 days
4. Negotiate flexibility clauses in cloud contracts allowing capacity scaling down as utilization changes — ongoing
5. Separate experimental AI workloads from production workloads to optimize costs across environments — immediate