AI Capex Cycle: The Largest Tech Build-Out in History

Key Takeaways

• AI capex cycle aggregate hyperscaler spending quadrupled from roughly $75 billion per year in 2022 to an estimated $300+ billion in 2025, with combined Amazon, Alphabet, Microsoft, and Meta capex approaching $200 billion in 2024.
• Power has emerged as the binding constraint ahead of chips for many 2025–2026 projects; a single gigawatt of AI data center capacity can cost over $15 billion and requires years of utility coordination.
• A common mistake is ignoring useful-life assumption changes; extending GPU depreciation from 4 to 6 years lowers reported depreciation expense and flatters operating margins without changing underlying cash economics.
• Training capex is lumpy and front-loaded; inference capex scales with revenue, investors should track incremental AI revenue versus incremental capex over time to assess whether the build-out is earning a return.

What It Is

AI capex refers to the capital spent on the compute, power, and real estate required to train and serve large AI models. The main buyers are the hyperscalers: Microsoft Azure, Google Cloud, Amazon AWS, and Meta. Oracle has become a meaningful fifth buyer. The main suppliers include NVIDIA for accelerators, TSMC for chip manufacturing, memory vendors for high-bandwidth memory, networking vendors, utilities and turbine makers for power, and developers for data-center real estate.

Aggregate hyperscaler capex quadrupled from roughly 75 billion dollars per year in 2022 to an estimated 300-plus billion in 2025, according to Epoch AI tracking and company disclosures. Combined spending across Amazon, Alphabet, Microsoft, and Meta approached 200 billion dollars in 2024, with guidance and analyst forecasts pointing to further increases of 40 percent or more in 2025. Dell'Oro Group measured record hyperscaler data center capex in Q2 2025.

The Intuition

Two phases of AI demand drive the spend. Training builds the model, running for weeks or months on tens of thousands of accelerators in tightly coupled clusters. Inference serves queries to users, running continuously at lower intensity but at a much larger footprint if usage scales. Training capex is lumpy and front-loaded. Inference capex scales roughly with revenue, which is the number investors care about most.

The cycle looks different from past tech build-outs because the unit economics are still being discovered. A training run might cost hundreds of millions of dollars for a single model. Inference margins depend on workload mix, hardware utilization, and pricing discipline. Bulls point to early margin data from mature models. Skeptics point to depreciation expense embedded in annual capex running multiples of cash operating costs.

How It Works

The capex dollar moves through a tight supply chain:

1. Hyperscaler places an order for accelerators, often 12 to 24 months forward. Orders are firm, with deposits.
2. Chip makers (NVIDIA is the largest) design the accelerator, which TSMC manufactures on leading-edge nodes, with HBM stacks from memory vendors co-packaged.
3. Server OEMs assemble accelerator trays with networking, liquid cooling plates, and interconnects.
4. Real estate developers build or lease data center shells with pre-negotiated power. Power procurement has become the binding constraint, with some developers securing nuclear or gas baseload years in advance.
5. Hyperscalers deploy the clusters as training or inference capacity, amortized over a service life that was recently extended from 4 or 5 years to 6 years at multiple firms, a change that materially flatters near-term earnings.

Guidance into 2026 suggests further acceleration. Public commentary from large hyperscalers points to combined capital plans of 400 to 600-plus billion dollars for 2026, with roughly three-quarters tied to AI infrastructure. IoT Analytics projects data center infrastructure market spend toward 1 trillion dollars by 2030 under optimistic scenarios.

Worked Example

A single hyperscaler plans to deploy 1 GW of new AI data center capacity in 2025.

Capex buckets, approximate industry ratios:

• Accelerators and related silicon: roughly 50 to 55 percent of total, or 8 to 10 billion dollars per GW depending on GPU mix.
• Networking and interconnect: about 10 percent.
• Servers, racks, cooling, storage: about 15 percent.
• Data center shell, power distribution, substation, cooling plant: about 15 to 20 percent.
• Land and site work: about 5 percent.

Total capex for 1 GW can exceed 15 billion dollars, equivalent to what the entire semiconductor industry spent in a typical year before 2020. The accelerators depreciate over 5 to 6 years, the buildings over 20 to 30. When a company extends useful life assumptions, depreciation expense falls and reported operating margins improve, even as cash capex keeps growing.

Common Mistakes

1. Assuming capex equals durable competitive advantage. Money buys capacity, not necessarily the best model or the best application layer. A trillion dollars of spend does not stop open-source models from eroding pricing or startup cloud providers from undercutting hyperscaler inference prices.

2. Ignoring depreciation drift. Extending useful life from 4 to 6 years lowers annual depreciation expense and boosts reported earnings with no change in underlying cash flows. If actual utility is shorter, future impairments follow. Compare cash capex to operating cash flow, not just reported EPS.

3. Overlooking the power constraint. Many 2025 to 2026 projects are not chip-limited. They are power-limited. A GW of new capacity requires substation build-out, transmission upgrades, and sometimes dedicated generation. Schedule slips hit revenue, not capex.

4. Treating AI revenue as proven at scale. In 2024 disclosures, reported AI-specific revenue was small relative to the capex committed. That may be early-cycle normal, or it may be a warning. Track the ratio of incremental AI revenue to incremental capex over time.

5. Conflating training and inference economics. Training is a fixed cost that amortizes over whatever the model earns. Inference is a marginal cost per query that scales with usage. A business that sells inference for less than its marginal compute cost is not saved by scale.

Frequently Asked Questions

Q: What is the AI capex cycle in simple terms? The AI capex cycle is the wave of spending by large technology companies on the compute, power, and real estate needed to train and serve large AI models. Hyperscaler capital expenditure quadrupled from roughly $75 billion per year in 2022 to an estimated $300+ billion in 2025, creating a multi-year demand surge for accelerators, networking equipment, data centers, and electricity infrastructure.

Q: How does the AI capex cycle affect investment decisions? The cycle creates direct beneficiaries, semiconductor manufacturers, power utilities, data center REITs, and networking vendors, and raises questions for the hyperscalers themselves about whether the massive capital spending will generate commensurate revenue returns. Investors need to track incremental AI revenue against incremental capex to determine whether the build-out is on track to earn a reasonable return on invested capital.

Q: What is a real-world example of AI capex cycle analysis? In the worked example, deploying 1 GW of AI data center capacity costs over $15 billion, with roughly 50–55 percent going to accelerators and related silicon. Extending GPU useful-life assumptions from 4 to 6 years lowers annual depreciation by hundreds of millions of dollars at large hyperscalers, flattering reported operating margins even as cash capex keeps increasing.

Q: How can investors use AI capex cycle analysis? Compare cash capex to operating cash flow rather than reported EPS, since useful-life extensions mask the true cash burden. Separately track training versus inference economics, training is a front-loaded investment while inference revenue should scale with usage. Also watch power procurement and utility partnership news, since power is now the binding constraint for many projects, not chip availability.

Q: How is AI capex different from prior technology build-out cycles? Previous tech build-outs like fiber or cloud were characterized by known unit economics before the spending peaked. The AI capex cycle is unusual because the economics of training and inference at scale are still being discovered. The 2000 fiber overbuild had clear unit cost metrics; AI model training cost trajectories, utilization rates, and inference pricing are all evolving, making the return on the current wave harder to evaluate in real time.

Sources

1. Dell'Oro Group. "Hyperscaler AI Deployments Lift Data Center Capex to Record Highs in 2Q 2025." https://www.delloro.com/news/hyperscaler-ai-deployments-lift-data-center-capex-to-record-highs-in-2q-2025/
2. Epoch AI. "Hyperscaler Capex Has Quadrupled Since GPT-4's Release." https://epoch.ai/data-insights/hyperscaler-capex-trend/
3. IoT Analytics. "Data Center Infrastructure Market: AI-driven CapEx Pushing IT and Facility Equipment Spending Toward $1 Trillion by 2030." https://iot-analytics.com/data-center-infrastructure-market/
4. TrendForce. "Hyperscalers Ramp Up Capex Amid AI Boom, Risks Lurk: Microsoft, Meta and Alphabet in Spotlight." https://www.trendforce.com/news/2025/10/30/news-hyperscalers-ramp-up-capex-amid-ai-boom-risks-lurk-microsoft-meta-alphabet-in-spotlight/

Disclaimer

This article is educational content only and is not financial advice. Nothing here is a recommendation to buy, sell, or hold any security. Consult a licensed advisor before making investment decisions.