Reading the Shift in AI Infrastructure Investment Cycles

TL;DR

AI infrastructure may be moving from pure expansion toward more maintenance, replacement, and upgrade planning.
This matters because capex, cash, depreciation, and replacement cycles can move differently across servers, networks, and facilities.
Review capex and cash together, then separate server, network, and power plans into distinct replacement schedules.

$43.2 billion, $11.94 billion, and 3 to 5 years frame this market more clearly than revenue growth alone. NVIDIA held $43.2 billion in cash, cash equivalents, and marketable securities on January 26, 2025. Micron held $11.94 billion on August 28, 2025. Capacity is still expanding. Cash is also accumulating. That combination suggests the market is not driven only by a "keep expanding" view.

Example: A data center team sees strong demand, but avoids treating every asset as one budget line. It reviews servers, networks, and cooling separately. It then times upgrades around support windows, downtime risk, and facility limits.

The AI infrastructure investment cycle can be read in two phases. In the first phase, GPUs and data center expansion drive financial performance. In the next phase, maintenance, replacement, software upgrades, network refreshes, and power and cooling operations matter more. Corporate disclosures and operating documents already include figures and language that can help gauge that transition.

Current state

Corporate disclosures show that major semiconductor and AI infrastructure companies are still increasing investment. In its 2025 annual report, TSMC said annual production capacity increased by about 0.5 million 12-inch equivalent wafers through capital expenditures. In its 2025 Form 10-K, Micron said cash and marketable investments totaled $11.94 billion as of August 28, 2025. That was up from $9.15 billion on August 29, 2024. In its 2025 Form 10-K, NVIDIA said it held $43.2 billion in cash, cash equivalents, and marketable securities on January 26, 2025.

The key point is simple. Expanded investment and cash accumulation are appearing at the same time. If the market had higher certainty, companies might hold less cash and expand more aggressively. However, research summaries related to AMD, Micron, and NVIDIA mention risks. Those risks include sharp demand declines, order delays, changes in AI investment timing, delays in new technology ramps, and export controls. Even with strong results, capacity does not rise in a straight line.

Operating documents show a similar pattern. ASHRAE indicates a service life of 3 to 5 years for datacom equipment. It indicates 15 to 20 years for mechanical and electrical infrastructure. Servers and GPUs are replaced on a shorter rhythm. Power and cooling infrastructure runs on a longer rhythm. NVIDIA documentation says the OEM hardware warranty for vGPU-supported GPUs is typically 3 years. It also describes Extended Full Support for 3 years and Maintenance Support for 3 years.

Networks move differently as well. Cisco documentation ties switch replacement to EoX events, such as End of Life and End of Support. NVIDIA AI Enterprise and DPU documents also describe upgrades through maintenance windows and staged rollout. Demand does not come only from buying all-new equipment at once. The ability to replace generations without interrupting service also matters.

Depreciation and useful life provide another signal. Meta reported depreciation expense for server and network assets of $13.36 billion in 2025. It reported $11.34 billion in 2024 and $7.32 billion in 2023. In separate disclosures, Meta changed the useful life of servers from five to six years. It later adjusted the estimated useful lives of certain server and network assets. This does not support only one interpretation. One reading is delayed replacement. Another reading is selective replacement around generational transition points.

Analysis

From a decision-making perspective, the focus may be shifting. The question may be less about "whether to build more." It may be more about "what to replace, and how often." If demand keeps growing and model gains support faster payback, GPU server and network upgrades may move forward. If customer AI spending is delayed, the response may differ. Export controls, ramp delays, and order delays can also overlap. Suppliers may then slow expansion and hold more cash. Even within one AI cycle, semiconductor, server, network, and power companies may move on different schedules.

A common misunderstanding appears here. Some assume a larger installed base makes the market more stable through maintenance and replacement demand. That view is only partly supported. Maintenance and replacement may be easier to forecast than new capacity builds. But upgrade economics still depend on performance gains, power efficiency, software compatibility, and downtime costs. Servers may follow a 3 to 5 years rhythm. Mechanical and electrical infrastructure is designed for 15 to 20 years. Data center operators therefore manage IT and facilities budgets on different cycles, even within one campus. Investors should not treat all of this as one category called "AI infrastructure."

Suppliers face a trade-off as well. Expanding capacity early may help secure share. If demand weakens, utilization and margins can come under pressure. Accumulating cash can help absorb downside risk. During supply shortages, it can also mean missed opportunities. TSMC's capacity expansion and the large cash positions of Micron and NVIDIA show both choices at once. This can be read as carrying offense and defense together.

Practical application

The useful question now may not be the "total AI budget." It may be the replacement calendar. If GPU servers, network equipment, and power and cooling facilities sit in one budget line, decisions can blur. Servers should be evaluated by performance, power efficiency, and software support periods. Networks should be evaluated by EoX events and interruption risk. Power and cooling should be evaluated by longer asset lives and expansion headroom. Even within one AI investment plan, the approval logic can differ.

If you are an investor, three checks stand out in earnings releases. First, did capex increase? Second, did cash-equivalent assets also increase? Third, do risk terms appear in the text? Examples include order delays, ramp delays, export controls, and useful life adjustments. Looking only at capex can miss one side of the cycle. Looking only at cash can miss the other side.

Checklist for Today:

Read the capex section and the cash-equivalent assets section side by side in each relevant disclosure.
Separate GPU servers, networks, and power and cooling assets, then define replacement triggers for each category.
Review depreciation and useful life changes for signs of either demand pressure or replacement strategy changes.

FAQ

Q. Can we already say that the AI infrastructure market has shifted to being centered on maintenance and replacement?

It is difficult to conclude that from this material alone. Corporate disclosures and market documents do show upgrade and replacement signals. However, they do not fully establish the exact market-wide mix change.

Q. Does having a lot of cash mean a company is investing less?

Not necessarily. In this material, investment expansion and cash accumulation appear together. That pattern may reflect both demand response and downside protection.

Q. Can we set a single number for the data center replacement cycle?

That is not recommended. Datacom equipment such as servers and GPUs follows a shorter cycle. Mechanical and electrical infrastructure is designed for a longer service life. Networks should also be assessed with EoX events, not age alone.

Aionda

Reading the Shift in AI Infrastructure Investment Cycles

TL;DR

Current state

Analysis

Practical application

FAQ

Further Reading

References

Get updates