Depreciation Accounting and Power Constraints Break Artificial Intelligence Infrastructure Spending

The sudden evaporation of nearly $500 billion in market value from Nvidia in a matter of sessions was widely reported as market jitter or macroeconomic anxiety. That interpretation misses the underlying mechanics.

Market capitalization swings of this magnitude are rarely about sentiment alone. They signal that a foundational assumption—that unconstrained capital expenditure on compute guarantees compounding competitive advantage—has collided with accounting realities and the laws of thermodynamics.

To understand why the market flinched, strip away the narrative of artificial intelligence as an epochal transformation and examine the physical and financial pipelines feeding it.

1. Whose Cash Flow Is Actually Improving?

Whenever an industry undergoes an investment supercycle, the primary question is not what the technology can do, but whose cash flow improves.

Over the past twenty-four months, cash flowed in a clean vector: from enterprise IT budgets and venture treasuries into hyperscalers (Microsoft, Alphabet, Amazon, Meta), and from hyperscalers straight into Nvidia’s gross margins, which hovered near 75%.

Enterprise / VC Treasuries
         │
         ▼
  Hyperscalers (Capex surge)
         │
         ▼
Nvidia / Foundries (75% gross margins)
         │
         ▼
Physical Constraints (Power grids, thermal limits)

Nvidia’s cash flow expanded at an unprecedented rate because it sold physical picks and shovels into an arms race. But for the buyers, the transaction represents a severe cash drain. Hyperscalers have poured well over $150 billion annualized into capital expenditures, overwhelmingly concentrated in data center infrastructure and advanced accelerator clusters.

Now, follow the return loop. Where does the cash flow back to the buyers?

It must come from end users paying for enterprise software, automated workflows, and software-as-a-service (SaaS) seat upgrades. Yet enterprise software margins are under pressure, not expanding. Deploying copilots and internal language models has generated marginal productivity gains and internal white papers, but it has not driven a commensurate surge in enterprise revenue or structural cost elimination.

The cash flow circuit has failed to close. Nvidia captured the cash upfront; the hyperscalers are left holding the equipment.

2. Why Now? The Depreciation Clock

The conditions supporting the hardware boom have existed since late 2022. What changed in 2024 is the calendar.

When an enterprise purchases an H100 or B200 server rack, the cash outflow is immediate, but the expense does not hit the income statement all at once. It enters the balance sheet as property, plant, and equipment (PP&E) and begins a relentless countdown called depreciation. For server hardware, that amortization window typically runs between three and five years.

[Year 0: Hardware Capex Deployed]
   │
   ├─ Year 1: Narrative phase. Depreciation hits; software monetization lags.
   ├─ Year 2: Fixed cost compounding. Hardware approaches mid-cycle.
   └─ Year 3–4: Obsolescence / Replacement cycle looms.

In 2023, hyperscalers enjoyed the narrative phase: massive capex announcements were rewarded by equity markets as signs of strategic dominance. In 2024, the depreciation from those early procurement batches began compounding on the income statement.

Depreciation is a fixed cost that does not negotiate. Whether an accelerator cluster runs at 95% utilization serving high-margin commercial inference or sits at 30% utilization running internal experiments, the monthly accounting charge remains identical.

As the depreciation expense rises quarter over quarter, the bar for incremental software revenue climbs with it. The moment revenue growth decelerates relative to the depreciation schedule, operating margins contract. That inflection point is no longer theoretical; it is visible in the forward guidance of major cloud providers.

3. The Physical Limits: Power, Transformers, and Marginal Capex

The second constraint is grounded in physics. The compute expansion is running out of cheap physical headroom.

A standard enterprise server rack historical power density sat between 5 and 15 kilowatts (kW). High-density AI clusters deploying modern architectures demand 40 to 100+ kW per rack. This leap fundamentally alters data center engineering:

• Cooling limits: Air cooling fails at these densities. Facilities are forced to retrofit for liquid cooling—circulating coolant directly to the cold plate on the silicon—multiplying facility capital costs per megawatt.
• Grid interconnections: In major data center hubs like Northern Virginia or Dublin, utility interconnections face lead times stretching from three to seven years.
• Component scarcity: High-voltage step-down transformers and backup switchgear face order-to-delivery delays of 100 to 150 weeks.

The marginal cost of adding a single unit of compute is no longer just the price of the GPU. It is the cost of securing scarce substation capacity, building liquid distribution loops, and competing for limited industrial power allocations.

Meanwhile, the economic value of the output is deflating. The price per million tokens across leading foundational models has dropped substantially over the past eighteen months, driven by model optimization and open-source alternatives.

When the marginal cost of physical deployment rises while the unit price of the digital output falls, returns diminish rapidly.

4. The Engineering Parallel

From the vantage point of automotive powertrain engineering, this structural pattern is instantly recognizable.

In powertrain development, pursuing headline performance figures—such as chasing peak horsepower or extreme 0–60 times via oversized battery packs and aggressive inverter setups—inevitably hits a wall. Beyond a certain threshold, the thermal rejection requirements scale non-linearly. You need heavier cooling loops, larger radiators, thicker busbars, and reinforced structural casing.

Pursuing Peak Output Metric
         │
         ▼
Non-Linear Escalation of Thermal & Electrical Load
         │
         ▼
Structural Mass / Cooling Penalties Compound
         │
         ▼
Diminishing Real-World Efficiency & Margin Compression

Eventually, the vehicle becomes too heavy, the marginal efficiency plummets, and the cost of the cooling system consumes the vehicle’s gross margin. Development teams are forced to abandon peak metrics and return to system efficiency: optimizing kilowatt-hours per mile, downscaling pack sizes, and improving thermal management at steady-state loads.

The AI infrastructure buildout is arriving at its own system-efficiency reckoning. Brute-force hardware scaling is yielding diminishing returns relative to the infrastructure footprint required to support it.

5. Falsification Conditions

This structural thesis argues that hyperscaler capex growth on AI hardware must decelerate toward existing depreciation absorption rates, constraining high-end accelerator demand.

This hypothesis fails if the following data emerges over the next two to four quarters:

1. Direct monetization breakout: The combined AI-driven cloud revenue run rate across Microsoft Azure, AWS, and Google Cloud grows at an accelerating sequential rate (>25% quarter-over-quarter), visibly outpacing their combined depreciation and amortization expense increases.
2. Structural power breakthrough: Grid interconnection timelines compress significantly via distributed modular power (e.g., behind-the-meter small modular reactors or dedicated utility-scale solar/storage deployments) coming online ahead of the standard 36-month timeline.

If hyperscalers prove that their enterprise clients are generating durable net margin—not merely running exploratory pilot projects funded by temporary innovation budgets—the capex cycle can sustain its pace. Until that evidence appears in audited financial statements, the physical and accounting drag remains active.

6. Value Flows and Structural Realignment

As capital allocation shifts from speculative cluster expansion to operational efficiency, the distribution of value changes across multiple sectors:

• Winners: Utilities with dedicated industrial base-load generation, grid infrastructure suppliers (switchgear, transmission components, high-voltage transformers), and software frameworks specialized in model quantization and inference efficiency on existing hardware.
• Losers: Pure-play semiconductor hardware providers trading at valuations priced for indefinite 40%+ capex compounding, and marginal data center operators unable to finance liquid-cooling retrofits or secure primary power allocations.
• Impacted Asset Classes: Multiples in the semiconductor complex (especially equipment and design firms reliant on leading-edge data center expansion) face structural compression toward standard industrial cyclicality. Concurrently, long-duration utility debt and electrical infrastructure equities will reflect sustained capital competition for physical grid assets.

The issue facing the technology sector is not that artificial intelligence lacks utility. It is that capital allocation ignored the friction of physical infrastructure and the discipline of depreciation accounting. When the cost of moving heat and electrons exceeds the near-term economic utility of the tokens produced, capital expenditure stops compounding. It rationalizes.

— Garryu

Source: 計算資源を買い続ける自動車エンジニアが直面した物理と採算の壁 | 日本経済新聞 https://www.nikkei.com/article/DGXZQOGN04BUH0U6A101C2000000/