Capital Intensity and Depreciation: The Margin Story Nobody Tells

TransDigm Group reported an operating margin of 47.2% in its most recent fiscal year. General Electric, an industrial conglomerate that also makes aerospace components, posted 21.4%. The difference, 26 percentage points, is larger than the full operating margin of most manufacturing companies. The two firms occupy adjacent squares on the industry chessboard, yet their profit profiles diverge so sharply that the comparison reads as a category error. As Motley Fool noted in February, the global backlog of unfilled aircraft orders tops 17,000 jets, so demand is not the binding constraint. The binding constraint is how each company chose to allocate capital across the cycle, and what that choice did to the depreciation line on the income statement.

TransDigm does not build engines or airframes. It acquires and manages proprietary aftermarket parts: valves, pumps, actuators, cockpit displays, the components that airlines cannot substitute when a jet is already in service. The original engineering and tooling were amortized years ago, often by a different company, and TransDigm arrives as the consolidator that collects the remaining decades of high-margin replacement revenue. Its capital expenditure runs at roughly 2% to 3% of revenue. The depreciation charge is correspondingly small. Most of what flows through as gross profit drops almost directly to operating income. There is no heavy factory depreciation schedule eating 10 or 12 cents of every revenue dollar before the company even reaches the operating line.

GE Aerospace, by contrast, carries the legacy of a balance sheet built around turbine engineering, casting facilities, and test cells that cost billions to build and decades to depreciate. Even after the restructuring that split off GE HealthCare and GE Vernova, the remaining aerospace business retains capital intensity that acts as a permanent headwind to margin expansion. None of this is hidden. It appears in plain sight on the cash flow statement, in the gap between net income and free cash flow. But it is remarkable how rarely the depreciation variable itself is treated as the primary explanatory factor rather than a footnote to a story about pricing power, scale, or brand.

The same variable is now the under-examined centre of gravity in the technology sector's largest capital cycle. On April 29, Reuters reported that Meta Platforms raised its annual capital spending forecast again, signalling plans to pour billions more into artificial intelligence infrastructure. Meta's Q1 2026 earnings reached $7.31 per share, rising 13.7% year over year on advertising growth and AI momentum, according to Zacks. Revenue growth was real. But the capex number kept climbing, and it will keep climbing through at least the remainder of the fiscal year. Every dollar of that spending eventually becomes a depreciation charge.

The same day, Microsoft guided to $190 billion in capital spending for its 2026 fiscal year, a figure that landed well above Wall Street estimates, CNBC reported. The driver was not just data centre construction but soaring memory prices, a reminder that AI infrastructure inflation is not limited to GPUs. Taken together, the hyperscalers are now on a trajectory where aggregate capex could exceed $1 trillion by 2027, CNBC reported in early May. The revenue flowing from these investments is beginning to show up. Azure grew faster than expected. Meta's ad business accelerated. But revenue recognition and cost recognition operate on different clocks, and the cost clock is the one that eventually determines whether the investment cycle was wise.

Depreciation is the mechanism that synchronises those two clocks. When a company buys a server, it does not expense the full purchase price in the quarter of acquisition. It capitalises the cost on the balance sheet and then charges a portion to the income statement each year over the asset's estimated useful life. The longer the useful life, the smaller the annual charge, and the higher the reported operating income. This is not a quirk of Generally Accepted Accounting Principles. It is the core design of accrual accounting, meant to match the cost of an asset with the revenue it generates over time. The problem, as with any forward estimate, is that the chosen useful life is a forecast, and forecasts can be wrong.

In late 2025, hedge fund manager Michael Burry began arguing publicly that the hyperscalers had extended the assumed useful lives of their server fleets in ways that flattered near-term operating margins. Amazon, Alphabet, and Microsoft had all lengthened server depreciation schedules over the preceding two years, some from four years to six. The accounting change was disclosed in the footnotes of their 10-K filings, as required, but the effect on reported earnings was not trivial. Extending a server's depreciable life from four years to six reduces the annual depreciation charge by roughly one-third. Operating income rises. Free cash flow does not change, because the cash was already spent. Only the timing of the income statement recognition shifts. Motley Fool summarised the thesis in November 2025: Burry argued that hyperscalers were using depreciation periods to inflate the appearance of AI profitability.

Nvidia responded explicitly. In a disclosure document released in late 2025, the chipmaker quoted Burry by name and corrected his characterisation of GPU longevity, TheStreet reported in December. Nvidia's position was that its newest accelerators were designed for longer operational lives than the prior generation, and that the hyperscalers' accounting adjustments reflected genuine engineering improvements, not financial engineering. The exchange was unusually pointed for a debate conducted through SEC filings, and it revealed how much was at stake. If the useful-life assumptions were too generous, then billions of dollars of capitalised assets would eventually require accelerated depreciation or impairment charges, compressing margins at precisely the moment when investors were expecting them to expand.

Brian Anderson, the CEO and founder of Nacelle, laid out the structural dimensions of the debate in a Forbes Business Council column published in April 2026. He described depreciation as the hidden variable in the AI rally, a line item that most investors glance past but that will determine whether the current capex cycle generates genuine returns or merely shifts costs into future reporting periods. Anderson pointed out that when innovation outruns the spreadsheet, the useful-life assumption is the first place the tension becomes visible. The physical reality is that AI training clusters run hot, cycle relentlessly, and face obsolescence pressure from each successive GPU generation. Accounting schedules assume a steady, linear decay. Engineering reality is lumpier.

The capital-intensity problem extends well beyond the hyperscalers to the semiconductor supply chain itself. Zacks reported on May 13 that Entegris, a supplier of materials and consumables to chip fabricators, is positioned at the intersection of rising materials intensity and the 2-nanometre process node ramp expected to accelerate through the second half of 2026. Entegris does not build fabs. It sells the specialty chemicals, filtration systems, and wafer-handling products that fabs consume continuously. Its capital intensity is lower than that of the foundries it supplies, which must invest tens of billions of dollars in a single facility that then depreciates over five to seven years. The asymmetry is instructive: Entegris captures the volume growth of the 2nm transition without carrying the depreciation burden of the fab itself.

The foundries, by contrast, face an escalating capital-intensity curve with every process node. Taiwan Semiconductor Manufacturing Company's 2nm ramp requires gate-all-around transistor architecture, extreme ultraviolet lithography at higher numerical aperture, and new materials integration at nearly every step. Each wafer starts at a higher cost and each fab costs more to build. Depreciation becomes a structural component of the cost of goods sold, and the margin profile of the semiconductor manufacturing segment is dictated less by volume leverage than by the arithmetic of how quickly a $30 billion fab must be amortised against the wafers it produces.

The Entegris thesis, as Zacks described it, is that a 2026 consumables rebound plus rising wafer starts at leading-edge nodes will drive revenue growth even as the capex cycle creates margin pressure elsewhere in the semiconductor ecosystem. The timing risk is real. If the hyperscalers eventually slow their spend, the foundries cut orders, and the consumables pipeline contracts. But the structural argument is that materials intensity per wafer rises with each node, so Entegris benefits from a secular trend that is partially decoupled from the quarterly rhythm of equipment orders. That decoupling is rare in a supply chain where almost everyone else is tightly yoked to the capex cycle of their largest customers.

A different capital-intensity story has been unfolding in industrial manufacturing. Graphic Packaging, a producer of paperboard packaging, saw its share price fall roughly 50% from its peak through early May 2026, Motley Fool reported on May 13. The company operates in a sector where the fixed-cost base, including paper mills, coating lines, and converting equipment, is enormous relative to the variable margin on each unit of output. When demand softens, as it did through late 2025 and into 2026 amid consumer packaged goods destocking, the depreciation and maintenance costs persist. Revenue declines flow almost directly to operating income, a phenomenon that capital-light businesses can avoid but that capital-heavy manufacturers cannot.

The Graphic Packaging case is a useful counterpoint to the current AI euphoria. Investors are pricing the hyperscalers as though their infrastructure spend will generate durable, software-like margins once the build phase matures. But a data centre is a physical asset. Servers degrade. Network switches reach end-of-life. Cooling infrastructure requires maintenance. The depreciation that is being deferred in the income statement today through extended useful-life assumptions will eventually flow through, and when it does, it will compress margins at the same time that the assets themselves need replacement. The double hit, depreciation plus refresh capex, is the scenario that capital-intensive industries have always had to manage, and technology companies have rarely been tested by it at the scale now approaching.

TransDigm's model offers a clarifying lens. The aerospace aftermarket generates 47% operating margins not because the parts are technologically extraordinary but because the capital was sunk long ago, the depreciation is negligible, and the customer cannot switch. The AI infrastructure build has none of those attributes. The capital is being sunk right now, at historically unprecedented scale. The depreciation is being spread across an assumed useful life that is the subject of active debate. And the customer, whether an enterprise buying cloud compute or a consumer using an AI-assisted application, absolutely can switch if the price does not reflect the value. There is no regulatory mandate to run workloads on Azure rather than Google Cloud, and there is no airworthiness directive that grounds a chatbot if a particular inference chip is unavailable.

The margin question for the hyperscalers is therefore not whether AI revenue will grow. It almost certainly will, given the enterprise adoption curves visible in Microsoft's 20 million paid Copilot seats and Meta's accelerating ad yields. The question is whether the depreciation schedule will be the friend it has been for the past two years or the headwind it becomes when the assets age faster than the accounting assumes. In the semiconductor supply chain, the same question applies in a different register: will the foundries' depreciation burden at 2nm be offset by high enough wafer pricing, or will the margin improvement accrue disproportionately to the capital-light companies, like Entegris, that sell into the ramp without building it?

What to watch for, across the next two earnings seasons, is not the revenue line. Revenue will almost certainly continue to benefit from the AI buildout. What to watch is the gap between depreciation and capital expenditure in the cash flow statements of the hyperscalers. When depreciation is running well below capex, the company is under-earning relative to the replacement cost of its asset base. When depreciation catches up to capex, operating margins will face a squeeze that no amount of revenue growth can fully offset. And for the semiconductor supply chain, watch the consumables-to-equipment revenue ratio at companies like Entegris. A rising ratio in the back half of 2026 would confirm that the capex cycle is maturing from build to operate, which is typically when margin expansion begins. A flat or falling ratio would suggest the cycle still has further to run, and the depreciation reckoning remains ahead rather than behind.