Goldman Sachs’ $7.6 Trillion AI Infrastructure Ledger: Chip Depreciation Lifecycle Becomes the Wild Card in a Profit-Driven Gamble

Capital expenditure on artificial intelligence infrastructure is swelling at an unprecedented pace, but the accounting assumptions underpinning these trillion-dollar commitments are becoming a flashpoint on Wall Street. A major report from Goldman Sachs argues that the final scale of AI investment is not merely a function of demand, but is heavily swayed by a set of easily overlooked, yet massively consequential, supply-side assumptions—with the economic lifespan of AI chips being the most critical variable.

In its report titled Tracking Trillions, Goldman Sachs Global Research, using NVIDIA’s forward-looking data center revenue as a baseline, estimates that total cumulative global capital expenditure on AI infrastructure from 2026 to 2031 will reach approximately $7.6 trillion. This massive bill breaks down as follows: roughly $5.1 trillion for computing chips, $2.15 trillion for data centers, and $358 billion for power. However, the report's core thesis is not the figure itself, but its extreme fragility. The market habitually views AI capital spending as a demand-side issue, but Goldman argues that supply-side uncertainties are equally enormous and dangerously underappreciated.

Chip Depreciation: A Profit-and-Loss Numbers Game

The report identifies the economic lifespan of AI chips as the most influential of the four key assumptions driving the total. Currently, major hyperscale cloud providers typically depreciate GPU servers over a period of four to six years. However, NVIDIA has shifted to an annual product release cadence: from Hopper (2022), Blackwell (2024), to Rubin (2026) and Rubin Ultra (2027), with each generation delivering orders-of-magnitude gains in energy efficiency and performance. This makes a five-to-six-year depreciation cycle increasingly economically unsound.

Goldman Sachs’ sensitivity analysis reveals that shortening the chip lifespan from five to three years would cause the implied annual depreciation from 2026-2031 to surge from roughly $3 trillion to nearly $4 trillion. Conversely, extending it to seven years would lower the figure to $2.2 trillion. Adjusting this single parameter is enough to shift the depreciation burden on the ecosystem by hundreds of billions of dollars.

Prominent investor Michael Burry, when publicly shorting NVIDIA and Palantir in the second half of 2025, used this exact argument as his core thesis. He estimated that from 2026 to 2028, hyperscale operators would understate cumulative depreciation by roughly $176 billion due to overestimated chip lifespans, inflating profits by over 20%. He argued that the true economic life of an AI chip is closer to two to three years, making current accounting practices a form of earnings management.

Actual practices among major firms show clear divergence. Amazon in early 2025 shortened the depreciation period for some of its servers from six to five years, absorbing an operating profit hit of roughly $700 million and recognizing $920 million in accelerated depreciation for a batch of prematurely retired equipment. Microsoft CEO Satya Nadella has publicly stated the company is deliberately staggering its procurement cycles for different chip generations to avoid being burdened by four-to-five-year depreciation obligations on a single product. In contrast, Meta has extended its server lifespans three times in three years, most recently in January 2025, converting the resulting depreciation reduction into a $2.9 billion profit boost for a single quarter.

However, the CEO of CoreWeave provides contrary evidence: A100 chips purchased in 2020 are still operating at full capacity, and a batch of H100 chips coming off contract was re-leased at 95% of their original price. The Goldman Sachs report also concedes that older chips retain economic value in lower-sensitivity use cases like inference, edge computing, and synthetic data generation, and this tiered deployment model could support longer lifespans.

Surging Data Center Costs: From Warehouses to Specialized Fortresses

The second key assumption is data center construction costs. Goldman's baseline assumption is $15 million per megawatt, but the report notes significant upward pressure. Traditional cloud data center costs are around $10 million per megawatt. AI-era data centers are fundamentally different: rack power density has soared from the past range of 5-15 kilowatts to 130-200 kW in the Blackwell era and over 500 kW in the Rubin era. Cooling has shifted from air-based to full liquid cooling, requiring compute, memory, networking, and power to be designed as a complete, integrated system.

NVIDIA’s Vera Rubin platform, unveiled at GTC 2026, pushes this pressure to an extreme. The NVL72 rack packages 72 Rubin GPUs and 36 Vera CPUs into a standard 42U cabinet, consuming as much power as 40 average U.S. households. It requires direct liquid cooling with a 45°C inlet water temperature and 800V DC power supply—demands most existing facilities simply cannot meet.

Goldman’s sensitivity analysis shows that adjusting data center costs from $15 million to $19 million per megawatt would raise total cumulative data center capital expenditure over six years from $2.15 trillion to $2.72 trillion, an incremental increase of over $570 billion. The report also highlights an uncomfortable reality: "transitional AI data centers" built less than two years ago may already be unable to meet the power and cooling requirements of next-generation chips. When a data center is designed for a 20-year lifespan, but core technical requirements can fundamentally change within two years of operation, a long lifespan itself becomes a liability.

The Rise of ASICs: Cheaper Compute or a Catalyst for More Demand?

The third assumption concerns compute architecture choices. Beyond GPUs, an increasing share of computing power is being delivered via ASICs (Application-Specific Integrated Circuits): Google's TPU, AWS Trainium, Meta's MTIA, and custom chips developed by OpenAI in collaboration with Broadcom. These chips offer lower cost and power consumption per unit of effective compute for specific tasks.

Recent contracts illustrate the scale: Anthropic in October 2025 announced a deal to procure up to one million TPUs from Google, valued at "hundreds of billions of dollars." By April 2026, this collaboration expanded to 5 gigawatts of TPU capacity and $40 billion in Google investment. Broadcom’s AI ASIC revenue reached approximately $20 billion in fiscal 2025, with a backlog of orders worth $73 billion. Morgan Stanley has raised its 2027 TPU shipment estimate to 5 million units and its 2028 estimate to 7 million units.

But the report's central question remains: Will these cheaper chips ultimately reduce the total scale of construction, or will they be absorbed by a new wave of usage? Goldman frames this as a question of the elasticity of compute demand. In one scenario, demand is relatively fixed, and cheaper chips directly shrink the capital expenditure pie. In another, demand follows price: cheaper compute leads teams to train larger models, run longer contexts, and embed AI into more applications. The report's baseline leans toward the latter, arguing that in a phase where compute demand is far from saturated, cheaper computing breeds more usage, not less investment.

Lengthening Construction Cycles: From 'How to Build' to 'Should We Build'

The fourth assumption involves lengthening construction cycles. Backlogs for power grid connections, permitting delays, shortages of specialized labor, and extended lead times for transformers and cooling equipment (GPU delivery lead times have already stretched to 36-52 weeks) are widening the gap between capital committed and operational capacity. Goldman argues that in a baseline scenario, these bottlenecks only slow the pace of deployment, not the total volume. However, if bottlenecks become severe and persistent enough, the narrative can shift from the supply side to the demand side. When a large number of projects are simultaneously stalled, market focus can turn from "how do we build it" to "should we be building this much at all."

The report judges the current environment to be closer to the baseline scenario, but with a narrow margin for error. The combined 2026 capital expenditure guidance for the five largest hyperscale operators has risen to roughly $700 billion, more than tripling from $200 billion-plus in 2024. Capital intensity has reached 45% to 57% of revenue, making them resemble utility companies more than tech firms. In 2025 alone, these companies raised over $108 billion from bond markets, with projected bond issuance of $1.5 trillion over the coming years. At these leverage levels, execution delays can easily translate into demand-side skepticism.

The Trillion-Dollar Closed Loop: Who Ultimately Pays for AI Compute?

Adding to the complexity highlighted by Goldman’s report, another force is intensifying this cycle. According to 36Kr, global venture capital surged to roughly $300 billion in the first quarter of 2026, an all-time high, with 80% flowing to AI companies. OpenAI, Anthropic, and xAI collectively raised $173 billion, accounting for nearly 60% of global VC in the quarter.

This money flows toward one primary destination: compute infrastructure. A self-reinforcing cycle is forming: Big Tech invests in model companies, which then use that same capital to purchase compute from the investor. The money exits one pocket, circles around, and returns to the same pocket. For instance, Google is investing up to $40 billion in Anthropic, with the model company committing to pay Google Cloud $200 billion over the next five years for cloud services and TPU compute. Amazon's terms with Anthropic are similar: a commitment to invest up to $25 billion, with Anthropic pledging over $100 billion in spending on AWS over the next decade.

According to The Information, contracts from OpenAI and Anthropic now account for over half of the combined $2 trillion-plus order backlog across the four major cloud platforms. The International Monetary Fund has already issued a warning that this circular AI financing model could breed systemic risks.

The Predicament of Business Arithmetic

More pressingly, even setting aside these structural questions, the most basic business arithmetic is already failing. GitHub Copilot is a textbook example. As early as 2023, Microsoft was losing an average of over $20 per user per month, with losses on heavy users reaching $80. By April 2026, GitHub announced a complete shift to usage-based billing starting in June.

OpenAI's financials are even more challenging. ChatGPT has 900 million weekly active users and over 50 million paid subscribers, but a conversion rate of only 5.5%. Sam Altman stated publicly as early as January 2025 that even the $200-per-month Pro subscription was losing money. Market estimates project OpenAI will lose $14 billion in 2026, with cumulative losses potentially reaching $44 billion by 2028.

In contrast, Anthropic appears to be charting a different curve: roughly 80% of its revenue comes from enterprise clients, and the company’s first-quarter annualized revenue grew approximately 80-fold year-over-year, giving it a lead over OpenAI in the enterprise generative AI market share. Some market analysts believe Anthropic could approach cash-flow positivity sooner than OpenAI.

But whether it is OpenAI’s consumer scale or Anthropic’s enterprise revenue, both paths face the same fundamental problem: AI companies must prove not just that their products are used, but that the unit economics improve as usage grows. This is the most fragile point in the current AI boom. If AI agents truly embed themselves into core enterprise workflows, then cloud contracts, data centers, and chip procurement will become investments in next-generation infrastructure. But if enterprise ROI materializes slower than expected, or if AI companies can only subsidize inference costs through continuous fundraising, then the massive compute orders being placed today could quickly become a long chain of risk tomorrow.

Goldman’s report acknowledges that its analysis rests on a circular logic: if construction succeeds—if infrastructure is built, bottlenecks are cleared, and compute prices continue to fall—then the outcome may not be oversupply, but rather a new wave of demand being activated at lower price points. The very scale of today's construction, which seems sufficient for today's AI ambitions, will ultimately be the reason it falls short of tomorrow's opportunities.