Big Tech "AI Cash Burn Battle": Current scale underestimated, future depreciation underestimated, earliest price war to break out in 2027

U.S. tech giants are in an unprecedented arms race for AI infrastructure, with their capital expenditure intensity approaching the peak levels seen during the internet bubble.

According to news from the trading desk, recent research from Bank of America and Morgan Stanley shows that the market severely underestimates the true scale of current AI investments, while being inadequately prepared for the impact of future depreciation costs. The supply-demand imbalance could trigger a cloud service price war as early as 2027.

Morgan Stanley's research indicates that "hyperscale" players, including Amazon, Google, Meta, Microsoft, and Oracle, are expected to have their capital expenditure as a percentage of sales revenue reach 26% by 2027, close to the peak level of 32% during the internet bubble and exceeding the 20% seen during the shale oil boom. More critically, these publicly reported capital expenditure figures do not fully reflect the entirety of the investments, as off-balance-sheet tools like financing leases are increasingly being used to accelerate data center expansion, leading to an underestimation of the current true investment scale.

Bank of America's analysis focuses on the long-term impacts of these investments. The research report shows that the market generally underestimates future depreciation costs. By 2027, the projected depreciation for just Google, Amazon, and Meta could be nearly $16.4 billion lower than the actual situation. Bank of America also stated that if supply growth continues to outpace demand, the industry may see more aggressive pricing strategies as early as 2027.

Capital Expenditure Race: An Underestimated "Arms Race"

Morgan Stanley's report compares the current AI investment wave to two historical capital frenzies: one during the internet bubble in the telecommunications sector's fiber optic construction, and the other in the energy sector's drilling during the shale oil revolution. The report points out that the current capital intensity is nearing the peak of the former. Unlike in the past, tech giants are accelerating expansion through increasingly complex financial maneuvers, making traditional capital expenditure (Capex) data unable to fully capture the entirety of their investments.

Morgan Stanley emphasizes that two major factors have led to the underestimation of actual investment scale:

First, the rise of financing leases. Companies like Microsoft and Oracle are increasingly using financing leases to build data centers. This method is economically similar to borrowing to purchase assets, but its initial investment is typically not counted in traditional capital expenditures, thus bypassing the cash flow statement. The report found that the capital intensity of Microsoft and Oracle significantly increases when financing leases are accounted for. For example, according to Morgan Stanley's estimates, Microsoft's capital expenditure to sales ratio for fiscal year 2026 will jump from 28% to 38%, while Oracle's will soar from 41% to 58%. Additionally, these giants have signed but not yet initiated lease commitments exceeding $335 billion, indicating that this trend will continue.

Secondly, the delayed effect of "Construction in Progress": Huge investments are being accumulated on the balance sheet in the form of "Construction in Progress (CIP)." These assets will not incur depreciation until they are officially put into use, so their costs have not yet impacted the income statement. According to Morgan Stanley, the balance of construction in progress for Google, Amazon, Meta, and Oracle has seen a sharp increase over the past year; for example, Amazon grew by about 60% (USD 17 billion), and Google grew by about 40% (USD 15 billion). This means that a large amount of capital has already been spent, but its impact on profitability is just beginning.

The "time bomb" in financial reports: Wall Street underestimates future depreciation costs

If Morgan Stanley revealed the "iceberg" beneath the scale of investment, then Bank of America pointed out how these investments will translate into real cost pressures in the future. Its core view is that Wall Street is "slow to react" to the growth rate of future depreciation expenses.

Bank of America analyst Justin Post noted in the report that as Google, Meta, and Amazon increase their capital expenditures by 56% and 63% respectively in 2024 and 2025, their depreciation and amortization (D&A) expenses will inevitably accelerate in 2026 and beyond. Data shows that by 2027, Bank of America's depreciation expense forecasts for the three giants differ significantly from the general market predictions:

• Alphabet: A difference of about USD 7 billion
• Amazon: A difference of about USD 5.9 billion
• Meta: A difference of about USD 3.5 billion

A total expected difference of nearly USD 16.4 billion means that these companies' actual profitability in the future may be far lower than the current market consensus.

The report also pointed out another factor exacerbating depreciation risk: the "short lifespan" issue of AI assets.

Unlike traditional servers, hardware used for AI computing, such as GPUs, faces faster technological iterations and higher workloads, with an effective lifespan that may only be three to five years.

Bank of America noted that Amazon has shortened the expected lifespan of some servers and network equipment from six years to five years in the first quarter of 2025, citing the accelerated technological development in AI and machine learning. This is contrary to the trend in recent years where tech giants generally extended equipment lifespans to smooth expenses. Once this trend reverses, it will lead to accelerated recognition of depreciation expenses, impacting short-term profitability.

Risks and Returns: Price wars may erupt as early as 2027

Bank of America warns that the AI infrastructure market may repeat a historical pattern where aggressive investments lead to overcapacity and price pressure. As major tech companies continue to accelerate investments in AI infrastructure, there is a risk of overbuilding, meaning that the supply of computing power may exceed the demand for high-value AI services.

Additionally, the performance of large language models becoming increasingly consistent may weaken product differentiation, leading to the commoditization of infrastructure services. Meta is building multiple gigawatt-level data centers, expected to be operational between 2026 and 2029; the $500 billion Stargate project proposed by Oracle and OpenAI is expected to bring significant AI capacity between 2028 and 2029. If demand does not keep pace with the scale of supply deployment, hyperscale vendors may resort to aggressive pricing strategies to maintain utilization, thereby compressing profit margins.

Bank of America believes that if supply exceeds consumption (which it sees as likely to occur no earlier than 2027), hyperscale vendors may adopt more aggressive pricing strategies to maintain utilization, thus eroding profitability.