Is AI infrastructure burning through money so quickly—can it actually make a return?

AI infrastructure investment is entering a critical validation period—money is being burned faster and faster, but return signals remain mixed.

According to Wind Chasing Trading Desk, Bernstein’s latest quarterly tracking report on hyperscale cloud service providers points out that by Q1 2026, the four major hyperscale cloud providers (Microsoft, Amazon, Google, Oracle) will have a combined capital expenditure of about $130 billion. Bernstein estimates full-year capital expenditure will reach $623 billion, and if Meta is included, it approaches $760 billion. Meanwhile, the four companies’ combined operating cash flow this year is expected to be about $635 billion—capital expenditure and cash flow are nearly balanced, meaning most hyperscale cloud providers now need external financing to maintain their expansion pace.

Acceleration on the revenue side is not yet enough to fully dispel market doubts.

Cloud revenue growth has improved somewhat in recent quarters, but Bernstein notes that it needs to further accelerate to match the pace of capital expenditure expansion. By contrast, contract backlog (RPO) data is much brighter—combined backlog for the four major hyperscale providers has reached $2 trillion, with a stunning 176% year-on-year growth, providing some visibility support for investors. However, operating cash flow growth remains volatile, and clear signs of strong acceleration have yet to appear, providing pessimistic investors with grounds for rebuttal.

Out-of-control capital expenditure? Financing pressure has arrived

The expansion of capital expenditure by hyperscale cloud providers has reached a subtle critical point.

Bernstein estimates that in 2026, the four hyperscale cloud providers’ operating cash flow will be about $635 billion, while capital expenditure is expected to reach $623 billion, nearly offsetting each other. If capex continues to grow strongly in 2027 and cash flow lags, except for Microsoft with a bit of surplus, the other hyperscale providers will need to rely on capital market financing.

This trend has already been confirmed in practice. Amazon, Google, and Oracle have all undertaken financing activities recently. Oracle raised about $48 billion in FY2026 through combined debt and equity financing and plans to refinance about $40 billion in FY2027. Microsoft raised its 2026 calendar year capex guidance to $190 billion (including about $25 billion impact from component price increases).

Notably, the bottleneck currently restricting cloud provider expansion has shifted from GPU supply to the powering capacity of physical data centers. This shift means whoever can bring computing power online faster and at lower cost is becoming a new competitive dimension.

Diverging results; Google's unexpected surge

In Q1 2026, performance among hyperscale cloud providers showed clear differentiation.

Google Cloud, with 63% year-on-year growth, became the biggest highlight, with quarterly revenue surpassing $20 billion for the first time, and it outpaced AWS in incremental revenue (Google about $2.4 billion, AWS about $2.0 billion). Google Cloud order backlog nearly doubled quarter-on-quarter to about $462 billion, with over 50% expected to convert to revenue in the next 24 months. Google also announced direct sales of TPU hardware to client data centers, with related revenues mainly reflected in 2027.

Amazon AWS grew 28%, with annualized operating scale about $150 billion, accelerating by about 400 basis points quarter-on-quarter—the fastest pace in 15 quarters. AWS order backlog increased to $364 billion, up about 50% year-on-year, and this figure does not yet include the $100+ billion contract signed with Anthropic. Trainium chip business revenue commitments reached about $225 billion, with chip business annualized revenue exceeding $20 billion.

Microsoft Azure grew 39% at constant currency (reported at 40%), slightly beating guidance, with AI annualized recurring revenue (ARR) topping $37 billion, up 123% year-on-year. Management expects next quarter’s growth rate to stay at 39%–40% and guides for 'mild acceleration' in the second half. Oracle’s OCI grew by 93%, total RPO climbed to $638 billion, up 363% year-on-year, most of the increase from customer prepayments or hardware contracts, with limited cash consumption for Oracle itself. Alibaba Cloud revenue grew 38.2% year-on-year; AI-related products posted triple-digit growth for the 11th consecutive quarter, raising the share of external cloud revenue to 30%.

AI: Drag or Catalyst?

Profit margin trends are currently one of the most hotly contested topics in the market.

Bernstein believes that AI, at least during the infrastructure build-out phase, will depress IaaS/PaaS gross margins. The actual lifespan of GPU servers and the balance between computing supply and demand are two core uncertainties.

Looking at each company’s actual data: AWS operating margin is about 38%, up 270 basis points quarter-on-quarter, but management warns of pressure in Q2 from annual equity incentives and increased depreciation costs. Google Cloud operating margin rose to about 33%, up around 300 basis points quarter-on-quarter, and a big improvement year-on-year (last year about 17.8%), but the Wiz acquisition will drag margin by a low single-digit percentage in 2026. Microsoft doesn’t separately disclose Azure margin, Bernstein estimates its GAAP gross margin at about 56.4%, down year-on-year, though management guides for operating margin to remain flat for the year.

Oracle’s new CFO has made return on invested capital (ROIC) its core metric, stating OCI AI data center ROIC is expected to reach above 20%. Alibaba Cloud’s adjusted EBITA margin is 9.1%, management guides it will rise to low double-digits in the coming quarters and maintain a long-term EBITA margin target of 20%.

CoreWeave: Structural concerns behind high growth

As the newest emerging cloud provider tracked this quarter, CoreWeave (CRWV)’s performance reflects both the opportunities and limitations of the GPU-as-a-Service (GPUaaS) model. In Q1 2026, CoreWeave revenue reached $2.1 billion, up 112% year-on-year; order backlog expanded to $99.4 billion (about 4 times increase year-on-year), and new orders exceeded $40 billion in a single quarter.

However, profitability is seriously lagging. Adjusted operating margin is just about 0.5%, mainly because costs for data center infrastructure leasing, electricity, and depreciation are incurred once the facility is delivered, but revenue recognition can only happen after a deployment cycle of about 6–8 weeks. Management calls this a timing issue, but Bernstein points out that at the current pace of high-speed expansion, this lag could become a sustained drag. The company expects steady-state margins of 25%–30%, but the path to achieving this depends heavily on execution.

On financing, CoreWeave now has total debt over $25 billion, mainly secured by GPU hardware and client contracts. This quarter, the company completed $8.5 billion in investment-grade financing (DDTL 4.0), with an effective rate of about 7%, backed by Meta’s "as agreed" contract. Bernstein believes this financing is more a proof of concept than a replicable standard—after all, the number of hyperscale cloud providers with Aa2+ ratings is exceedingly limited.

Bernstein maintains an "underperform" rating on CoreWeave, with a target price of $67, believing that the company lacks structural advantages for long-term competition in the cloud market and doubting the long-term sustainability of its business model.

How to judge the outcome of this big bet?

Bernstein notes that current market judgments on the return of AI infrastructure investment depend largely on which set of data investors choose to believe.

Revenue growth is accelerating, but not yet enough to fully match the expansion of capital expenditure; backlog orders are surging, providing visibility on demand, but contract terms and client concentration bring certain distortions; operating cash flow growth is still volatile, and decisive strong acceleration has yet to be seen.

Looking at each company’s rating, Bernstein maintains "outperform" ratings on Microsoft (target price $646), Amazon ($315), and Oracle ($325); it rates Google ($390) and Alibaba ($180/176 HKD) as "market perform" and "outperform" respectively.

The core logic: AI-driven cloud demand is real, rapid backlog growth and accelerating AI revenue penetration provide medium-term support; but with capex still above operating cash flow and depreciation pressure increasingly evident, investors must see revenue growth accelerate further in the coming quarters to truly confirm whether this unprecedented scale of AI infrastructure investment will deliver the expected returns.

~~~~~~~~~~~~~~~~~~~~~~~~

The above content is from Wind Chasing Trading Desk.

For more detailed interpretation, including real-time commentary and frontline research, please join [Wind Chasing Trading Desk Yearly Membership]

Risk Warning and DisclaimerThe market has risks, investment requires caution. This article does not constitute personal investment advice and has not taken into consideration individual users' specific investment objectives, financial condition, or needs. Users should consider whether any opinions, viewpoints, or conclusions herein fit their personal circumstances. Invest accordingly, at your own responsibility.