Amazon spent roughly $83 billion on capital expenditures in fiscal 2025, according to its annual report filed with the SEC. That is more than the entire GDP of Luxembourg. And CEO Andy Jassy wants to spend even more this year, with the bulk of new investment flowing into data centers, custom silicon, and AI partnerships that Amazon believes will define the next decade of its business. The problem, as of May 2026, is that no one outside the company can say with precision how much of that money is actually paying off.
The scale of the bet
Amazon’s 2025 10-K filing tells the story in raw numbers. Property and equipment purchases surged year over year, depreciation expenses climbed in lockstep, and the company’s risk-factor disclosures now flag “new and emerging technologies” as a source of material uncertainty. Those disclosures are worth reading carefully. They represent a legal and accounting judgment that the AI-driven buildout carries outcomes Amazon itself cannot reliably predict.
The spending goes well beyond servers and real estate. Amazon has committed up to $8 billion to Anthropic, the AI safety company behind the Claude family of models. Anthropic, which was valued at roughly $18 billion at the time of Amazon’s initial investment, competes directly with OpenAI and Google DeepMind in the race to build frontier AI systems. The deal, structured across multiple tranches starting in late 2023, gives Amazon Web Services a privileged relationship with one of the most capable foundation-model developers in the world. It also ties a significant sum to a startup that, despite rapid growth, does not yet generate revenue at the scale traditional investors would demand to justify the price tag.
Amazon is hardly alone in this race. Microsoft has disclosed plans to spend more than $80 billion on AI-capable data centers in its fiscal 2025 alone. Alphabet and Meta have each outlined tens of billions in AI infrastructure commitments. But Amazon’s position is distinct because it is funding two enormous cost centers at once: the cloud computing business (AWS) and a retail and logistics operation that is absorbing AI tools for everything from demand forecasting to warehouse robotics.
Signs of a selective pullback
Even as Jassy has publicly defended the pace of investment, there are signs Amazon is being choosier about where the money goes. In April 2025, Wells Fargo analysts reported that the company had paused some data center leasing negotiations, according to Reuters. The report landed just as Jassy was telling shareholders at Amazon’s annual meeting that the capital outlays were necessary to secure long-term competitiveness.
Amazon has not publicly confirmed or denied the characterization. The reasons behind any pause could range from updated demand forecasts for cloud capacity to hardball negotiation tactics in a data center real estate market where landlords have been pushing rents higher. It could also reflect a genuine reassessment of how quickly enterprise AI workloads will materialize at the scale Amazon has been planning for.
What the report does suggest is that the buildout is not running on autopilot. Decisions are being made inside Amazon about where to accelerate and where to slow down, and those decisions are happening in real time as the company watches how customers actually adopt AI services.
Regulators are watching the partnerships
The spending has also attracted federal scrutiny. In January 2024, the Federal Trade Commission launched a formal inquiry into generative AI investments and partnerships, issuing compulsory 6(b) orders to several major technology companies. Amazon and Anthropic were among those named.
The FTC’s concern is straightforward: whether large cloud providers are using investment deals to gain outsized influence over foundational AI model development, potentially distorting competition before the market has had a chance to take shape. The agency’s use of compulsory orders, rather than voluntary requests, signaled it considered the matter serious enough to invoke statutory authority. In early 2025, FTC staff published a report summarizing initial findings, noting that the partnerships it examined involved complex webs of financial and commercial arrangements that could affect competitive dynamics.
As of spring 2026, no enforcement actions have been announced. That leaves the regulatory risk in a gray zone: procedurally real and documented, but with no visible resolution. For investors, the practical takeaway is that Amazon’s Anthropic relationship and similar arrangements could face constraints that are impossible to price until the agency acts or formally closes the matter.
What investors still cannot see
The core frustration for anyone trying to value Amazon’s AI push is a disclosure gap. The 10-K reports total capital expenditures and depreciation, but it does not break out how much flows specifically to AI infrastructure versus fulfillment centers, logistics hubs, or traditional cloud capacity that would exist regardless of the current AI boom. Analysts are left building estimates from fragments: earnings call commentary, third-party supply chain data, and lease filings in local jurisdictions where Amazon is constructing new facilities.
On the revenue side, AWS reported $107.6 billion in net sales for 2025, a 19% increase over the prior year, according to Amazon’s 10-K. When Amazon released its Q1 2026 earnings in late April 2026, Jassy highlighted accelerating demand for AI services on AWS, though the company again declined to break out AI-specific revenue as a standalone figure. Company leadership has attributed a meaningful portion of recent growth to AI-related services, including Bedrock, its managed platform for foundation models, and custom Trainium and Inferentia chips designed to reduce the cost of training and running AI workloads. But “meaningful portion” is not a number. Until Amazon provides segment-level disclosure that isolates AI-driven revenue and margin contribution, investors are underwriting a thesis rather than verifying a result.
Jassy’s public comments offer directional guidance. On the Q1 2026 earnings call, he pointed to growing enterprise adoption of Bedrock and cited customers using Amazon’s platform to deploy AI agents for tasks like customer service automation and internal knowledge retrieval. But shareholder presentations are designed to build confidence, not to substitute for audited financials. The gap between executive optimism and measurable outcomes is exactly where investor risk concentrates.
Where the receipts have to come from
Amazon’s AI infrastructure push fits a pattern that has repeated across technology cycles: massive upfront spending on capacity the company believes demand will eventually fill. The difference this time is the sheer dollar volume, the speed of the buildout, and the fact that regulators are examining the competitive structure of the AI market while it is still forming.
The strongest evidence available, drawn from Amazon’s own SEC filings, confirms that capital expenditures are large and rising, that the company’s legal team views the trajectory as carrying material risk, and that depreciation costs will weigh on earnings for years regardless of how quickly AI revenue scales. The FTC inquiry adds a regulatory dimension that remains unresolved. And the Wells Fargo report on paused leasing talks suggests the company is making tactical adjustments even as it maintains a strategically aggressive posture.
For shareholders, the near-term picture depends on a handful of concrete signals: incremental disclosures in future SEC filings, any public updates from regulators, AWS growth rates in upcoming quarters, and whether Amazon begins to break out AI-specific financial metrics. The company has placed one of the largest bets in corporate history on the premise that artificial intelligence will reshape how people shop, how businesses run, and how computing itself works. The spending is visible. The returns are not, at least not yet.
