Amazon's $33 billion commitment to Anthropic signals a fundamental shift in how the tech industry operates: compute has become so scarce that companies are now pre-paying for infrastructure years before it's built. The deal guarantees Anthropic access to 5 gigawatts of computing capacity over the next decade, but here's the catch: most of those chips haven't been manufactured yet, the power plants don't exist, and the data centers themselves are still on drawing boards.

This represents a structurally different kind of deal from anything the tech industry has seen at scale. Amazon isn't simply funding Anthropic's operations. Instead, it's pre-committing capacity it plans to construct, betting that Anthropic's demand will absorb it once it comes online. The money moves now because the capacity won't exist later unless someone agrees to pay for it before it's built.

Why Is Computing Capacity So Hard to Build?

The demand for AI computing has shattered the traditional cloud economics model. For most of the past decade, compute was abundant and treated like a commodity. Companies rented capacity from hyperscalers, and supply grew incrementally to match demand. That model broke when artificial intelligence arrived.

Anthropic's revenue tripled in just four months, reaching a $30 billion annualized run rate by April 2026, up from $9 billion at the end of 2025. Over 1,000 enterprise customers now spend more than $1 million annually on Claude, the company's AI assistant, up from 500 customers in February. Eight of the Fortune 10 are now customers. And Anthropic isn't even the largest AI player in the market.

The supply side of computing infrastructure faces four major constraints, each taking years to resolve:

• Chip Fabrication: Taiwan Semiconductor Manufacturing Company (TSMC), which produces advanced chips for Nvidia, Amazon, Google, and Microsoft, has its advanced packaging capacity booked into 2027, creating a bottleneck for every AI chip manufacturer competing for the same wafers.
• Power Generation: A single 1-gigawatt data center requires the power output of a nuclear reactor, and utilities measure their response times in years, not quarters, making rapid expansion nearly impossible.
• Data Center Construction: Even with chips and power secured, the physical buildings take 18 to 24 months minimum to construct at scale, creating unavoidable delays.
• Custom Silicon Design: Amazon's Trainium chips, Google's TPUs (Tensor Processing Units), and Microsoft's Maia all require multi-year design cycles before volume production can begin.

When demand doubles every few months and supply takes three to five years to meet it, the economics change completely. Compute stops being a commodity and becomes a scarce resource. Scarce resources get locked in by whoever has the capital and credibility to commit early.

How Are Hyperscalers Reshaping AI Distribution?

Anthropic's strategy reveals something equally important: the company isn't hedging its bets across cloud providers. Instead, it's pursuing a deliberate distribution strategy by securing infrastructure relationships with all three major hyperscalers simultaneously.

A Fortune 500 company running on Amazon Web Services (AWS) won't switch to Microsoft Azure just to access Claude. Its security reviews, compliance audits, data residency requirements, identity layers, and existing deployment pipelines are all built around AWS. The same applies to Microsoft and Google Cloud customers. The cloud provider isn't a vendor you swap; it's a platform you build on top of.

Anthropic's infrastructure commitments span:

• Amazon/AWS: $33 billion investment, $100 billion AWS spending commitment, 5 gigawatts of Trainium capacity, and native Claude integration into AWS Bedrock for over 100,000 existing enterprise customers.
• Google/Broadcom: 3.5 gigawatts of next-generation TPU capacity coming online in 2027, extending an earlier October 2025 agreement, with Claude available through Google Cloud's Vertex AI platform.
• Microsoft/Nvidia: Approximately $15 billion invested in January 2026 at a $350 billion valuation, with Anthropic committing around $30 billion of Azure capacity and Claude available through Microsoft 365 Copilot and Azure services.

Each hyperscaler has a distinct customer base that Anthropic cannot reach any other way. When Anthropic signs with AWS, it's buying distribution into thousands of enterprises locked into the AWS ecosystem. When it signs with Google, it gains access to Google Cloud's enterprise base. When it signs with Microsoft, it reaches the Microsoft 365 customer base, measured in hundreds of millions of seats.

Is the Demand for AI Infrastructure Real or Overstated?

While the infrastructure deals are massive, significant questions linger about whether the underlying demand will actually materialize. Between 30 and 50 percent of planned US data center builds for 2026 are projected to be delayed or canceled, driven by power shortages, equipment backlogs, and supply chain constraints.

Microsoft Chief Executive Officer Satya Nadella made a notable public admission that "there will be an overbuild" of AI infrastructure, a striking statement from a company that had simultaneously announced $80 billion in data center spending for 2025. Satellite imagery of major announced projects, including OpenAI-linked campuses in Texas, shows construction running months or years behind announced schedules.

Data center developers routinely announce many projects in parallel to test which can clear local regulations and secure power fastest. This has created a forecasting problem: utilities and grid operators struggle to determine how much load will actually come online, flooded with speculative or "phantom" load requests that may never materialize. Industry analysts have coined the term "bragowatts" for announced capacity that never materializes as actual load.

The governance problem runs deeper. Utilities in most regulated markets earn a fixed rate of return on capital investment, giving them financial incentive to build infrastructure whether or not the demand justifies it. When Duke Energy Carolinas announced that customer energy needs over the next 15 years would grow at eight times the growth rate of the prior 15 years, it was responding to developer requests, not measured consumption. When Southern Company extended the operating lives of coal plants it had committed to retiring, it cited data center demand projections it had no independent way to verify.

"When those projections prove wrong, when the data centers don't come, or come at half the projected scale, or shift to a different region after local opposition blocks the original site, the cost of the overbuilt infrastructure is not absorbed by the companies whose announcements drove the investment, but is passed to ratepayers," noted Dan Vermeer, energy analyst.

Dan Vermeer, Energy Analyst

Community opposition alone blocked or delayed at least $156 billion in data center projects in 2025. Every project that gets announced, drives grid investment, and then gets canceled or delayed leaves a trail of infrastructure costs that someone has to pay, but it is not the hyperscalers.

What Does This Mean for the Power Grid and Clean Energy Goals?

The infrastructure buildout is creating genuine challenges for grid operators. If the demand is real and the delays are structural, transformer lead times have stretched from two years to five, while data center deployment cycles run under 18 months, then the infrastructure won't be ready when demand arrives, and reliability consequences fall on everyone connected to the grid.

Meanwhile, every major hyperscaler has announced ambitious renewable energy goals. Google's 24/7 carbon-free energy initiative attempts to match consumption with clean generation on an hourly basis in the same location, representing a genuine advance over standard practice. Typically, companies match annual electricity consumption against renewable energy certificates that may have been generated at a different time, in a different place, and on a different grid.

Yet the same Southern Company that extended coal plant lifetimes to serve data center demand also has hyperscaler customers claiming to run on clean energy. These facts can both be true simultaneously under current accounting standards, but they cannot both be true simultaneously on the actual grid.

A structural irony sits at the center of this story. When DeepSeek released its AI model and claimed it required dramatically less energy per query than comparable American systems, the industry's response wasn't to scale back infrastructure plans. It was to announce more of them, on the logic that cheaper queries enable more applications, which enable more demand, which requires more compute. This is what energy experts call the "Jevons Paradox" applied to artificial intelligence: efficiency gains that reduce resource consumption per unit often lead to increased total consumption because lower costs drive higher demand.

How Are Equipment Manufacturers Responding to the Boom?

Equipment manufacturers are racing to capitalize on the infrastructure surge. GE Vernova, the power and electrification company spun off from General Electric, reported extraordinary growth in its first quarter of 2026, with total orders reaching $18.3 billion, up 71 percent year over year.

The company's backlog hit $163 billion, including a sequential increase of $13 billion, with management targeting $200 billion in backlog by 2027, a year earlier than previously expected. Free cash flow nearly doubled previous annual results in just one quarter, reaching $4.8 billion, exceeding the prior full year's free cash flow of $3.7 billion.

Data center orders are driving much of this growth. GE Vernova's Electrification segment, which supplies power distribution equipment for data centers, booked $2.4 billion in data center orders in the first quarter alone, exceeding the entire full-year 2025 orders for this vertical. The company's Gas Power segment signed 21 gigawatts of new agreements in the quarter, raising total gigawatts under contract to 100, spanning 90 customers in 24 countries, with 20 percent tied directly to data centers.

GE Vernova raised its 2026 adjusted EBITDA (earnings before interest, taxes, depreciation, and amortization) margin outlook to 12 to 14 percent, with Power segment EBITDA margin guided to 17 to 19 percent and Electrification margin to 18 to 20 percent, all reflecting pricing and productivity improvements driven by strong demand.

The company also noted progress on small modular reactors (SMRs) and energy management systems, with regulatory progress on Ontario Power Generation's Darlington Unit 1 and new North American and European pipeline developments. SMR-related government support now totals up to $40 billion for projects in the US and Japan.

Key Takeaways for Understanding AI Infrastructure

• Compute Scarcity: AI demand is growing so fast that supply cannot keep pace, transforming computing from a commodity into a scarce resource that companies must lock in years in advance.
• Distribution Over Funding: Amazon's $33 billion deal with Anthropic is fundamentally about securing distribution into enterprise customers, not simply funding the AI company's operations.
• Forecasting Uncertainty: Between 30 and 50 percent of planned US data center builds for 2026 face delays or cancellation, suggesting announced capacity commitments may significantly exceed actual demand.
• Grid and Accountability Gaps: Utilities lack independent verification of demand projections and face financial incentives to build infrastructure regardless of whether demand materializes, creating cost risks for ratepayers.
• Equipment Manufacturer Opportunity: Companies like GE Vernova are experiencing extraordinary growth, with data center orders alone exceeding prior full-year performance in single quarters.