A single number buried in Amazon’s financials has Wall Street rethinking the AI hierarchy. Morgan Stanley analyst Brian Nowak points to the company’s ballooning capital expenditures—projected to top $100 billion in 2025—as the critical indicator that AWS could soon narrow Microsoft’s generative AI lead. While investors fixate on Azure’s revenue sprint and its exclusive OpenAI pipeline, Nowak’s call bets on a less glamorous truth: in the infrastructure games, massive, sustained spending often wins.

Why the Market Thinks Amazon Is Losing the AI War

Two fears have punished Amazon’s stock. First, Microsoft locked in a zeitgeist-grabbing alliance with OpenAI and plastered its models across Copilot, Azure OpenAI Service, and Teams. The result is a clear, recurring revenue line tied directly to gen AI—visibility that Amazon simply hasn’t matched. Second, AWS’s top-line growth, stuck in the mid-teens, has lagged the faster acceleration Google Cloud and Azure posted in recent quarters. Even when Amazon beat core earnings, disappointing AWS guidance triggered a sell-off. In that vacuum, the narrative hardened: Amazon is behind, and its blizzard of AI announcements—Bedrock, Titan, Trainium, Anthropic—hasn’t yet gelled into a monetization machine.

The Capex Thesis: Spend Big, Then Grow Big

Nowak’s contrarian lens flips the worry. He isn’t looking at today’s AI service catalog; he’s measuring the physical layer beneath it. Hyperscalers are racing to erect AI-optimized data centers—facilities packed with custom networking, dense GPU and accelerator clusters, and cooling systems tailored for sustained training runs. Amazon’s capex acceleration in 2025, with the vast majority earmarked for AWS, suggests a buildout so large that it implies management sees a tidal wave of future workloads. When Morgan Stanley overlays projected data center square footage additions—roughly 8.5 million square feet in 2025 and 10 million in 2026, per filings and third-party datasets—the math hints at a revenue re-acceleration once that capacity comes online and customers shift AI from pilots to production.

Crunching the Numbers: $100B+ Amazon vs. $80B Microsoft

The dollar sums are staggering. Amazon has signaled its 2025 capex will surge well into the tens of billions, with several industry reports and company commentary placing the figure north of $100 billion when counting property, equipment, logistics, and cloud assets. Microsoft’s own fiscal 2025 plan for AI infrastructure hovers around $80 billion, by its own disclosures. Both numbers are independently reported across financial outlets and earnings calls—they’re not analyst fantasy. The difference is that Azure’s AI revenue is already surfacing in per-seat subscriptions and consumption models, while AWS’s return on its spend is largely future-tense.

Warehouse-sized capacity data matters, too. Analysts parsing AWS project filings, campus proposals, and lease activity forecast 8.5 million square feet of incremental data center space this year, followed by 10 million square feet in 2026. Those are modeled estimates, not guarantees, but they align with the aggressive build-out pattern visible in Virginia, North Carolina, and other key regions. Meanwhile, AWS’s trailing growth rate sits in the mid-teens, compared with the higher double-digit spurts Microsoft and Google have flaunted during the gen-AI boom. That gap is precisely the opening Nowak thinks capex can close.

Inside Amazon’s AI Arsenal: More Than Just Capacity

Amazon’s strategy is infrastructure-first but feature-rich. The centerpiece is Amazon Bedrock, a managed service that exposes multiple foundation models—including third-party options—through a single API, letting enterprises pick the best model while keeping data within their AWS environment. Bedrock supports Amazon’s own Titan models and the newer Nova family, both positioned for customization. On the silicon front, AWS’s homegrown Trainium and Inferentia chips promise lower cost-per-epoch than GPU-heavy alternatives, with Amazon touting significant price-performance gains for specific model types. The multibillion-dollar anchor investment in Anthropic further tethers a leading model provider to AWS’s ecosystem, driving both training and inference workloads. Developer tooling—CodeWhisperer for code suggestions, QuickSight generative BI features, and tight integration with existing data services—rounds out a portfolio designed to pull enterprise data into gen-AI applications without forcing customers to leave the AWS fold.

Strengths That Make the Thesis Plausible

Amazon’s hand holds several durable cards. The balance sheet can underwrite multi‑year capex at a scale only a handful of companies can match, enabling capacity growth without immediate payback pressure. Owning the full stack—chips, data centers, cloud services—cuts reliance on external vendors and can compress inference and training costs over time. Bedrock’s multi-model approach appeals to enterprises wary of single-provider lock-in, while AWS’s entrenched position with Fortune‑scale companies gives it a captive migration audience once AI hits production. Additionally, the data center pipeline isn’t theoretical; public filings and project plans show concrete, region-by-region expansion that underpins analyst models.

Real Risks That Could Torpedo the Bet

For all its muscle, AWS faces obstacles that even $100 billion can’t instantly solve. Timing is the biggest wild card: physical floors can be poured in quarters; customer migration to production-scale AI workloads can take years. Underutilized infrastructure becomes a capital drag. Supply chain snarls—chip lead times, rack availability, regional power constraints—can delay ramp-ups or reduce yield per square foot. Software lock-in poses another hurdle; the majority of ML workloads today are CUDA-optimized for NVIDIA GPUs, and convincing developers to retool for Trainium requires robust libraries, tooling, and a frictionless path that hasn’t fully materialized. Meanwhile, Microsoft’s product bundling is already sticky, Google’s Gemini and end-to-end tooling add pressure, and massive capex can squeeze free cash flow in the near term. Regulatory headwinds around AI safety and data privacy further complicate enterprise adoption, particularly in regulated industries.

Amazon vs. Microsoft: Two Radically Different Playbooks

Comparing the rivals reveals a divide in go-to-market philosophy rather than a simple “who spends more” contest. Microsoft leans on an exclusive OpenAI relationship, deep integration into Office and Teams, and a subscription‑based revenue model that makes AI consumption predictable and visible. Amazon, by contrast, champions model neutrality via Bedrock, invests heavily in custom silicon to drive down unit economics, and positions itself as the infrastructure layer for a multi‑model world. Both are building aggressively, but Microsoft anchors customer value in product productivity, while AWS bets that owning the cost curve will win with enterprises that prioritize flexibility and long‑term total cost of ownership.

The Investor’s Dilemma: Short‑Term Pain, Long‑Term Payoff

Wall Street’s current math punishes uncertainty. Microsoft’s AI revenue is tangible today; Amazon’s resides largely in a future that might not arrive on schedule. The Morgan Stanley framework treats capex as a leading indicator—a signal that AWS’s growth could re-accelerate to the high-teens or even 20%+ once capacity is live and enterprises graduate from proofs of concept. But that optimism is coupled with a bifurcated risk. If demand ramps on the timeline Nowak envisions, AWS reclaims momentum and validates the spend. If enterprises consolidate around Microsoft’s packaged AI or Google’s ecosystem, Amazon could face a prolonged drag on margins and returns. The market will withhold the benefit of the doubt until that utilization becomes visible.

What to Watch: Lead Indicators for a Reversal

Investors and IT decision‑makers should track a handful of concrete signals beyond capex headlines:

  • Enterprise production deployments: public case studies citing Bedrock or AWS as the backbone for large-scale AI workloads, not just pilots.
  • Trainium adoption: developer tooling releases, library support, and migration stories that demonstrate practical performance parity or cost advantages.
  • AWS infrastructure margins: rising gross margins on cloud services would suggest better pricing power or higher utilization of that added capacity.
  • Revenue per square foot: any disclosures reconciling added data center space to realized cloud revenue would validate the Morgan Stanley framework.
  • Strategic tie‑ups: new long‑term investments or exclusivity deals that lock major model builders onto AWS will be strong tailwinds.

Cautious Optimism, Hard Execution

Amazon’s AI story isn’t a fable of instant triumph. It’s a long‑duration bet on the idea that infrastructure depth matters more than product-led flash in the early innings of generative AI. The bull case is credible precisely because the capex numbers are too big to ignore and because few rivals can fund a buildout of this magnitude. Yet the path from square footage to recurring, high‑margin revenue is littered with execution risks—developer friction, competitive bundling, and the stubborn reality that capacity alone doesn’t create demand. For now, the $100 billion signal is real: a starting gun, not a finish line. The next 12 to 24 months will determine whether it becomes a durable moat or an expensive lesson in overinvestment.