In a market where cloud margins are generally compressing, Dylan Patel makes a startling claim: Amazon Web Services has not only stabilized but accelerated its profitability by betting on a specific, high-risk distribution model for artificial intelligence. While competitors are stuck in the low-margin business of renting raw compute power, Patel argues that the executive branch of the tech industry—specifically Amazon—has cracked the code on "Token-as-a-Service," turning AI inference into a high-margin revenue stream that dwarfs traditional infrastructure deals. This is not just a quarterly earnings beat; it is a structural shift in how the most powerful cloud providers are monetizing the AI boom.
The Margin Inflection
Patel's analysis cuts through the noise of generic AI hype to focus on the hard economics of cloud profitability. He notes that while peers like Oracle and Coreweave have disappointed investors with shrinking profits, and Azure is seeing a downward trend, AWS has achieved a rare inflection point. "While other CSPs have seen declining-to-flat operating margins over the last several quarters, Amazon's AWS margins inflected this past quarter driven primarily by customer spending growth on Claude through Bedrock," Patel writes. This is a crucial distinction. The author isn't just reporting a number; he is identifying a specific product mix—Bedrock and its partnership with Anthropic—as the engine of this growth.
The core of Patel's argument rests on the idea that Amazon's strategy is fundamentally different from its rivals. He points out that AWS has a unique "risk appetite," securing massive power purchase agreements (PPAs) to ensure energy access, a constraint that has become the primary bottleneck for the entire industry. This connects to broader industry dynamics where energy availability, not just chip supply, dictates market share. Patel observes, "Amazon has secured more power than any other cloud provider besides Google, understanding before others that energy drives market share in this constrained environment, and that requires capital and multibillion dollar PPAs." This foresight allowed Amazon to build capacity faster than anyone else, a move that pays dividends now as demand outstrips supply.
The only CSP with a true rising trend is AWS – a remarkable achievement considering their server depreciation (5yrs) is the lowest of all CSPs.
Patel's framing here is sharp. He highlights a counterintuitive fact: Amazon uses a five-year depreciation schedule for its servers, the shortest among major providers. Typically, this would suggest lower margins due to higher capital expenditures. Yet, because of the high-margin nature of their new AI business mix, they are outperforming despite this accounting reality. Critics might note that aggressive depreciation can mask long-term asset health issues, but in the context of rapid technological obsolescence in AI, Patel's point holds: speed and volume are currently more valuable than long-term asset stability.
The Economics of Distribution
The most distinctive part of Patel's coverage is his deep dive into the "Token-as-a-Service" (TaaS) model. He explains that the economics of selling AI tokens are vastly different depending on whether the cloud provider owns the model or distributes it. In the distribution model, which Amazon has perfected with Anthropic, the cloud provider acts as a marketplace. "As Seller, Anthropic books full revenue of the sold tokens. As computer and marketplace provider, AWS gets both an infrastructure fee (akin to an EC2 IaaS fee) and a distribution or revenue share fee," Patel explains. This revenue share is the secret sauce that boosts margins.
This structure allows AWS to avoid the heavy capital burden of owning the intellectual property while still capturing a significant slice of the value chain. Patel argues that this is a superior position to the traditional Infrastructure-as-a-Service (IaaS) model, where providers are essentially renting out metal at thin margins. "The margin mix and distribution advantage these hyperscalers now possess is significantly widening the gap amongst the top hyperscalers and the rest of the field," he writes. This is a powerful insight for investors and strategists: the winner of the AI race may not be the one with the best model, but the one with the best distribution deal.
Patel also touches on the vertical integration of hardware, noting that Amazon's custom chips, Trainium and Graviton, are now deeply embedded in this ecosystem. "Our AWS Trainium chips, designed in-house for AI workloads, now power more than 50% of Amazon Bedrock token usage," he quotes AWS CEO Matt Garman. This integration reduces reliance on external suppliers like Nvidia and improves cost efficiency. The author connects this to a broader trend of CPU constraints, referencing previous deep dives on how Reinforcement Learning is driving a surge in CPU demand. By controlling both the software platform (Bedrock) and the underlying silicon, Amazon creates a moat that competitors struggle to cross.
The Anthropic Advantage
Patel identifies the partnership with Anthropic as the linchpin of AWS's success. He details how Anthropic's revenue has exploded, driven largely by enterprise adoption of Claude Code. "Anthropic was the first AI Lab to roll this out with AWS and Google, recently followed by OpenAI with AWS," Patel notes. The deal structure is key: it involves a flat infrastructure fee plus a revenue share, creating a symbiotic relationship where both parties benefit from growth. Patel estimates that Bedrock is now a $5.5 billion run-rate business, with the vast majority of customers using Anthropic models.
The author contrasts this with Microsoft and Google. While Microsoft has a strong position with OpenAI, its AI business remains heavily skewed toward IaaS, which dilutes margins. Google, meanwhile, has a strong model with Gemini but faces accounting complexities where training costs are not fully reflected in their cloud segment margins. "Google Cloud has had a great upwards climb recently, but margins are inflated since they do not include training costs from DeepMind in the GCP segment," Patel writes. This nuance is vital for anyone trying to compare the financial health of these giants. It suggests that AWS's reported margin expansion is more organic and sustainable than it might appear on the surface.
Bedrock has steadily grown as a % of total AWS AI revenue to 37% today, up from 9% in 1Q25 when IaaS dominated the AI business.
Patel's data shows a dramatic shift in the revenue mix. In just one year, the high-margin TaaS business has gone from a niche offering to the dominant force in AWS's AI portfolio. This rapid transition is what has driven the 213 basis point increase in EBIT margins quarter-over-quarter. The author's use of the Tokenomics 2.0 model to triangulate these figures adds a layer of credibility that goes beyond standard analyst estimates. He breaks down the revenue per megawatt of compute, showing that Anthropic's performance on Bedrock is generating significantly higher returns than traditional IaaS deals.
Bottom Line
Dylan Patel's analysis provides a compelling, data-rich argument that AWS has successfully pivoted from a commodity infrastructure provider to a high-margin AI distributor. The strongest part of this argument is the detailed breakdown of the "Token-as-a-Service" economics, which explains why AWS is outperforming peers despite a shorter depreciation schedule and intense competition. However, the argument relies heavily on the continued success of the Anthropic partnership; if that relationship sours or if Anthropic's growth stalls, the margin boost could evaporate. Readers should watch closely for how Amazon replicates this success with other model providers and whether the custom silicon strategy can keep pace with the rapidly evolving demands of AI inference.