Anthropic Growth and Bedrock Mix Drive AWS Margins Higher While Peers Lag

Redundant with selected · 4

• evidenceBedrock is a $5.5B run-rate business overwhelmingly powered by Anthropicc 0.80 · sim 0.83

  Bedrock has grown from 9% of AWS AI revenue in 1Q25 to 37% today, with Q/Q growth of 60% in 4Q25 and 170% in 1Q26. 80–90%+ of Bedrock customers are using Anthropic models.

  overlapped with: Anthropic-on-Bedrock is the dominant driver of the margin step-up

• evidenceHyperscaler TaaS businesses are now $10B+ while Oracle and neoclouds have nothingc 0.70 · sim 0.85

  AMZN, MSFT, and GOOGL each run $10B+ ARR token-as-a-service businesses, whereas Oracle and the neoclouds are at essentially zero. This mix gap is what's widening the margin spread between top hyperscalers and the rest.

  overlapped with: AWS is uniquely token-as-a-service rather than IaaS-led

• caveatTaaS deals carry more demand risk than take-or-pay IaaSc 0.55 · sim 0.86

  Unlike a 5-year take-or-pay IaaS contract, Bedrock revenue depends on actual token consumption. AWS accepts that volume risk in exchange for materially higher margins — a bet that has paid off because of Anthropic's growth.

  overlapped with: Bedrock's distribution model gives AWS infra fee plus revenue share

• contextAWS achieves rising margins despite the most conservative depreciationc 0.50 · sim 0.84

  AWS's server depreciation life is the shortest among CSPs at 5 years, which mechanically depresses reported margins — making the rising trend more striking than it appears on the surface.

  overlapped with: AWS margins inflected while peer CSP margins stalled or fell

Below top-k · 16

• contextGoogle Cloud's margin gains are inflated by accounting bucketingc 0.80

  GCP margins look great partly because DeepMind/Gemini training costs sit in Alphabet-Level Activities (~$5.4B in 1Q26, up from $3.0B), while Gemini API revenue flows into GCP. The result is effectively an 'EBTIT' margin that overstates true cloud profitability.

• claimAccess to frontier LLMs is the real differentiator in API endpointsc 0.75

  Model library breadth and headline pricing get most of the marketing attention, but Tokenomics data shows frontier LLMs generate the vast majority of API industry revenue. Whoever can resell OpenAI, Claude, and Gemini wins.

• evidenceAWS will out-build every rival on capacity through 2027c 0.70

  No other provider will add more datacenter capacity than AWS in 2025, 2026, or 2027. Microsoft is close in 2024–26 but is dwarfed by 2027, especially after its year-long datacenter pause.

• mechanismCustom silicon compounds the Bedrock margin advantagec 0.70

  Graviton CPUs already give Amazon better economics than merchant chips, and Trainium is strong for inference and RL workloads — the exact workloads Bedrock runs. Because Bedrock abstracts hardware from customers, AWS can route token traffic to its in-house chips with no porting friction for the buyer.

• evidenceBedrock EBIT margins around 55% at current Anthropic ARR densityc 0.70

  At an estimated $26M of Anthropic Bedrock revenue per MW in 1Q26, Bedrock EBIT margins land near 55%, with 2Q tracking toward $42M per MW and Bedrock rising to 53% of AWS AI revenue — adding roughly 9 points to total AWS revenue growth.

• claimGoogle is supply-constrained across too many fronts to mirror AWSc 0.70

  Google is trying to compete with AWS, Nvidia, Anthropic, OpenAI, Meta, and Tesla simultaneously. Its capacity buildout is enough to feed an AI lab but not enough to scale a Bedrock-style third-party token business — Gemini Enterprise Agent Platform is getting far less capacity than Bedrock.

• mechanismPower procurement as the real AI infrastructure moatc 0.65

  Amazon has secured more power than any cloud provider except Google, recognizing earlier than peers that energy — not GPUs — is the binding constraint and that capturing it requires aggressive multibillion-dollar PPAs with IPPs like Talen, Vistra, and NiSource.

• contextMicrosoft's compute is largely pre-sold to OpenAIc 0.65

  Most of Microsoft's AI capacity is committed to OpenAI under long-term contracts — OpenAI's backlog alone is 2.5x Azure's annual revenue — leaving little room for Azure to scale a Bedrock-equivalent TaaS business on its own customer base.

• contextOnly three CSPs sell all three frontier model familiesc 0.60

  AWS now resells OpenAI alongside Claude, while Microsoft added Claude alongside OpenAI, and Google has Claude and Gemini. No CSP outside this trio can sell OpenAI, Claude, and Gemini tokens.

• evidenceTrainium already powers most Bedrock token usagec 0.55

  AWS CEO Matt Garman stated in November 2025 that in-house Trainium chips power more than 50% of Amazon Bedrock token usage, validating the custom-silicon-meets-TaaS flywheel.

• caveatMicrosoft will have to rent from neoclouds to keep up, eroding marginsc 0.55

  After a year-long datacenter pause and slow large-cluster builds in Wisconsin, Microsoft's only path to matching AWS's 2027 capacity is contracting from neoclouds at higher cost — which directly compresses the margin advantage it needs to defend.

• implicationGCP functions as an upsell funnel rather than a standalone TaaS enginec 0.55

  Google's pattern, exemplified by Meta, is to use a large GPU deal as the entry point that leads to Gemini adoption and then a large TPU hardware purchase. GCP's role is distribution for hardware and models, not the high-margin third-party token marketplace AWS has built.

• exampleAnthropic's coding-driven API surge has no Gemini equivalentc 0.50

  Anthropic API revenue is up roughly 13x year-over-year on the back of coding-agent demand. Google's Gemini API has done well but has not captured the same coding wave, and Microsoft's CoPilot products have not gained traction.

• contextAmazon was late to AI but has reversed its trajectoryc 0.45

  Amazon was widely written off as an AI loser as recently as 2023, but is now combining accelerating revenue growth with outperforming margins — a combination no other CSP currently has.

• contextAnthropic pioneered the model-distribution structure with hyperscalersc 0.40

  Anthropic was the first AI lab to roll out this hosted-distribution structure, first with AWS and Google and now followed by OpenAI with AWS. The design keeps model weights tightly controlled while letting the CSP monetize compute and distribution.

• caveatGCP margin lift may include one-off Anthropic TPU royaltiesc 0.40

  Some of Google Cloud's recent margin strength may reflect one-time royalty payments from the sale of TPUs to Anthropic via Broadcom, with Google acting as an IP vendor rather than a recurring cloud customer relationship.