Small print.
Wide open.

Table of Contents

This customer momentum is driven by significant usage expansion as customers realize the value of deploying 
Claude  across  their  organizations.  The  value  that  Claude  delivers  to  customers  has  helped  drive  our  Net 
Revenue Retention of approximately 900% as of September 30, 2026.

The  growth  in  our  business  reflects  the  degree  to  which  enterprises  are  embedding  Claude  directly  into  their 
business-critical  workflows.  Despite  these  trends,  adoption  across  the  global  economy  remains  nascent.  As 
Claude’s capabilities continue to advance, we believe the addressable opportunity will expand, and Claude will 
continue its rapid diffusion across more enterprises.

Compute Efficiency

Compute is the most significant operating cost in our business, and compute efficiency is a critical determinant 
of  our  operating  efficiency  and  revenue.  Our  research-driven  algorithmic  improvements  in  training  and 
inference  optimization  and  multi-chip  compute  architecture have  combined  to  dramatically  increase  our 
compute  efficiency  as  our  revenue  has  scaled.  The  result  is  that  we  are  able  to  continue  increasing  the 
intelligence  our  research  can  produce  and  the  revenue  our  business  can  generate  per  dollar  of  compute 
investment. 

We  define  compute  efficiency  as  maximizing  the  intelligence  produced  and  revenue  generated  per  dollar  of 
compute invested. Three pillars drive our compute efficiency:

•

•

Training Efficiency. Our researchers regularly innovate compute multipliers. Increasingly, Claude 
itself assists in the research and engineering that produce each successive version of our models, which 
is  a  recursive  dynamic  we  refer  to  as  “closing  the  loop,”  with  successive  versions  of  our  models 
accelerating the development of the next and compounding these gains across model generations. 

Inference Efficiency. Our researchers have delivered meaningful compute multipliers for inference
—reducing the compute required to serve a given level of model performance—while our engineering 
teams have driven significant optimizations in how efficiently that compute is utilized in production. 
This allows us to provide greater value for customers and generate greater revenue from our compute 
base  over  time,  which  is  critical  as  we  have  consistently  operated  in  a  compute-constrained 
environment where customer demand exceeds available compute.

• Multi-Chip  Compute  Approach.  We  train  and  serve  Claude  natively  across  four  major  chip 
families  (AWS  Trainium,  Google  TPUs,  NVIDIA  GPUs,  and  AMD  GPUs)  using  a  proprietary 
orchestration  layer  that  enables  us  to  optimize  workloads  for  price  and  performance.  This  flexibility, 
which  took  years  and  significant  investment  to  build  and  we  believe  is  difficult  to  replicate  at  scale, 
expands  our  available  compute,  reduces  dependence  on  any  single  vendor,  and  enables  us  to 
continuously  route  workloads  to  the  most  optimal  hardware  for those  workloads.  This  foundation, 
paired with Claude’s increasing understanding of chip and model architectures, allows us to onboard 
new chip platforms in a fraction of the time this once required.

135