Small print.
Wide open.

Table of Contents

•

Our Safety Research Is a Commercial Benefit. As models take on agentic roles with access to critical 
business  systems,  trust  and  reliability  are  paramount.  And  as  capabilities  increase,  the  ability  to 
operate  safely  becomes  one  of  the  key  determining  factors  for  commercial  viability.  Models  that  are 
well-aligned  are  better  at  following  a  user’s  instructions—for  example,  they  avoid  taking  potentially 
destructive actions without fully understanding user intent. They are also more robust against attacks 
such  as  prompt  injection,  where  malicious  actors  attempt  to  hijack  the  model  by  feeding  it  illicit 
instructions. Our work on alignment, interpretability research, safeguards, and red-teaming, as well as 
the use of external reviewers who can assess our safety practices, helps us create safer, more reliable 
models, and therefore positions us to meet our customers’ rigorous requirements. Safety and capability 
research are integrated, which enables us to deliver the trust and performance required for enterprise 
workflows.

Our Approach to Compute

Our  approach  to  compute  starts  with  training  models  that  are  designed  to  be  more  capable  and  run  more 
efficiently on any given type of compute.

Our ability to build more capable models and serve them more efficiently reflects the compounding returns of 
an iterative cycle between our pre-training, post-training, and inference research, all of which is enabled by our 
fungible  compute  pool.  Pre-  and  post-training  research  produces  proprietary  compute  multipliers  that  make 
our  models  more  capable.  Research  and  inference  engineering  drive  sustained  reductions  in  the  compute 
required to serve those models. 

In  addition  to  enabling  more  capable  and  efficient  models,  our  multi-chip  compute  architecture  across  AWS 
Trainium,  Google  TPUs,  NVIDIA  GPUs,  and  AMD  GPUs  allows  us  to  orchestrate  workloads  to  optimize  for 
price  and  performance.  We  have  years  of  experience  effectively  training  and  running  our  models  across 
multiple chip families, which we believe is differentiated and difficult to replicate. 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.

Our  compute  pool  is  fungible  by  design.  We  have  made  substantial  engineering  investments  to  build  the 
internal  orchestration,  scheduling,  and  capacity  management  systems  that  allow  accelerator  capacity, 
regardless  of  provider,  chip  architecture,  or  physical  location,  to  be  directed  across  workloads.  The  same 
infrastructure  that  serves  customer  workloads  is  also  directed  to  research,  model  development,  and  internal 
usage of Claude. These systems manage the underlying hardware as a unified capacity pool, monitor utilization 
and demand across that pool, and direct compute to wherever it is expected to be most productive at any given 
time.

Our  multi-chip  compute  approach  provides  us  with  substantially  greater  chip  access  than  relying  on  one 
platform  and  enables  us  to  opportunistically  capture  the  price-performance  benefits  of  each  chip  family  and 
generation. Securing this access requires significant advance commitment. Given the long lead times between 
contracting for compute and that compute becoming available to serve workloads, we secure compute well in 
advance  of  anticipated  demand. In some cases, we may also make payments in advance of the delivery of the 
equipment or infrastructure. We anticipate making significant investments for the foreseeable future, including 
in  compute  and  infrastructure,  to  ensure  we  have  the  technological  resources  to  maintain  and  expand  our 
position at the frontier of AI.

Enterprise Strategy

Our  enterprise  strategy  aligns  with  our  mission.  We  build  our  most  capable  models  for  the  problems  that 
demand the highest intelligence. That same intelligence is also applied efficiently to everyday work—calibrating 
effort to the difficulty of each task and routing workloads to the models within our model family that are best 
suited for them. As a result, our customers around the world are adopting Claude across a broad range of work, 
from routine knowledge tasks to their most important problems. Today, Claude is deployed for knowledge work 
across major industries, including financial services, insurance, healthcare, life sciences, law, education, retail 
and  commerce,  telecommunications,  professional  services,  technology,  and  cybersecurity.  Beyond  practical 
enterprise  challenges,  Claude  is  also  being  applied  to  pressing  global  problems  in  scientific  research,  drug 
discovery, education, and climate analysis.

188