Small print.
Wide open.

Table of Contents

as  research,  drafting  assignments,  and  supporting  administrative  work,  all  with  the  security  and 
compliance  built  for  institutional  requirements.  Claude  for  Teachers  gives  verified  K-12  educators  in 
the United States free access to premium Claude capabilities, a library of teaching skills, and a direct 
connection to evidence-based curricula, mapped to academic standards in all 50 states. Claude helps 
with the paperwork, so teachers can spend more time with their students.

•

Claude  for  Government.  Brings  Claude  models  to  U.S.  government  agencies  for  drafting,  research, 
analysis,  and  coding  in  a  FedRAMP  High  authorized  environment,  designed  to  support  the  data 
handling and auditability requirements of government customers. Government customers outside the 
United  States,  and  customers  with  workloads  in  classified  environments,  can  access  Claude  models 
through Amazon Bedrock, subject to the authorizations and availability of that platform. 

Our Infrastructure

Our  infrastructure  underpins  our  ability  to  train  frontier  models,  serve  inference  at  scale,  manage  internal 
usage  of  Claude,  and  deliver  Claude  to  our  customers  with  the  performance  and  reliability  that  enterprise 
deployments  require.  We  have  built  a  differentiated  infrastructure  layer  that  allows  us  to  train  and  deploy 
frontier models more efficiently across multiple hardware platforms. This infrastructure advantage compounds 
over time: as we bring optimizations to each chip family, every layer of our system benefits accordingly.

Multi-Chip Compute Approach 

We  have  built  a  deliberate  and  diversified  compute  sourcing  strategy  across  multiple  chip  platforms  that 
provides the following distinct advantages: 

• We can secure compute from multiple providers rather than being constrained by any single provider.

• We  can  select  chips  based  on  the  best  price-performance  and  efficiency  characteristics  for  each 

workload, benefiting from the differentiated strengths of each platform.

• We  maintain  collaborative  relationships  with  each  of  our  providers,  and  we  influence  roadmaps  and 

features of future chip generations. Our chip platforms include:

◦

◦

◦

◦

AWS  Trainium.  Our  strategic  collaboration  with  Amazon  provides  us  with  access  to  AI 
accelerators,  including  custom-designed  Trainium  chips,  and  compute  infrastructure,  as  well  as 
the opportunity to influence future generations of Trainium chips. 

Google  TPUs.  Our  partnerships  with  Google  and  Broadcom  provide  us  with  access  to  compute 
capacity powered by Google TPUs across multiple generations of the TPU platform.

NVIDIA GPUs. Our strategic partnership with NVIDIA Corporation (“NVIDIA”), together with our 
arrangements with Microsoft, SpaceXAI, Nscale, Lambda, and other cloud providers, provides us 
with  access  to  compute  capacity  powered  by  NVIDIA  GPUs,  including  NVIDIA  Grace  Blackwell 
and Vera Rubin GPU systems, as well as the opportunity to collaborate directly with NVIDIA on 
design and engineering to optimize future NVIDIA hardware for our workload profile. 

AMD  GPUs.  Our  relationship  with  AMD,  together  with  our  arrangements  with  approved 
manufacturers  of  AMD-based  Helios  rack-scale  systems  and  third-party  service  providers, 
provides  us  with  access  to  compute  capacity  powered  by  AMD  Instinct™  GPUs,  including  AMD 
Instinct™ MI450 Series GPUs, as well as the opportunity to collaborate directly with AMD through 
a multi-year engineering collaboration.

• We can onboard new chip platforms in a fraction of the time it once took. Historically, bringing a new 
chip  platform  into  production  required  intense  engineering  effort  over  many  months.  By  leveraging 
Claude, our engineering teams can now bring new chip platforms into production more quickly. This 
speed  and  efficiency  also  relies  on  the  robust  foundation  we  have  already  built  across  multiple  chip 
families,  including  a  proprietary  orchestration  layer  designed  to  seamlessly  test  and  measure 
performance against our actual workloads. As a result, we are able to adopt new AI accelerators as they 

204