Small print.
Wide open.

Table of Contents

performance  characteristics  across  workload  types,  the  composition  of  our  committed  compute  portfolio, 
including  the  mix  of  accelerator  architectures  procured  from  our  partners,  reflects  our  demand  forecasts  and 
development  roadmap  assumptions  at  the  time  those  commitments  are  made.  If  the  magnitude  or  mix  of 
workloads  we  need  to  support  differs  materially  from  those  forecasts,  whether  due  to  changes  in  model 
architecture,  customer  demand,  internal  operations,  or  research  priorities,  the  composition  of  our  available 
compute may not optimally match our needs in a given period. This could affect our ability to serve customers, 
develop models, and support internal operations. 

If we fail to secure sufficient compute, or if we fail to effectively direct the limited compute we have across our 
workloads to develop and deploy models with capabilities at least comparable to those of our competitors, we 
risk  falling  behind  the  technological  frontier.  The  pace  of  AI  development  is  extremely  rapid,  with  successive 
versions  of  models  emerging  in  compressed  timeframes,  often  measured  in  months  rather  than  years.  As  a 
result, even a brief period without access to sufficient compute could cause us to fall behind quickly. With AI 
development, the gap between frontier models and those one or two versions behind can be significant in terms 
of  capability,  performance,  and  commercial  applicability.  Customers,  partners,  and  software  developers  may 
shift  to  competitors  whose  models  offer  superior  capabilities,  and  we  may  be  unable  to  attract  or  retain  the 
technical talent necessary for frontier research if our researchers lack access to adequate compute or if we are 
not perceived as operating at the forefront of model development. Any inability to maintain our position at the 
frontier could result in loss of customers and revenue, diminished brand and reputation, and a decline in our 
competitive position, each of which may be difficult to regain. If we are unable to manage the foregoing risks, 
our business, financial condition, and results of operations could be adversely affected.

If  our  actual  compute  requirements  differ  materially  from  our  projections,  we  may  fail  to 
secure sufficient compute to meet our needs or we may be committed to excess compute that 
remains underutilized.

The  competitive  and  structural  dynamics  of  the  compute  market  require  us  to  make  binding  financial 
commitments for compute months or years in advance of our day-to-day delivery and use. Chip manufacturers, 
CSPs, and data center developers require advance commitments in order to allocate production capacity, plan 
infrastructure  investments,  and  manage  their  own  supply  chains.  As  a  result,  we  must  forecast  our  compute 
requirements well in advance of their deployment and commit capital accordingly, often with limited ability to 
adjust commitments if our needs change.

These forecasts for future compute usage are inherently uncertain. Our actual compute requirements depend on 
numerous factors that are difficult to predict, including the pace of our research progress, the compute intensity 
of  training  techniques  we  develop  or  adopt,  changes  in  the  competitive  landscape  that  may  require  us  to 
accelerate  or  modify  our  development  plans,  customer  demand  for  inference  compute,  and  the  evolution  of 
hardware or other technological efficiencies and changes. If we underestimate our compute requirements, we 
may  be  unable  to  secure  sufficient  compute  to  support  our  forecasted  customer  demand  and  model 
development roadmap. In the past, we have significantly underestimated our compute requirements, which left 
us  capacity  constrained  and  impacted  both  product  development  and  deployment.  Further,  the  timing  of 
compute  coming  online  may  not  align  with  the  timing  of  demand  growth,  which  could  result  in  periods  of 
excess or underutilized compute and committed cash outflows that do not produce immediate revenue.

Our ability to forecast and secure the compute we need is further challenged by our ongoing transition from a 
cloud-only, fixed-time arrangement compute model to a hybrid model that includes directly leased chips that 
are  hosted  in  dedicated  data  center  facilities  leased  and  operated  on  our  behalf,  to  our  specification.  This 
transition  introduces  additional  variables  into  our  procurement  planning,  including  the  timing  of  facility 
completion,  the  pace  at  which  newly  commissioned  facilities  achieve  full  operational  capacity,  our  ability  to 
operate,  maintain,  and  integrate  on-premise  capacity  with  cloud-based  infrastructure,  the  operational 
execution of our suppliers, and the performance of hardware in new facility environments. Delays in securing 
energy,  water,  and  other  critical  services  for  these  facilities  could  further  slow  the  pace  at  which  newly 
commissioned  compute  becomes  available  for  use.  Errors  in  forecasting  any  of  these  transition-related 
variables  could  compound  our  risks  related  to  securing  sufficient  compute  and  result  in  either  capacity 
shortfalls or underutilized capacity during the transition period.

We  have  entered  into,  and  expect  to  continue  to  enter  into,  long-term  agreements  for  compute  and  other 
compute resources that involve minimum purchase commitments, take-or-pay provisions, and other terms that 

31