AI has a power problem with foundation model training costs approaching $200M. Cassia.ai developed semiconductor IP to increase the TOPS/watt efficiency of AI chips. We do math better.
Current Status
1) 3-way university-foundry license: multiple free test-chip fabrications (~$1M value)
2) $20k safe investment from a customer who asked for a custom solution
3) three evaluations licenses with companies who are building AI chips.
4) more that 20 NDAs signed
Problem or Opportunity
AI has a power problem. Data center power demand is driven by AI with Meta and Microsoft investing in wind power and nuclear power plants so they can support training of foundational models. Moreover, requests for power in data centers are increasing ten-fold to 2GW or more while it is now becoming very difficult, or even illegal to build data centers in certain locales. As AI takes on a bigger role in technology, it will be limited by our power capabilities.
Solution (product or service)
Cassia.ai does math better. We developed core math IP blocks for AI chips that can improve the TOPS/watt efficiency by as much as 23x in some cases. When implemented on chip, this can increase the performance while reducing power depending on the parameters used to build a chip.
Business model
Semiconductor IP licensing. We charge an up-front fee for the license to build the chip and a percentage royalty for each chip manufactured after moving to production.
