GPU clusters bottlenecked Kakao's AI. Switching to Cloud TPUs, the team upcycled a 2.1B model into a 13.4B parameter MoE system.
South Korea's leading technology company operates the country's dominant messaging platform along with a vast ecosystem spanning digital payments, mobility, and entertainment.
Developing a proprietary Korean-language AI model required massive computing power, but scaling GPU clusters beyond single server units caused...
“현재 카카오의 대규모 언어 모델은 모두 TPU Trillium으로 학습되고 있습니다. GPU는 익숙하고, 다양한 접근이 가능하긴 하지만 필요한 만큼의 성능을 확보하는 데에 근본적인 어려움이 이어졌습니다. 구글 클라우드의 TPU는 언제든 필요한 만큼 자원을 할당받을 수 있었고 성능이나 모델의 확장도 손쉬웠습니다.”
Kakao's AI model training is part of this use case:
Mobile messaging platform and internet services for the South Korean market.
Cloud computing services, AI infrastructure, and data analytics platforms for enterprises.
Supported Kakao's AI and cloud transition to develop its 'Kanana' language model.
Related implementations across industries and use cases
Local servers caused latency bottlenecks. Cloud AI now translates complex docs in days, cutting months from market entry timelines.
Early outputs felt like awkward translations. Now, slang-fluent personas drive a lifestyle AI platform for 6.5M users.
Engineers manually provisioned servers during spikes. Automated GPU scaling now cuts user waits from days to seconds.
Engineers spent 90% of time on data prep. New pipelines flipped that to 90% modeling and cut tuning from 7 days to 1 hour.
Coding requirements locked out domain experts. Now, they build computer vision models in weeks without writing a line.
Scattered spreadsheets couldn't catch AI hallucinations. Now, automated LLM judges evaluate every prompt change to block regressions.
Moderation couldn't keep pace with 600M users. AI agents now filter toxicity while models recognize 2.5B objects to refine search.
Repetitive coding slowed R&D. Now 80% of engineers use agentic tools to automate work, saving up to 2 hours weekly per person.
Manual testing consumed 20% of developer time. Now, 1,500 engineers use AI agents to auto-generate tests and prototype solutions.
GPU clusters bottlenecked Kakao's AI. Switching to Cloud TPUs, the team upcycled a 2.1B model into a 13.4B parameter MoE system.
South Korea's leading technology company operates the country's dominant messaging platform along with a vast ecosystem spanning digital payments, mobility, and entertainment.
Developing a proprietary Korean-language AI model required massive computing power, but scaling GPU clusters beyond single server units caused...
“현재 카카오의 대규모 언어 모델은 모두 TPU Trillium으로 학습되고 있습니다. GPU는 익숙하고, 다양한 접근이 가능하긴 하지만 필요한 만큼의 성능을 확보하는 데에 근본적인 어려움이 이어졌습니다. 구글 클라우드의 TPU는 언제든 필요한 만큼 자원을 할당받을 수 있었고 성능이나 모델의 확장도 손쉬웠습니다.”
Kakao's AI model training is part of this use case:
Mobile messaging platform and internet services for the South Korean market.
Cloud computing services, AI infrastructure, and data analytics platforms for enterprises.
Supported Kakao's AI and cloud transition to develop its 'Kanana' language model.
Related implementations across industries and use cases
Local servers caused latency bottlenecks. Cloud AI now translates complex docs in days, cutting months from market entry timelines.
Early outputs felt like awkward translations. Now, slang-fluent personas drive a lifestyle AI platform for 6.5M users.
Engineers manually provisioned servers during spikes. Automated GPU scaling now cuts user waits from days to seconds.
Engineers spent 90% of time on data prep. New pipelines flipped that to 90% modeling and cut tuning from 7 days to 1 hour.
Coding requirements locked out domain experts. Now, they build computer vision models in weeks without writing a line.
Scattered spreadsheets couldn't catch AI hallucinations. Now, automated LLM judges evaluate every prompt change to block regressions.
Moderation couldn't keep pace with 600M users. AI agents now filter toxicity while models recognize 2.5B objects to refine search.
Repetitive coding slowed R&D. Now 80% of engineers use agentic tools to automate work, saving up to 2 hours weekly per person.
Manual testing consumed 20% of developer time. Now, 1,500 engineers use AI agents to auto-generate tests and prototype solutions.