00:00Korea's startup funding landscape is shifting.
00:03Brands are chasing their most loyal fans.
00:05Here's the number that matters most.
00:0771% of GRIP's annual gross transaction value comes from just 1.3% of its users.
00:13And AI agents are consuming up to 3,000 times more compute than a chatbot.
00:18If you're building in Korea right now, that changes your playbook.
00:22It's Thursday, November 19th.
00:24Let's look at what this means for startups and entrepreneurs in Korea.
00:28Three stories today cover the same underlying shift.
00:31Korean platforms are turning superfans into revenue engines.
00:35Microsoft says AI agents are reshaping compute demand.
00:38And Korean startups are finding ways to make AI models cheaper to run.
00:43Together, they show where the next wave of opportunity sits.
00:46Here's what the numbers show.
00:48GRIP, a live commerce platform, now streams its broadcasts simultaneously on YouTube,
00:53Instagram, and Facebook.
00:55The reason is simple.
00:56Its most valuable users, the top 1.3%, buy 2.4 times more often than the average customer,
01:03and repurchase at a rate 36.9 percentage points higher.
01:07Its most valuable users buy 2.4 times more often than the average customer,
01:12and repurchase at a rate 36.9 percentage points higher.
01:15Shoulder Pick, an illustration platform, is launching Goods Pick this month, a limited edition merchandise service where artists sell signed
01:23prints and cards that will never be reissued.
01:26Meanwhile, B-Stage, the fan community platform that hosts fan communities for Big Bang, actor Park Ji-hoon, and esports
01:34team T1, has begun offering 300 vote tokens to 100 buyers who spend more than 10,000 won.
01:40Fan competition is now a conversion tool.
01:43On the AI side, Microsoft launched its Cobalt 200 server CPU on November 17.
01:50Its Azure infrastructure chief says agent workloads consume up to 3,000 times more compute than standard co-pilots or
01:57chatbots.
01:57Agents act on their own, so the bottleneck is shifting from GPUs towards CPUs.
02:03Meanwhile, Upstage used Nota's model compression to cut a workload from 8 NVIDIA H100 GPUs down to 2.
02:11SqueezeBits sped up NVIDIA's Cosmos world model by about 3.1 times.
02:16SqueezeBits sped up NVIDIA's Cosmos world model by about 3.1 times.
02:21These are startups that found a real margin by making AI cheaper to run.
02:25So what does this mean for founders weighing their next move in Korea?
02:29Here's the takeaway.
02:31First, loyalty is now a measurable asset, not a marketing slogan.
02:35A small group of superfans can carry an entire revenue line, so the question is which community you can own.
02:41Second, inference cost is becoming the core competitive variable.
02:46Agents and physical AI multiply compute demand,
02:49so startups that squeeze more performance out of fewer GPUs will hold better unit economics.
02:55Third, the market is shifting from asking how smart your model is to asking how economically smart it is.
03:01Earlier, we said that superfans and efficiency were the two levers that matter.
03:06Here's what that actually means for you.
03:08Build a loyal, measurable community, and design your product so it runs cheaply enough to scale.
03:14What to Watch
03:15Goods
03:16Goods
03:16Pick launches this month, and the agent compute race is still accelerating.
03:20That's today's AI Prism
03:22Founders
03:23This episode was produced with AI Assistants based on Seoul Economic Daily reporting.
03:28AI Prism is a won IFRA award-winning series.
03:31We'll be back tomorrow.
03:33AI Prism