0:00 SemiAnalysis just launched GDPx, a benchmark testing AI agents across eleven high-value industries where the top-scoring model in the world lands at only twelve percent.0:12 In episode two hundred twenty-one, we tracked how bot drama and policy noise were drowning out real technical signal on X.0:21 Today, the timeline finally cut through the theatre with cold data.0:26 Here is what moved while everyone else was arguing.0:30 Sam Altman sat down in South Korea with roughly fifty Samsung affiliate presidents, delivering a blunt warning: simply adopting AI provides zero lasting edge.0:41 Advantage belongs entirely to whoever builds faster, a point underscored by Altman bringing along Richard Ho, the head of OpenAI's in-house chip division, to talk next-generation memory.0:55 Nvidia quietly lowered the memory specs on its shipping Rubin silicon, dropping its high-bandwidth memory from twenty-two terabytes per second to nineteen point two.1:07 That leaves AMD's upcoming MI455X ahead on paper in raw bandwidth, shifting the enterprise GPU war strictly into memory constraints.1:17 DeepSeek open-sourced its super-agent architecture, using a decoupled router and worker system that handles GitHub tasks concurrently, cutting completion latency down from twenty-nine minutes to under sixteen.1:32 And BTCPay Server issued an urgent advisory regarding an attack vector hitting unpatched setups running version two point four point one or older.1:43 So what does it all add up to?1:45 Look at the quiet presentation from Revolut that everyone in engineering circles is passing around.1:52 David Abitbol, Revolut's head of AI operations, laid out how the digital bank slashed its AI coding bills by roughly ninety percent.2:02 They didn't do it with another frontier API.2:05 They ditched Anthropic and OpenAI token limits entirely, self-hosting open-weight GLM models on dedicated GPUs.2:13 Their blended cost dropped to thirty cents per million tokens.2:18 Once the meter stopped running, Revolut's two thousand engineers stopped rationing prompts.2:24 They pointed the models at configuration files, legacy codebase navigation, and daily plumbing.2:31 The result?2:32 Development lead time plunged by seventy percent.2:36 That is the exact thread connecting Revolut to the new SemiAnalysis GDPx benchmark.2:42 SemiAnalysis didn't test broad trivia or chat banter.2:46 They partnered with fourteen specialized firms across chip layout, biology, and materials science because the first trillion dollars in enterprise value won't disperse across the entire economy.3:00 It will concentrate where autonomous execution creates immediate leverage.3:05 Claude Opus five point five leads the pack, but at twelve percent accuracy, general models are still fumbling domain tasks.3:15 The takeaway across every thread today is unmistakable.3:19 The era of paying retail API margins for generic chat is ending.3:23 Competitive advantage now turns entirely on owning your inference pipeline and eliminating token rationing at the desk.3:32 Next quarter belongs to the teams that run their own weights and measure output in shipped software instead of hype.
