Early on, Huang Renxun was as naive and eager as a little girl with a sweet tooth—on his first day of entrepreneurship, the main topic of discussion was what to have for lunch? 😂 On his 30th birthday, the makeshift startup officially launched, and daily meetings mostly consisted of awkward small talk. Months later, the conversation evolved to: Should we keep donuts in the fridge? It’s hard to imagine he started from virtually nothing—reading books on entrepreneurship, learning what venture capital was, and how to register a company. Fortunately, Huang had great character and immense personal charm. VCs don’t invest in pitch decks—they invest in people. And he was exceptionally skilled at packaging complex, abstract concepts like CUDA architecture and AI computing power into compelling visions. Most importantly, he discovered that wet snow and long, steep slope.

Alibaba's Qwen3.7-Max just ran fully autonomous for 35 hours straight. Not chatting. Not answering prompts. Designing a hardware attention kernel for its own AI chip — from scratch. 432 kernel tests. 1,158 tool calls. Zero human intervention. The model had never seen this chip architecture during training. Started with: no docs, no sample code, no measurement data. Just a reference implementation in Triton. 35 hours later: optimized kernel. Compiled, tested, revised in loops. Caught its own errors. Tracked its own performance. This is what agentic AI looks like when you remove the guardrails and let it work. Not a demo. Not a benchmark. A real engineering task on real hardware. The line between "AI assistant" and "AI engineer" is getting very thin. — Robbie Jeo

Video Game Chair of the Day Folks 🎮🎮

CoWoS TSMC: Capacity will expand from 35K wafers/month at the end of 2024 to 130K wafers/month by the end of 2026 (nearly a fourfold increase), yet demand still exceeds supply. NVIDIA accounts for approximately 60% (595K wafers/year), with AMD, Google, and others competing for the remaining share. Outsourcing: ASE and Amkor handle subcontracted CoWoS subprocesses from TSMC; advanced packaging revenue is expected to double by 2026. Intel’s EMIB/Foveros serves as an alternative solution.

Data Centers Construction Speed: The five major North American CSPs (AWS, Microsoft, Google, Meta, Oracle) have deployed 800–900 global data centers, with AWS holding the largest share. CoreWeave: 850 MW of active power by end of 2025 (43 DCs); target exceeding 1.7 GW in 2026, with CapEx of $30–35 billion (doubling). Plans to add 5 GW by 2030. Delay Risk: Some CoreWeave projects have faced shareholder lawsuits due to construction delays.

Token Growth OpenAI: Claims 900 million weekly active users, with continued growth in API calls. Google: Gemini paid users increased 258% YoY, surpassing Claude’s 200%. Google Cloud backlog reaches $460 billion (nearly doubled). Inference Speed: Gemini 2.5 Flash Lite reaches 240 TPS, Claude Haiku 4.5 reaches 180 TPS, DeepSeek V3.2 reaches 110 TPS. Improved inference efficiency may partially offset hardware demands.

A lot of people seem to think that Huawei has side-stepped EUV and developed a method to scale to 1.4nm without it But as u can see from their graph, a huge density & clock speed jump happens between 2030 & 2031 After the 2019 sanctions, Huawei started project Mount Everest in 2020 The goal was to develop and mass-produce EUV machines within 10 years The tentative date of China's EUV is 2030 rn

🚨🚨 Microsoft reportedly told engineers to stop using Claude because AI bills were exploding, while Uber says its entire yearly AI budget was already destroyed by April.

three weeks = elon time, so probably september but the cursor data thing is actually interesting. that’s real dev workflow signal, not synthetic garbage.