🟣 EV Daily: Two Moonshots — one hit, one miss

Lead story: Two Moonshots — one hit, one miss

Two headlines over the weekend capture the AI industry’s divergent tempos. In Beijing, Moonshot AI open-sourced Kimi K2, a trillion-parameter mixture-of-experts model that tops GPT-4.1 on coding and math tests and is free to download. In San Francisco, OpenAI hit the brakes yet again, postponing its own open-weight release “indefinitely” for extra safety checks. The difference between approaches is stark. The insurgent treats cutting-edge capability as a public good, accelerating adoption and learning curves; the incumbent, wary of reputational and regulatory blow-back, keeps its crown jewels under wraps a little longer. The releases of DeepSeek R1 and Kimi K2 show how China’s open-source push is eroding the idea of durable moats. The only thing likely to slow that momentum now is the tightening grip of chip controls. [VentureBeat, TechCrunch] #AIModels

Key signals, quick scan

A 30-second scan of four secondary signals that hint at where the curve is bending.

• Despite US export controls, Beijing has acquired around 115,000 top-tier Nvidia GPUs (such as H100s) to power dozens of data centers in its new AI “desert hub.” [DigiTimes] #Geopolitics #Hardware

• Goldman Sachs is deploying hundreds of instances of Devin, a self-directing AI engineer capable of reading tickets, fixing bugs and committing code without human keystrokes. If successful, the model could reduce the need for junior developers (Goldman has 12,000 devs) and compress multi-sprint cycles into hours. [CNBC] #Economics

• Google’s $2.4bn cash-plus-acquihire of Windsurf, weeks after Meta’s recruitment blitz, shows Big Tech will now write multi-billion-dollar checks simply to lock down scarce LLM talent, usage data and codebases. [TechCrunch] #SoftwareApps

• Two-thirds of UK children aged 9-17 use AI chatbots regularly, and 35% of that group say their interactions feel like talking to a friend. “It’s not a game to me,” one 13-year-old boy told nonprofit Internet Matters, which conducted the survey of 1,000 people, “because sometimes they can feel like a real person and a friend.” [Futurism] #Society

Data of the week

Five data points that show how the world is changing.

• $750 billion is going into new and in-progress data centers, driven by hyperscalers, AI-native clouds, and sovereign funds betting on infrastructure. [X/ShanuMatthews] #Hardware

• Gemini traffic jumped 148% between December 2024 and June 2025, triple ChatGPT’s 46%, though ChatGPT still holds 78.6% of the market compared with Google’s 8.6%. [Similarweb] #Economics

• Anthropic’s annualized revenue has quadrupled to $4 billion since the start of the year, nearly half that of OpenAI’s. [X/Deedydas] #Economics

• Nvidia’s 2025 GPU output is expected to fall by 10% (500,000 units) because of ongoing production bottlenecks. [X/Rwan07] #Hardware

• Paid AI adoption among US firms fell 0.5% this month, the first decline since the Ramp AI Index began tracking. [RampAI] #Economics