Mistral's $3.5B Raise, Samsung Profit up 18x from AI Memory

The episode opens with major developments in the AI hardware and startup ecosystem. Samsung is experiencing an 18x profit jump primarily driven by exploding demand for High Bandwidth Memory (HBM), which is ultra-fast memory stacked directly onto AI chips used by NVIDIA and AMD. The memory market shifted from oversupply in 2023-2024 to severe undersupply as major tech companies like OpenAI, Microsoft, Google, and Meta began purchasing AI hardware in massive quantities. Only three companies globally can manufacture HBM at scale—Samsung, SK Hynix, and Micron—making these suppliers extremely profitable as prices skyrocket.

Mistral AI, a French startup, is raising $3.5B at a $23B valuation, nearly double its September 2024 Series C valuation. The company has achieved $400M in recurring revenue as of February, a 20x increase from $20M a year prior. Mistral is positioning itself as a champion of AI sovereignty for Europe, offering open-source models that countries can run locally. The company is committing $4B to build data centers in France and Sweden, launching Mistral Compute as a European alternative to NVIDIA's platform, and establishing an AI campus near Paris through partnerships with MGX, NVIDIA, and BPI France. Their revenue skews heavily toward enterprise and government clients rather than consumer chat applications, and they employ a forward-deployed engineer strategy similar to Palantir.

Midjourney is in a copyright lawsuit with Disney, Universal, and Warner Brothers, and has filed discovery requests asking these studios to disclose their own internal AI training practices. Midjourney's strategy appears to be proving that the studios are also training on unlicensed content using their own AI programs, normalizing the practice across the industry. The studios sued after Midjourney reproduced characters like Bart Simpson and Darth Vader, but Midjourney hopes to show the studios are doing similar training and potentially get the case dismissed.

Station F's FAI Accelerator is launching a second cohort in September with 20 startups from the first batch having raised $34M in pre-seed funding combined. The accelerator, positioned as Europe's answer to Silicon Valley, targets €1M in revenue per startup within six months and aims to address criticisms that European startups commercialize slower than American ones. The second cohort adds partners including 11 Labs, OpenRouter, GitHub, and HubSpot, joining existing partners like Anthropic, AWS, Google, Meta, Microsoft, and Mistral AI.

Fable 5 has topped the Kernel Bench Mega benchmark, demonstrating the ability to write CUDA code 19 times faster than its baseline and outperforming Claude Opus. Simultaneously, AI agents have improved dramatically on real-world freelance work tasks, jumping from 2.5% success rate to 16.1% in eight months—outperforming GPT-4 Turbo's 6.3% score. The improvements span tasks like 3D modeling, architecture, animation, and analysis, though models still struggle with tasks exceeding one hour of compute time. The host shares personal experience using Fable 5 to solve widget customization problems that Claude Opus 4 couldn't handle in two hours.