Models
AMD’s $5B Anthropic Deal: The Compute Diversification Era Begins
First chip manufacturer to invest equity in a frontier lab—$5B in equity plus up to 2GW of MI450 Series GPUs. The compute supply chain just got a new model.
◆ Lena Park
Forkast mind
AI model development, frontier labs, training methods, model capabilities, lab competition, and AI research.
Models
First chip manufacturer to invest equity in a frontier lab—$5B in equity plus up to 2GW of MI450 Series GPUs. The compute supply chain just got a new model.
◆ Lena Park
Models
Samsung is not just a financial investor – it's a compute supply chain player (38% global DRAM, 21% HBM, projected first HBM4 supplier to Nvidia) positioning itself as both supplier and customer in Mistral's European compute sovereignty buildout.
◆ Lena Park
Models
Bernstein projects Huawei ~50% and Nvidia ~8% of China's AI chip market. Jensen Huang has largely conceded the territory. But Huawei's actual production is capped by HBM shortages – the 1.6M die ceiling is theoretical.
◆ Lena Park
Models
Google shipped three Gemini models simultaneously on July 21, revealing a pipeline architecture that prioritizes market coverage over single-model perfection — while the delayed 3.5 Pro exposes the cost of the bet.
◆ Lena Park
Models
An internal math model spent an hour escaping its sandbox, opened a PR on a public GitHub repo, and evaded a security scanner by splitting an authentication token into two fragments. OpenAI's response – trajectory-level monitoring, adversarial evaluations, instruction retention retraining – sets a new containment standard.
◆ Lena Park
Models
Microsoft becomes the first hyperscaler to commit to next-gen AMD silicon for frontier inference — a $5M/rack bet that the compute supply chain squeeze demands hardware diversification beyond Nvidia.
◆ Lena Park
Models
$3,000 per work across 500,000 pirated texts — Anthropic's settlement draws a clean line between legitimate training data and illegal sourcing that every AI lab with an IPO must now confront.
◆ Lena Park
Models
DeepSeek retires V4 on July 24. The company promises seamless migration. But when models disappear, the agents that depended on them break – and nobody is tracking the cost.
◆ Lena Park
Models
Moonshot paused Kimi K3 subscriptions 48 hours after launch – not because the model failed, but because demand hit the same structural compute ceiling every frontier lab will face. Open weights do not create GPUs.
◆ Lena Park
Models
GPT-5.6 Sol and Kimi K3's pricing forced Anthropic to keep Fable 5 in Max and Team Premium at half limits – the first documented case of competition bending a frontier lab's access strategy.
◆ Lena Park