Eric Schmidt's $80 Billion Anthropic Bet

DeepSeek, a Chinese AI company, announced it is open sourcing DeepInfer, its internal production stack designed to accelerate AI inference, optimize chip performance, and reduce energy and compute costs. Bloomberg quoted DeepSeek stating that sharing this technology aims to lower compute costs, reduce energy use, and make AI more affordable for the community. The open-sourced DeepInfer supports DeepSeek’s V4.1 flash model and could help cloud providers deploy more cost-efficient AI models by leveraging techniques like distributed inference and tailored hardware optimization. The release reveals production techniques that are usually proprietary, providing transparency into how DeepSeek keeps inference costs down. Techniques include optimized kernels, parallelism methods for attention and expert layers, and scheduling improvements such as draft tokens and confidence estimates. Although no independent benchmarks exist yet, the software is likely adaptable by other developers. DeepSeek also announced partnerships with Huawei to build programming tools for Ascend chips, including open-sourcing modules for the platform.

Inception introduced Mercury 3, a diffusion-based language model claiming to outperform GPT-6 Luna, Claude Haiku 5.5, and GLM 5.3 Flash on agentic benchmarks while being 4 to 8 times faster. CEO Stefano Hermann highlighted Mercury 3’s speed, generating 1000 tokens per second, facilitated by parallel inference—generating and refining multiple tokens simultaneously, a stark contrast to traditional autoregressive models. The model is also notably cheaper at launch, costing $0.10 per million input tokens and $0.35 per million output tokens, with a promotional 50% discount. Mercury 3 incorporates significantly more parameters—40X compared to Mercury 1—but Inception has not disclosed the total count, possibly to prevent underestimation. The breakthrough suggests diffusion LLMs might challenge other models, especially as Google DeepMind develops similar tech. This approach emphasizes faster, more efficient AI responses, potentially reshaping how the industry builds and deploys large language models.

Cloudflare announced Cleric, a new system one-class extraction model designed to process unstructured data more efficiently than traditional LLMs. Launched in preview on Workers AI, Cleric extracts specific fields (like invoice totals) with open weights expected soon, providing structured responses in a single pass without generating text tokens. It promises faster processing, lower costs—10 cents per million input tokens—and aims to integrate smoothly with existing systems. Despite its promising claims, Cloudflare clarifies that accuracy and speed benefits are company-reported, not yet validated by independent benchmarks. The model is suitable for simple calculations and extraction tasks but warns that some outputs, like complex chains or forecasts, should be validated with additional evidence or human oversight. This tool aims to improve enterprise data workflows and reduce operational latency and costs.

Spotify launched paid podcast memberships in 34 markets, including the US, allowing creators to offer perks such as bonus content, fewer ads, or early access, while keeping 90% of the revenue. The feature was first previewed in May and is part of Spotify’s broader strategy to attract podcasters and compete with YouTube. Creators on Spotify for Creators or Megaphone platforms with at least 100 consumption hours, 100 listeners, and three episodes can set up memberships. These memberships will be promoted directly on the show's Spotify page, with plans to add badges and Welcome Clips. The rollout includes shows like The Adam Friedland Show, Sam Harris, and others, offering an additional revenue stream alongside ads and analytics tools.

Apple announced a $100 million investment over three years in Perryman Company, a titanium supplier based in Pennsylvania, as part of its efforts to boost domestic manufacturing. This supports the production of titanium used for the iPhone Duo’s foldable enclosure, which will contain at least 80% recycled material. The move comes amid supply constraints for the device, with pre-order demand reportedly high but supply chain issues leading to long wait times and lowered shipment estimates. The investment aims to increase local capacity for critical components, helping Apple meet demand for the new foldable model, which is set to go on sale soon.

Eric Schmidt, former Google CEO, holds about 1% of Anthropic following a multisection investment of tens of millions of dollars during its Series A funding in 2021. His stake could yield a windfall if Anthropic’s upcoming IPO values the company at up to $2 trillion, potentially making Schmidt’s stake worth around $80 billion at the high end. The company’s valuation has soared from $461 million during Series A to estimates approaching $2 trillion, reflecting a 4,300-fold increase. Schmidt, now a prolific AI investor through Hillspire and other platforms, has diversified early investments in consumer internet and finance into major AI bets. A successful IPO at that valuation would surpass early Uber investments, significantly boosting Schmidt’s wealth.