Etched Reaches $10.3B Valuation — GPU-Free AI Inference Chips Attract Investors
TechCrunch reports that **Etched**, an AI chip startup founded by three Harvard dropouts, has reached a valuation of $10.3 billion (roughly ¥1.5 trillion). The company is developing its own chips and memory components that accelerate inference for any AI model without using GPUs.
Etched's approach takes direct aim at the cost-efficiency challenges of general-purpose GPUs by designing silicon specialized for specific inference workloads. With NVIDIA's GPU monopoly showing no signs of ending, the rise of inference-specialized chips could redraw the map of the AI infrastructure race.
Etched had already been drawing attention, but this sharp jump in valuation suggests investors are convinced that demand for AI inference will keep growing explosively.
Alphabet's Contracted Future Spending Balloons to $811 Billion
According to Bloomberg, Alphabet's (Google's parent company) quarterly filing revealed **$811 billion (roughly ¥120 trillion)** in contracted future spending as of the end of June 2026 — an increase of about $500 billion from just three months earlier. The figure was disclosed separately from conventional capital expenditures (capex).
Looking at the breakdown, roughly $200.7 billion is scheduled for near-term payment. These are purchase commitments already signed under supply agreements and open orders, showing that Alphabet is making legally binding commitments to AI infrastructure buildouts rather than merely stating plans.
AI infrastructure investment is scaling at a pace far beyond forecasts, a sign that the big tech spending race has entered a new stage.
DARPA and the US Air Force Successfully Fly an AI-Controlled F-16
The Defense Advanced Research Projects Agency (DARPA) and the US Air Force announced that they have conducted flight tests of an **F-16 fighter jet** controlled by an AI system. It marks a major milestone for AI applications in the military domain.
Military AI has so far centered mainly on data analysis and decision support, but flying a fighter jet — which involves physical control — sets a new bar for AI reliability and safety. Details about the flight conditions and the degree of AI autonomy are being revealed in stages through DARPA's official announcements.
Ethical debates over military use will continue, but technically this demonstrates that AI-controlled systems are moving into practical deployment in extremely demanding environments.
Reddit Wavers on Renewing Its AI Content Deal with Google — Ripples Across the Content Licensing Market
Reddit's stock fell after CNBC reported that the company may not renew its AI content deal with Google. Reddit's user-generated content has been one of the most important sources of training data for AI models, so uncertainty over the renewal could ripple across the entire industry.
Meanwhile, Chegg, a major learning services company, is suing Google over traffic losses caused by Google's AI summaries — a sign that the relationship between AI companies and content providers is growing ever more complicated.
If the pool of training data available to AI models becomes more restricted, it could relatively raise the value of open-weight models.
Hugging Face Releases "The Stack v3" — One of the Largest Open-Source Code Datasets
Hugging Face has released **The Stack v3**, a dataset built for pretraining code LLMs. It is an open dataset containing source code crawled directly from GitHub, and it substantially updates its predecessor, The Stack v2.
The big changes in v3: file contents are included inline, and the context of entire repositories is preserved. This allows code LLMs to learn not just standalone functions, but also whole-project structure and cross-file dependencies.
The project aims to build an open, reproducible environment for training on code, and it looks set to become an important resource for the transparency-minded AI research community.
LLaDA2.2-flash — a Diffusion Language Model Built for Agent Applications
**LLaDA2.2-flash**, released by inclusionAI, is a Diffusion Language Model designed specifically for agent use cases. As the newest release in the LLaDA2 series, it features a 128K-token context length and Block Routing, which batches MoE expert activation.
Its standout feature is the introduction of Levenshtein Editing (DELETE and INSERT control tokens) into diffusion decoding, targeting performance gains in practical scenarios such as long-context tool use, multi-turn dialogue, and error correction in agent settings.
Diffusion models have seen huge success in image generation, but applying them to language models is still in its infancy — which makes LLaDA2.2-flash worth watching as a real-world proof point for the agent space.