DeepSeek designs its own AI chip; Gemma 4 raises the open-model bar

Reuters reports that DeepSeek has spent about a year secretly recruiting chip engineers and courting manufacturing partners to build an in-house processor for AI inference — the stage where a trained model answers user queries. The goal is to cut its dependence on Nvidia and Huawei as US export controls keep tightening. It follows OpenAI's Broadcom-built Jalapeno inference chip and Anthropic's own reported chip ambitions — a sign that frontier labs increasingly want to own their silicon.

Google released Gemma 4, a family of open-weight models from 2.3 billion to 31 billion parameters that natively handle text, images, and audio. The larger models add a built-in reasoning mode and, Google says, rival much bigger frontier open models on human-rated tasks. Because the weights are downloadable and run locally, Gemma 4 gives developers a capable option that does not depend on a paid API.

Bloomberg reports that Microsoft has begun routing a share of Excel and Word prompts to its own in-house MAI models instead of paying for OpenAI and Anthropic. Microsoft still uses those partners, but the shift follows its Build launch of seven new MAI models, including an agentic coder and an image generator. It puts Microsoft alongside Amazon, Meta, and others tightening AI spending as infrastructure bills climb.

Meta launched Muse Image, a free AI image generator built into the Meta AI app, Instagram, and WhatsApp. The controversy is a feature that lets users transform other people's public Instagram photos into new AI images without notifying them — and it is on by default, so users must opt out. Given Meta's Cambridge Analytica history, critics called the tool a privacy landmine.

SambaNova, which designs chips tuned to run large models efficiently, raised $1 billion in a round led by General Atlantic that values it at $11 billion — just five months after its last mega-round. Backers include Intel, BlackRock, and T. Rowe Price. The company pitches itself as an inference-focused alternative to Nvidia, and its CEO signaled that a public listing, not a sale, is the likely path.

Morgan Stanley began covering the newly public SpaceX with an Overweight rating and a $300 price target, arguing the launch business is a small slice of the story. The bulk of the value sits in Starlink and, increasingly, AI compute: SpaceX cites multibillion-dollar compute deals with Anthropic and Google, and Morgan Stanley expects its AI-related revenue to climb from about $3 billion in 2025 toward roughly $190 billion by 2030.

In a six-month study across 10 major US cities, Google tweaked Maps to steer a small share of drivers onto alternate routes with similar travel times, spreading traffic out instead of piling everyone onto the fastest road. The result was measurable: modestly higher speeds and lower fuel burn on targeted roads, adding up to thousands of tons of avoided carbon per city each year — a quiet example of AI optimizing a shared system rather than replacing anyone.