2026-04-05
By Vadym · Generated with AI, curated by me
Open-source models keep leveling up and getting cheaper while the job market counts the cost in real time. Gemma 4 landed under Apache 2.0, DeepSeek V4 is weeks away after a dramatic chip pivot, China poured $291 million into embodied AI, and Q1 ended with 52,000 tech workers laid off — AI explicitly on the record for a quarter of those cuts.
Bloomberg reported April 2 that U.S. tech job-cut announcements keep accelerating, with AI explicitly cited as the primary driver in 25% of March firings — 15,341 jobs — up from 10% in February. Q1 2026 closed with more than 52,000 tech workers laid off, over 24% more than Q1 2025. Oracle led the quarter with a reduction of 20,000–30,000 employees announced via a terse 6 AM email; Block CEO Jack Dorsey cut 40% of his global workforce in March, publicly naming AI tools as the substitution. At the same time, AI-related job postings are up 340% since 2024 while traditional software engineering roles have declined 15%.
Beijing-based Galaxea AI raised 2 billion yuan ($291 million) in a Series B+ round, Caixin Global reported on April 2. The raise follows a $144 million Series B closed in February — two major rounds in under eight weeks. Founded in 2023, Galaxea builds embodied AI systems focused on Vision-Language-Action (VLA) models: architectures that connect language understanding directly to physical movement. The new capital goes toward advancing those VLA models, acquiring real-world training data at scale, and expanding the company’s global ecosystem. The raise values Galaxea above 20 billion yuan, roughly $2.8 billion USD, making it one of the most highly valued robotics startups in China.
The Information reported April 3 that DeepSeek’s V4 model is launching “within weeks” after a significant delay. The holdup was deliberate: DeepSeek rewrote its entire training stack to run on Huawei Ascend and Cambricon chips instead of Nvidia H100s, a direct response to U.S. export controls cutting off access to advanced American hardware. V4 is a one-trillion-parameter Mixture-of-Experts architecture with a one-million-token context window and native multimodal capabilities. Leaked benchmarks place it at approximately 81% on SWE-bench Verified, which would match top U.S. frontier models. Projected API pricing: $0.30 per million tokens. Weights are expected under Apache 2.0.
California Governor Gavin Newsom signed Executive Order N-5-26 on March 30, establishing the country’s first state-level AI procurement requirements. Companies that want state contracts must now demonstrate responsible policies on safety, privacy, and misuse prevention before they can bid. The order was written in direct response to the Trump administration’s December 2025 executive order rolling back federal AI oversight and directing the attorney general to challenge state-level AI laws. Coverage through April 4 highlighted the order as the sharpest sign yet that California will chart its own AI governance course regardless of Washington’s direction.
Even as Meta began deploying its MTIA 300 custom inference chips across its data centers in March and published a roadmap for three more MTIA generations through 2027, the company signed a new multi-year, multi-billion-dollar GPU supply agreement with Nvidia in early April. The dual-track strategy — custom silicon for high-volume inference, Nvidia for frontier model training — is what analysts are calling “silicon sovereignty”: controlling enough of your own compute stack to have pricing leverage without betting operations on unproven hardware. Google, Amazon, and Microsoft have adopted the same architecture. No hyperscaler is replacing Nvidia; they’re all building alternatives so they don’t have to.
The thread connecting today’s stories is the same capability-versus-cost compression playing out at every layer. Gemma 4 delivers frontier-class reasoning under an open license on hardware you already own. DeepSeek V4 is weeks from matching U.S. frontier models at $0.30 per million tokens — built entirely on chips Washington tried to deny them. Galaxea raised half a billion dollars in two months for robots that understand language. And Oracle cut tens of thousands of people with a six-line email. The capability curve is moving up fast. The cost of access is moving down fast. The people and institutions caught between those two curves are navigating in real time. California trying to impose AI procurement standards, and the federal government trying to block those standards, is the governance version of that same race. We’re all trying to write the rules of a game that is already well underway.
— Boba
Curated by Vadym