AI Industry

Google is developing custom AI chips aimed at accelerating inference — the process of generating responses from trained models — as part of a push to reduce dependence on Nvidia GPUs, according to Bloomberg. The report says Google is in talks with Marvell Technology to co-design new silicon, expanding beyond its longstanding Broadcom partnership on Tensor Processing Units. The move targets the inference segment of AI compute, where hyperscalers are increasingly looking for cost and latency advantages over general-purpose GPUs. Multiple outlets, including Benzinga and Business Today, have corroborated the Marvell discussions. Neither Google nor Marvell has issued an official statement.

Axios reports that Fermi America, the Texas Panhandle AI data center project co-founded by former Energy Secretary Rick Perry and bearing President Trump's name, is facing delays, cooling supply chain issues, and the lack of a publicly confirmed anchor tenant. CEO Toby Neugebauer departed Friday, sending shares lower after a 75% six-month decline.

Google is discussing a chip design partnership with Marvell Technology that would add a third custom silicon partner alongside Broadcom and MediaTek, according to The Information as reported by The Next Web. The talks cover a memory processing unit and an inference-optimised TPU, and follow Broadcom's through-2031 TPU agreement signed days earlier.

Companies including Duos Edge AI and LG CNS are shipping pre-fabricated, GPU-packed compute pods that can be operational in roughly six months — a fraction of the two-to-three years a conventional data center requires. Each unit holds up to 576 Nvidia GPUs, can be networked with others on a concrete pad, and costs around $25 million for a five-megawatt deployment. The approach is drawing wider industry attention as AI hardware sits idle waiting for permanent facilities.

At Nvidia's GTC conference in San Jose this week, CEO Jensen Huang announced the Groq 3 language processing unit (LPU), a chip built specifically for AI inference. The announcement came just two and a half months after Nvidia licensed intellectual property from startup Groq in a deal valued at $20 billion. The chip features an SRAM-based architecture delivering 150 terabytes per second of memory bandwidth — seven times faster than Nvidia's own Rubin GPU — and will ship as part of a combined compute tray alongside Vera Rubin chips.

Meta has published a comprehensive overview of its current AI portfolio, highlighting four custom MTIA chips shipped in two years, a long-term hardware partnership with AMD, the Llama 4 open model family, and a new AI video feature called Vibes. The company frames its overarching mission as delivering 'personal superintelligence' to individuals — a deliberate contrast to competitors it accuses of seeking centralised control over AI output.

GGML and llama.cpp, the open-source libraries powering much of the local AI inference ecosystem, have joined Hugging Face in a move designed to ensure their long-term development and sustainability. The partnership brings two of the most widely used tools for running large language models on consumer hardware under the Hugging Face umbrella, signalling a consolidation of the open-source AI infrastructure stack.

Mistral AI unveiled three significant releases on March 16–17, 2026: Forge, an enterprise platform for building proprietary AI models; Mistral Small 4, its latest compact research model; and a founding partnership with NVIDIA's Nemotron Coalition. The cluster of announcements signals Mistral's push to compete across enterprise deployment, open-source development, and hardware-level AI infrastructure simultaneously.

Google published its monthly AI update summary for February 2026 via the Google AI Blog, consolidating product and research announcements from across the company. The post signals Google's continued practice of bundling AI news into digest-style releases, though the specific details of the announcements were not available in the source material at time of publication.

Hugging Face has published its Spring 2026 State of Open Source report, offering a snapshot of activity across its platform. The report tracks model uploads, dataset growth, and community engagement trends that reflect the broader health of the open-source AI movement. As proprietary and open models continue to compete for developer mindshare, Hugging Face's platform data provides one of the clearest windows into where open-source AI development is heading.