Tags: TPU

Google has introduced Multi-Token Prediction (MTP) drafters for its Gemma 4 open models, promising up to a two‑fold reduction in response time for locally run AI. The experimental feature uses speculative decoding to guess future tokens, allowing a lightweight draft model to fill idle processing cycles. Built on the same architecture as Gemini, Gemma 4 can run on a single high‑power accelerator or, when quantized, on consumer‑grade GPUs. A shift to an Apache 2.0 license also makes the models more permissive, encouraging broader adoption of edge AI. Read more

Alphabet's Google will invest $10 billion in AI startup Anthropic now, with an option to add $30 billion if the firm hits performance targets. The agreement also secures five gigawatts of Google‑provided TPU capacity by 2027 and mirrors a similar multi‑billion‑dollar pact the startup struck with Amazon earlier this year. Anthropic, fresh from a $30 billion funding round, will rely on Google’s silicon and cloud infrastructure as part of the deal, underscoring the growing trend of circular investments in the artificial‑intelligence sector. Read more

Alphabet’s Google Cloud announced a multi‑phase deal that could total $40 billion for AI lab Anthropic. The first tranche of $10 billion values the startup at $350 billion, with an additional $30 billion tied to performance milestones. The partnership expands Anthropic’s access to Google’s tensor processing units and adds roughly 5 gigawatts of compute capacity over five years. The move follows Anthropic’s rollout of its latest model, Mythos, and comes as the firm also secures funding from Amazon and eyes a possible IPO later this year. Read more

San Francisco‑based Anthropic is weighing the development of its own artificial‑intelligence chips, according to three sources familiar with the effort. The move comes as the company’s annualized revenue run rate for its Claude models surged past $30 billion, up from roughly $9 billion at the end of 2025. Anthropic still runs workloads on a mix of Google‑Broadcom TPUs, Amazon‑custom silicon and Nvidia GPUs, and has just secured a long‑term deal for 3.5 gigawatts of TPU capacity beginning in 2027. The firm has not yet formed a dedicated chip team and may continue buying off‑the‑shelf silicon. Read more

Anthropic announced on April 6 that it will tap roughly 3.5 gigawatts of next‑generation Google Tensor Processing Unit (TPU) compute through Broadcom starting in 2027, adding to the 1 GW already supplied for 2026. The move backs the AI lab’s $50 billion pledge to expand U.S. AI infrastructure and comes as the company reports a revenue run‑rate exceeding $30 billion—more than triple its figure at the end of 2025. Broadcom’s role as the silicon‑to‑workload bridge and the scale of the deal underscore the accelerating compute arms race among AI firms. Read more

Google has elevated longtime AI infrastructure architect Amin Vahdat to the newly created role of chief technologist for AI infrastructure, reporting directly to CEO Sundar Pichai. The move underscores the importance of AI compute as Alphabet plans to spend up to $93 billion on capital expenditures through 2025. Vahdat, a former professor with a PhD from UC Berkeley, has driven key projects such as the seventh‑generation TPU "Ironwood," the high‑speed Jupiter network, the Borg cluster manager, and the Axion Arm‑based CPUs. His promotion signals Google’s commitment to maintaining a competitive edge in the fast‑evolving AI hardware landscape. Read more

Google announced Project Suncatcher, a moonshot initiative to explore placing its Tensor Processing Units (TPUs) on solar‑powered satellite constellations in low‑Earth orbit. The goal is to scale machine‑learning compute in space by creating swarms of satellites equipped with AI accelerators for tasks such as training, content generation, synthetic speech, vision, and predictive modeling. Google’s senior director Travis Beals highlighted growing AI demand as a driver, while CEO Sundar Pichai noted early tests show TPUs can survive intense radiation, though thermal management and on‑orbit reliability remain challenges. Read more

Google is researching the concept of placing AI hardware in low‑earth orbit through a project called Suncatcher. The plan envisions solar‑powered satellites carrying Tensor Processing Units (TPUs) to run machine‑learning models using continuous, clean energy. While the idea promises higher power efficiency and reduced carbon emissions, Google acknowledges significant technical hurdles such as radiation exposure, high‑speed inter‑satellite data links, and precise formation flying. Economic analysis suggests comparable power efficiency to Earth‑based data centers by the mid‑2030s, and the company aims to launch prototype satellites by 2027 to test the concept. Read more

Google is developing Project Suncatcher, a plan to place AI‑focused data centers on a free‑fall satellite constellation. The design calls for tightly spaced satellites—within a kilometer or even several hundred meters—to maintain power links, a formation tighter than any existing constellation but deemed feasible by Google’s models. To keep costs down, Google intends to reuse Earth‑based hardware, testing its durability by exposing its latest Cloud TPU to intense radiation. Prototype satellites could launch by early 2027, with broader deployment targeted for the mid‑2030s when launch costs may fall dramatically, offering a potential solution to the environmental and community challenges of terrestrial data centers. Read more

Google has unveiled Project Suncatcher, a research effort to place AI‑focused Tensor Processing Units on solar‑powered satellites, creating data centers in orbit. The company argues that space could offer near‑continuous solar energy, potentially making compute more sustainable. Key hurdles include ultra‑high‑speed inter‑satellite links, tight formation flying, radiation tolerance, and cost competitiveness. Google plans a joint launch with Planet to test prototype hardware by 2027, hoping the approach could become comparable to Earth‑based energy costs by the mid‑2030s. Read more

Google Cloud is focusing on early‑stage AI companies, offering $350,000 in cloud credits, technical assistance, and go‑to‑market support. The firm promotes an open AI stack that spans custom TPUs, foundation models and applications, aiming to win future unicorns before they grow large. Partnerships include TPU deployments with Fluidstack and collaborations with startups such as Loveable and Windsurf, while Google also hosts Anthropic’s Claude and provides TPUs to OpenAI. The strategy reflects Google’s broader commitment to open‑source tools and comes amid regulatory scrutiny of its search dominance. Read more

Google introduced its seventh‑generation Tensor Processing Unit, dubbed Ironwood, at a recent Hot Chips event. The dual‑die chip delivers 4,614 TFLOPs of FP8 performance and pairs each die with eight stacks of HBM3e, providing 192 GB of memory per chip. When scaled to a 9,216‑chip pod, the system reaches 1.77 PB of directly addressable memory—the largest shared‑memory configuration ever recorded for a supercomputer. The architecture includes advanced reliability features, liquid‑cooling infrastructure, and AI‑assisted design optimizations, and is already being deployed in Google Cloud data centers for large‑scale inference workloads. Read more