Microsoft unveiled Maia 300, its third-generation inference accelerator, built on TSMC's N2P process with eight HBM4 stacks for 288GB and 18TB/s of memory per chip, 2.5x the bandwidth of Maia 200. A 96-accelerator liquid-cooled rack draws 167 kW, double Maia 200's density, and Microsoft claims up to 7x the rack-scale token throughput and more than 2.5x the performance per watt, with Copilot, GitHub Copilot and Foundry as the target products. Reuters reported that Microsoft plans to deploy potentially more than 1M of the chips next year. The rack, the Ethernet-based scale-up network that links up to 12,288 accelerators in a superpod, and the memory are specified together, which narrows what merchant GPUs and networking can address in Microsoft's inference fleet; the performance figures are Microsoft's own and no outside party has tested the chip. (Microsoft)
Private Companies
Decartother
Amazon is in talks to acquire Decart for about $7B, the Wall Street Journal reported Sunday, about 75% above its May valuation of nearly $4B, set when it raised $300M led by Radical Ventures with Nvidia joining. Decart builds real-time video models, including Lucy for live editing and Oasis for simulated training environments. Anthropic explored a $6B purchase and withdrew after diligence. Live generation is a continuous inference workload, so an owner with its own chips and capacity changes its unit cost (Reuters)
Oxide Computerfundraise
Oxide Computer raised a $445M Series D led by Eclipse, with AMD Ventures, Jane Street and others participating, after a $200M round in February. Its racks arrive pre-integrated, with 24 sleds that each carry an AMD CPU of up to 144 cores, and the company says it scaled manufacturing capacity 20x in 12 months while demand still exceeds supply. Proceeds fund materials, manufacturing scale and a sled that can host a GPU, which would extend a CPU-only on-premises rack into AI workloads (Oxide)
Public Markets
Applied unveiled its EPIC Center, a $5B facility where chipmakers, equipment partners and universities will develop process technology and equipment alongside its engineers, with TSMC, Samsung, SK hynix, Micron and Intel as founding chipmaker partners. (Applied Materials press release)
Nvidia is in early talks to buy Reflection AI or deepen its investment, the Financial Times reported Saturday, and one option is an acqui-hire without an acquisition. It has already put $800M into the open-weight model developer, and the talks could still fall apart. (Reuters)
Emerging
Diffusion LLMs: A paper from Minghan Jiang, Jiayi Wang, Shuaiting Li, Haibin Shen and Kejie Huang, DynaTE: Accelerating Diffusion LLMs via Dynamic Token Execution, designs an accelerator that skips computation on low-utility tokens and uses a reconfigurable processing-element array to follow the shifting set of active tokens. Across two diffusion LLMs it reports 2.05x to 2.78x speedup and 2.99x to 3.93x energy efficiency over prior diffusion-LLM accelerators, and 2.55x speedup and 6.07x efficiency over Nvidia's Jetson AGX Orin. Diffusion decoding changes the compute pattern that autoregressive inference chips were built around, so accelerator design for it is still open (arXiv)