AMD agreed to acquire Taalas, a two-year-old Toronto startup that etches AI model weights directly into a chip's metal layers instead of shuttling them in and out of memory — its early test chip served Meta's Llama 3.1 8B model at nearly 17,000 tokens per second. It's AMD's third AI acquisition in nine months, and the clearest signal yet that specialized, model-baked-into-silicon inference is where the next round of compute-stack consolidation is headed, the same logic pulling Qualcomm toward Tenstorrent and Marvell toward Celestial AI. For the technical case behind burning a model into transistors, The Next Platform's breakdown of Taalas' HC1 chip is worth the ten minutes. (CNBC)
Private Companies
WEKApartnership
WEKA signed Ingram Micro as its exclusive AI data platform distributor in Australia, putting NeuralMesh software and WEKApod storage appliances on Ingram's Xvantage platform for the first time. The deal targets the bottleneck that stalls AI deployments after the pilot stage — data movement and memory, not raw compute. It's evidence that channel reach, not just product depth, is becoming a contested layer in AI storage as vendors race into every regional systems-integrator relationship. (ARN)
Acrabfundraise
Acrab raised a $130M Series B led by existing backers Vertex Ventures SEA & India and Vertex Growth, pushing the Singapore startup's cumulative funding past $350M. The company pairs purpose-built silicon with an edge-AI software stack — its GΞLIX 1-powered Agent Box — to run autonomous AI agents outside the data center, with expected revenue this year as it commercializes. It's a bet that agentic AI workloads need dedicated compute at the edge, not bigger clusters in the cloud. (Yahoo Finance)
Public Markets
Revenue jumped 372% year-over-year to $8.97B and the board approved a $14B buyback, but Q1 FY27 guidance of $10.3B–$10.8B missed the market's bar, dragging SK Hynix and Samsung down in sympathy and triggering a KOSPI sidecar halt. (KED Global)
BNP Paribas warned SpaceX's Nvidia-exclusive supply deal could crowd neoclouds out of scarce GPU allocation, and Nebius fell hardest as CoreWeave and IREN also slid — evidence Nvidia's biggest new customer is now a competitive threat, not just a demand signal. (Seeking Alpha)
Emerging
Photonic inference: A new paper proposes MDTransformer, a photonic transformer accelerator that uses mode-division multiplexing and an inverse-designed coherent crossbar to avoid the costly multi-wavelength light sources and bulky phase-shifter arrays that have made earlier optical accelerators impractical to manufacture at scale. It's the first design to combine mode-division multiplexing with inverse-design methods for this class of chip — a manufacturability fix that matters as photonic compute moves from lab demos toward the co-packaged-optics products Lightmatter and Ayar Labs are now shipping. (arXiv)