An AI accelerator is a processor or hardware system optimized for the highly parallel mathematical operations used by machine learning models. GPUs, tensor processing units, neural processing units, and purpose-built inference chips are common examples. Compared with general-purpose processors, accelerators can increase throughput, reduce latency, or lower the energy required to train and run models. Their design may emphasize matrix multiplication, low-precision arithmetic, fast memory access, or communication between many chips. Choosing an accelerator depends on the model architecture, batch size, software support, deployment environment, power budget, and whether the workload involves training, real-time inference, or both.
King Slide Works, once known for cabinet hinges, now dominates the market for AI server rack rails, making founder Lin Tsung-chi Taiwan’s richest person.
OpenAI’s first custom inference chip, Jalapeño, delivered up to 1.9 times more work per watt and far lower latency than rival systems, with AI accelerating its own design.
Apple launched new Mac Mini and Mac Studio models with its M6 and M5 Ultra chips, explicitly built for developers running AI models and agents locally instead of in the cloud.
OpenAI paused reinforcement learning on its newest models after it could not rule out that an upcoming model, Astra, reached the top cybersecurity risk tier in its safety framework.
Google is designing Frozen v2, a chip that hardwires parts of its Gemini model into silicon and could serve six to ten times more tokens per watt, targeting deployment by 2028.
OpenAI launched Codex Micro, a $230 mechanical keypad built with Work Louder that gives developers physical controls for its Codex coding agent, its first hardware to reach market.
SK Hynix raised $26.5 billion in its Nasdaq debut, the largest US listing ever by a foreign company, as investors rushed to buy into the AI memory chip boom.
Meta plans to begin manufacturing its in-house Iris AI chip in September as it aims to double computing capacity to 14 gigawatts next year and cut its reliance on Nvidia.
Apple expanded its partnership with Broadcom in a multi-year deal worth more than $30 billion, its largest US manufacturing commitment, to produce over 15 billion American-made chips.
Nvidia’s Kyber rack for its 2027 Rubin Ultra chips has slipped more than a year to 2028 over a hard-to-manufacture circuit board, SemiAnalysis says, with no proven fallback.