AI-chips

Huawei's rotating chairman Eric Xu says Ascend has surpassed Nvidia in China by Huawei's own data, while DeepSeek released an Ascend port of its DeepGEMM kernels.

Anthropic discussed buying AI chip startup MatX for roughly $7 billion, then walked away; the two sides are now said to be exploring a supply partnership instead.

NVIDIA's Groq-derived Groq 3 LPX inference accelerator is now shipping, promising ultrafast token generation for agentic coding workloads, with Nebius first to deploy it.

UK inference-chip startup Fractile is in talks to raise about $600M at a $6.5B pre-money valuation, driven by an initial $250M chip deal with Anthropic.

AMD is buying Toronto-based Taalas, whose chips etch model weights directly into silicon instead of storing them in memory, to strengthen its AI inference lineup.

Meta plans to start manufacturing its in-house Iris AI chip in September, aiming to double computing capacity to 14 gigawatts in 2027, an internal memo reviewed by Reuters shows.

Meituan released LongCat-2.0, a 1.6T-parameter MoE model under an MIT license, claiming it is the first trillion-parameter model trained and served entirely on domestic ASICs.

At its June 24 Investor Day, Qualcomm detailed a multi-generation data center chip line built for agentic AI inference, anchored by the Dragonfly C1000 CPU and AI300 accelerator, and named Meta as a server-CPU customer.

Reports say Qualcomm is examining a full acquisition of Tenstorrent, the RISC-V AI accelerator startup led by Jim Keller, at a valuation between $8 billion and $10 billion.

OpenAI's first custom silicon, the Broadcom-built Jalapeño inference accelerator, went from design to tape-out in nine months and targets deployment by the end of 2026.

The startup says its 3nm Napier inference chip turns matrix multipliers into adders, with commercial sales of a 72-chip system slated for the second half of 2027.

d-Matrix says its Corsair inference accelerator is in full production, pairing on-chip SRAM with GPUs to cut AI response times by more than 10x.