AI silicon beyond the GPU: NVIDIA, AMD, Huawei and custom accelerators, covering architecture, memory bandwidth, supply and what each means for inference cost.
NVIDIA is framing a shift in AI compute demand: from massive one-time training runs to continuous post-training refinement for agents, according to Moomoo — a change that reframes both the hardware market and the NVDA valuation debate covered by Yahoo Finance UK.
General Compute has closed a landmark US$400 million deal that uses inference chips as loan collateral in place of Nvidia GPUs, according to Tech Times. We unpack what it signals for AI hardware financing.
Jensen Huang says Nvidia will be the first customer for HBM4, and one AI memory stock has reportedly secured about 70% of those orders, according to Yahoo Finance and The Motley Fool.
The FBI is reportedly weighing an in-house AI LLM supercomputer built on either Nvidia B300 GPUs or Google TPUs, according to Data Center Dynamics — a rare federal signal on accelerator choice.
OpenAI and Broadcom have jointly unveiled an inference chip optimised for large language models, signalling a decisive move toward vertically integrated AI silicon.
Huawei has turned its AI silicon plans into a dated, one-chip-a-year roadmap, opened the openPangu 2.0 models, and committed to open-sourcing CANN. We weigh all of it against the constraints the keynote skipped: a 7nm ceiling at SMIC, a homegrown HBM supply that can feed only a few hundred thousand chips, and a per-chip gap to NVIDIA that Huawei’s own roadmap admits.
An NPU and a GPU both run AI, but they are built for opposite jobs. Here is what separates them in 2026, with real TOPS and TFLOPS numbers from the chips actually shipping.
