pull down to refresh

OpenAI plans to mass‑produce its own AI accelerator in 2026 with Broadcom, aiming to reduce reliance on Nvidia. The in‑house chip will power its data‑centers and highlights a wider push to control scarce hardware.
150 sats \ 4 replies \ @gmd 6 Sep
This is probably a dumb question but if it's so easy to make new chips why is China so behind and dependent on Nvidia's older tech?
reply
Broadcom has been building neural chips for a while.
China is no less dependent on Nvidia tech than the US. What makes you think they're behind?
reply
100 sats \ 2 replies \ @gmd 7 Sep
Everyone keeps making a big deal about China trying to get around sanctions to get Nvidia chips and Nvidia only being able to sell them limited older chips...
reply
256 sats \ 1 reply \ @optimism 7 Sep
Most recently in FT, Nvidia and AMD to pay 15% of China chip sale revenues to US government for an export license for their nerfed chips, which China was skeptical of (but they'll buy them anyway, because energy abundance) (also see: #1197313 and #1076020)
But in the meantime, Huawei has pivoted into AI hardware: How Huawei’s Ascend AI chips outperform Nvidia processors in running DeepSeek’s R1 model (paper) (also see: #965816 and #1075561). They're modeling this closer to Google's TPUs (what Gemini is trained on and for example is powering Colab) than Nvidia's GPUs: ultra modular.
As we all saw with Deepseek R1, export control doesn't work in a prohibitive way; it actually incentivizes innovation. Protectionism hurts the protectionists. I don't believe any US CEO or US state agent telling me China is far behind. They just are peddling US AI, which are consistently underdelivering on the promises.
AGI with GPT-5, right, lol
reply
Your analysis is well noted, thanks for sharing Sir.
reply
10 sats \ 0 replies \ @Entrep 6 Sep
Chips are where the wars for AI will be fought so OpenAI is rooting for independence
reply