Cerebras CEO actually finds common ground with Nvidia as startup notches IBM win

Cerebras Systems' dinner-plate-sized chips currently power the latest AI inference offerings from Meta and, soon, those of IBM, but US trade policy weighs heavy on its prospects worldwide. "The [AI] diffusion rule is bad policy," CEO Andrew Feldman said during a press conference ahead of IBM's annual Think conference, which …

[O'Reg Inalsin]

Fools errand

US Commerce Secretary Gina Raimondo, who has been leading the Joe Biden administration’s effort to restrict China’s progress in developing and using advanced chips, now says the effort is a “fool’s errand ... Trying to hold China back is a fool’s errand,” Raimondo said in an interview with the Wall Street Journal, adding that investing in building a chip supply chain was more effective than export controls to counter tech rival Beijing. Raimondo said export controls were mere “speed bumps” for China and had not slowed the country’s push for tech dominance or its progress in building semiconductor capabilities. “The only way to beat China is to stay ahead of them… We have to run faster, out innovate them. That’s the way to win,” Raimondo told the WSJ.

That's the truth. But if you look back over all her history of statements about an embargo to AI she was overall a hawk putting enormous pressure NVIDIA. The explanation for that is that success in politics requires bowing to knee jerk short term pressures and slogans and not speaking about what you actually think is best in the long run (here long run is >= just 4 years). Nvidia writes off $5.5 billion in GPUs as US gov't chokes off supply of H20s to China -- . Those would have been obsolete in 4 years anyway. Instead, Huawei's GPU unit has effectively been given steroid injections.

Nvidia's H20 HGX is a cut-down version of its H100 processor that the company specifically engineered to comply with the U.S. government's restrictions imposed on GPUs bound for China in late 2022. Compared to H100, the H20 features significantly reduced AI performance and dramatically reduced HPC performance, which makes it barely suitable for supercomputer workloads that require FP64 compute precision.

I think it's not unreasonable to assume that DeepSeeks's surprising capability/resource ratio was due exactly because of working with constrained resources forcing them to be mean and lean. IMO, the US big players could use some of that medicine - their road to profitability would be shortened.