Chip design, Huawei software and a blocked energy bill: the accelerator week in review

THE DECODER reported the chip design deal on 30 September. The two companies will market GPT-Synopsys together and share the revenue. OpenAI is licensing Synopsys' EDA tools for the project, and the model will run on OpenAI's infrastructure. Both companies say customer data will not be used for training and will be stored encrypted.

Early tests with semiconductor customers are already underway.

Synopsys CEO Sassine Ghazi says AI could significantly speed up the design process. OpenAI co-founder Greg Brockman frames the deal as a path to better chips and better AI. THE DECODER also notes OpenAI's existing chip work with Broadcom, including the recently unveiled Jalapeno inference chip.

Washington votes, and the bill falls three short

The Senate blocked the Ratepayer Protection Act that same day, by 57 votes to 43. That left it three short of the 60 needed to advance. The Guardian reported that four Democrats, Maggie Hassan, Amy Klobuchar, Jon Ossoff and Raphael Warnock, joined Republicans in support. The House had approved the bill earlier in September by an almost unanimous vote.

The bill would have required electric utilities to consider standards so that datacenters pick up the cost of transmission, generation and distribution upgrades that can otherwise land on households. Senate Democratic leader Chuck Schumer called it a "toothless datacenter bill" that "completely misses the mark". He also criticised a Tuesday meeting between Trump and tech executives that the president described as yielding a "morally binding" agreement on AI controls. Ohio Republican Jon Husted, who championed the bill, said big tech "should be paying their own way". The vote was one of the last expected before the 3 November midterms.

Deepseek hands Huawei a software stack

China's side of the stack moved too.

THE DECODER reported on 30 September that Deepseek is open-sourcing programming tools for Huawei's Ascend chips, with TileLang at the centre. Deepseek says Huawei "fully supported" the work, and the two companies optimised a supernode of 128 Ascend 950 chips. TileLang was originally developed at Peking University, and Deepseek has used it for about a year. The New York Times, cited in that report, says TileLang is now Deepseek's main tool for its AGI work.

The context matters more than the code. Nvidia's position rests partly on an estimated four million CUDA developers, a moat AMD has struggled to cross. SemiAnalysis, quoted in the same piece, tested OpenAI's Jalapeno chip and called that moat "potentially dead". It tested only easy-to-optimise scenarios, roughly 8,000 input tokens and 1,000 output tokens, and has not run its AgentX benchmark. In August SemiAnalysis found Nvidia still well ahead on agent workloads.

Hardware trackers add the price dimension.

Flopper.io lists H100 SXM5 80GB at a $1.79/hr lowest rate against a $3.63/hr median, and B200 SXM 180GB at $4.09/hr lowest versus $6.64/hr median, with B200 availability marked as a waitlist. The same database lists Instinct MI355X OAM at $5.99/hr, a single unconfirmed listing, which is a thin basis for a price comparison.