SoftBank Group Corp. and Intel Corp. held talks in recent months about developing an artificial intelligence chip, the Financial Times reported today.

The negotiations, which have since fallen apart, are believed to have been part of a multibillion-dollar push on SoftBank’s part to challenge Nvidia Corp. in the AI accelerator market. The Tokyo-based conglomerate reportedly hopes to not only build AI chips but also develop the software that will power them. Furthermore, it intends to play a role in supplying electricity to customers’ data centers.

SoftBank is the majority stakeholder in Arm Holdings plc, the world’s top provider of central processing unit designs. Arm’s latest CPU blueprint includes optimizations that allow it to run AI models significantly faster than earlier silicon. However, the company has not yet produced a dedicated AI processor with compatible capabilities to Nvidia’s graphics processing units.

Last July, SoftBank expanded its presence in the chip market by acquiring venture-backed semiconductor startup Graphcore Ltd. The latter company developed an AI chip called the IPU Bow that it positioned as a direct alternative to Nvidia GPUs. According to the Financial Times, SoftBank’s plan to enter the AI accelerator market involves combining Arm processor designs with Graphcore’s expertise in “bringing a chip into production.”

SoftBank had reportedly considered partnering with Intel to manufacture its AI chips. The negotiations are believed to have failed because the chipmaker couldn’t meet the project’s “volume and speed” requirements. It’s unclear whether speed refers to chip performance or the amount of time it would have taken Intel to bring the first SoftBank-developed processor to market.

SoftBank has also held talks with Taiwan Semiconductor Manufacturing Co. about making its chips. However, the companies have reportedly not yet inked a supply agreement in a delay that may be connected to the limited capacity at TMSC’s fabs. The Financial Times’ sources said that SoftBank could restart the discussions with Intel if it struggles to find another manufacturing partner.

The Graphcore acquisition bought SoftBank a chip technology called wafer-on-wafer stacking. It provides the ability to place a silicon module with power management components directly atop an AI processor. Using electricity stored in the power management module, the processor can run at faster frequencies than would otherwise be possible.

SoftBank may seek to use the technology to give its planned chips an edge over Nvidia’s graphics cards. Today’s report indicates that the first prototype of the company’s AI processor may be ready in a matter of months.

The development effort could be led by the AI chip design business that Arm is rumored to be setting up. In May, Nikkei Asia reported that the unit was expected to build a prototype machine learning accelerator by spring 2025. At the time, executives were reportedly weighing spinning off the AI chip design business to move it under the SoftBank corporate umbrella.

The chips that the company plans to develop would compete with not only Nvidia graphics cards but also Intel’s growing lineup of machine learning accelerators. The latest addition to the lineup debuted earlier this year in the form of a processor called the Gaudi 3. The chip, which is geared toward data centers, features two core clusters that are each optimized for a different set of machine learning workloads.

Next year, Intel plans to launch a more advanced AI processor codenamed Falcon Shores. The chip was originally set to combine a GPU and CPU circuits in a single package, but it’s believed that the company may have since revised the design. Falcon Shores is expected to consume 1,500 watts of power, significantly more than Nvidia’s fastest data center chips. 

Image: Graphcore