SoftBank and Graphcore – AI Armoury

More than meets the eye.

• Contrary to my expectations (see here), SoftBank has acquired AI chipmaker Graphcore for around $600m which is way below its peak valuation meaning that the shares of many employees and investors will end up being worthless.
• Graphcore has been in trouble from the time it lost its two biggest customers (Microsoft and Google) and has been struggling for traction and burning a lot of cash ever since.
• My understanding is that these losses stemmed from a design decision that failed to take into account the requirement for high bandwidth memory caused by ever-increasing model sizes.
• This was a blow from which the company has never recovered and in October last year it admitted to “material uncertainty” which clearly meant “we need more money or we are out of business”.
• The subsequent raise did not go well, leaving the company having to seek a trade sale earlier this year.
• At the time, both OpenAI and SoftBank were in line to buy it and I thought that OpenAI made more sense as it is actively investigating making its silicon and Graphcore would have accelerated that process.
• However, it looks like SoftBank has ended up acquiring the company which I suspect in the long term could easily be merged with Arm.
• The transaction is being financed by SoftBank itself and not the Vision Fund indicating that this acquisition is opportunistic and will serve as part of Mr. Son’s grand plans for AI.
• This is a lease of life for Graphcore as it will now have funds to continue developing its technology and to continue hunting for a chink in Nvidia’s armour.
• The problem is that this chink is not going to come for some time because as long as CUDA remains the preferred development environment of AI developers, I don’t think anyone else is going to get a foot in the door.
• However, this may not always be the case and at that point, Graphcore will have a much better chance.
• SoftBank allows Graphcore to survive until that moment.
• However, what is more interesting is the potential for Arm and Graphcore to work more closely together or even merge which would greatly increase Arm’s position in AI.
• At the moment it is doing well in AI that is executed on edge devices and is gaining a lot of traction in the cloud as part of Nvidia’s offerings, but it does not have a complete training offering of its own.
• Graphcore could fill that gap, but creating a fully integrated solution will take some time.
• This works well because I suspect it will be a while before anyone can dislodge Nvidia.
• Unfortunately, SoftBank’s gain is Graphcore investors’ loss as liquidity preference in some share classes will mean that those lower down the table will get wiped out and receive nothing.
• For once, it appears that SoftBank might not have overpaid.