Graphcore – The have nots

Sale to OpenAI makes sense. SoftBank less so.

• It makes far more sense for OpenAI to acquire Graphcore than SoftBank as OpenAI clearly wants to become more vertically integrated and it can give Graphcore the one thing it really lacks, customers.
• Despite early success, AI chipmaker Graphcore is in trouble and having failed to raise more money, is seeking a trade sale.
• Graphcore is one of the have-nots in the AI training ecosystem where along with several others it shares 10% – 15% market share of the AI training chip market with the other 85% – 90% belonging to Nvidia.
• Things started well with Open AI, Google and Microsoft all using its chips but in 2022 revenues fell by 46% to $2.7m from $5m as its customers switched to Nvidia.
• The reason is simple: CUDA.
• CUDA is Nvidia’s software development platform for training AI algorithms, and it has become an industry standard.
• This is because it has a 10+ year lead over everyone else meaning that it has the best and most mature tools and everyone in the industry knows how to use them.
• However, CUDA is only available for Nvidia and so if one wants to use CUDA, then one has to buy Nvidia chips.
• GPT-3 was originally trained using a combination of Graphcore and Tensorflow (Google) chips but subsequent versions and its other products like Sora are all trained on Nvidia.
• The other advantage Nvidia has is the speed of its product cadence which everyone else large and small struggles to match.
• Hence, Nvidia always has the fastest, most cost-effective (it claims) chips in the market which combined with its ownership of CUDA has allowed it to crush its competitors.
• Nvidia is currently earning 70%+ gross margins on its AI training chips and demand is so great that it can sell as much as it can make, and its clients remain pretty price insensitive.
• None of them are particularly happy with this state of affairs which is why there is a lot of interest in both in-house and merchant market alternatives but so far no one has been able to lay a glove on Nvidia.
• This is why despite a massive boom in AI investments, Graphcore has struggled to raise money and its valuation has fallen more than 80% from its peak of $2.8bn in 2020.
• The loss of its cornerstone clients was a massive blow from which I suspect it will not recover without a trade sale which may have already been agreed upon.
• This is because according to the Daily Telegraph (see here), Baille Gifford has doubled the valuation of its stake in Graphcore which is something that can’t happen unless a corporate action such as a sale or fundraising has been agreed.
• It is possible that a new round with only existing investors could have triggered the upgrade in valuation, but this is considered to be a pretty dubious practice.
• Hence, a sale may have already been agreed in principle with some of the finer details yet to be agreed before the deal is announced.
• Of all the mooted acquirors, OpenAI is the one that makes the most sense because Graphcore would be a good fit for both companies.
• OpenAI is actively considering designing its own training chip (along with everyone else who hasn’t already done so) while Graphcore has a chip that nobody seems to want to use.
• Hence, by owning Graphcore, OpenAI would be able to standardise its growing stable of foundation models on an in-house chip which would help reduce its dependency on Nvidia.
• This would allow OpenAI to fully optimise its foundation models for Graphcore, which in all likelihood make it a much stronger competitor to Nvidia than it currently is.
• This also plays into the very early emergence of the AI ecosystem where developers base their services on an easily available foundation model and care less about the silicon that is used to train it.
• This is at a very early stage, and I suspect that it will take years to materialise but it’s the only way I can see Nvidia’s grip on AI training being weakened.
• A sale to SoftBank might result in a much better-than-expected exit for investors, but it does nothing to solve Graphcore’s customer problem and its ability to compete against Nvidia.
• Hence, if Graphcore goes anywhere except to a large foundation model owner (Google, OpenAI, Meta Platforms etc), this will be good news for Nvidia.
• Nvidia’s valuation is not stratospheric at 32.7x 2024 PER and 28.3x 2025 PER but this depends on the AI bubble remaining inflated.
• This is not a chance I would particularly want to take.