Artificial Intelligence – The Dynamo Effect

Dynamo is CUDA for inference.

• Meta’s failure to acquire FuriosaAI is a sign that the focus of AI is moving towards inference where rival chip companies have a better chance of competing against the incumbent, but as usual, Nvidia has a fix for that.
• FuriosaAI is a Korean chip company that offers data centre chips for inference which would have helped accelerate Meta’s plans to support its AI efforts with in-house silicon.
• However, FuriosaAI rejected the $800m offer preferring to remain independent, raise money separately and eventually seek an IPO.
• The fact that FuriosaAI thinks that it can survive on its own is a sign that the market is shifting from training to inference which is something that Nvidia alluded to at GTC 2025 last week.
• This makes sense because these “reasoning” models work by either “thinking” for longer or generating many answers and then having a separate model select the best one.
• Either way, the compute consumption by inference increases by many orders of magnitude when using this type of inference meaning that demand growth for inference is likely to significantly outstrip training going forward.
• The problem for Nvidia here is that the CUDA platform that has served it so well for training is less sticky for inference as many players already train on Nvidia but then run inference on something else.
• This is where FuriosaAI would have helped Meta which has said that it will spend $60bn to $65bn on capital expenditures in 2025, the vast majority of which will be spent on data centres for AI.
• FuriosaAI would have helped Meta accelerate its silicon independence and allowed it to equip its data centres with in-house inference silicon much more quickly.
• Instead, Meta will be forced to source more silicon from outside to build the data centres in the planned time frame and Nvidia will be a lead contender to take the slot.
• This is where Nvidia’s new product Dynamo comes in which I think was the most important announcement at GTC 2025 which hardly anyone is talking about.
• Instead, most of the focus was on Blackwell Ultra and Rubin but with inference really taking off, it is Dynamo that will help Nvidia maintain its market share as CUDA becomes less important in the purchase decision of customers.
• Dynamo is a software layer that sits in the data centre and ensures that the data centre outputs as many tokens as possible by optimising GPU, memory and communications within the data centre.
• According to Nvidia, Dynamo increased the output for DeepSeek R1 by 30x implying that the economics of a data centre can be greatly improved by using Dynamo.
• However, Dynamo has been designed to run only on Nvidia silicon and while it will be possible to port it to other chips as it is available in open source, I suspect that all of the savings will disappear when it is on non-Nvidia silicon.
• This means that it makes no sense to run Dynamo on non-Nvidia silicon despite the software being available in open source.
• Hence, if Dynamo proves popular, it will recreate the lock-in that CUDA has created for training and make it harder for other players to take share from Nvidia.
• This is why I think Dynamo was the most significant announcement made at GTC 2025 and is a similar strategy to Nvidia Inference Microservices (NIMs) which also aims to keep clients on Nvidia silicon as opposed to competitors.
• These are medium to long-term strategies to keep market share as the market evolves but in the short term, CUDA looks to be as sticky as ever.
• This combined with its rapid product cadence makes the next 12-18 months look pretty safe from the point of view of market share and gross margin.
• Hence, I think that Nvidia is likely to continue to grow in line with the capex spend for data centres which remains very healthy.
• This is why Nvidia is the only direct AI company I would own as its valuation is still in the realm of sanity as it is already making money and generating cash from AI.
• However, I still prefer the adjacencies of inference at the edge and nuclear power.