A Quick Take on DeepSeek

A Sputnik moment?

• We won't be a buyer of either TSSI or NVDA until the dust settles.

• The DeepSeek model (developed apparently on a shoestring by a hedge fund in Hong Kong) may very well lead to some reevaluation of the speed of hyperscaler build-out of data centers for AI.

• It seems that a new type of model can achieve an order of magnitude of efficiency gains without suffering performance, which seems on a par with (and even exceeds them in some cases) the best Western models.

• While we don't know for sure what kind of resources went into training the model, it’s not implausible it produces significant economies here as well:

  • The DeepSeek model was developed by a Hong Kong hedge fund, not known to splash tens of billions on AI data centers.

  • Even if they wanted to and provided they had the means, there are all kinds of restrictions on getting the latest gear, more especially the latest generation Nvidia chips.

  • Being faced with these restrictions, they had incentives to do more with less and look for alternative architectures and designs, apparently with success.

• We do know it's very efficient in inference (producing results) and we can look under the hood as it's largely an open-source model. From Gavin Baker:

r1 costs 93% less to *use* than o1 per each API, can be run locally on a high end work station and does not seem to have hit any rate limits which is wild. Simple math is that every 1b active parameters requires 1 gb of RAM in FP8, so r1 requires 37 gb of RAM. Batching massively lowers costs and more compute increases tokens/second so still advantages to inference in the cloud.

• Note: o1 is the latest reasoning model from OpenAI.

• As hyperscalers have pending investments of tens of billions of dollars with questionable returns and very high depreciation (servers' economic life is limited), some of them might get back to the drawing board and scale back some of their efforts at least a little.

• Marc Andreessen has called this a Sputnik moment, who are we to differ..

• Too many unknowns for firm conclusions, which has us moving to the sideline with respect to AI infrastructure stocks like TSS International (TSSI).