Microsoft – The pitfalls of success

Generative AI could hurt Office 365 margins.

• Everyone is much too busy adding generative AI to their products in order to be part of the new AI crowd and so no thought is being given to how much it might cost or whether it is a good idea at all.
• This is typical of a craze, and I suspect that when the dust settles and the cloud computing bills start arriving, people might start to think twice about whether this is a benefit to their business.
• Microsoft is a typical example which has just added a variable cost to a fixed revenue stream which in my opinion is never a good idea.
• Microsoft has demonstrated an AI “co-pilot” for Office software that will create draft documents, generate artwork and generate graphs that aims to take the drudgery out of daily Office tasks.
• This is something that GPT-3, ChatGPT and I assume GPT-4 are all very good at and so this represents a good use commercial use case.
• The problem is whether it is a good use case for generating profit and here I am far from convinced.
• The current estimate by the market is that GPT-4 uses 100tn parameters which is 571x the number that OpenAI used for GPT-3 (see here).
• It is important to remember that the 100tn is only an estimate because OpenAI has not disclosed how many parameters or processors it used which it says is for competitive reasons.
• Given how badly Microsoft has rattled Google which is what it has been trying to do for over 10 years since the infamous Scroogled campaign, I can see its point.
• However, OpenAI does concede that GPT-4 is bigger than GPT-3 which I suspect means at least 10x bigger if not more.
• That means that 10x more processors, memory and compute capacity are going to be needed to support it meaning that even including volume discounts, it is going to be much more expensive to run.
• This would be fine if there was a 10x increase in value creation or utility for GPT-4 over GPT-3 but this is not what OpenAI has published.
• Instead, the improvements look linear to me meaning that a business supported by GPT-4 will be less profitable than one supported by GPT-3 unless the service is much more expensive when it uses GPT-4.
• This will be hard to justify given the linear improvement in the performance of the service.
• Furthermore, the costs to provide this service are variable if outsourced and a combination of variable and fixed if built in-house.
• I think that very few people are going to build one of these in-house especially when there are plenty of hyper-scalers who are worried about growth and already at operating scale.
• If we assume that the AI “co-pilot” is a massive success, then the GPT-3 or 4 model will be inundated with requests to create all sorts of Office documents meaning that even more compute, processors and electricity will be needed to support these requests.
• Office 365 today is a subscription where one pays a fixed price to subscribe to the Office apps and updates for one year.
• Consequently, I think that Microsoft could find that it has to increase expenses to support a service that is bringing in no extra revenue.
• Given how resource-intensive generative AI is, this could have an impact on Office 365 margins which investors are not going t like.
• Hence it would require an extra subscription or pay-as-you-go package to be put in place for the co-pilot.
• This is why I don’t think that the case for generative AI has been made from an economic perspective which at the end of the day is all that really matters when it comes to matters of business.
• Consequently, huge success of generative AI as a tool in productivity software may not be all it is being cracked up to be as it may turn out to be very expensive to provide meaning that users have to pay a lot for it.
• Currently, nobody cares or is paying attention, but I suspect this may become a significant issue as servers, processors and electricity do not grow on trees.
• The best way to play this craze remains AMD and Nvidia as the suppliers of the silicon and Nvidia in particular but be ready to get out the minute there are signs of generative AI not living up to the hype.
• The stock is not cheap, but I think it is ripe for another run.