Microsoft & OpenAI – Fudge Pie

Still space for major problems.

• OpenAI has clarified both its corporate structure and its relationship with Microsoft and while there is more clarity, the problems of a potentially destructive lawsuit with Microsoft and internal conflicts of interest have not been fully addressed.
• OpenAI has reorganised into a public benefit corporation (PBC), clearing the way for SoftBank’s $40bn investment, but the real issues have not been completely solved.
• This means that when things get difficult, there is likely to be a bitter legal fight with Microsoft, as well as internal strife, as the non-profit objectives of the foundation remain at odds with the PBC.
  • First, Governance, where the business will now become a PBC, which on top of the usual fiduciary duty to shareholders and debt holders, the company must also demonstrate that its activities also benefit the public and society more widely.
  • This structure is pretty rare, but one I have heard of is Black Rifle Coffee, which is a coffee roasting company that also supports US veterans, law enforcement and first responders as its public benefit.
  • OpenAI’s, of course, will be its current statement, which is to “ensure that artificial intelligence benefits all of humanity”, but unlike Black Rifle Coffee, many of OpenAI’s actions remain in direct contradiction to this statement.
  • The non-profit entity will now be known as the OpenAI Foundation and will hold a 26% stake in the company, and have the exclusive power to hire and fire the board directors of the PBC, which is how it can maintain oversight of the PBC.  
  • The foundation will be responsible for using OpenAI’s IP to further disease research, as well as to work on industry-wide initiatives to ensure that everyone’s AI is not used for nefarious purposes.
  • One can immediately see scope for conflict, as the use of AI in drug discovery is already thought to be a big financial opportunity, and the PBC would now appear to be precluded from making the most of that opportunity due to the conflict of interest with the OpenAI Foundation.
  • However, now that there is a PBC, SoftBank, Microsoft, and everyone else can have equity stakes in it rather than a share in profits, which is a key condition on SoftBank’s $40bn investment.
  • The way to an IPO is now much clearer, and OpenAI stated that while it has no immediate plans for an IPO, this is now the most likely outcome given how much extra capital it thinks it is going to need.
  • Second, OpenAI Group (PBC), which is clearly aiming to become both a consumer and an enterprise ecosystem.
  • This is how it will be able to earn the big revenues it needs to cover the investment costs.
  • Hence, the real focus is moving from 1^st party products like ChatGPT or the Atlas browser to providing the platform and the tools for 3^rd parties to build their own services.
  • Apple, Google and Tencent are great examples of the kind of value that can be created when one becomes the go-to place to buy and sell 3^rd party services, and it was very clear that this is precisely where OpenAI intends to go.
  • This also reduces the potential for conflict with the OpenAI Foundation as OpenAI will be taking a cut of revenues from services that others create, rather than creating the services itself.
  • Infrastructure commitments to date are just over 30GW, which will cost around $1.4tn to roll out, amounting to roughly $47bn / GW.
  • OpenAI reckons that in time it can get this cost down to $20bn/GW, which in my mind makes little sense.
  • A GW is a unit of power consumption, not compute and does not take into account the fact that the number of tokens generated per GW is changing all the time.
  • Hence, I think that OpenAI is looking at it from the point of view that it will roll out that much power consumption and then look to maximise the tokens it can generate with that power.
  • The 30GW is expected to be rolled out over “many years”, which I assume means 5 to 10, and the company “hopes to do much more than this”.
  • These commitments are not firm in that OpenAI can delay or even cancel them as per the prevailing market conditions, which is why I am not concerned that if there is a correction, OpenAI sinks under the weight of its obligations.
  • What will sink OpenAI in a correction is the fact that it is still burning billions of dollars every quarter servicing 700m customers who don’t pay it a cent.
  • A correction would very likely result in the money drying up, which could cause an existential crisis at OpenAI and the realisation of some of the risks and conflicts that I have highlighted.
  • Third Microsoft: where the position is clearer, but the core issues have not been fully tidied up, which could easily lead to serious legal entanglements down the road.
  • Microsoft will own a 27% stake in the PBC valued at $135bn after SoftBank has put its money in as part of the latest round.
  • Microsoft’s IP rights for models and products (excluding hardware) are extended to 2032, and are independent of AGI, while its access to research in development is extended to 2030 but will come to an end if AGI is declared before that time.
  • AGI will now be decided by an independent expert panel (rather than the OpenAI Foundation), which does address some of the potential conflict, but in my opinion, not all of it.
  • I can see a scenario where AGI is declared, which disadvantages Microsoft, and so it challenges the declaration, leading to a big legal battle.  
  • Microsoft has given up its right of first refusal to provide compute to OpenAI, and in return, OpenAI has committed to spending a further $250bn of Azure compute
  • OpenAI can now provide API access to the US government national security customers and is also able to release models into the open source.
• The net result is that this restructuring clears up enough of the problem that allows SoftBank to invest, but with the non-profit still pulling the levers and Microsoft still somewhat subject to an arbitrary declaration of AGI, leaves the door open to serious strife.
• The real winner here is Microsoft, which can now book a 10x return on OpenAI at $135bn and should see strong revenue growth from OpenAI with most of the previous risks and dependencies eliminated.
• The real question now is where the $1.4tn that OpenAI needs to build 30GW of capacity is going to come from and what returns the capacity will offer when it does.
• In the race to the AI ecosystem, there are many runners, and while OpenAI has an early lead, its competitors are formidable.
• Hence, I remain far from convinced with regard to the business case for OpenAI, which could be further sunk by a correction in demand and valuation that sees the flow of money dry up.
• Hence, I continue to prefer the adjacencies of nuclear power and inference at the edge.