• The Decoding
  • Posts
  • What People Are Missing About Microsoft's $10B Investment In OpenAI

What People Are Missing About Microsoft's $10B Investment In OpenAI

Sam Altman just pulled the coup of the decade

Author

Gregor von Dulong
February 02, 2023

What People Are Missing About Microsoft's $10B Investment In OpenAI

Sam Altman just pulled the coup of the decade!

Microsoft is investing $10B into OpenAI!

There is a lot of frustration in the community about OpenAI not being all that open anymore. They appear to abandon their ethos of developing AI for everyone, free of economic pressures.

The fear is that OpenAI's models are going to become fancy MS Office plugins. Gone would be the days of open research and innovation.

However, the specifics of the deal tell a different story.

To understand what is going on, we need to peek behind the curtain of the tough business of machine learning. We will find that Sam Altman might have just orchestrated the coup of the decade!

To appreciate better why there is some three-dimensional chess going on, let's first look at Sam Altman's backstory.

Let's go!

A Stellar Rise

Back in 2005, Sam Altman founded Loopt and was part of the first-ever YC batch. He raised a total of $30M in funding, but the company failed to gain traction. Seven years into the business Loopt was basically dead in the water and had to be shut down.

Instead of caving, he managed to sell his startup for $43M to the finTech company Green Dot. Investors got their money back and he personally made $5M from the sale.

By YC standards, this was a pretty unimpressive outcome.

However, people took note that the fire between his ears was burning hotter than that of most people. So hot in fact that Paul Graham included him in his 2009 essay about the five founders who influenced him the most.

He listed young Sam Altman next to Steve Jobs, Larry & Sergey from Google, and Paul Buchheit (creator of GMail and AdSense). He went on to describe him as a strategic mastermind whose sheer force of will was going to get him whatever he wanted.

And Sam Altman played his hand well indeed!

He parleyed his new connections into raising $21M from Peter Thiel and others to start investing. Within four years he 10x-ed the money [2]. In addition, Paul Graham made him his successor as president of YC in 2014.

Within one decade of selling his first startup for $5M, he grew his net worth to a mind-bending $250M and rose to the circle of the most influential people in Silicon Valley.

Today, he is the CEO of OpenAI - one of the most exciting and impactful organizations in all of tech.

However, OpenAI - the rocket ship of AI innovation - is in dire straights.

OpenAI is Bleeding Cash

Back in 2015, OpenAI was kickstarted with $1B in donations from famous donors such as Elon Musk.

That money is long gone.

In 2022 OpenAI is projecting a revenue of $36M. At the same time, they spent roughly $544M. Hence the company has lost >$500M over the last year alone.

This is probably not an outlier year. OpenAI is headquartered in San Francisco and has a stable of 375 employees of mostly machine learning rockstars. Hence, salaries alone probably come out to be roughly $200M p.a.

In addition to high salaries their compute costs are stupendous. Considering it cost them $4.6M to train GPT3 once, it is likely that their cloud bill is in a very healthy nine-figure range as well [4].

So, where does this leave them today?

Before the Microsoft investment of $10B, OpenAI had received a total of $4B over its lifetime. With $4B in funding, a burn rate of at least $0.5B a year, and eight years of company history it doesn't take a genius to figure out that they are running low on cash.

It would be reasonable to think: OpenAI is sitting on ChatGPT and other great models. Can't they just lease them and make a killing?

Yes and no. OpenAI is projecting a revenue of $1B for 2024. However, it is unlikely that they could pull this off without significantly increasing their costs as well.

Here are some reasons why!

The Tough Business Of Machine Learning

Machine learning companies are distinct from regular software companies. On the outside they look and feel similar: people are creating products using code, but on the inside things can be very different.

To start off, machine learning companies are usually way less profitable. Their gross margins land in the 50%-60% range, much lower than those of SaaS businesses, which can be as high as 80% [7].

On the one hand, the massive compute requirements and thorny data management problems drive up costs.

On the other hand, the work itself can sometimes resemble consulting more than it resembles software engineering. Everyone who has worked in the field knows that training models requires deep domain knowledge and loads of manual work on data.

To illustrate the latter point, imagine the unspeakable complexity of performing content moderation on ChatGPT's outputs. If OpenAI scales the usage of GPT in production, they will need large teams of moderators to filter and label hate speech, slurs, tutorials on killing people, you name it.

Alright, alright, alright! Machine learning is hard.

OpenAI already has ChatGPT working. That's gotta be worth something?

Foundation Models Might Become Commodities:

In order to monetize GPT or any of their other models, OpenAI can go two different routes.

First, they could pick one or more verticals and sell directly to consumers. They could for example become the ultimate copywriting tool and blow Jasper or copy.ai out of the water.

This is not going to happen. Reasons for it include:

  1. To support their mission of building competitive foundational AI tools, and their huge(!) burn rate, they would need to capture one or more very large verticals.

  2. They fundamentally need to re-brand themselves and diverge from their original mission. This would likely scare most of the talent away.

  3. They would need to build out sales and marketing teams. Such a step would fundamentally change their culture and would inevitably dilute their focus on research.

The second option OpenAI has is to keep doing what they are doing and monetize access to their models via API. Introducing a pro version of ChatGPT is a step in this direction.

This approach has its own challenges. Models like GPT do have a defensible moat. They are just large transformer models trained on very large open-source datasets.

As an example, last week Andrej Karpathy released a video of him coding up a version of GPT in an afternoon. Nothing could stop e.g. Google, StabilityAI, or HuggingFace from open-sourcing their own GPT.

As a result GPT inference would become a common good. This would melt OpenAI's profits down to a tiny bit of nothing.

In this scenario, they would also have a very hard time leveraging their branding to generate returns. Since companies that integrate with OpenAI's API control the interface to the customer, they would likely end up capturing all of the value.

An argument can be made that this is a general problem of foundation models. Their high fixed costs and lack of differentiation could end up making them akin to the steel industry.

To sum it up:

  • They don't have a way to sustainably monetize their models.

  • They do not want and probably should not build up internal sales and marketing teams to capture verticals

  • They need a lot of money to keep funding their research without getting bogged down by details of specific product development

So, what should they do?

The Microsoft Deal

OpenAI and Microsoft announced the extension of their partnership with a $10B investment, on Monday.

At this point, Microsoft will have invested a total of $13B in OpenAI. Moreover, new VCs are in on the deal by buying up shares of employees who want to take some chips off the table.

However, the astounding size is not the only extraordinary thing about this deal.

First off, the ownership will be split across three groups. Microsoft will hold 49%, VCs another 49%, and the OpenAI foundation will control the remaining 2% of shares.

If OpenAI starts making money, the profits are distributed differently across four stages:

  1. First, early investors (probably Khosla Ventures and Reid Hoffman’s foundation) get their money back with interest.

  2. After that Microsoft is entitled to 75% of profits until the $13B of funding is repaid

  3. When the initial funding is repaid, Microsoft and the remaining VCs each get 49% of profits. This continues until another $92B and $150B are paid out to Microsoft and the VCs, respectively.

  4. Once the aforementioned money is paid to investors, 100% of shares return to the foundation, which regains total control over the company. [3]

What This Means:

This is absolutely crazy!

OpenAI managed to solve all of its problems at once. They raised a boatload of money and have access to all the compute they need.

On top of that, they solved their distribution problem. They now have access to Microsoft's sales teams and their models will be integrated into MS Office products.

Microsoft also benefits heavily. They can play at the forefront AI, brush up their tools, and have OpenAI as an exclusive partner to further compete in a bitter cloud war against AWS.

The synergies do not stop there.

OpenAI as well as GitHub (aubsidiary of Microsoft) e. g. will likely benefit heavily from the partnership as they continue to develop GitHub Copilot.

The deal creates a beautiful win-win situation, but that is not even the best part.

Sam Altman and his team at OpenAI essentially managed to place a giant hedge. If OpenAI does not manage to create anything meaningful or we enter a new AI winter, Microsoft will have paid for the party.

However, if OpenAI creates something in the direction of AGI - whatever that looks like - the value of it will likely be huge.

In that case, OpenAI will quickly repay the dept to Microsoft and the foundation will control 100% of whatever was created.

Wow!

Whether you agree with the path OpenAI has chosen or would have preferred them to stay donation-based, you have to give it to them.

This deal is an absolute power move!

I look forward to the future. Such exciting times to be alive!

As always, I really enjoyed making this for you and I sincerely hope you found it useful!

Thank you for reading!

References: