This is Expensive Coffee
Jessica E. Lessin, Natasha Mascarenhas, & Aaron Holmes
Just to follow-up on Microsoft's deal to acquire decision to hire Inflection's AI team, The Information's Jessica E. Lessin, Natasha Mascarenhas, & Aaron Holmes were seemingly able to get a lot of the details on the deal:
The software giant has agreed to pay Inflection approximately $650 million, mostly in the form of a licensing deal that makes Inflection’s models available for sale on the software giant’s Azure cloud service, according to a person involved in the transaction. The startup is using the licensing fee to help provide its investors with a modest return on their capital, according to a second person who was briefed on the arrangement.
While less than Inflection was said to have raised this makes sense and is in line with what I noted a couple days ago:
While, as noted, Inflection had raised over $1.5B, if you assume that a large chunk of that was from Microsoft, and that a large chunk of that was in cloud credits (and perhaps NVIDIA's deal was structured in a similar manner for deals on GPUs), there's potentially less to make whole here. Still, there's likely hundreds of millions that investors paid into the company.
Impressively, The Information pulled even more granular detail about the payout:
Investors in the company’s first major round of funding, a $225 million investment from venture firm Greylock, hedge fund Dragoneer Investment Group and others, will receive one and a half times their investment, according to the person involved in the deal. Investors in a subsequent $1.3 billion funding round last year will receive 1.1 times their investment, the person said. Microsoft invested in both rounds, but the vast majority of the funding came from other investors such as former Microsoft CEO Bill Gates, former Google CEO Eric Schmidt, Nvidia and Inflection co-founder Reid Hoffman, which together led the later round. (Inflection likely hasn’t spent much of that capital, which means it could also use that cash to give the investors a return.)
That latter point was something else I surmised may have been the case – especially since, as reported, Inflection will run a much leaner operation now, basically just as a sort of shop to sell its models to Microsoft (and others, they claim, but probably just Microsoft). There's a new CEO, but who else is left at the company? It's entirely unclear – some reports have said that "essentially all" of the Inflection team is going to Microsoft.
Inflection was also said to be out fundraising ahead of this outcome, so they clearly needed more capital. But given the $1.3B raised just nine months ago, and presuming they weren't raising with $0 left in the bank, but more likely at least 6 months of runway and hopefully more, they still undoubtedly had quite a bit of money on the balance sheet (though a seemingly incredible burn rate, as many of these AI startups do).
In addition to a $620 million licensing fee, Microsoft has also agreed to pay Inflection about $30 million to waive any legal rights related to the mass hiring. And it renegotiated a $140 million line of credit that aimed to help Inflection finance its operations as well as pay for Microsoft services, said the person involved in the deal.
CYA – meaning here not "see ya" but "cover your ass". The credit line is amusing though, getting pretty ouroboros here with "pay for Microsoft services".
One reason Microsoft opted to pay for the rights to Inflection’s models was to protect itself from potential lawsuits from Inflection investors if ex-Inflection staff do similar AI development work at Microsoft, even if it doesn’t end up using the software, according to a Microsoft source briefed on the situation.
Yes, of course. Are they going to get any value from the actual models they're buying? Maybe, but also they undoubtedly would have just not licensed those models if they thought they could get away with this whole charade without doing so.
One more thing: Not mention by The Information, but called out by Ben Thompson and others, Inflection had "around 22,000 H100s", as Inflection co-founder Mustafa Suleyman boasted to Reuters at the time of their last funding, in which NVIDIA participated. How many NVIDIA chips is that exactly? "This is approximately three times more compute than what was used to train all of GPT4. Speed and scale are what's going to really enable us to build a differentiated product," Suleyman went on to say.
Those chips, of course, are valuable by themselves. Each H100 can cost something in the range of $30,000 (undoubtedly different deals for bulk purchases, etc). And they're incredible hard to get right now. Let's break out that calculator...
22,000 times $30,000 equals... $660 million. What a strange coincidence! Or not.