OpenAI's Leaked Financial Data

Channel: Alex Kantrowitz

Published at: 2024-10-11

YouTube video id: uDq6xEBUeBc

Source: https://www.youtube.com/watch?v=uDq6xEBUeBc

this story that you broke Cory actually
has some internal information from open
aai and it goes to show some like pretty
draw dropping projections that the
company is making I mean specifically
around the losses which we've been
talking about uh at length here on the
show recently but then also in terms of
where it expects its Revenue to go so
for me the top line was that they're
expecting their losses to get as high as
14 billion in 2026 remember the company
just raised 6.6 billion right in a round
so its losses could be that high uh but
we also took a look at some of the C
some of the re the places where expects
the revenue to come from and we'll get
into this uh maybe this is actually a
good place to start because one of the
things that astonished me was that chat
PT is what open AI expects to lead its
Revenue uh which it expects to increase
expects it re its Revenue to increase
100 fold by 2029 and the bulk of that
revenue is going to be coming from chat
GPT according to open ai's projections
Which flies completely in the face of
everything we've been discussing on this
podcast and everything we've expected
which is that the API and the other
companies using this GPT technology was
going to be The Driver of open AI
Revenue but it does look like the
company's expecting chat GPT to be the
bulk of that Revenue I mean it already
is but but throughout the upcoming
history what do you make of that Corey
it surprised me too on the top line I
mean I think first of all you need to to
view these projections and these numbers
like you view a lot of startup
projections which is this is in in many
ways the best case scenario especially
on on the top line like right I'm
assuming this is what they sold
investors on which means that that big
round sort of predicated on investors
believing the story yeah 100% And I
think look when you you have uh growth
so far that puts them um you know sort
of in Google and Facebook territory in
terms of Revenue growth
um you know early in their history you
know yeah they're going to lean into it
and be aggressive and say yeah by the
end of the decade we're going to
generate as much revenue as Nvidia or
Tesla like they they definitely went big
with it and then yeah when you dig into
that it's like okay maybe you can view
those Revenue figures skeptically and I
think we all should but I also think you
know let's let's actually take some of
their forecast at face value and I do
think the actual breakdown of that
Revenue was pretty revealing and
interesting I mean you you H have a
company that says their main product
right now in Chachi PT which is both the
consumer conscript uh subscription
product also there's you know an
Enterprise product for Chachi PT um I
don't quite know how that breaks down
between the two of them in that in that
Revenue bucket but yeah they're saying
this is going to be the money maker for
or at least the revenue driver for for
years to come and it seems like the API
revenue is is you know maybe they're
kind of giving away giving it away more
than uh than they um uh are actually
like charging the full value of it and
it's also more easily commoditized um so
yeah I did think that was pretty
interesting R what did you think when
you took a look at this uh projection
and saw that basically the company is
expecting to be really it's a chatbot
company I mean that's what it is if the
majority of its revenu is coming from
chat GPT it's going to be a chatbot
company and not just a chatbot company
but a paid chatbot company I mean yeah
no no that that one jumped out at me and
Corey just kind of hit it the most
important point that the API Revenue
could be commoditized because actually I
mean I have been arguing for a long time
that that's actually where the money is
going to be it's in the Enterprise it's
in the API access side of things and
maybe they they laid out the case
they're arguing it's going to be on the
actual consumer facing side which was
interesting to me because even though
I've so strong believe the other way
maybe it is in the user interface the UI
the ux side that they launch great
products that everyday people will use
and that's where they're going to win do
I think it's correct or right I don't
know but it's an interesting way to
raise $6.6 billion is there a limit to
the way that this generative AI
phenomenon can continue to grow if the
leader which is open AI uh ends up not
really making a lot of money on the API
like at some point it's like why are you
developing these foundational models if
it's I guess maybe it's just to make the
chat pop better what do we think about
this Corey well I think if you just
think about it if I as a user of these
products as someone who is also trying
to cover you know these companies that
are growing PR big as a reporter it's
hard for me to keep up with sort of how
each these Foundation model companies
are um uh actually moving ahead of of
each other in different aspects of how
they actually can can do math do science
do you know sort of legal stuff like
like it seems like between mraw versus
open AI versus anthropic like you know
it it's getting somewhat these
Foundation models are you know in very
tight competition with with each other
and the value uh I think open AI is
saying is in the consumer and Enterprise
brand the the not sort of what you're
developing for other application
developers but a they're kind of casting
themselves I think as like an essential
service that they can um you know
provide a lot of value to Consumers on
and can drive up subscription prices and
that's like an easier Story to Tell in a
fundraising so maybe that's why they're
leaning into it and you know no one's
going to necessarily hold them to
account to it if if it turns out any
differently but yeah uh I I I think it's
an acknowledgement that they think the
API business is is is more of a
commodity right which is amazing because
that is effectively a tested
acknowledgement by open AI admitting
that it's not going to be able to stay
ahead of its competition and building
the state-of-the-art
models yeah I think that it's possible
there's also there's other products they
say they're going to develop that are
quite big we don't know what that'll
look like but obviously you know we know
there's a lot that's not out in the
world that they've teased like Sora and
you know kind of a perplexity type of
competitor too yes the old future quite
big potential product we know always
comes
through say something yeah it is
interesting though I was just thinking
about like a company like Google create
the killer consumer facing product and
then build off of that build an
advertising business off of that build
Google cloud and Google workspace and
all these other more Enterprise focused
efforts maybe that is the way they're
pitching this because I think like
versus some boring Enterprise API type
of company that just you know is people
are trying to build on and where the
cost of those actions go less and less
and less building being the true
consumer facing brand that introduces
the whole world to generative AI I know
I've been incredibly bearish on open Ai
and suddenly I sound like I'm pitching
them for a fundraising but this deck
really worked for you man you're like
take my money
[Laughter]
s r looks at one slide all of a sudden
he's knocking at the door yeah show me a
masayoshi sun deck and uh I'm first in
line but that's okay so that's just
looking at the revenue and then things
get really interesting because uh Corey
and you published this this is again
coming from open a eyes internal
projections and financials uh it's a
there I think this is my chart I've ever
looked at from a tech company uh this is
their revenue versus their costs and it
is one of those things where like oh you
look at total revenue and I'm like Ah
that's pretty nice $4 billion and then
it starts getting into the costs so
immediately and this this goes right to
our question about we've talked about in
the past couple weeks sub Prime AI
crisis and can AI uh these AI companies
sustain themselves and you look at these
costs and you go hm so Microsoft's uh
Revenue share right off the top of that
4 billion is 700 million so it looks
like what they're making something like
20% of all open AI Revenue uh the
compute to train the models then takes
another $3 billion uh off so you know
thinking about profitability with those
two alone you're just barely profitable
so it's then you get into compute to run
the models and that's another two
billion now you're in the red and then
it just adds uh research compute
amortization which is Ron Jon's favorite
outline item down a billion employee
salaries down 700 million General and
administrative 600 million data 500
million hosting 400 million sales in
marketing 300 million and you go from
this very nice looking 4 billion in
Revenue after you add in all these costs
to a $5 billion loss and that excludes
stock-based compensation by the way
which is going to be another major cost
I mean Corey when you first came across
this chart what did you what was your
reaction uh you know at first a bit of
confusion and uh mostly trying to
understand I mean look I think the th
this is like a new type of business and
so how they run their numbers how they
do their accounting um is is going to be
an interesting thing to watch right like
I remember I I like I interviewed some
startup uh one time like maybe four or
five years ago about like why they
hadn't hit their revenue projections or
why their numbers were slightly off or
whatever and he was like well look we're
not running Alcoa here where where're
you know some like giant you know old
company where like we know how the
business runs it's run the same way for
a long time blah blah so I think just
first like understanding how open AI is
sort of thinking about its business and
where the costs are coming from um that
you know in itself is really interesting
it also is not surprising I think like
Sam Alman has said himself like like
open AI is going to be one of the most
Capital intensive companies in history
uh I think he said the most Capital
intensive startup in history so like if
we take him at his word which as we know
maybe you shouldn't always do that I do
think in this case like he's probably
telling the truth this he he is not you
know uh saying that um uh you know this
company is not going to cost a ton of
money to build um and that is you know
by far the biggest cost is is compute
it's the it's the data centers it's
actually being able to mostly train
these models um the smaller cost of the
compute bucket is um uh to actually run
them which is called inference um and
that is uh they're saying they're
they're lowering those costs but
training is sort of it's a huge huge
expense yeah not much lower two billion
to run versus three billion to train so
but yeah yeah they're but that's going
to expand over time essentially the the
training cost yeah and I think the idea
of these are completely new businesses
there are two numbers that jumped out
which I think were that i' not seen
anywhere else before and I thought were
really interesting first sales and
marketing at 300 million out of 4
billion in Revenue that's around 7% a
traditional SAS company is spending 40
to 50% on sales and marketing and we
have talked about this a lot that and
people if you've ever used them they
don't really have Enterprise customer
success they don't have the old school
SA sales guy out there and maybe they'll
create that all with agentic AI but at
least in the near term if they're really
trying to move in that direction the
spend is so low on the marketing side
it's amazing to me that they think
unless they're going to really
understand how to scale that then the
other was gross margin which You'
reported 41% where typically software is
I mean 70% is the the Baseline so the
the actual like you know profit for
revenue or margin relative to your
revenue is so much lower because of all
the actual Foundation model building
costs and how that actually scales I
mean how they even I I am curious what
these discussions are like with
investors like do they actually present
genuine plans around what these graphs
look like over the next three years or
it's just kind of we'll figure it out no
I mean I think they they do I mean if
you uh you know at the risk of insulting
you know very um
well
going like look I I mean I wrote about
this like so but like the if you look at
the big checks that made up this round
um you know it's a lot of people that
had reasons Beyond just like uh a clear
Financial uh rationale for investing in
open AI you know you you had Thrive
Capital uh and soft Bank um writing
really big checks like Thrive is already
very much they are in the the Sam Alman
business they are in the open AI
business they are deep in it they they
very much have hinged the reputation I
think on a relationship with Sam and
with with open Ai and open is gonna open
an office in one of Kisner's New York
building building yeah just like any
Council days Josh Kushner had made money
off of open AI it's is any part of this
industry not round tripping you have
Microsoft investing in open Ai and open
a using Microsoft aure credits you have
Kushner investing in open a open AI uh
opening an office in his building it's
wild but no totally but more to the
point R I'm curious when you when you
saw the numbers did you you know we've
been talking again about like is this
business sustainable now we've looked at
the numbers uh what do you think do you
have answers to the
question so this is a tough one because
again as I seem to be like the ultimate
open AI Fanboy five minutes ago um it's
I actually got less of a clear picture
of where this company's going bearish or
bullish after reading this and the
reason is the thing it made me most
excited is to one day read the S1 filing
of this company for the IPO because it's
so weird and complex I guess that's the
only the only consistency from the last
few few months is everything gets
weirder even Microsoft it was reported I
think the cut of Revenue was 20% Which
was higher than previously thought
there's still there you had some
information in there around they've been
able to expense last year $500 million a
cloud comput I think it was first half
of this year yeah oh so first half of
this year already so Microsoft's
relationship with open AI is bananas
like it I cannot think of another
company certainly at this scale but even
smaller where I've seen something that
contorted and then even just the way
from the accounting side and it's fair
this is a completely new type of
technology and I actually what you're
saying Corey I do kind of agree with
that that the idea that no one knows
there's no standard accounting practice
for how to amortize the value of a large
language model because it's so different
in the past you build software you have
like essentially zero marginal cost can
go forever that's this is a much more
cost intensive like marginal cost
intensive way of doing technology so so
basically no one it became more clear to
me no one has any idea how to actually
do the accounting for this to do the
financial projections for this and and
we're all in the same ultimate business
basically and even and I think you're
right to put your finger on the um you
know the Microsoft open AI relationship
being really key to all of this not just
in assessing it it does muddy the waters
on its financial statements I I'll tell
you like I you look at their you look at
sort of um the actual cash leaving open
AI um and it's a lot lower than you
would think compared to their income
statement losses because so much of the
expense is tied up in Microsoft compute
credits which is not technically a cash
item um so that muddies the waters a
little bit I don't know how long they'll
be able to sustain that sort of pace um
and then we have no idea how was it ever
finalized what of the initial 10 billion
doll was Cloud compute I don't know if
it's ever been fully reported I think
it's it's Mo it is the majority of it is
my understanding um it whether that
means it's seven or 8 billion or I think
it's in that range is my understanding
but my colleague uh had a great story
Also earlier in the week um basically
saying that Sarah frier open AI CFO has
told employees that look we're going to
spend some of the6 A5 billion dollars um
you know racing also to to develop data
centers with other uh potential uh uh
sort of companies and essentially like
they're in a race to kind of get compute
and to get data center space and
Microsoft has one company isn't
necessarily able to get a lock on all of
it so that relationship is I wouldn't
necessar it that is slowly they are very
much joined at the hip but both are
trying to figure out how do we not
become too dependent on each other I
think we should put this all together
because we're now we're getting some
projections in terms of what they're
expecting to spend right they're
expecting to spend 9.5 billion
potentially on training alone training
alone right up from three this year and
2026 that's just in 2 years uh and Corey
you report that their loss excluding
stock comp could be 14 billion in
2026 that's a time where they're
expecting again Chachi PT to be the lead
in terms of their um in terms of their
revenue so let's just take our head out
of the spreadsheet for a second and like
think about this logically I mean is
there a chance they could potentially
Rec coup all that money from chat PT
subscriptions like how much better would
chat PT have to get cuz I don't think
it's an awareness thing at this point
like Ronan mentioned it's they are
getting as much word of mouth as they
could potentially get and they still are
where they are in terms of chat PT
adoption so how much better does chat
GPT need to get to sort to start to
justify those costs I would say like it
would basically have to be
AGI Raja why don't you take this and
then we go to Corey I I agree that it's
interesting that I was just thinking
about like is it the consumer or the
Enterprise side that gets commoditized
and you can make a case for against both
and on the consumer side that idea of
like how much better does it have to be
I would say I mean even out of people I
know most people aren't paying 20 bucks
I think I'm sure we all are to some one
of the services but like for the every
everyday average user is this something
you even need to pay for want to pay for
given the current state of how these
things work I don't think so so I think
like it would have to
completely transform or revolutionize I
don't know like entire new products and
entire new problems being solved for
everyday people versus it can just write
you some stuff or it'll help you with
code from that'll hit a certain segment
or I I'm trying to even picture what
that product would be that every single
person would be ready to pay 20 bucks a
month or more yeah Cory is this feasible
yeah I think it's going to have to be
also just way deeper in the Enterprise I
think like just having a consumer
product is you
know uh it's going to have to be you
know pretty pretty insane it's going to
have to be very agentic you know where
it's doing things for you to use their
jargon um but I think if it's able to
revolutionize like you know actual uh
you know there's Mone there's there's
there's a lot of money in the Enterprise
I would say like if it can actually like
improve uh you know sort of company's
efficiencies and bottom lines there but
like look I think right now the bare
case is like a lot of this revenue is
from early adopters who are you know
sort of playing around with the stuff
like that's how I would characterize my
spending on on Chachi WT it's like I
haven't really figured you know like
yeah I've listened to like plenty people
say like oh here's how you really get
into it here's how you do it really well
and it just hasn't quite stuck and and
I'm humble enough to you know maybe to
say like um maybe I just don't get it
and like maybe I just haven't integrated
it into my practices very well um but
yeah there's definitely like some early
um adopter sort of effect going on that
will need to expand beyond that yeah and
on the Enterprise side in particular
we've been talking about it like if
enterpr Enterprises all went in because
they had to uh and if they're not seeing
a return they could just as easily pull
out pretty fast so we're going to see
what's going to happen on that point
yeah last point for me uh so you
mentioned in your story we talked about
this a little bit we don't know what's
going to happen we don't know when their
Microsoft credit is going to run out and
do they Factor the fact that that credit
could be completely gone uh maybe that's
what's blowing up their costs and you
know do they then because you mentioned
they're basically either giving away or
giving at Cost what's coming through
their API so does do they then need to
start charging a lot more to developers
building on top of this Stu stuff in
order to sort of stem the losses they're
already expecting to see right well I
don't think well there's nothing in the
documents that really answers either of
those questions necessarily um they um I
do think they when you see the the
compute amortization it it is a line
that is going up uh in in the they're
they're going to be uh expending more on
Research compute in the coming years and
I would imagine it is with my my guess
is that it's with the assumption that um
uh they won't be able to get as many
like research credits for Microsoft for
it or they'll have to go to another
vendor and have to pay more or something
like that so one quick followup on that
on that front when are they going to
need to raise again that's next year
yeah yeah I think that is like the
question I think um look they had a
billion dollars on their balance sheet
when they went out to raise this $6.6
billion uh and they also have like a few
billions of like a credit revolver that
they raised on top of it so my very back
of the end and they still have I think
billions in research compute left from
Microsoft um my back of the envelope
would say 2026 at the latest but you
know this is their competitive Advantage
is over like anthropic is clearly Sam
Alman being one of the best fundraisers
in the world I think they're going going
to try to keep leveraging that and keep
losing money to stay ahead like if no
one's pushing them to like get
profitable you know um they're going to
keep spending money on that
compute and uh on that I'm glad we got
to compute amortization one thing I had
to I had to bring up listeners you
should see the smile on Ron John's face
when he says the words compute
amortization by the way it's as happy as
I've seen him in months I know because
it it I mean from like a pure student
business it's interesting it's like
something completely the way they're
approaching it and how large language
models like should be thought of in a
company's financial statements is a
completely new thing but my favorite
part of uh your piece was is that open
AI is emphasizing to investors a metric
of profitability that excludes some
major expenses such as the billions it's
spending annually on training it's large
language models and and I have to give
credit there's a Bjorn Jeffrey who
tweeted back
it's time for llm adjusted iida and for
those listeners who remember the famous
wew Community adjusted iida I I actually
don't think it's crazy that they're
going to say we are profitable and
create a completely new type and maybe
it's going to be the future or maybe
it'll be like the thing we all remember
but llm adjusted iida is uh I don't
think it's un I don't think it's
impossible that that we see that phrase
in a in a filing one day I I think
you're 100% right my brand went there I
covered we work like like it's an it is
very much a a place where um you're like
H I've seen this movie before I don't
think we know all the answers yet in
terms of what are the real drivers of
the business that's like an area I'd
like to like still be reporting on um
and actually try to answer is it legit
to kind of strip a lot of your training
compute costs out of your cost of good
Soul like like these are the accounting
questions and the business nerd
questions that I think um are going to
be relevant to like how do we actually
think about whether this is a good
business or not like that's kind of like
the overarching question around AI like
is this okay the technology seems cool
so far you know it's really expensive is
this going to be a good business can I
weigh in here on the um should you
include training cost in your
profitability statement yeah gpt3
becomes obsolete the second GPT 4 comes
out right so the idea that training
costs are eventually going to like you
know taper off or level out or you don't
need them and they're not part of your
overall mix to me is is just so crazy
and you know I know that we're the
cool-headed and nuanced show but I just
need to say you cannot take training
costs out of your profitability
statement there's no basis in reality to
do that it's
crazy I love it get that off my chest I
think it can be cool-headed and nuanced
and factual like I'm losing my chill
here I came into this very's end but I'm
no longer training costs don't include
are not include in your profitability
we're we're profitable except for our
number one costs I mean come on yeah you
nailed it I think
yeah anything else R
John um I think
that that captured it the uh the fiery
passionate statement you cannot remove
your goddamn training cost from your
profitability statement Cory thank you
for joining
us thanks guys have a great weekend all
right everybody we'll be back on the
second half to talk about Tesla's Robo
taxi event and a couple other big
stories from the week back right after
this