72 comments

[ 3.0 ms ] story [ 134 ms ] thread
So what is the end game here? IPO on the stock market for over $500 billion or stay private like Stripe, SpaceX or Bytedance?

Nobody can buy them without bankrupting themselves many times over, unless the big companies buy them (very unlikely)

Or is OpenAI just going to keep raising and raising more and more?

Is OpenAI even profitable?

Allegedly yes (excluding all of the existing AI researcher costs).
Wait - they're 'profitable, excluding all of their employee costs'? Does that just mean they're charging more than the cost of their electricity to generate the tokens?
Presumably this would mean also earning more than the amortized cost of all the hardware they’ve purchased too which is a much higher bar than just electricity costs.
If they were why would they raise? They’ve already achieved scale.
they need upfront $$ to fund GPT6 and beyond.
IPO is one option. Merger is another. Companies don't need to buy them in cash but in stocks.
https://www.theinformation.com/articles/why-openai-could-los...

If you take this at face-value, they probably aren't profitable (to the tune of a $5 billion loss). I don't like idle speculation, but having actually read this their analysis seems pretty sound. Their numbers on training costs and hosting actually came from an anonymous insider. Impossible to say how reliable this anonymous insider is, but planting false, exaggerated numbers to outwit competitors seems a bit too much like a hare-brained scheme, hence why I opt to go with the numbers. If anything, companies tend to downplay costs when lying.

The one thing that caught me as a little bit crazy was spending $1.5 billion on employees. They have ~1500 employees, so that's $1 million per head. That seems insane, but then again, AI researchers are a valuable commodity these days.

Even if the cost of employees was $0, you'd still be looking at estimated losses in the billions.

Kind of a moot question. They’re not building rockets or payment rails.
Sometimes I think some companies are playing a different game. Where profit was never really a concern. The goal is just to keep showing growth and raising. You just have to cash out before the music stops or The Singularity happens.

One aspect of The Singularity is the rapid acceleration into an unpredictable point of no return. Maybe the financial model mirrors this?

> One aspect of The Singularity is the rapid acceleration into an unpredictable point of no return.

One aspect of reality is that governments actually like to govern predictably and those with sufficient military power make sure they can.

Regulatory capture, monopoly prices when the earth is dependent on them and no new companies can form
> Nobody can buy them without bankrupting themselves many times over, unless the big companies buy them (very unlikely)

There are lots of M&A structures. A company can in fact buy another that is bigger than itself. Plus Microsoft will easily be able to afford a $100B acquisition with some creative financing, or maybe even outright cash.

Heaven forfend. It would be Nadella's equivalent of Ballmer's attempt to buy Yahoo.
I don't understand the market. By most metrics they don't even have the best LLM, Anthropic do, so why are investors valuing them so highly? Is it because of the strength of their brand?
We've started using meta.ai and the 200B llama model on a couple servers we run at my company. I really have to go with Zuckerberg's approach on this one. I'm not sure OpenAI and Anthropic really have anything that can't be cloned much more cheaply.
There's a 200bn parameter llama model?
Bleh you're right. I meant the 70B model. I think there's some other context where the number 200 makes sense. Thanks for the correction.
Zuckerberg's approach may be good for you but it's costing Meta astronomical amounts of money for pretty nebulous benefit to themselves, so I wouldn't count on their current approach continuing forever. Enjoy it while it lasts.
Most of FAANG is effectively in competition with each other; OpenAI might be joining that cohort soon; as long as Facebook can kill their business model, Facebook is keeping a major player out of the running. Zuckerberg will keep doing this for as long as it takes to kneecap or bankrupt OpenAI.
> for pretty nebulous benefit to themselves

All of Meta's products are increasingly dependent on AI for engagement and ads. Their investments in the space are first and foremost for the benefit of the company itself. So no, they are not going away any time soon.

OpenAI's revenue is more diversified than just having the "best" LLM (ChatGPT+, enterprise deals, etc).

Only developers care about the academic LLM benchmarks and even fewer will switch all their code to a different LLM provider and spend the time implementing it.

They have the highest number of registered and concurrent users of all LLM providers, the best apps and they have deals with Microsoft AND Apple.

So I would say they have a solid grip on the market, and a good position to keep it with minimal improvements.

Most users don’t care about the best LLM, they will not switch. I mean they probably are not even aware they are using OpenAI because it is just integrated into Office or Siri…

Anyone feel like they missed the boat?
Nope. I don't wanna do alchemy, even if it is more successful in applications than chemistry.
The boat isn’t even here yet. We’re watching it maybe getting built.
you miss 100% of the cars if you chase them on the street like a dog.
If AI is a bubble then the value is going to go down to ~$0. If AI is real then the value of OpenAI and related companies is going to go up into the trillions of dollars and join/replace all the current established tech players. Either way, there's plenty of time to join the fun.
And why, exactly, would the other players not be able to simply do the same? Since all the important parts of the tech are widely understood and largely published?
Why has no one else been able to replicate Google, Apple, Facebook, Amazon, Salesforce, Oracle, IBM, Cisco, HP even though all of the tech is largely out there?
Because half of those have a real moat. Facebook has a strong network effect. Apple controls a platform and is deeply integrated between devices. There are cases to be made to a lesser degree for Amazon and Google and Salesforce.

Oracle, IBM, Cisco and HP have all strong competitors and hardly a unique product.

Anthropic and other models are roughly par with OpenAI.

I would value each of those companies' practical revenue potential as equal for now.

The operating costs themselves are eye watering, and the hardware is deprecating. So that has to be subtracted from revenue projections to get something like a real valuation.

I would invest in anthropic today, they seem to understand use cases better at the moment.

I also think LLMs are wildly over hyped, and they will be shunted to a niche in the next few years, essentially providing autocomplete or search for business tasks for the most part, with strong inroads on video analysis, automated translation, and other classic machine learning & neural net tasks.

Why would they go backwards in capabilities and get shunted to a niche instead of growing?
Probably as copyright lawsuits settle and they will have to pay for their training data
that would increase costs, not downgrade capabilities though
Data holders often ban others from using their data and then create a competitor using it. See Netflix etc. So it is possible that capabilities go down instead of costs going up, simple because data holders would rather create their own service than sell their data.
What's more likely is both. Capabilities go down because companies see their data as their moat, and costs go up as people price their data closer to the real value.

See reddit

> I also think LLMs are wildly over hyped, and they will be shunted to a niche in the next few years, essentially providing autocomplete or search for business tasks for the most part, with strong inroads on video analysis, automated translation, and other classic machine learning & neural net tasks.

There are many useful tasks that LLMs can automate and they will provide great value for humanity. The problem (for investors) is that it's hard to make people pay for it or make money out of it. This is the same problem Google Search had, and ads seem to be the only way forward here as well.

It's hard to get people to pay for just AI in a way I think is similar to the dotcom companies like WebVan and Pets.com. They couldn't just spin up some software and a website and hope that was enough to create a sustainable business, they still needed to be good at their core competency.

I think AI will be just like the web and software: there are a thousand little niches it could be applied to at any business but at the end of the day, it's just an enabling tool. The current hype around LLMs hasn't reached that point yet. Everyone is still selling AI for the sake of AI, which will inevitably fail because they're not good at their core business.

Why is that? There are many thinkable use cases for the API. Their revenue isnt limited to consumers.
Don't need them. Ever since Llama 3.1 came out I can control my data and not give it to OpenAI.
This is kind of a HN dropbox comment. Very few non programmers would go through that
They won't have to because a thousand developers like me will offer them simple solutions. Facebook absolutely destroyed anyone's "moat" in this area.
True, but enterprises wanting to make LLMs available to their non-technical employees would.
(comment deleted)
Thing is, I don't think openAI has that dropbox mindset of 'do what the user wants and make it super easy' either.

Sure, they have a great chat interface... But they have tech that can do so much more, and they aren't getting it directly into the hands of users... For example, where is "openAI translate"? Where is "openAI photo backup"? Where is "openAI phoneOS"?

With a burn rate of $1B/month, they could hire tens of thousands of regular developers and rebuild a whole suite of AI powered tools.

Do you host your own 405B parameter model?
I've tried it on Lambda but it doesn't seem any better than 70B or even 8B for most tasks.
"Above $100B" is > 30x annualized revenues reportedly in excess of $3B.

The company is still unprofitable, reportedly burning around $1B every month.

The risk for investors is vertigo-inducing high. The potential payoff is vertigo-inducing uncertain.

Meanwhile, Meta is doing its darnedest to turn humongous SOTA models into commodities.

I think the play for these companies was/is to erect regulatory barriers to entry by exploiting AI fear-mongering to regulate or outlaw open source or upstart competitors. That seems to be failing for the most part.
Meanwhile all the “safety” fears are really just reframing content moderation.
For a growing company, the PEG ratio [1] is more appropriate than P/E. What it asks is how many years of sustained growth brings the P/E ratio in line with population. (As is customary in finance, it's inverted from its intuitive form.)

OpenAI reportedly "doubled its annualized revenue to $3.4 billion in the past six months" to June from "$1.6 billion in late 2023" [2]. That implies a CAGR of 350% [a].

$100bn is a 29x revenue multiple to that ARR. S&P 500's average is 2.9x [3]. For OpenAI to match that multiple its sales would need to grow at 350% for 3 years. That would put OpenAI near the top 100 companies in America by sales [4]. Huge. But not unprecedented.

[1] https://www.investopedia.com/terms/p/pegratio.asp

[2] https://www.theinformation.com/articles/openais-annualized-r...

[a] (3.4 / 1.6) ^ (1 / 0.5) - 1

[3] https://www.multpl.com/s-p-500-price-to-sales

[4] https://en.wikipedia.org/wiki/List_of_largest_companies_in_t...

Information on openai changes quickly and is very unreliable, but:

https://www.maginative.com/article/openais-annualized-revenu...

This site says 3.2B of that comes from subscriptions. Omitting usage via API and assuming 1 user = 240$ annually that makes 133.3 million subscriptions.

https://backlinko.com/chatgpt-stats (says Jun. 04, 2024 but this might be outdated a bit more)

Here it says openai has roughly 3.9 million subs in USA. On that basis I would assume roughly 8 million in Europe. Even if we add India and China(but they don't operate in China currently from what I gather) with numbers raised linearly along their population this doesn't add up. Maybe they rake in a lot via their enterprise plans or am I dumb and missing something ?

> Information on openai changes quickly and is very unreliable

Agree. My napkin analysis is dependent on the reliability of those data.

> Omitting usage via API

That's a big assumption. I'd assume larger purchasers have negotiated a minimum buy in exchange for discounts.

> sales would need to grow at 350% for 3 years

I assume you mean 350% per year over 3 years. That works out to sales growing (1+3.5)^3 = 91-fold over three years.

Sure, it could happen. What's more, if OpenAI achieves AGI, growth could be even faster. But what are the odds?

For investors counting on those crazy insane growth rates to justify a present valuation "above" $100B, the potential payoff is vertigo-inducing uncertain.

I did my math wrong. It’s grow 350% for about 19 months. 10x.
Thanks. That's still a pretty high hurdle, and it justifies only the current valuation. To justify a greater future valuation (say, 5x the current valuation, or "above" $0.5T), even faster growth would be required. That's a bold gamble.
> The company is still unprofitable, reportedly burning around $1B every month.

Where do these reports come from? I'm not saying they're wrong (or right), but I've never seen anyone link to an American equivalent of Companies House in the UK, e.g.

https://find-and-update.company-information.service.gov.uk/c...

Closest is newspapers, but I have no way to tell which are reputable in financial accounting, vs. which are rumour-mongers repeating what ultimately turns out to be someone tweeting a guess based on nothing.

What Meta is doing seems borderline illegal...

It basically funds a different area with ads money to destroy competition.

> funds a different area with ads money to destroy competition

They’re commoditising their complement. That isn’t anticompetitive per se.

$1B a month is almost unfathomable.

I know the world does work this way, but there are so many things we could do for humanity with $1B a month.

As a reference point, Intel's market cap is $83bn. Crazy if OpenAI pull it off.
You're comparing a dinosaur to a caterpillar. OpenAI still has to bloom.
A butterfly still isn't much compared to a dinosaur.
Fwiw, OpenAI was already valued at $86B at the end of last year.
$100B. With Ilya Sutskever gone?

With competitors catching up.

It is the brains to make the next leap that is important.

They'll probably get it. Investors still have money to burn. And despite stocks making all time highs, the Fed is going to cut rates. Easing credit = even more money to burn.
> With Ilya Sutskever gone?

By all accounts he wasn't the secret sauce, far from it, he was just around from the beginning.