Why OpenAI's $157B valuation misreads AI's future (Oct 2024)
foundationcapital.com
foundationcapital.com
Deepseek followed llama and will be followed by others in the usual mushroom fashion of open source. People really dont appreciate the magnitude of the disruptive force that is unleashed by the open source paradigm. In a year from now the landscape will be brimming with new initiatives. In a few years nobody will even remember "open"ai.
Conventional economic theory will always misread the future of computing (and thus "AI"). The zero marginal cost and infinite replicability is not a bug, its a feature. But so far we dont really have a good model how to think about it and merge it with mainstream business models. Something must pay the bills eventually but these are very different bills from those of conventional scarcity based businesses. Ironically in the end the main scarcity is human ingenuity. Read the interview of the Deepseek founder on why their models are open source.
The killer app had better start giving better value, or I'd gladly pay the same amount of DeepSeek for unlimited access if they decided to charge.
(I thought you had this exactly right when I read it, but I kept noodling it while I brushed my teeth and now I'm not so sure llms won't just prove hard to build durable margins at meaningful volume on?)
VCs (esp those who missed out on OAI) are heavily incentivized to root for OAI to fail, and commoditize the biggest COGs item (AI models).
This guy is just talking his book.
you can both talk your book and also sincerely believe what you say. ad hominem (or whatever the latin equivalent is of ad bookinem) is not as substantive a criticism as you make it out to be. he can be both biased and correct.
If your grand dream is to dominate the market through sheer massive scale and that's what you're selling to capital, you're not exactly looking for reasons to buy less hardware and your vendor is hardly going to talk you out of it.
"It's hard to get a man to understand something when his fat valuation depends on his not understanding it"
The inevitable haircut all the funds are going to take in OpenAI and other AI startups when revenue fails to materialize[1] will herald a bust cycle and lot more circumspection in large investments like it happened few years back when a large number of Soft Bank investments did not pan out, notably most of them relied on big funding rounds to muscle out other players, not all of them have failed but all of them lost of enterprise value for investors.
---
[1] This is inevitable regardless of success of the space, either because cost of inference keeps dropping in combination with competing with high quality open weight models as DeepSeek, Stable Diffusion and others have shown. It will have strong downward pressure on pricing impacting both revenue and profit.
The AI hype might have as well setback a lot of other products that never got the necessary funding to get up from the ground, it's not looking pretty.
IMO the equivalent of moores law for AI (both on software and hardware development) is baked into the price, which doesn’t make the valuation all too crazy.
If anything, it’s an downward adjustment in the cost implications but could actually unlock exponential improvements on a shorter time horizon than expected because of that. Investors getting scared probably is a good opportunity to buy in.
If (big if!) I understand correctly, the ceiling for edge/local/offline AI has just blown off.
There's a huge motte and bailey thing with DeepSeek conversation, where the bailey is "It only took $5.5 million!*" (* for exactly one training run for one of several models, at dirt-cheap per-hour spot prices for H100s) and the motte is all sorts of stuff.
Truth is one run for one model took 2048 GPUs fulltime for 2 months, and my experience with FAANG ML, that means it took 6 months part-time and another 1.5-2.5 runs went absolutely nowhere.
- fp8 instead of fp32 precision training = 75% less memory
- multi-token prediction to vastly speed up token output
- Mixture of Experts (MoE) so that inference only uses parts of the model not the - entire model (~37B active at a time, not the entire 671B), increases efficiency
- PTX (basically low-level assembly code) hacking in old Nvidia GPUs to pump out as much performance from their old H800 GPUs as possible
Then, the big innovation of R1 and R1-Zero was finding a way to utilize reinforcement learning within their LLM training.
So the real question is why does anyone believe that OpenAI will bring AGI when actual innovation was happening in some hedge fund in China while OpenAI was going on an international tour trying to drum up a trillion dollars.
Valuations for most large companies have been crazy for a while now. No one values a company based on fundamentals anymore, its all pure gambling on future predictions.
This isn't unique to OpenAI by any means, but they are a good example. Last I checked their revenue to valuation multiplier was in the range of 42X. That's crazy.
And here comes DeepSeek and takes the steam out of this and the cost arguments that follow it.
Even if the whole story about the training cost was fake, R1 and the distilled models are still very efficient at inference.
Now there's another supplier to match the (potential?) consumer or corporate demand that's diffused among more competitors, and open source.
How does OpenAI get paid for a use case that can easily be run locally on an iPhone?
Some of the distilled models we're seeing are very good.
I also still don’t believe their cost figures, and think they’re leaving out the capital to acquire their secret GPU stash and the cost of pre training their base model (DeepSeek-V3-base). I also suspect their training corpus, which they’ve only vaguely described, would reveal the savings came from working off other foundational models’ work without counting those costs in their figure.
For now, I treat the cost claim as simply a calculated strategy for China to not look like they’re behind in the most important race, to prevent investors from continuing to boost US technology by causing them to doubt the ROI, and to take value out of the US stock market as they did today.
But as for your first paragraph: even if the "big AI players" have some secret sauce that will make their products better (and that they can actually keep secret), it seems unlikely it would be enough to command higher prices durably.
A model would have to be incredibly superior to justify paying for it, when there are so many free (or dirt cheap) alternatives that are simply good enough.
I’m curious about the Llama3 bit - do you have a source for that? I’ve been hearing they trained using OpenAI outputs (not sure how that would work).
DeepSeek V3/R1 architecture isn't anything like Llama 3. Llama 3 isn't even a mixture of experts, not to mention the various other differences like attention compression etc
One point I’ll agree on is his final one: that the true big players haven’t even been founded yet. Right now, the AI hype seems to still revolve around the dream of replacing humans with machines and still magically making Capitalism work in the process, which is something I (and other “contrarians”) have beaten to death in other threads. That said, what these companies have managed to demonstrate is that transformer-based predictive models are a part of the future - just not AGI.
If I were a VC, I’d be looking at startups that take the same training techniques but apply them in niche fields with higher success rates than general models. An example might be a firm that puts in the grunt work of training a foundational model in a specific realm of medicine, and then makes it easier for a hospital network to run said model locally against patient data while also continuously training and fine-tuning the underlying model. I wouldn’t want to get into the muck of SaaS in these cases, because data sovereignty is only going to become an ever-thornier issue in the coming decades, and these prediction models can leak user data like a sieve if not implemented correctly. Same goes for other narrow applications, like single-mode logistics networks or on-site hospitality interfaces. The real money will be in the ability to run foundational models against your own data in privacy and security, with inference at the edge or on-device rather than off in a hyperscaler datacenter somewhere.
Then again, I could be totally wrong. Guess we’ll all find out together.
If you mean developing a model from scratch just for your niche - the bitter lesson is that scale is everything and that a finetune from an internet-scale model will outperform you easily.
The (regrettably temporary) ousting of Sam Altman looks like the right call, in hindsight. Of course some amount of showmanship is expected, but the extreme nature of this self-serving BS is just laughable.
6 months from now we may be looking at Sam Altman the way we look at Adam Neumann.
"Even as some U.S. tech stocks plunged on Monday after it appeared that DeepSeek could produce similar results as rival models with a system that was cheaper to build, Trump projected confidence, calling it “very much a positive development.” He reasoned that American companies would be able to adapt and evolve based on DeepSeek’s demonstration that effective systems can be developed more easily than some assumed."
I'm not seeing anyone at OpenAI abandon the static weights model and yet they have audacity to claim that they just need to scale more?
I feel like the author's concluding point contradicts himself. There is a gold rush and OpenAI is selling shovels.
I think the author argues that OpenAI is not the only one selling shovels, and their shovels won't be always better that others'.
The vertical specific companies, though, are harder to clone as the invest in the product offering around/on top of AI
But Nvidia is also selling steroids (inference hardware) that everyone will need to use their new free shovels.
This analogy may have gotten out of hand.
That’s where we are in the AI journey in 2025. The year 2000.
and that valuation would crumble because of deepseek
well since most of that money comeback anyway to MS since OAI use Azure heavily but it still a lot of money and stock value of OAI would tank sooner or later when competitor like deepseek come
Van Goghs have it; QBASIC doesn't. Anyone can download QBASIC for free.
I can't admit to myself there's any open question as to if there is any long-term value.
I expect within 2 years, this will seem like a non-controversial idea, and it won't bring in a ton of assumptions about the speaker.
I have invested much time and effort making sure local models are a peer to remote ones in my app, and none, including DeepSeek's local models, are remotely close to the things needed to make that flow work.
EDIT: Reply-throttled, so answering replies here:
- The machine is building the machine: Telosnex, a cross-platform Flutter app
- it can do 90% of the scope, especially after I wrote precanned instructions for doing e.g. property-based testing.
- Things it's done mostly wholesale: -- secure iframe environment, on all 6 platforms, to: execute JS in, or render react components it wrote. -- completely refactoring my llama.cpp inference to use non-deprecated APIs.
- Codebase is about 40K real lines of code. (I have to think this helps a lot I doubt that ex. from scratch it would be able to build a Flutter app that used llama.cpp.)
- $30/day!?! -- Yeah, it's crazy, its up an order of magnitude from my most busy days when I just copy-pasted back and forth. It reads as much code as it wants, and you're doing more work literally, so it adds up.
- $20/day is realistic average
- Lines added per day +55%, lines deleted per day +29%, files changed per day 9 -> 21 https://x.com/jpohhhh/status/1881453489852948561
Unfortunately, the current available version doesn't have the agent stuff yet.
Hopefully in a week, realistically two.
I had the existing client app I've released-but-not-released-out-loud. Couple days before Christmas, for fun, I spent a couple hours wiring up the Anthropic Model Context Protocol filesystem server example. Within an hour it was clear this was special and I needed to get it out ASAP. Stunning stuff in action.
Wouldn't it be just a few prompts to get it done?
Maybe consider something like https://www.gyan.dev/ffmpeg/builds/
Nonetheless I will try it out
Granted, I ask it very specific questions that generate short answers (many of which are incorrect, btw), but still, it's difficult to imagine what kind of tasks, done by a single person, would generate such amounts?
First, OpenAI’s valuation is a bit wild—$157B on 13.5x forward revenue? That’s Meta/Facebook-level multiples at IPO, and OpenAI’s economics don’t scale the same way. Generative AI costs grow with usage, and compute isn’t getting cheaper fast enough to balance that out. Throw in the $6B+ infrastructure spend for 2025, and yeah, there’s a lot of financial risk. But that said... their growth is still insane. $300M monthly revenue by late 2023? That’s the kind of user adoption that others dream about, even if the profits aren’t there yet.
Now, the “no moat” argument... sure, DeepSeek showed us what’s possible on a budget, but let’s not pretend OpenAI is standing still. These open-source innovations (DeepSeek included) still build on years of foundational work by OpenAI, Google, and Meta. And while open models are narrowing the gap, it’s the ecosystem that wins long-term. Think Linux vs. proprietary Unix. OpenAI is like Microsoft here—if they play it right, they don’t need to have the best models; they need to be the default toolset for businesses and developers. (Also, let’s not forget how hard it is to maintain consistency and reliability at OpenAI’s scale—DeepSeek isn’t running 10M paying users yet.)
That said... I get the doubts. If your competitors can offer “good enough” models for free or dirt cheap, how do you justify charging $44/month (or whatever)? The killer app for AI might not even look like ChatGPT—Cursor, for example, has been far more useful for me at work. OpenAI needs to think beyond just being a platform or consumer product and figure out how to integrate AI into industry workflows in a way that really adds value. Otherwise, someone else will take that pie.
One thing OpenAI could do better? Focus on edge AI or lightweight models. DeepSeek already showed us that efficient, local models can challenge the hyperscaler approach. Why not explore something like “ChatGPT Lite” for mobile devices or edge environments? This could open new markets, especially in areas where high latency or data privacy is a concern.
Finally... the open-source thing. OpenAI’s “open” branding feels increasingly ironic, and it’s creating a trust gap. What if they flipped the script and started contributing more to the open-source ecosystem? It might look counterintuitive, but being seen as a collaborator could soften some of the backlash and even boost adoption indirectly.
OpenAI is still the frontrunner, but the path ahead isn’t clear-cut. They need to address their cost structure, competition from open models, and what comes after ChatGPT. If they don’t adapt quickly, they risk becoming Yahoo in a Google world. But if they pivot smartly—edge AI, better B2B integrations, maybe even some open-source goodwill—they still have the potential to lead this space.
Then the market blew up and a few big winners ate up all the profit and everyone else died on the long tail.
Today, iPhone exists, but people don’t even think about them much. They just use it.
Pestering someone now about your iPhone app idea is like pestering someone about your website idea back in 2015.
VR movies was a fad. You can date a TV by it's "VR mode" feature on the remote to a few years. No one is trying to sell me VR TVs anymore. That's what a fad looks like.
Impact isn't inherently positive.
You're holding it wrong.
What process did you use with the AIs, any prompting insights - context, agentic prompts etc?
What tech stack did you use that you found AIs were familliar enough with, ive found them woefully misinformed about most libraries and technologies ive tried them with in game development, often confidently mixing out of date and new information
I experimented today with Aider (to get R1 involved) and had less success, but it might be that I don't have the workflow down.
I have found cursor can handle a .NET C# back-end using highly standard code structures very well. SignalR for networking.
I've created servers and very basic HTML visualization for three projects - a fairly simple autobattler (took a day), a web-based beat-em-up (2 days), and now a bit more ambitiously my dream RTS-MMO (3rd weekend running).
I started with concise MVP specifications including requirements for future scaling, and from these worked with the AI to make dot point architectural documents. Once we had those down I moved step-by-step, developing elements and tests simultaneously then having the agent automatically run the tests and debug. The test-driven debugging is the part that saved the most frustration, as the initial implementation was almost always broken, but leaving the agent to its own devices (tabbing in and typing "continue" when hitting Cursor's 25 tool call limit, sometimes for hours) the tests guided bug fixing and amazingly it got there fairly consistently, though occasionally it will go off the rails and start modifying the tests to pass or inventing unwanted functionality.
The code is as standard as possible, with the servers all organized identically API -> Application -> Domain <- Infrastructure, and well separated between client/server. Getting basic HTML representations wasn't an issue, but it does begin to struggle and requires a lot more direction when it comes to client-side code that expands beyond initial visualization. I had a lot more success with Monogame C# than Phaser or other web formats (e.g. I quickly gave up on SFML, same issues you were having).
I'm a professional game developer but without formal CS/programming training, so I'm aware of my requirements but not always how to implement them cleanly. I understand the code it writes which feels vital when it occasionally rolls a critical miss, but these projects would have taken me months without AI.