That seems like a lot of money. How quickly can sustainable capacity be built up in terms of building power plants, data center construction, silicon design and fabrication, etc.? Are these industries about to experience stratospheric growth, followed by a massive and painful adjustment, or does this represent a printing press or industrial revolution like inflection point?
Would anyone like to found a startup doing high-security embedded systems infrastructure? Peter at my username dot com if you’d like to connect.
> Even if you think that OpenAI’s growth is impressive — it went from 700 million to 800 million weekly active users in the last two months — that is not the kind of growth that says “build capacity assuming that literally every single human being on Earth uses this all the time.”
I’d argue the other way around: 100M growth in two months suggests literally every single human being on Earth would benefit from using this all the time, and it’s just a matter of enabling them to.
Beware the sigmoidal curve, though. Growth is exponential till it’s not.
> 100M growth in two months suggests literally every single human being on Earth would benefit from using this all the time, and it’s just a matter of enabling them to.
This doesn’t make any sense. Popular is not the same as useful. You’d have a more compelling argument if you included data showing that all this increased LLM usage has had some kind of impact on productivity metrics.
Instead, some studies have shown that LLMs are making professionals less productive:
The number of users is irrelevant. Revenue is only slightly relevant. The only thing that matters is profit. It would even be a decent thing if they could show marginal profit per user.
I just stopped paying for ChatGPT this month, once I found out that they made the projects available to free users too. The free version is just as good and I can shuffle between Grok, Mistral, Deepseek and Gemini when I run out of free quota.
So maybe giving away more and more free stuff is good for growth? The product is excellent, ChatGPT is still my favorite but the competition isn't that behind. If fact, I started liking Grok better for most tasks
> 100M growth in two months suggests literally every single human being on Earth would benefit from using this all the time, and it’s just a matter of enabling them to.
Only about 5% see enough value to drop $20 a month... It's like VR and AR, if people get a headset for free they'll use them every now and then, but virtually nobody wants to drop money on these.
Elon can give them 400B and he would still have a net worth of 100B and from that he can again get back to his former net worth in a few years. The world is not fair.
Vast sums of money that are being invested in it (obviously in the hope of AGI) but I'm not sure the world would notice if OpenAI/current LLM products just disappeared tomorrow.
Why are you -not- angry at all of this insanity? I feel the same way as him, hype has blown the bubble bigger and bigger, and it's just a matter of time until it poops out and causes huge amounts of pain.
It's going to perhaps sound nuts, but I'm beginning to wonder if America is actually a giant Ponzi scheme.
I've been thinking about American exceptionalism - they way it is head and shoulders above Europe and the developed world in terms of GDP growth, market returns, start up successes etc. and what might be the root of this success. And I'm starting to think that, apart from various mild genuine effects, it is also a sequence of circular self-fulfilling prophecies.
Let's say you're a sophisticated startup and you want some funding. Where do you go? US of course - it has the easiest access to capital. It does so presumably because US venture funds have an easier time raising funds. And that's presumably because of their track record of making money for investors - real, or at least perceived. They invest in these startups and they exit at a profit, because US companies have better valuations than elsewhere, so at IPO investors lap up the shares and the VCs make money. It's easy to find buyers for US stocks because they're always going up. In turn, they're going up because, well, there's lots of investors. It's much easier to raise billions for data centres and fairy dust because investors are in awe of what can be done with the money and anyway line always go up. Stocks like TSLA have valuations you couldn't justify elsewhere. Maybe because they will build robot AI rocket taxis, or maybe because the collective American Allure means valuations are just high.
The beauty of this arrangement is that the elements are entangled in a complex web of financial interdependency. If you think about these things in isolation, you wouldn't conclude there's anything unusual. US VC funding is so good because there's a lot of capital - lucky them. This thought of circularity only struck me when trying to think of the root cause - the nuclear set of elements that drive it. And I concluded any reason I can think of is eventually recursive.
I'm not saying America is just dumb luck kept together by spittle, of course there are structural advantages the US has. I'm just not sure it really is that much better an economic machine than other similar countries.
One difference to a Ponzi scheme is that you might actually hit a stable level and stay there rather than crash and burn. So it's more like a collective investment into a lottery. OpenAI might burn $400bn and achieve singularity, then proceed to own the rest of the world.
But I can't shake the feeling that a lot of recent US growth is a bit of smoke and mirrors. After adjusting for tech, US indices didn't outperform European ones post GFC, IIRC. Much of its growth this year is AI, financed presumably by half the world and maintained by sky-high valuations. And no one says "check" because, well, it's the US and the line always go up.
I don't think you're describing a Ponzi scheme. More like a merry go around. Seems pretty common, just turned to 11.
Of course it can stop, an a little history reading will show that it has always stopped, but it can take a long time.
If anything, I fear the AI hype + the orange idiot you put in charge, can fuck you up much faster than otherwise. OTOH, Trump is also a symptom, showing that things were not going great.
His "$400B in next 12 months" claim treats OpenAI as paying construction costs upfront. But OpenAI is leasing capacity as operating expense - Oracle finances and builds the data centers [1]. This is like saying a tenant needs $5M cash because that's what the building cost to construct.
The Oracle deal structure: OpenAI pays ~$30B/year in rental fees starting fiscal 2027/2028 [2], ramping up over 5 years as capacity comes online. Not "$400B in 12 months."
The deals are structured as staged vendor financing:
- NVIDIA "invests" $10B per gigawatt milestone, gets paid back through chip purchases [3]
- AMD gives OpenAI warrants for 160M shares (~10% equity) that vest as chips deploy [4]
- As one analyst noted: "Nvidia invests $100 billion in OpenAI, which then OpenAI turns back and gives it back to Nvidia" [3]
This is circular vendor financing where suppliers extend credit betting on OpenAI's growth. It's unusual and potentially fragile, but it's not "OpenAI needs $400B cash they don't have."
Zitron asks: "Does OpenAI have $400B in cash?"
The actual question: "Can OpenAI grow revenue from $13B to $60B+ to cover lease payments by 2028-2029?"
The first question is nonsensical given deal structure. The second is the actual bet everyone's making.
His core thesis - "OpenAI literally cannot afford these deals therefore fraud" - fails because he fundamentally misunderstands how the deals work. The real questions are about execution timelines and revenue growth projections, not about OpenAI needing hundreds of billions in cash right now.
There's probably a good critical piece to write about whether these vendor financing bets will pay off, but this isn't it.
Let's assume that estimate is good. For some perspective an context, the last finalized DOD budget (2023) was $815B, and, plus supplementals, turned into about $852 billion.
AGI is absolutely a national security concern. Despite it being an enormous number, it'll happen. It may not be earmarked for OpenAI, but the US is going to ensure that the energy capability is there.
> but the US is going to ensure that the energy capability is there.
We're doing a pretty shit job of ensuring that today. Capacity is already intensely strained, and the govt seems to be decelerating investment into power capacity growth, if anything
At this point we all know this is just a massive bubble. I'm done paying attention to it really. I'm prepared for all my investments to go down in the next 1-5 years. If you're nearing retirement now is the time to cash out. Yes, investments could go up in a value a lot until the correction, but I don't really think that is worth the risk.
Can someone explain why we measure these datacenters in Gigawatts rather than something that actually measures compute like flops or whatever the AI equivalent of flops is?
To put it another way, I don't know anything but I could probably make a '1 GW' datacenter with a single 6502 and a giant bank of resistors.
Mh, in my recently slightly growing, but still tiny experience with HW&DC-Ops:
You have a lot more things in a DC than just GPUs consuming power and producing heat. GPUs are the big ones, sure, but after a while, switches, firewalls, storage units, other servers and so one all contribute to the power footprint significantly. A big small packet high throughput firewall packs a surprisingly high amount of compute capacity, eats a surprising amount of power and generates a lot of heat. Oh and it costs a couple of cars in total.
And that's the important abstraction / simplification you get when you start running hardware at scale. Your limitation is not necessarily TFlops, GHz or GB per cubic meter. It is easy to cram a crapton of those into a small place.
The main problem after a while is the ability to put enough power into the building and to move the heat out of it again. It sure would be easy to put a lot of resistors into a place to make a lot of power consumption. Hamburg Energy is currently building just that to bleed off excess solar power into the grid heating.
It's problematic to connect that to the 10kv power grid safely and to move the heat away from the system fast.
It simplifies marketing. They probably don't really know how much Flops or anything else they will end up anyway. So gigawatts is nice way to look big.
My search habits have evolved quite fast - when I search for something now I first ask for quick results from Chatgpt, which gives me pointers I then drill down on. Google's revenue for 2024 was 350 billion. I know it's not all Google ads, but s lot of it is. When you follow a link from Chatgpt, it always has a utm_source=Chatgpt in it, so companies are quickly learning how important getting linked there is.
I'm not saying there's no bubble, and I personally anticipate a lot of turmoil in the next year, but monetisation of that would be the most primitive way of earning a lot of money. If anyone is dead man walking it's Google. For better or worse, Chatgpt has become to AI what Google was to search, even though I think Gemini is also good or even better. I also have my own doubts about the value of LLMs because I've already experienced a lot of caveats with the stuff it gives you. But at the same time, as long as you don't believe it blindly, getting started with something new has never been easier. If you don't see value in that, I don't know what to tell you.
If you're reading this article and wondering "When is this house of cards going to collapse!?", a little advice, gained at a high price to myself: you can waste years waiting for it to collapse, 95% of the time, it never will. I never thought Uber or Tesla would survive COVID. I'd have $450K in bitcoin if I held onto the "joke" amount I bought for $200 in 2013.
Thing that make me skip this specific narrative:
- There's some heavy-handed reaching to get to $400B next 12 months: guesstimate $50B = 1 GW of capacity, then list out 3.3 gigawatts across Broadcom chip purchases, Nvidia, and AMD
- OpenAI is far better positioned than any of the obvious failures I foresaw in my 37 years on this rock. It's very, very, hard to fuck up to the point you go out of business.
- Ed is repeating narratives instead of facts ("what did they spend that money on!? GPT-5 was a big let down!" -- i.e. he remembers the chatgpt.com router discourse, and missed that it was the first OpenAI release that could get the $30-50/day/engineer in spend we've been sending to Anthropic)
Yeah, I've re-evaluated my misconceptions around the same time. As another datapoint - infamous Tether/Bitfinex corporation which was blatantly cheating and insolvent (per normal definition) for years and every time reasonable people were giving them less than a year to bankruptcy and jail. But here were are, Tether is still printing tokens like madman, no one is in jail and every analyst was proven both wrong and right at the same time. Apparently company can be insolvent and cheating and breaking law for a looong time, with zero repercussions. And no one bats an eye.
I suspect OpenAI is deeply in red today and any normal small company would went out of business years ago. But they are too big to fail and will continue working this way. They may even transition to being profitable later on, and people will retcon that they always have been so.
At this point I just wonder, what would be exact mechanism with which Sam will redirect his negatives to a regular public like us here. Bailouts? Irresponsible government investments? Banks overreaching and them getting bailouts? Something new?...
How much does the capex model of a datacenter change when the goal is 100% utilization, with no care for node uptime beyond capex efficiency/hardware value mainenance?
I wouldn't be surprised if the cost came down by at least one order of magnitude, two if NVidia and others adjust their margin expectations. If the bet is that OpenAI can ship crappy datacenters with crappy connectivity/latency characteristics in places with cheap/existing power - then that seems at least somewhat plausible.
OpenAI burning 40 billion dollars on datacenters in the next 1 year is almost guaranteed. Modern datacenter facilities are carefully engineered for uptime, I don't think OpenAI cares about rack uptime or even facility uptime at this scale.
57 comments
[ 2.7 ms ] story [ 68.1 ms ] threadWould anyone like to found a startup doing high-security embedded systems infrastructure? Peter at my username dot com if you’d like to connect.
I’d argue the other way around: 100M growth in two months suggests literally every single human being on Earth would benefit from using this all the time, and it’s just a matter of enabling them to.
Beware the sigmoidal curve, though. Growth is exponential till it’s not.
This doesn’t make any sense. Popular is not the same as useful. You’d have a more compelling argument if you included data showing that all this increased LLM usage has had some kind of impact on productivity metrics.
Instead, some studies have shown that LLMs are making professionals less productive:
https://metr.org/blog/2025-07-10-early-2025-ai-experienced-o...
So maybe giving away more and more free stuff is good for growth? The product is excellent, ChatGPT is still my favorite but the competition isn't that behind. If fact, I started liking Grok better for most tasks
Only about 5% see enough value to drop $20 a month... It's like VR and AR, if people get a headset for free they'll use them every now and then, but virtually nobody wants to drop money on these.
LLMs have already been commodified
That aside, his math is wrong
Why can't OpenAI keep projecting/promising massive data centre growth year after year, fail to deliver, and keep making Number Go Up?
I've been thinking about American exceptionalism - they way it is head and shoulders above Europe and the developed world in terms of GDP growth, market returns, start up successes etc. and what might be the root of this success. And I'm starting to think that, apart from various mild genuine effects, it is also a sequence of circular self-fulfilling prophecies.
Let's say you're a sophisticated startup and you want some funding. Where do you go? US of course - it has the easiest access to capital. It does so presumably because US venture funds have an easier time raising funds. And that's presumably because of their track record of making money for investors - real, or at least perceived. They invest in these startups and they exit at a profit, because US companies have better valuations than elsewhere, so at IPO investors lap up the shares and the VCs make money. It's easy to find buyers for US stocks because they're always going up. In turn, they're going up because, well, there's lots of investors. It's much easier to raise billions for data centres and fairy dust because investors are in awe of what can be done with the money and anyway line always go up. Stocks like TSLA have valuations you couldn't justify elsewhere. Maybe because they will build robot AI rocket taxis, or maybe because the collective American Allure means valuations are just high.
The beauty of this arrangement is that the elements are entangled in a complex web of financial interdependency. If you think about these things in isolation, you wouldn't conclude there's anything unusual. US VC funding is so good because there's a lot of capital - lucky them. This thought of circularity only struck me when trying to think of the root cause - the nuclear set of elements that drive it. And I concluded any reason I can think of is eventually recursive.
I'm not saying America is just dumb luck kept together by spittle, of course there are structural advantages the US has. I'm just not sure it really is that much better an economic machine than other similar countries.
One difference to a Ponzi scheme is that you might actually hit a stable level and stay there rather than crash and burn. So it's more like a collective investment into a lottery. OpenAI might burn $400bn and achieve singularity, then proceed to own the rest of the world.
But I can't shake the feeling that a lot of recent US growth is a bit of smoke and mirrors. After adjusting for tech, US indices didn't outperform European ones post GFC, IIRC. Much of its growth this year is AI, financed presumably by half the world and maintained by sky-high valuations. And no one says "check" because, well, it's the US and the line always go up.
Of course it can stop, an a little history reading will show that it has always stopped, but it can take a long time.
If anything, I fear the AI hype + the orange idiot you put in charge, can fuck you up much faster than otherwise. OTOH, Trump is also a symptom, showing that things were not going great.
The Oracle deal structure: OpenAI pays ~$30B/year in rental fees starting fiscal 2027/2028 [2], ramping up over 5 years as capacity comes online. Not "$400B in 12 months."
The deals are structured as staged vendor financing: - NVIDIA "invests" $10B per gigawatt milestone, gets paid back through chip purchases [3] - AMD gives OpenAI warrants for 160M shares (~10% equity) that vest as chips deploy [4] - As one analyst noted: "Nvidia invests $100 billion in OpenAI, which then OpenAI turns back and gives it back to Nvidia" [3]
This is circular vendor financing where suppliers extend credit betting on OpenAI's growth. It's unusual and potentially fragile, but it's not "OpenAI needs $400B cash they don't have."
Zitron asks: "Does OpenAI have $400B in cash?"
The actual question: "Can OpenAI grow revenue from $13B to $60B+ to cover lease payments by 2028-2029?"
The first question is nonsensical given deal structure. The second is the actual bet everyone's making.
His core thesis - "OpenAI literally cannot afford these deals therefore fraud" - fails because he fundamentally misunderstands how the deals work. The real questions are about execution timelines and revenue growth projections, not about OpenAI needing hundreds of billions in cash right now.
There's probably a good critical piece to write about whether these vendor financing bets will pay off, but this isn't it.
[1] https://www.cnbc.com/2025/09/23/openai-first-data-center-in-...
[2] https://w.media/openai-to-rent-4-5-gw-of-data-center-power-f...
[3] https://www.cnbc.com/2025/09/22/nvidia-openai-data-center.ht...
[4] https://techcrunch.com/2025/10/06/amd-to-supply-6gw-of-compu...
AGI is absolutely a national security concern. Despite it being an enormous number, it'll happen. It may not be earmarked for OpenAI, but the US is going to ensure that the energy capability is there.
> Despite it being an enormous number, it'll happen.
Care to share your crystal ball ?
We're doing a pretty shit job of ensuring that today. Capacity is already intensely strained, and the govt seems to be decelerating investment into power capacity growth, if anything
They Don't Have the Money: OpenAI Edition
https://news.ycombinator.com/item?id=45545236
To put it another way, I don't know anything but I could probably make a '1 GW' datacenter with a single 6502 and a giant bank of resistors.
You have a lot more things in a DC than just GPUs consuming power and producing heat. GPUs are the big ones, sure, but after a while, switches, firewalls, storage units, other servers and so one all contribute to the power footprint significantly. A big small packet high throughput firewall packs a surprisingly high amount of compute capacity, eats a surprising amount of power and generates a lot of heat. Oh and it costs a couple of cars in total.
And that's the important abstraction / simplification you get when you start running hardware at scale. Your limitation is not necessarily TFlops, GHz or GB per cubic meter. It is easy to cram a crapton of those into a small place.
The main problem after a while is the ability to put enough power into the building and to move the heat out of it again. It sure would be easy to put a lot of resistors into a place to make a lot of power consumption. Hamburg Energy is currently building just that to bleed off excess solar power into the grid heating.
It's problematic to connect that to the 10kv power grid safely and to move the heat away from the system fast.
I'm not saying there's no bubble, and I personally anticipate a lot of turmoil in the next year, but monetisation of that would be the most primitive way of earning a lot of money. If anyone is dead man walking it's Google. For better or worse, Chatgpt has become to AI what Google was to search, even though I think Gemini is also good or even better. I also have my own doubts about the value of LLMs because I've already experienced a lot of caveats with the stuff it gives you. But at the same time, as long as you don't believe it blindly, getting started with something new has never been easier. If you don't see value in that, I don't know what to tell you.
Thing that make me skip this specific narrative:
- There's some heavy-handed reaching to get to $400B next 12 months: guesstimate $50B = 1 GW of capacity, then list out 3.3 gigawatts across Broadcom chip purchases, Nvidia, and AMD
- OpenAI is far better positioned than any of the obvious failures I foresaw in my 37 years on this rock. It's very, very, hard to fuck up to the point you go out of business.
- Ed is repeating narratives instead of facts ("what did they spend that money on!? GPT-5 was a big let down!" -- i.e. he remembers the chatgpt.com router discourse, and missed that it was the first OpenAI release that could get the $30-50/day/engineer in spend we've been sending to Anthropic)
I suspect OpenAI is deeply in red today and any normal small company would went out of business years ago. But they are too big to fail and will continue working this way. They may even transition to being profitable later on, and people will retcon that they always have been so.
At this point I just wonder, what would be exact mechanism with which Sam will redirect his negatives to a regular public like us here. Bailouts? Irresponsible government investments? Banks overreaching and them getting bailouts? Something new?...
I wouldn't be surprised if the cost came down by at least one order of magnitude, two if NVidia and others adjust their margin expectations. If the bet is that OpenAI can ship crappy datacenters with crappy connectivity/latency characteristics in places with cheap/existing power - then that seems at least somewhat plausible.
OpenAI burning 40 billion dollars on datacenters in the next 1 year is almost guaranteed. Modern datacenter facilities are carefully engineered for uptime, I don't think OpenAI cares about rack uptime or even facility uptime at this scale.