Excellent article Devansh. What stands out to me is how this isn’t just capital consolidation but cognitive consolidation. Once CUDA becomes the default mental model for compute, the monopoly extends far beyond hardware. The next fault line may not be chips, but energy and memory bandwidth. I think whoever cracks those constraints will shape the next architecture cycle. Keen to hear your thoughts. Once again, great article. Got me thinking!
You outdid yourself on this one Devansh - really excellent.
I may be completely off here, but the only way I can see an alternative commercial architecture gaining traction is if someone like AWS (and honestly, probably only AWS) puts their weight behind it - like they did with Graviton.
The migration with AI compute is more complex than the move to Graviton was, but with Amazon's scale, the payoff could potentially be huge for them even with relatively small cost savings from an alternative architecture. And it would also give AWS something that might be even more valuable to them long-term than the cost savings - lack of dependence on NVIDIA.
The other potential disruptor here is China. AI tech export restrictions are already pushing them to develop alternatives, and I only expect it to continue. Even with the NVIDIA sales carve-out, China is highly allergic to dependence on NVIDIA (or any non-Chinese company), and they're likely very willing to pour national resources into it. So, I expect they'll eventually have their own alternative, and to push it internationally once they do. But I have no idea how long that will take.
Regarding the topic of the article, this piece is spot on. The bubble talk totally ignores the fundamental technological shift what's happening. The massive infrastructure investments, like with Oracle and Nvidia, are real and fundational. Keep up the sharp analysis, it's needed.
maybe needed add to the head title.. in the west. China, an Russia are applying AI where it worst, agriculture, science, research, engineeristics, communications, medicine, prospection, extraccions, chains production, logistics, transports, and so on, the west its using AI on a hollybollywoodian mind asset. Even if as fkr EU racism is killing them, both, them picked up advanced and so well SGEM students, technics, processkrs to fullfill Silicon Valey advance, inEurope the opposite, advances were reserved for only the rich sonvs as collider particles Laboratory in Switzerland.. and even expulsing or cancel foreignets from the science places as US did. Result: sinical, China in 2 years did own orbital Station, and working at home finished first the own "Sun" project and are ahead of mist science. Payenting in sny sector more thsn the west.
AI is a tool, if you put stupid people to handle..or use it for fun...
I enjoyed the article, but I'm left wondering. Is there a cuda bubble?
Politically, I can see a forced break of global monopolies, which could lead to a break up of truly global monopoly companies... Leading to a collapse in their future values?
You can't really break up cuda because it's dominance is due to an ecosystem around it, not the technology itself. Regulating that without overreach would be a shit show. Especially with how many ml frameworks are always being built on that
That’s a really interesting take Timmy. I don’t think it’s a CUDA bubble as much as a kind of architectural lock-in. The real fragility in my mind is dependency not valuation. When so much of the AI economy, from national R&D to private capital, runs through one compute stack, any disruption whether political, supply-chain, or regulatory becomes systemic. The next realignment might not come from policy, but from geopolitics or a new compute paradigm that breaks the current mental model.
Good question, Timmy. I probably over-compressed that thought. CUDA is the specific tech layer (Nvidia’s programming framework for GPUs). Architectural lock-in is what happens because of it. Once everything from research code to data centers is built around CUDA, switching off it becomes economically irrational. So CUDA is the tool; the lock-in is the trap. Hope this clarifies?
Thanks I get that. I'm still not picking up the difference between "it's not a cuda bubble", and "the lock-in is the trap". My point was that a trap is a bubble, IF global forces break open the trap.
Yes, exactly. I think we are describing the same fragility from different angles. If the dependency ever snaps, it would unwind like a bubble. I just see that risk as slower-moving, because the dependency itself is still deepening.
The cognitive lock-in angle is brillant - most people focus on market share but miss how CUDA has become the default mental framework for parallel computation. Your point about the porting tax being organizational rather than just technical is spot-on. Companies aren't just switching chips, they're asking teams to unlearn years of accumulated knowledge and rebuild from scratch.
I think the tech can be great and there can be a bubble. Internet, rails come to mind. But perhaps you are just being more stringent with your definition
The difference between bubble and not isn't utility (tulips still have utility) but on the basis of returns. Essentially do you buy because you think this asset will be priced higher than you can sell to another person (without it generating more value per se) or are you buying because you think the tech itself and what it does will become more useful over time.
With dotcom and real estate the Internet was useful but a lot of the specific companies didn't have plan to capture value. Investment was made on the assumption that it was good for its own sake. That's a bubble. I think there are companies that are doing this even in AI but the biggest players with the deals have clear business cases (possibly faulty if you disagree with the assumptions but still a clear set of cases).
I agree with basically everything you just said. I think if you go into the private space you find a low more of these but the hyper scalers are obviously not in this realm. I think the debate really ends up on how the economics shake out. Vibe coding platforms are already commodities and will have horrible economics (I presume they do already).
That said, for a lot of these use cases it is years away till we have certainty.
Devansh you are flat out wrong. It’s one thing for AWS to give discounts to start ups. It’s another for massive investments in prime customers. Look up Global Crossing and Lucent for clear examples of the same fraud play book.
The issue is revenue recognition if I pay you $100 and you “invest” $90 back to me, that is not a true sale. Transactions like that cook financial statements. You did not address that, which is the real issue, not the monopoly lock in of CUDA, or any othe tech issue. Mark my words.
1. Companies are relying on financial engineering (something we've discussed at length in the live streams/other posts).
2. There are companies primed to bust.
3. The technology itself has very strong long term potential.
Let's take your example of lucent etc. Unlike those cases, Gen AI has very clear growing demand from outside. The companies with the biggest bets in it have massive revenue streams from outside it and very clear floors for how much they will use it. So this is structurally very different.
Also this article is specifically about AI as a technology. Going back to the dot com bubble, yes many internet companies then were wiped out, but the Internet has more than paid for itself. Not sure why you think ai will be any different.
I do wonder though does the cuda monopoly exclude an AI bubble? I suppose there is still a chance that the revenue of AI cannot justify the extreme spendings on data centers and infrastructure?
It seems like you are using a definition of “bubble” different than everyone else. Relevant Example: dot-com bubble. No one would argue that dot-com technology didn’t have real fundamental value during that bubble. Only that the market prices decoupled from that value during the bubble then corrected back BUT NOT TO ZERO like you seem to be using in your definition. Bubble or not is a question about the FUTURE fundamental value, if future value end up high enough to justify today’s market values then no bubble, if you think it won’t and today’s market valuations are too high then you are saying there is a bubble.
This article is excellent in general for pointing out that even if there IS a correction coming, that the monopoly position can still be cemented. Also by pointing out that vendor financing IS a way for them to activate accelerate the tech uptake and grow the future fundamental value faster than it would have (I.e. preventing a bubble by making today’s speculation come true).
It's not about going back to zero (my feet pics, tulips etc aren't free).
Bubbles are set by the purchase of the good as a resellable. Essentially it would be a bubble if people were buying ChatGPT licenses to resell them (or doiing something similar with APIs).
People buying AI are either trying to use it directly (use ChatGPT) or shape it into somethign else (an AI Apps company like Iqidis). In either case, that's a very different situation than a bubble.
Now you could argue that there is some speculation on the equities space, but I talk to a lot of investors and most of them are pricing on various models and assumption. Now whether you agree with thgeir exact strategies (which is where the argument for overvalued would come from), you can't really say that there is matches the bubble dynamic
Interesting, but I am still not understanding this resellable distinction you are using. Can you clarify?
I agree that AI tokens aren’t a resellable resource. However some of these companies are certainly buying AI HW infrastructure in the hopes of future sales of compute to other companies, is that resale?
I agree things like NFT and feet pictures match the resellable pattern. But what about the dot-com bubble example? What exactly was being resold there in a way that is distinct from current AI situation? Or you not calling that a bubble?
Shares in public AI companies are certainly resellable. This I think is exactly what you mean by speculation in the equity space. I agree with you here, and I do think this is exactly the sense that most people are calling this a bubble. And this is exactly what the dot-com bubble was related to. I also agree that no one knows for sure if this is a bubble or not yet as speculation is fundamentally about the future and thus won’t be settled till it happens. If market corrects then it was a bubble. If real revenue continues to grow fast enough that it’s not a bubble.
The difference is in the actual process of profitability. During the.com bubble people were getting websites and.coms because it was seen as an inherently Superior thing to do so. The.com itself are evaluations and that is true for a lot of AI startups as well. So that's why I'm saying that parts of the market might be overvalued and parts of the industry might be called bubble. But on the big players like the ones we mentioned the bet is on unit economics ironing out. So basically the goal is you make AI cheap for now and you invest a lot into infrastructure. Okay and infrastructure is going to create much cheaper production of tokens in the future and as the tokens got cheaper and cheaper people will continue to use them more. But because the cost of tokens will fall much more than the cost of actually using them. So cost of production is lower than the cost of consumption. Your margins will still be higher so you might be making a 20% profit on tokens wire. Right now you can only make a 10% profit on tokens and because it's cheaper more people will be buying it. Cost will be much better. Joao that really is ai's foundational better right now and that's what makes it not rhyming with a bubble just like there is a very clear assumption here now. Whether you agree with this assumption is another story and you can argue against it, but calling it a bubble on the basis that you don't agree with. The assumption is bad. Faith argument.
Can I take it another step back? The AI infrastructure spend literally did not happen yet. Altman and these 2030 projections are about 2030. Last I checked, we are in 2025. If anyone talks to people in the supply chain and infra side, they will quickly find out that OpenAI is pie in the sky. People get all worked up about the spend but what they are really saying is 1) some of the stock prices of certain stocks are deranged (or some of the people are haters bc they missed out and are too lazy to understand the industry and rather rely on confirmation bias); or 2) these are really big numbers that I cannot understand (and am scared of).
Think its healthy to have strong perspectives, not so much about any right/wrong takes on it, but that our depth of understanding is that much more - illuminated ⭐️🔥….🤞
If the argument you are driving is NVIDIA = CUDA = AI and this is not in a bubble. It is a fair argument. This can't be extended across the whole AI hype that is driving valuations. The hard reality is that GenAI failed miserably on enterprise adoption is not because of technology itself but enterprises need low or zero entropic processes to reliably operate which GenAI is not good at. Of course there are other usecases for GenAI and again most of it is slop and it will take lot more time to figure out where it really fits in. Kind of similar to what we experienced with dot com bubble.
Loved this! There’s so many great points, from breaking down how ponzi & bubble language fail to match the actual model of future architectural power getting locked in, to calling out how nvidia is dominating mindshare… but the best one was this:
Call me a hater, but I’ve never seen a single tech company succeed because they hired more MBAs.
Lmao, great work! I understand the bubble discourse, however a far-travelling headline is also a solid tool for misconception. You have expressed a lot of the gaps I didn’t have the language or understanding for.
Slightly misleading "Every major AI deployment defaults to CUDA" - Deepseek did not use CUDA, programmed assembly at PTX level iirc, and made the whole lot open source...
They also demonstrated that brute force, NVidia et al's model, was not the only approach.
And many people have many reasons to look at that precedent.
Excellent article Devansh. What stands out to me is how this isn’t just capital consolidation but cognitive consolidation. Once CUDA becomes the default mental model for compute, the monopoly extends far beyond hardware. The next fault line may not be chips, but energy and memory bandwidth. I think whoever cracks those constraints will shape the next architecture cycle. Keen to hear your thoughts. Once again, great article. Got me thinking!
That's my hope too
You outdid yourself on this one Devansh - really excellent.
I may be completely off here, but the only way I can see an alternative commercial architecture gaining traction is if someone like AWS (and honestly, probably only AWS) puts their weight behind it - like they did with Graviton.
The migration with AI compute is more complex than the move to Graviton was, but with Amazon's scale, the payoff could potentially be huge for them even with relatively small cost savings from an alternative architecture. And it would also give AWS something that might be even more valuable to them long-term than the cost savings - lack of dependence on NVIDIA.
The other potential disruptor here is China. AI tech export restrictions are already pushing them to develop alternatives, and I only expect it to continue. Even with the NVIDIA sales carve-out, China is highly allergic to dependence on NVIDIA (or any non-Chinese company), and they're likely very willing to pour national resources into it. So, I expect they'll eventually have their own alternative, and to push it internationally once they do. But I have no idea how long that will take.
All great points.
I'm really glad you liked this. Thank you. Your kind words mean a lot given all the insight you've shared over the many posts
Regarding the topic of the article, this piece is spot on. The bubble talk totally ignores the fundamental technological shift what's happening. The massive infrastructure investments, like with Oracle and Nvidia, are real and fundational. Keep up the sharp analysis, it's needed.
Thank you
maybe needed add to the head title.. in the west. China, an Russia are applying AI where it worst, agriculture, science, research, engineeristics, communications, medicine, prospection, extraccions, chains production, logistics, transports, and so on, the west its using AI on a hollybollywoodian mind asset. Even if as fkr EU racism is killing them, both, them picked up advanced and so well SGEM students, technics, processkrs to fullfill Silicon Valey advance, inEurope the opposite, advances were reserved for only the rich sonvs as collider particles Laboratory in Switzerland.. and even expulsing or cancel foreignets from the science places as US did. Result: sinical, China in 2 years did own orbital Station, and working at home finished first the own "Sun" project and are ahead of mist science. Payenting in sny sector more thsn the west.
AI is a tool, if you put stupid people to handle..or use it for fun...
That too. But these critics wouldn't be able to appreciate all this nuance
Constructive critics would be always wellcome. It’s up on us ourselves find the diverses nuances for each temathics, its called..be informed.
I enjoyed the article, but I'm left wondering. Is there a cuda bubble?
Politically, I can see a forced break of global monopolies, which could lead to a break up of truly global monopoly companies... Leading to a collapse in their future values?
You can't really break up cuda because it's dominance is due to an ecosystem around it, not the technology itself. Regulating that without overreach would be a shit show. Especially with how many ml frameworks are always being built on that
That’s a really interesting take Timmy. I don’t think it’s a CUDA bubble as much as a kind of architectural lock-in. The real fragility in my mind is dependency not valuation. When so much of the AI economy, from national R&D to private capital, runs through one compute stack, any disruption whether political, supply-chain, or regulatory becomes systemic. The next realignment might not come from policy, but from geopolitics or a new compute paradigm that breaks the current mental model.
Thanks. Can you explain the difference between architectural lock-in and CUDA, I'm a bit lost on your point here.
Good question, Timmy. I probably over-compressed that thought. CUDA is the specific tech layer (Nvidia’s programming framework for GPUs). Architectural lock-in is what happens because of it. Once everything from research code to data centers is built around CUDA, switching off it becomes economically irrational. So CUDA is the tool; the lock-in is the trap. Hope this clarifies?
Thanks I get that. I'm still not picking up the difference between "it's not a cuda bubble", and "the lock-in is the trap". My point was that a trap is a bubble, IF global forces break open the trap.
Yes, exactly. I think we are describing the same fragility from different angles. If the dependency ever snaps, it would unwind like a bubble. I just see that risk as slower-moving, because the dependency itself is still deepening.
The cognitive lock-in angle is brillant - most people focus on market share but miss how CUDA has become the default mental framework for parallel computation. Your point about the porting tax being organizational rather than just technical is spot-on. Companies aren't just switching chips, they're asking teams to unlearn years of accumulated knowledge and rebuild from scratch.
I think the tech can be great and there can be a bubble. Internet, rails come to mind. But perhaps you are just being more stringent with your definition
I think I was just unclear.
The difference between bubble and not isn't utility (tulips still have utility) but on the basis of returns. Essentially do you buy because you think this asset will be priced higher than you can sell to another person (without it generating more value per se) or are you buying because you think the tech itself and what it does will become more useful over time.
With dotcom and real estate the Internet was useful but a lot of the specific companies didn't have plan to capture value. Investment was made on the assumption that it was good for its own sake. That's a bubble. I think there are companies that are doing this even in AI but the biggest players with the deals have clear business cases (possibly faulty if you disagree with the assumptions but still a clear set of cases).
I agree with basically everything you just said. I think if you go into the private space you find a low more of these but the hyper scalers are obviously not in this realm. I think the debate really ends up on how the economics shake out. Vibe coding platforms are already commodities and will have horrible economics (I presume they do already).
That said, for a lot of these use cases it is years away till we have certainty.
Thanks for the reply
Devansh you are flat out wrong. It’s one thing for AWS to give discounts to start ups. It’s another for massive investments in prime customers. Look up Global Crossing and Lucent for clear examples of the same fraud play book.
The issue is revenue recognition if I pay you $100 and you “invest” $90 back to me, that is not a true sale. Transactions like that cook financial statements. You did not address that, which is the real issue, not the monopoly lock in of CUDA, or any othe tech issue. Mark my words.
Multiple things can be true--
1. Companies are relying on financial engineering (something we've discussed at length in the live streams/other posts).
2. There are companies primed to bust.
3. The technology itself has very strong long term potential.
Let's take your example of lucent etc. Unlike those cases, Gen AI has very clear growing demand from outside. The companies with the biggest bets in it have massive revenue streams from outside it and very clear floors for how much they will use it. So this is structurally very different.
Also this article is specifically about AI as a technology. Going back to the dot com bubble, yes many internet companies then were wiped out, but the Internet has more than paid for itself. Not sure why you think ai will be any different.
AI is here to stay.
The tech isn’t going anywhere.
No question there.
Some CFOs might go to jail for booking bogus sales.
Bold take, respect.
I do wonder though does the cuda monopoly exclude an AI bubble? I suppose there is still a chance that the revenue of AI cannot justify the extreme spendings on data centers and infrastructure?
As I've mentioned in the work, an overvalued market is not the same as a bubble.
I do believe there will be a massive market correction as a lot of money goes into useless stuff. But that doesn't make the technology a bubble.
It seems like you are using a definition of “bubble” different than everyone else. Relevant Example: dot-com bubble. No one would argue that dot-com technology didn’t have real fundamental value during that bubble. Only that the market prices decoupled from that value during the bubble then corrected back BUT NOT TO ZERO like you seem to be using in your definition. Bubble or not is a question about the FUTURE fundamental value, if future value end up high enough to justify today’s market values then no bubble, if you think it won’t and today’s market valuations are too high then you are saying there is a bubble.
This article is excellent in general for pointing out that even if there IS a correction coming, that the monopoly position can still be cemented. Also by pointing out that vendor financing IS a way for them to activate accelerate the tech uptake and grow the future fundamental value faster than it would have (I.e. preventing a bubble by making today’s speculation come true).
It's not about going back to zero (my feet pics, tulips etc aren't free).
Bubbles are set by the purchase of the good as a resellable. Essentially it would be a bubble if people were buying ChatGPT licenses to resell them (or doiing something similar with APIs).
People buying AI are either trying to use it directly (use ChatGPT) or shape it into somethign else (an AI Apps company like Iqidis). In either case, that's a very different situation than a bubble.
Now you could argue that there is some speculation on the equities space, but I talk to a lot of investors and most of them are pricing on various models and assumption. Now whether you agree with thgeir exact strategies (which is where the argument for overvalued would come from), you can't really say that there is matches the bubble dynamic
Interesting, but I am still not understanding this resellable distinction you are using. Can you clarify?
I agree that AI tokens aren’t a resellable resource. However some of these companies are certainly buying AI HW infrastructure in the hopes of future sales of compute to other companies, is that resale?
I agree things like NFT and feet pictures match the resellable pattern. But what about the dot-com bubble example? What exactly was being resold there in a way that is distinct from current AI situation? Or you not calling that a bubble?
Shares in public AI companies are certainly resellable. This I think is exactly what you mean by speculation in the equity space. I agree with you here, and I do think this is exactly the sense that most people are calling this a bubble. And this is exactly what the dot-com bubble was related to. I also agree that no one knows for sure if this is a bubble or not yet as speculation is fundamentally about the future and thus won’t be settled till it happens. If market corrects then it was a bubble. If real revenue continues to grow fast enough that it’s not a bubble.
Voice to texting this, so forgive minor errors-
The difference is in the actual process of profitability. During the.com bubble people were getting websites and.coms because it was seen as an inherently Superior thing to do so. The.com itself are evaluations and that is true for a lot of AI startups as well. So that's why I'm saying that parts of the market might be overvalued and parts of the industry might be called bubble. But on the big players like the ones we mentioned the bet is on unit economics ironing out. So basically the goal is you make AI cheap for now and you invest a lot into infrastructure. Okay and infrastructure is going to create much cheaper production of tokens in the future and as the tokens got cheaper and cheaper people will continue to use them more. But because the cost of tokens will fall much more than the cost of actually using them. So cost of production is lower than the cost of consumption. Your margins will still be higher so you might be making a 20% profit on tokens wire. Right now you can only make a 10% profit on tokens and because it's cheaper more people will be buying it. Cost will be much better. Joao that really is ai's foundational better right now and that's what makes it not rhyming with a bubble just like there is a very clear assumption here now. Whether you agree with this assumption is another story and you can argue against it, but calling it a bubble on the basis that you don't agree with. The assumption is bad. Faith argument.
Can I take it another step back? The AI infrastructure spend literally did not happen yet. Altman and these 2030 projections are about 2030. Last I checked, we are in 2025. If anyone talks to people in the supply chain and infra side, they will quickly find out that OpenAI is pie in the sky. People get all worked up about the spend but what they are really saying is 1) some of the stock prices of certain stocks are deranged (or some of the people are haters bc they missed out and are too lazy to understand the industry and rather rely on confirmation bias); or 2) these are really big numbers that I cannot understand (and am scared of).
Yea,
Referring to some Mark Ross - a recent find
This time market is going to be an inflationary one, unlike 2008's deflationary one
Yes excellent and fair take by @Devansh 🫡👏🙏. Great Read. I agree with many aspects. Wrote quite a bit about Nvidia’ stack (extreme design) here: https://interestingengineering.substack.com/p/nvidias-extreme-co-design?utm_source=share&utm_medium=android&r=223m94
Also read this (if you have the time):
Think its healthy to have strong perspectives, not so much about any right/wrong takes on it, but that our depth of understanding is that much more - illuminated ⭐️🔥….🤞
https://interestingengineering.substack.com/p/all-in-ais-house-of-cards?utm_source=share&utm_medium=android&r=223m94
Will read it soon.
If the argument you are driving is NVIDIA = CUDA = AI and this is not in a bubble. It is a fair argument. This can't be extended across the whole AI hype that is driving valuations. The hard reality is that GenAI failed miserably on enterprise adoption is not because of technology itself but enterprises need low or zero entropic processes to reliably operate which GenAI is not good at. Of course there are other usecases for GenAI and again most of it is slop and it will take lot more time to figure out where it really fits in. Kind of similar to what we experienced with dot com bubble.
So 4th Industrial Revolution confirmed?
I keep wondering this rhetoric
Loved this! There’s so many great points, from breaking down how ponzi & bubble language fail to match the actual model of future architectural power getting locked in, to calling out how nvidia is dominating mindshare… but the best one was this:
Call me a hater, but I’ve never seen a single tech company succeed because they hired more MBAs.
Lmao, great work! I understand the bubble discourse, however a far-travelling headline is also a solid tool for misconception. You have expressed a lot of the gaps I didn’t have the language or understanding for.
Banks are looking at the fall out of the AI bubble popping, check this out:
https://substack.com/profile/315148915-staplestack/note/c-174045959?r=57mqgj&utm_source=notes-share-action&utm_medium=web
Slightly misleading "Every major AI deployment defaults to CUDA" - Deepseek did not use CUDA, programmed assembly at PTX level iirc, and made the whole lot open source...
They also demonstrated that brute force, NVidia et al's model, was not the only approach.
And many people have many reasons to look at that precedent.