Can AI Keep Growing Exponentially? Let’s Ask SemiAnalysis — With Dylan Patel and Jordan Nanos
Alex KantrowitzDylan PatelJordan Nanos
- SemiAnalysis argues that viral charts materially understate the scale of the AI buildout. Dylan Patel estimates next-year US capital spending near $2 trillion across data centers, chips, and the supporting supply chain—roughly 5–6% of a $30-trillion-plus economy, not 3.6%. America is also building training and inference capacity for overseas users, making this “much greater than in any other cycle of physical infrastructure investment.”
- Triple-digit capex growth must slow, but SemiAnalysis does not yet see the demand signal that would mark a bust. Hyperscaler capex was growing 116% year over year in Q3 2026, while engineering capacity, cooling, chips, capital, and the ability to organize another gigawatt are becoming constraints. Demand still “significantly exceeds supply,” and nobody can tell whether the excess is 10%, 50%, or effectively unbounded until enough capacity comes online.
- The payback math looks less absurd when infrastructure life and revenue lag are included. Dylan models GPUs over six years and data centers over 15: $1 trillion invested today might require about $1.5 trillion of cumulative revenue, or $250 billion annually over six years—not immediate matching revenue. He thinks only $3 trillion of AI revenue in 2031 would be a problem, while suggesting OpenAI and Anthropic together might reach $600–700 million by December of next year and perhaps $1.5 trillion by the end of 2027.
- AI spending is expected to draw on labor, R&D, and cost of goods sold while shifting value toward compute-based capital. Dylan’s formulation is that tokens are “essentially just capital” embodied in chips and data centers; companies can keep headcount stable while producing and selling much more. His chip-design example is the template: US engineering headcount stayed broadly flat for two decades while tools helped create far more valuable chips across phones, sensors, robots, and AI systems.
- Anthropic could have software-like—or better—economics, but its upside case implies profound social disruption. Dylan says Anthropic already earns roughly 75% gross margins on output, has almost no customer-acquisition cost, and may be a “better business than software” if it sustains model leadership. Yet a $20 trillion valuation at perhaps 5× sales implies several trillion dollars of revenue and an AI economy so concentrated that, in his deliberately pessimistic framing, “society as a whole could be in trouble.”
- Neo-cloud financing is real financial plumbing, not sufficient evidence that NVIDIA fabricated demand. Jordan Nanos says SemiAnalysis tracks $588 billion of NVIDIA-backed off-balance-sheet guarantees, about 2.5% of US M2, which can make projects financeable by giving lenders an investment-grade fallback buyer. NVIDIA often sells the capacity to customers willing to pay more, while the broader constraint spans construction, memory, optics, factories, and industrial equipment—not merely GPU purchases.
- The near-term fault line is operational quality and security, while frontier labs increasingly reserve their strongest models for themselves. SemiAnalysis found only a small recommended set among 323 tracked providers and described cases where customers could see other tenants’ data and government workloads; “most of these clouds are terrible when it comes to security.” Meanwhile Anthropic and OpenAI have delayed or limited models including Mythos, Fable, Glasswing, Astra, and Bell, suggesting public releases may become “minimally viable models” that preserve an edge while higher-value capabilities stay internal.
1. The AI buildout is larger than the headline comparisons
Dylan’s reaction to the viral infrastructure chart was that “people’s numbers are undercounted.” He puts next-year US spending near $2 trillion once data centers, chips, and the surrounding supply chain are included—about 5–6% of GDP rather than the chart’s 3.6%.
Railways and fiber were largely built for the countries recording the investment. US AI infrastructure trains models domestically and serves inference abroad: Dylan cites European Anthropic and OpenAI usage, plus Asian demand, as reasons America’s build carries economic capacity beyond its borders.
Alex’s benchmark was 116% year-over-year hyperscaler capex growth in Q3 2026, with consensus falling below 100% in Q4 and toward 70% in the first quarter of next year. Jordan says limits should arrive toward the end of next year and beyond because EPC capacity, industrial equipment, cooling, chips, and financing cannot keep compounding at triple-digit rates indefinitely.
2. Slower capex growth would not itself mean the boom is over
The arithmetic is already becoming harder: Dylan describes revenue rising from below $10 billion to above $100 billion in one year, but says it will not become $1 trillion by the end of next year. NVIDIA can still double while visibly slowing from earlier multiples.
SemiAnalysis’s distinction is crucial: infrastructure spending growth must eventually slow, while revenue and profit growth must accelerate to justify the installed base. Current AI revenue reflects prior years’ construction, so comparing today’s sales directly with today’s capex mismatches two different periods.
Jordan’s demand test is concrete: do not forecast cancellations until usage slows. Every model presently shows demand “significantly exceeds supply,” but constrained GPU and data-center availability prevents anyone from knowing whether unmet demand is 10%, 50%, or far larger.
The host’s neo-cloud concern prompted an important correction: major customers generally sign five- or six-year capacity contracts at fixed hourly rates, not cancellable monthly commitments. Meta can leave its own campuses empty or delay buying chips for them, but it cannot simply abandon signed third-party capacity without consequence.
3. The payback case depends on asset life and much broader monetization
Dylan uses six years for GPU depreciation and roughly 15 years for data centers. On his simplified math, $1 trillion invested today might need $1.5 trillion over six years—around $250 billion annually—while the revenue curve could start low and rise sharply later.
That is why he dismisses the claim that Meta committed “corporate suicide.” Meta has contractual obligations, but it can stop authorizing additional campuses; its existing cash flow can cover what it has signed, leaving the real question as return on investment rather than solvency.
Dylan says only $3 trillion of AI revenue by 2031 would indeed create “a big problem.” He says OpenAI and Anthropic together could reach $600–700 million by December of next year and perhaps $1.5 trillion by the end of 2027. He says Anthropic is already profitable on revenue minus compute costs for training and inference, while OpenAI could follow within two or three quarters.
Dylan argues that token subscriptions are too narrow a revenue frame. Frontier intelligence could enter robotics, autonomous vehicles, drug development, healthcare, and life sciences; a lab using AI to produce a cancer treatment could capture economics closer to a breakthrough drug than a cloud API.
4. AI shifts value from labor toward capital without necessarily shrinking payrolls
Asked where hundreds of billions of customer spending would come from, Dylan points to expanding R&D budgets, cost of goods sold, and the labor force. Tokens are “essentially just capital”—chips and data centers performing work—so AI accelerates the decades-long decline in labor’s share of GDP.
His explicitly hedged hope is that rapid overall growth keeps labor income stable or even rising in dollars despite its falling percentage share. Companies might employ the same number of people while building more products, charging more, improving margins, and directing a larger portion of incremental value to capital.
Chip design is Dylan’s model for the transition: US R&D engineering headcount has barely changed for roughly 20 years, yet the sector’s value exploded as better tools amplified each developer. AI could repeat that pattern across chips for phones, AirPods, robots, GPS, Bluetooth, gas sensors, and AI itself.
Alex pressed Dylan on his earlier suggestion that intellectual work was ending. Dylan separated the productive period through roughly 2028—when unusually active people gain enormous leverage—from the “madness of 2035,” when systems better than humans at everything might make intellectual labor genuinely obsolete.
5. Anthropic’s upside is both commercially attractive and socially destabilizing
Dylan called his $2 trillion Anthropic IPO post “a complete trash post,” then defended its underlying paradox. Anthropic might become the first $10 trillion company, but $20 trillion at mature multiples near 5× sales could require $2–4 trillion of revenue and capabilities powerful enough to transform the global economy.
His pessimistic branch is extreme concentration: employment disruption and gains accruing to Anthropic shareholders or supply-chain owners could produce political unrest and “tear apart the fabric of society.” Jordan is less pessimistic, imagining value created through new medicines and invoking GLP-1 as an example of research generating both broad benefits and exceptional profits.
The operating case is already strong. Dylan says Anthropic earns around 75% gross margins on output, especially advanced models, with customer-acquisition cost “almost zero”; R&D is rising more slowly than revenue, while existing customers appear loyal and tend to expand spending.
Claims of voluntarily “slowing down the frontier” strike Dylan as signaling more than strategy. Dario cannot trust Sam to stop, Sam cannot trust Dario to stop, and neither can ignore China; serious security incidents may cause temporary restraint, but “they have very little opportunity to slow down.”
6. NVIDIA’s guarantees finance a constrained market rather than manufacture it
Alex described the report’s mechanism as including minimum-revenue, landlord, and lease guarantees. Jordan says SemiAnalysis tracks $588 billion of NVIDIA-backed off-balance-sheet guarantees, equivalent to about 2.5% of US M2. The mechanism gives lenders a credible fallback buyer, allowing neo-clouds to secure financing without relying exclusively on hyperscalers.
In practice, NVIDIA may agree to take capacity but never need to: another customer arrives willing to pay more, so NVIDIA sells the GPUs to that party. Jordan therefore sees the arrangement as pragmatic diversification, giving NVIDIA more customers and labs more alternatives to three or four hyperscalers.
NVIDIA also has legitimate internal demand through model research, software testing, and continuous integration, although Jordan says this consumes only a small share of its total business. More importantly, projects need financing for construction, memory, optics, factories, and equipment throughout the supply chain—not merely chips.
The hard costs are similarly misunderstood: concrete, steel, and walls matter operationally but represent a minority of a new cluster’s bill. HBM, optics, and GPUs dominate, so an owner can complete a campus yet postpone its largest expenditure by leaving it unpopulated.
7. Cloud quality, security, and private models are becoming competitive moats
ClusterMAX 3.0 evaluates the renter’s experience—reliability, security, storage, networking, lifecycle tooling, Slurm, and Kubernetes—not stock attractiveness. CoreWeave and Nebius rank highest, while Oracle can provide excellent managed clusters even as debt, fixed-margin contracts, and execution failures hurt its investment case.
Oracle’s New Mexico site illustrates the split: missing pipelines led to delays and eventually force majeure, while Wisconsin also faced slippage. Nebius, with the shortest average contract duration among leading providers, could raise prices quickly; Oracle and CoreWeave were more constrained by long-term agreements.
Of 323 tracked providers, only a few satisfy SemiAnalysis’s recommendation threshold. Security is the ugliest failure: Dylan says GLM 5.3 does a good job of hacking neo-clouds and that Chinese open-source models can hack dozens of them. He also recounts accidentally seeing other tenants’ data, Slurm activity, and GPU workloads belonging to agencies from a non-US top-10-GDP country.
Google faces a different allocation problem. Jordan says DeepMind currently has the least computing power among DeepMind, Anthropic, and OpenAI, even as Google’s total compute rises. Dylan warns that next year OpenAI and Anthropic may add 5–10 gigawatts while Google adds only a few. “Nobody needs the third, fourth, or fifth best coding model,” he says of Gemini 4 Argon.
Labs are already withholding frontier capability. Anthropic had Mythos ready in February; the public later got a limited Fable, while Glasswing was limited to some subsectors. OpenAI held Astra for months and displayed Bell’s math performance without releasing it. Dylan expects “minimally viable models” publicly—enough to stay ahead of Chinese and open-source systems—while months of progress remain internal.
Full transcript
How big does AI have to become to justify trillions in infrastructure spending? We'll talk about this with leading experts from SemiAnalysis right after. Welcome to the Big Technology podcast, a show for thoughtful and insightful conversations about the world of technology and beyond.
The numbers for spending on AI infrastructure seem to be only increasing, with no end in sight. So we'll talk about what the physics of this whole construction effort looks like and whether it can sustain itself—if not destroy the world, at least destroy the economy.
Today we're joined by Dylan Patel, CEO and founder of SemiAnalysis, and Jordan Nanos, a member of the SemiAnalysis technical staff, to discuss everything happening in infrastructure development. Dylan, Jordan, welcome to the show.
Let's start by discussing the scale of what we're seeing in the development of AI infrastructure. Dylan, there's a graph currently circulating on X that compares the development of AI infrastructure to past projects like railroads or highways. The graph shows that the share of AI infrastructure in GDP today is about 3.6%. Sorry, the median is about 3.6%.
The annual average rate is 2.2% for railways. Actually, the comment is that it's much bigger than that, right? This means that it is 50% more than railways, even in terms of its share of GDP. The comment is that this is beyond anything we've seen before, on a scale that is virtually unprecedented.
I'd like to hear your thoughts on how big this is, whether this is the right way to look at it, and what it means if we're building something this big.
1. The Infrastructure Spending Surge
When this graph started going viral, I thought, “Wow, people's numbers are undercounted.” In fact, they are already much higher.
Really? Okay, so share the real numbers.
If you look at next year, capital spending in the United States is going to be about $2 trillion, including not just data centers and chips but the entire rest of the supply chain that people are investing in. So you're approaching $2 trillion in capital spending.
If we divide this by the GDP, which is about $30 trillion or more, you have somewhere around 5% or 6%. Isn't that right? Not 3% or 2%.
My reasoning is that these numbers are actually lower than they really are because people are underestimating the amount of capital spending that's happening in the U.S., North America, and the world.
Part of it is because, for example, with railways, your capital expenditure is directed to the country where you deploy it. With fiber, your capital expenditure is directed to the country where you are laying it. But for AI infrastructure, your capital spending in America is largely focused on serving users outside of America.
If you look at Europe, for example, most of Anthropic's inference for Europe takes place in America. Most of OpenAI's inference for Europe also happens in America, and the same goes for many Asian countries.
There is some aspect of capital spending in America being much higher because of this. First, you conduct all your training here, and second, you actually serve other countries from data centers in America. Ultimately, you find yourself at a point where the economic impact of capacity is much greater than in any other previous cycle of physical infrastructure investment.
Okay, Jordan, you are the author of ClusterMAX 3.0, a ranking of GPU cloud systems from SemiAnalysis. You tier a lot of hyperscalers and neoclouds, talking a little bit about their performance, and we'll get to that later. But you look at the big picture in your report, and I'd like you to comment on that.
Basically, we're looking at where the quarterly growth in hyperscaler capex is going year over year, and in Q3 2026, the growth is 116% year over year. That's impressive.
The consensus forecast—what Wall Street expects after that—is a decline in the fourth quarter to below 100% annual growth, then to 70% in the first quarter of next year, and then it continues to fall.
I think it's a cycle that we're used to, where it seems like we're all on a bigger base, so it can't beat expectations again and it can't double again. I'm wondering, in your opinion and based on your research, are we actually going to see a slowdown in this trend? Or are we falling back into the same trap—or is Wall Street falling back into the same trap—of saying, “Okay, you've hit the biggest numbers imaginable, but you still need to think bigger”?
I think we're definitely going to hit the limits toward the end of next year and beyond.
Hold on, Jordan, before you answer. Seriously, it's like asking a kid what the biggest number they can think of is, and they say “100.” Now it's, “No, no, no, we're already at the trillion level, baby.”
Yes. The club of four.
So wait, Dylan, in this example, the child is Wall Street.
I'm not making any statements that people are children. I just understood.
Yes, go ahead. Sorry.
Dylan may have a habit of sometimes calling his employees and customers children, but putting that aside, I think I'm using a much more offensive word.
Okay, yes. I already like where everything is going.
We're running out of big numbers. There are things to pay attention to when it comes to deployment. This could be anything from industrial equipment to firms willing to do EPC work—that is, engineering design and construction—to chips, cooling systems, and all the components beyond that, including IT infrastructure.
Essentially, there's a lot of work being done to expand the supply chain so it can support all of this capital spending. A lot of this is working, so we're going to see massive growth. But you can't have triple-digit growth and accelerate the second derivative even more than it is now.
Resources are running out. The most important thing right now has 2 aspects. First, it's becoming really difficult for the largest neoclouds and hyperscalers to organize and attract the resources needed to build another gigawatt on top of what they're already trying to build, for example. It also becomes quite difficult for them to raise capital.
While many of these investments already have incredible returns, you still need to find the capital to put into these projects. That is hard to do when you're talking about trillions of dollars.
The other thing is that, by definition, the numbers should be decreasing. This year we have a tenfold growth where revenue went from less than $10 billion to over $100 billion, meaning it grew 10 times. By definition, they simply won't have $1 trillion in revenue by the end of next year.
So yes, we are seeing some slowdown in growth rates. Nvidia's revenue growth is still doubling year over year, but it's not as much as it used to be, simply because it's harder by definition to grow 2 or 3 times from the same base.
This applies to other companies as well. So this is certainly, by definition, a slowdown. Cost growth has to slow down at some point, but revenue and profit growth have to accelerate. This should outpace large-scale infrastructure deployment and perhaps accelerate further than we have seen in terms of spending.
You can imagine what it takes to facilitate all this AI activity. There have been further discussions about this. Dylan, let me just give you these numbers, okay?
This is the Goldman Sachs report. Sorry, Goldman Sachs is talking about Meta—that's a tweet, right? The tweet says, “Meta committed corporate suicide by taking on unsustainable debt that the company will have difficulty repaying.”
Then there's Bain. Bain says hyperscalers need $6 trillion in revenue by 2031 to break even. Columbia University professor Stijn Van Nieuwerburgh also said that America will need to spend about $3.5 trillion annually on AI services by 2032, which is 8.8% of GDP, to justify today's massive investments in data centers.
What do you think about these estimates of the costs and revenues needed to recoup the investment? Do they have a chance of justifying the investment?
2. AI Investment Needs Revenue
I think all estimates of the required income are quite low.
Good. Where is this construction going?
I think everything will grow so that by 2031, income will be significantly higher, for example.
There are a few points here. One of them is the relationship between how much people spend on capital investment and the income that follows. Income always lags behind capital investment.
These infrastructure investments have a 6-year life, at least as depreciable assets for accounting purposes. There's a lot of evidence that they actually last longer, but let's just keep using 6 years.
If I spend $1 trillion today on AI data centers and GPUs, the GPUs will last 6 years and the data centers will last 15. That means, by definition, in terms of income, if I assume a straight line, that in 6 years I need, let's say, $1.5 trillion in revenue.
That's $250 billion in revenue per year on that $1 trillion in capital investment. Or it's going to be some kind of sloping line where the first year is much lower and the last year is much higher to get the payback.
As for the statements about Meta, they're just kind of pointless because Meta could stop buying new capacity at any point, and its cash flow would easily cover everything it has signed up for. They don't necessarily have to build $100 billion data centers next year.
They can stop whenever they want. Obviously, there are contracts they have committed to, but there are also a lot of things they haven't signed on to yet.
As far as Meta’s solvency is concerned, they are fully solvent and doing well. The question becomes: what income do they get on top of that? As for the impact of AI on the U.S. economy, if AI revenue is only $3 trillion in 2031, then we’re going to have a big problem.
I’m inclined to think that we can reach $600–$700 million combined through OpenAI and Anthropic alone by December of next year. I’m of the opinion that it’s possible to get to $1.5 trillion, maybe even by the end of 2027, just with these 2 companies. So by the end of 2027, the payback is obvious, right? You know where they get that from.
Ultimately, the capacities that people build up correspond to what they expect from businesses. So there will always be a mismatch, if we are in the infrastructure development phase, between the costs and revenues of AI. After all, today’s revenue from AI is actually based on the infrastructure invested in during previous years.
Now capital expenditures are not increasing 10 times, but inference revenue has increased 10 times in a year for AI labs. We see that Anthropic is now profitable in terms of revenue minus compute costs, for both training and inference combined. OpenAI will get to that within the next few quarters, probably 2 or 3, at least based on what we expect from the tokenomics model. So I think those companies are already breaking even.
If you look, for example, at Amazon, the largest developer of AI infrastructure, all of its investments in AI are profitable today. Of course, there were customers that could have become insolvent for them, like OpenAI and Anthropic, but now they are solvent and have very high gross margins. In fact, AI boosted Amazon’s gross margin last quarter, thanks to Bedrock and its revenue.
So yes, AI revenue should grow significantly. But I would suggest that the ability to automate everything we’ve been able to do and build everything the world has been able to build will eventually generate substantial revenue and consumer surplus. I think we need to broaden our vision a little bit to understand what we will measure as revenue from AI in the future.
If these companies truly have the best models in the world, capable of doing any intelligent work efficiently, one can imagine them branching out into a bunch of different industries beyond programming and getting a return on investment similar to what they saw in coding. You can look at things like robotics and self-driving cars. You can look at drug development. You can look at healthcare and life sciences in general.
We can apply this to all of these industries and say that if they want to bring a product like a cancer cure or a driverless car to market, you can only imagine the revenue they could potentially generate from their current investments. If they have artificial intelligence—a system that can do the intellectual work required to create a product like this extremely efficiently—I think they are the only ones who can really compete at the forefront of these things.
I think people lack the creativity to figure out how to look at revenue outside of today’s cloud-subscription or token-payment models. There is a graph that shows how much we spend as a share of GDP. This is data for 2025 only. Healthcare accounts for 18% of GDP. That’s category number 1. Food accounts for 9.1%, and AI services account for 8.8%. That’s the third category in 2025, and I think it will overtake food spending in 2026.
3. Labor Funds The AI Boom
Guys, I’m just curious: where does this money come from? Let’s say OpenAI and Anthropic together will have $70 billion in revenue next year. Does this money come from the companies themselves, from payroll funds, or is it just additional funding because they can sell more? Where does the money come from—the customers of these companies—that they direct to AI companies?
There are several factors at play here, right? First, it’s clear that companies will be able to accelerate what they create. Accordingly, they will be able to invest more. An increasing share of R&D spending will go to artificial intelligence, and an increasing share of the cost of goods sold will also go to AI.
But a significant portion of the money will come from the labor force, right? I don’t think the workforce is shrinking as such. I think the rest of the economy is growing so fast, at least in the short term, that the labor share remains relatively stable.
We’ve had different cycles in the economy in the past, but one long-term trend is that the labor share of GDP has been falling for decades. It’s been a steady, slow decline since the 1970s, which is why many complain that the American economy is a mess, even though it’s getting bigger. Labor’s share of the economy has simply shrunk.
A much larger portion of the value of labor is transferred to capital. AI will only exacerbate this long-term trend of division between labor and capital, as AI tokens are essentially just capital—in the form of chips and data centers that work and produce results. So the labor share will decline even faster.
I hope—and this is my belief, perhaps even an irrational one—that we will develop the economy so quickly that the labor share in dollar terms will not decrease and perhaps will even slowly increase. It’s just that the entire rest of the economy is growing very rapidly.
That’s where it all comes from, isn’t it? All these companies that provide various services will improve them, be able to charge more money for them, sell more, and do it with fewer people. Or, to look at it another way, they will have the same number of employees and their businesses will become more efficient.
There are many industries where this is possible. Programming is an area where people are already quite active in using AI. It’s clear that office work is rapidly being automated, but companies can also do much more than before. There are many industries, like pharmaceuticals, trading, and chip development.
Chip design is a great example. The number of American R&D engineers in chip design has remained virtually unchanged for 20 years, but the economics of the sector have exploded. Of course, we don’t count production; we only count developers. But developers are some of the most expensive companies in the world, right? NVIDIA, Broadcom, Google, and all the other companies that make their own chips, along with many others.
The number of these developers hasn’t actually increased much. For the U.S., this is a stable industry, and yet the accumulated value has grown rapidly. Why is that? Many tools have appeared.
So what does AI chip design do now? All the chips in the entire economy are getting better and better. Your phones will get better. Your AirPods are going to get better. The chips in them, the chips for robots, the chips for AI, and the chips for everything in the world—Bluetooth, GPS, and many other things—will get better. Sensors, for example, for gas leaks or something else, manufactured by Bosch, will get better.
All of this will improve as AI helps design chips. You do it with the same number of people. Revenue is growing, margins are growing, and everything looks fantastic, right? This is something that, in my opinion, can be replicated in many industries.
Dylan, you said earlier that it seemed like intellectual work was over. So you don’t think that anymore?
I think most of the intellectual work—okay, let’s divide my belief in the “madness of 2035” from, say, 2028. I think that in the next few years, as long as AI is very powerful but not better than humans at everything, the most active people will be able to do the coolest things. They will be able to have a much greater impact on the world.
Now, if AI just gets better than everyone else, everything could change dramatically, right? Then intellectual labor will truly be over, and what will really matter is that people can play games, have fun, role-play, and do the things that people do and that they only appreciate in other people.
Unless AI kills us all, of course. But we’re not going to go into that today.
4. Contracts Face The Demand Test
Jordan, this is your area of expertise. I want to start by talking about the risks that could ruin this scenario. We’ve already talked about the fact that there are certain contracts that a company like Meta may not be required to fulfill.
You obviously watch a lot of the neocloud business, which depends on whether companies like Meta actually deliver on their promises to work with them. Is it an underestimated risk when things are looking pretty good right now? If you’re talking to one of these neocloud providers, they’ll tell you they have the best customers in the world. They have Meta, Meta will deliver on its promises, and they are raising debt financing on that basis.
But the real risk might be that a company like Meta might just decide, “Yeah, we’re not going to do that,” and that these contracts may not be as binding as neocloud providers assure us.
I want to clarify one thing that Dylan was talking about regarding contract termination and what’s happening with Meta and neoclouds. Generally speaking, these hyperscalers or large frontier labs signing neocloud contracts don’t have the right to cancel arbitrarily every month, so they can’t just stop everything and end the contract.
They usually sign contracts for 5–6 years and buy out the full capacity. They have a set hourly rate and a total contract price, which is very attractive to neocloud providers because these companies have already purchased the land and electricity, built the site, and sometimes even purchased the chips before they signed the contract.
What Dylan is talking about regarding potential slowdowns, suspensions, or cancellations of projects is more about their own sites. Meta has huge campuses that will serve as baseloads in the future; they can either buy chips for them or leave them empty for years if they want to slow down.
Overall, construction costs, while very high, are only a small percentage of the total cost of a new chip project. In other words, concrete, steel, and walls are certainly important, large-scale, and difficult to implement, but in monetary terms, HBM, optics, and the GPUs themselves make up a much larger part of the bill.
The situation right now is that all these big companies are having to deal with the huge growth expected of them because of demand from their customer base. Whether it's Meta, OpenAI, or Anthropic, they have consumer applications like ChatGPT, Claude, Codex, or even Muse by Meta.
To meet this demand, they need GPUs, and it's profitable. They need to build capacity to deploy this to the billions of active users across Meta's various platforms, as well as the hundreds of millions of users of ChatGPT, Claude, and so on.
With that in mind, what is the risk of project cancellations, a recession, or anything like that? For us to predict any shift in the supply-and-demand curve, we have to see a slowdown in demand. There are strict limits on how many chips can be produced, how many data centers can be built, and how quickly it can all be up and running.
At this point, all our models show that demand significantly exceeds supply. Is demand exceeding supply by 10% right now? Maybe 50%? Or is it limitless? We cannot answer this question because we are not able to keep up with demand with the number of GPUs that can be put into operation right now.
When this question is answered—when we find the limits of demand for intellectual labor and artificial intelligence in the economy—then we will understand who needs to slow down and when. But those limits have not yet been found.
5. The Anthropic Valuation Gamble
Before moving on to clusters, I want to discuss one more topic. A big IPO on the horizon, of course, is the Anthropic IPO. Dylan, you had a very interesting tweet about this. I'll just read it. You said, “You can't invest in the Anthropic IPO at a valuation of $2 trillion because the company could go under and I'm out of business, or it could be worth $20 trillion, but at $20 trillion we're all out of business.” Can you explain this and tell me a little more about what you meant?
Just to be clear, this is a complete trash post. I was like, “What was I listening to?”
I understand, but I think there's some truth to what you're saying. So tell us about it.
I am generally extremely optimistic about Anthropic. I think Anthropic is going to be a very successful company in the world. Anthropic may even become the world's first company with a $10 trillion valuation.
I'm so bullish on Anthropic, but at the same time, for it to become a company with a valuation of $10 trillion or $20 trillion—especially $20 trillion—its revenue multiples need to drop significantly. They used to raise money at a 30× multiple, and now, let's say, their IPO will be at $100 billion in revenue, and they might have a 20× multiple.
If they reach $10 trillion, I'm sure they'll need about $1 trillion in revenue. Something crazy, right? That could happen sometime in 2028 or 2029. It depends, of course. The stock market can also go crazy and make multiples skyrocket.
But in a world where they're worth $20 trillion, I don't think their revenue—I don't think they'll even have a 10× sales multiple. They will probably have somewhere around 5× or even less. So if they have a $20 trillion valuation, will they have $2 trillion, $3 trillion, or $4 trillion in revenue?
At that point, what will be the state of AI? If, in a $150 trillion global economy, they generate several trillion dollars in revenue and are by far the largest company in the world, the possibilities of AI will be simply insane.
In that case, all this worry about whether AI is going to kill us all, or whether AI is going to get so good that everyone loses their jobs—and because the distribution of earnings is going to be heavily skewed in favor of Anthropic stockholders or supply-chain constituents, and it's not going to be distributed properly—people are going to become discontented and maybe overthrow governments because all the growth is going to accumulate in a very small number of people.
These things can tear apart the fabric of society. If you believe that Anthropic will reach $20 trillion, society as a whole could be in trouble.
Jordan, do you have similar concerns?
Yes. I think there's still a lot of work to do to get to that point. I don't actually think I'm necessarily as worried as Dylan is. I just painted a pessimistic picture, so I'll go straight to the optimistic one.
The rosy, optimistic view is roughly what I described earlier: if they reach $20 trillion in revenue, it's not achieved through tokenized payments where people use Claude Code as a product. I don't think you're going to see Claude Code become 20 times better than it is now, in terms of quality, and provide that kind of revenue.
I think they'll go into other areas of business, and I would have quite a bit of confidence in Anthropic bringing to market, as I said before, a cancer drug or something like that. We saw similar things happen recently with GLP-1, which became so significant to so many people in the economy. As a result, companies like Lilly made huge profits from this incredible research that led to these incredible drugs that help people.
I would be more optimistic about the future in which Anthropic is theoretically valued at $20 trillion. But I really think there will be many obstacles along the way. Going back to your previous question, how can they even meet all this demand, train these models, and spread the intelligence across all these different parts of the economy so that people can benefit?
For this, we need to increase computing power, and it depends on the technical details of construction, the installation of GPUs, and their reliable operation.
In my opinion, the center of the world on the path from a $2 trillion IPO to a $20 trillion market cap on public markets is the neo-cloud.
Anthropic is quite an interesting business case, isn't it? They spend a lot of money on infrastructure. Even the S-1 showed large losses alongside their revenue. Although, Dylan, as you mentioned, it seems they've managed to turn things around a bit, and now they're earning more than they're spending on computing.
Tell me a little more, because I think there's a perception that a company like Anthropic is a low-margin business, or at least a much lower-margin business than software. But it seems like that's not the case, and you've done the math.
I would say if you look at Salesforce, which is probably one of the best SaaS companies, it has grown steadily and become one of the largest SaaS companies. Its gross margin is around 75%, and Anthropic gets a 75% gross margin on its data output, even a little more, especially on advanced models.
It seems like it took Salesforce a long time to get to this point, right? They didn't have a 75% margin at the time of the IPO. It took them a long time. In fact, in the mid-to-late 2000s, they were not yet at this level.
It took Anthropic significantly less time to reach such high gross-margin figures. If you look at many other SaaS companies, they don't even have such indicators.
One more thing: their customer-acquisition cost is just crazy. Isn't that right? If you look at a SaaS company, the cost of goods sold, or COGS, is very low, but the cost of customer acquisition can be as high as 30%. At Anthropic, the cost of customer acquisition is almost zero.
Sure, they have huge R&D costs, but those costs have been growing more slowly than their revenue. So now they have reached profitability, and there's no reason why this trend shouldn't continue, right? They continue to become more and more profitable, at least from an operating-margin perspective.
Scalability is another point. The great thing about many SaaS companies is that even if the cost of acquiring a customer is very high, as long as your margins are big enough to cover it initially, you don't have to reacquire that customer. Now you have this superprofitable business.
But in the case of AI labs, their business seems a bit more stable. There are all these messages like, “Meta is cutting costs,” or Microsoft is cutting costs, or something like that. But in the end, the revenue graph was just a straight line up, and for almost every client, their revenue was only increasing.
The switching costs, even when some other Chinese model comes out or OpenAI releases a better model, don't really change the cost structure for existing customers. You get a fairly loyal customer who potentially increases revenue at a certain pace, with no customer-acquisition costs and high margins.
So it can be argued that an AI lab is a better business than software, if you can do it successfully.
But how big is the risk that they're now engaged in this emerging movement of “slowing down the frontier”?
I think these are just signals.
Tell me about it. You don't think that's true at all?
I think slowing down the frontier is a responsible step if you believe we can destroy ourselves. But I don't necessarily think that they're going to slow down the frontier too much.
I think they have very little opportunity to slow down, given the race they’re in. People have done sketches about it, but Dario can’t slow down because he doesn’t trust Sam, and Sam isn’t going to slow down. Sam can’t slow down much because it’s the same with Dario. Of course, they’ll slow down a bit when they have these large-scale security incidents. None of them can slow down because China isn’t too far behind. So you could say it’s very difficult.
This is calming.
Okay. Listen, we're going to take a short break now, and when we come back, we'll talk a little bit about the state of AI clouds, of course, about ClusterMax, and then we'll continue with other questions about technology and business. So let's do this right after the break.
We're back on the Big Technology podcast with Dylan Patel and Jordan Nanos from SemiAnalysis. SemiAnalysis, I'm sure you're familiar with. This is the bible for everything related to AI infrastructure. These guys are the best. We are very grateful that you took the time. Both, and you know, I definitely advise you to at least check out their blog. You can also find SemiAnalysis on Substack. This is newsletter.semianalysis.com. This is the number one technology newsletter, and deservedly so. So, congratulations, guys.
6. NVIDIA Shapes The Neocloud Market
There are many areas we can discuss, Jordan and Dylan, but let’s start with NVIDIA’s role. While the report spends most of its time on the technical side, it also highlights the fact that NVIDIA plays a major role in supporting these clusters, these neoclouds—that is, companies that provide AI infrastructure, perhaps to some frontier labs, but generally to make this whole system work. NVIDIA is powering many of them through, as you write, minimum revenue guarantees, landlord guarantees, and leases that it plans to pass on to third parties, helping neoclouds and AI labs get investment-grade funding without involving hyperscalers.
ClusterMAX 3.0 is an interesting read, even if you’re not a techie. I talked to Jordan about it, and he talks about the state of infrastructure development. The subtitle aptly outlines it: all the details are there—reliability, performance, support, pricing, and security of cloud GPU providers.
Jordan, a question for you: There’s a criticism, similar to Ed Zitron’s, that NVIDIA is essentially funding the demand for all these neoclouds, and without that support, it would all just evaporate, NVIDIA would look smaller, and these neoclouds would become unviable. What are your thoughts on this after delving so deeply into the topic?
I’m referring to both ClusterMAX and another article we recently put together called “The NVIDIA Warranty Universe.” We delve a little deeper into the dynamics of what guarantees are and how they help neoclouds raise the necessary funds to deploy GPUs, which customers are currently quite actively renting.
The number that we have is that we’re currently tracking $588 billion in off-balance-sheet guarantees that NVIDIA is currently maintaining, which is another incredible number on the list that we’ve looked at so far in this podcast. NVIDIA’s revenue is approaching $400 billion out of that $500 billion. So we’re talking about really big numbers. For context, that’s about 2.5% of the M2 money supply in the U.S. It’s obvious that there are huge amounts of money circulating in this universe, as we call it.
The only thing I can say about this is that it looks very pragmatic on NVIDIA’s part. They really want to have a wide variety of customers. They don’t want to depend on hyperscalers for all of their revenue. Additionally, they don’t want their customers to be dependent on getting chips from only a select group of 3 or 4 different hyperscalers. Therefore, they have every incentive to create a healthy market with a variety of options and a wide selection for all customers to work with.
In order to make this happen, people have more problems than just raising the funds needed to buy GPUs. First of all, they need a guarantee that if they raise money, the lenders will make a profit, and that requires some sort of investment buyer, which NVIDIA can act as in many of these cases. Right now, this is how it works: NVIDIA signs up as such a buyer, and when it comes time to take those GPUs for its own internal research, someone else comes along who wants them more, at a higher price, so NVIDIA sells them to that party. They don’t take them for their internal research and move on to the next option.
To anyone who says that NVIDIA is doing this for some nefarious purpose or to support some kind of “bubble,” I advise you to look at the quality of NVIDIA’s research. They post more models on Hugging Face than anyone else. They have incredible continuous integration, or CI, for the entire suite of software they support on their GPUs, and they train models like NeMo Triton. We’re not the biggest fans of these models or this software—we think the ecosystem should be diverse. But the idea that NVIDIA doesn’t use its own GPUs for research that serves the company’s goals is itself false.
The dynamic that we’ve seen is that this compute volume is not really a significant part of NVIDIA’s revenue. This is a very small percentage of NVIDIA’s total revenue, which goes toward its own internal research workloads. So all we see is that they’re supporting the capacity building of neoclouds to meet customer demand, which is currently incredibly high and difficult to service.
The last thing to say about this is that it’s not just about the chips. We track a whole wide range of things that they provide. This includes everything from construction costs and memory suppliers to people building factories and those creating industrial equipment. It’s not just, “I need help raising funds for my GPUs.” It’s, “I need help raising funds for everything else in the supply chain,” which is the main problem for a particular project.
Let’s talk about safety, because the first image in this report is a quote from a post by Ilya Sutskever made right after the Hugging Face incident. Ilya says: “Neoclouds have limited cybersecurity. The next time the agents become uncontrollable, they will try to capture the neocloud to launch more copies. This is bad. So neoclouds need to significantly strengthen their cybersecurity, and every company with strong cyber models should help with that.”
This is a rare tweet from Ilya. Dylan, what do you think about this? It sounds like he’s talking about it openly, right? Remember, he creates these things himself.
I would say that if it’s superintelligence, then we might have problems. That’s why we published an article called “Most Neoclouds Are Terrible at Security.”
By the way, the illustration for this is simply incredible. We’ll put it on the screen for those who are watching. I think it’s a beautiful image.
Yes, exactly. Everyone is stabbing him with knives, and he supposedly meets the requirements of SOC 2. We published that report, and then, 8 hours later or something, Ilya wrote his tweet, because they’re very actively using different clouds.
They’re obviously very secretive, and they care a lot about the confidentiality of what they’re working on, but they also believe in superintelligence. I mean, SSI—Safe Superintelligence—they were the first, and Trump actually copied Ilya. Ilya was first again, right? First in reasoning, first in scaling large language models, and now first in branding superintelligence.
Ilya tweeted this once, and that kind of responded to what we wrote, right? Namely: Listen, most of these clouds are terrible when it comes to security. You can use open-source models to simply hack neoclouds or access data inside other neoclouds. GLM 5.3 in particular does a really good job of this. So Chinese open-source models can just hack dozens of existing neoclouds, and you can look at what other people are doing.
We’ve seen similar things in other clouds. There was a cloud where we didn’t even intend to, but happened to see all the other customers in that cluster. We saw what data they were storing, what was in Slurm, and so on. And this is madness.
Secret services, right?
For example, that’s why we immediately shut everything down, let them know, and said, “Hey.”
What secret services?
We can’t say. I don’t think it’s appropriate to voice it. National—federal security agencies of governments. These were not Americans. These weren’t Americans, but we said it was a country with a top-10 GDP. This is not some tiny country.
I’m just shocked that this was possible.
Yeah, and it’s like if we were the intruders—and, by the way, there were probably a dozen other clients there, and maybe one of them was the intruder—we could just see what they were training.
Yes?
They probably trained models and ran models for themselves. Because it was GPU infrastructure. So what else can you do with a bunch of Blackwell GPUs?
We often use this term—sovereign artificial intelligence—and it’s often intertwined with the concept of neoclouds. People, even when raising funds, talk about it a lot because they view computing power as infrastructure, and I think that’s absolutely right. This is similar to the development of telecommunications, where this infrastructure is perceived by the sovereign states that deploy it and participate in the process as an asset that they want to control and have sovereignty over.
They believe, similar to what you said at the beginning of the program, that they want to be net importers or exporters of tokens, and it depends on the infrastructure that they build, which is data centers with GPUs. So I think it’s not surprising that federal government research and intelligence agencies use some of these clusters, in conjunction with others.
You would like them to have an isolated data center, but if there is only 1 facility in the entire country with GB300 NVL72, then it’s a seller’s market.
Oh my God. I can’t sleep tonight. So, in your report, you also rank a lot of these hyperscalers by tier, as well as the neoclouds, and at the highest tier you have CoreWeave and Nebius. I’m more interested in the second tier, which is Oracle and Google Cloud.
Mostly, Jordan, because in the document that I prepared for our conversation today, I have a lot of themes and statements that I pulled from the report, along with questions. For Oracle, I just have “Oracle” written on it and a bunch of question marks, because that’s how I feel about this company. I don’t know what to think about them.
It’s obvious that their debt has a low rating from the rating agencies. They bet it all. They have quite high stakes, and in your report they are even recognized as pioneers in certain technologies. But I would like to know your thoughts on Oracle, whether it was a smart bet for them, and whether they will succeed or perhaps fail miserably, as many expect.
I think the main thing about ClusterMAX is that it’s not investment advice, right? It’s not about saying this new cloud provider will build a gigawatt next year or 200 megawatts. That’s much more important for stocks than, say, how well they manage the cluster, right?
This has nothing to do with the fact that a data center they said was supposed to be delivered with such-and-such capacity was delayed by 3 months. This is not what ClusterMAX evaluates. ClusterMAX openly measures how user-friendly a cluster is if you rent a GPU, how reliable it is, how secure it is, and how the storage, network, lifecycle, and Slurm on Kubernetes work.
In this case, we try to be more objective. Of course, creating a rating pyramid is what you have to do. It would be nice to just do a long report and have no one care about the lack of ratings. Actually, that’s how it would be. So ratings have to be made, certain boundaries have to be set, and you can always argue endlessly about why something ended up in one segment or another.
But this is an objective fact: We believe that Oracle is one of the best new cloud providers in terms of service quality. On the other hand, that doesn’t mean, “Hey, we at SemiAnalysis told customers in April that their data center in New Mexico was going to be delayed because they weren't getting the pipelines.”
This was a big problem for OpenAI, as they had contracted for a huge data center in New Mexico. But if the pipelines are not built, the data center will not be able to operate. Later, we repeated it in May and June, and just recently Oracle declared force majeure, and its shares fell because of it. Yes?
So we’re separating 2 things, right? Yes, Oracle is an amazing cloud for managed clusters—one of the best new cloud services. But at the same time, they may have problems executing their plans, whether it’s regulatory, like with the data center in New Mexico, where they can't lay the pipeline, or other problems, like their balance sheet.
So we’re trying to distinguish between what’s good or bad for stocks and what we’re trying to measure here. The goal of creating ClusterMAX was to provide a tool for those who rent GPUs. But it turns out that now there are many other uses that people are trying to pull it into, like, “Hey, we’re going to issue debt for this neo-cloud.” If their rating increases from bronze to silver, their rate will actually decrease.
Things like that, or, “Hey, I’m going to invest in stocks that are on top of ClusterMAX.” Of course, these are not the solutions. This is not what ClusterMAX was designed for. This is just a data point to use if you’re dealing with these companies.
But in terms of Oracle as a company, there are 2 points here, right? First, they signed a bunch of AI data center deals with OpenAI, I think last year?
Yes, that was last year. Oh my God, that was only last year. It seems like an eternity has passed.
Yes. I think it was exactly a year ago. But the interesting thing is that they signed these huge deals, and there are a number of risks associated with them.
First, they signed them essentially with a fixed margin. So now that GPU prices for cloud services have increased significantly, they can’t take advantage of that higher margin. They seem to be trapped.
Secondly, they raised a lot of debt, and the market started to worry about the companies’ debt levels, so investors got scared, even though Oracle is completely solvent. Thirdly, they have some implementation issues. I mentioned regulatory issues, but there are other data centers that are also experiencing delays, such as in Wisconsin.
The main problem is the facility in New Mexico, which could be delayed for years due to pipeline issues. While these things can have a very negative impact on the stock, I think the company is still doing okay.
Of course, this is not Nebius. Nebius can adapt quickly, and they had the shortest average contract duration among the top cloud providers in ClusterMAX. So Nebius was able to quickly orient itself, significantly increase prices, and take advantage of this situation. Isn’t that right?
While CoreWeave and Oracle also have high-quality cluster management, reliability, and security, they were unable to gain the same benefits because they were mostly tied to long-term contracts.
Oh, that’s exciting. Okay, Jordan, let me ask you a question about technology, then. You know, as Dylan pointed out, it’s a pyramid. There are top groups like CoreWeave, Nebius, Oracle, and Google Cloud, but the “not recommended” section is full of logos.
There are those who can’t cope and those who are unavailable. It seems to me that there are even more logos in the “unavailable” category than in the first 4 levels combined. What do you think it says about technology that so many companies don’t meet your standards?
Look, as people trying to represent the interests of buyers, we’re just upset about this. It’s sad to see that it’s been almost 1.5 years, and actually almost 2, since the first version of ClusterMAX was released.
And since we at SemiAnalysis introduced the term “neo-cloud,” many have never progressed to what we wrote about. We wrote exactly about what the customers of these neo-clouds with whom we communicate want. So I think the dynamics of a seller’s market mean that the priority for these companies is how to raise as much money as possible, how to build data centers as quickly as possible, and how to deploy chips as quickly as possible, instead of working on the real quality of the basic service.
It’s the same dynamic: People can have a great business without worrying about providing quality services that customers truly appreciate and recommend to others. Another thing is the constant dynamic that we will see over time: There can’t be many leaders, because there are only a few true innovators, and everyone else simply copies them.
When you have a market where, according to our calculations as of this morning, there are 323 providers—which we track in our cloud TCO model for SemiAnalysis subscribers—they have this complete list. Only a certain number of them are relevant, large enough to make the list, and only a few of them meet the criteria we consider necessary for a recommendation.
We think the threshold for recommendation is pretty low. For example, the basic things that fail in terms of security and reliability, which are the 2 main aspects, are, in our opinion, extremely simple. It’s sad to see that others don’t pay enough attention or care enough to implement them.
Okay, 2 quick questions before we wrap up. This is a broader question. Google obviously announced Gemini Argon, Gemini for Argon. Do you think that, given there were people calling for Google to essentially become a neo-cloud and lease TPUs from Google Cloud for compute power, you see them going down this path? Or do you think they will still use a large percentage of their compute for model training and inference?
I mean, they have by definition allocated less computing power to training and deploying models in the last 6 months than they did 2 or 3 years ago. This percentage of distribution has decreased over time.
This is one of the things we joked about—or maybe we weren’t even joking, but just stating—that Thomas Kurian and Demis Hassabis lost, right? Well, now probably Pichai. Google is producing significantly more computing power year after year, but the amount going to DeepMind is not growing as fast.
It’s not like, “Oh, DeepMind had a third of Google’s computing power, and now it still has a third.” In fact, this percentage is falling, which is very interesting, because right now, if you compare just DeepMind with Anthropic and OpenAI, DeepMind has the least computing power among the 3. And by the end of this year… So this is actually a very important matter.
So Google—will there be a neo-cloud? What does the term “neo-cloud” even mean? Is selling tokens of other models considered a neo-cloud? Isn’t this, in a sense, what the new cloud business is—selling AI model tokens, right? So maybe this is a neo-cloud. The definition is quite vague.
Buying chips and new construction projects to build out your GPUs and deploy them directly to customers outside of your main platform? Or selling chips, right? Is this a cloud?
No, I mean, they sell a lot of Anthropic chips. That is, this is a physical sale of chips, not just their rental. So I would say that Google’s business model is evolving, and they want to make money at every level of the stack.
But by definition, having such significant help from Anthropic with TPUs and other things hurts Google’s DeepMind business.
They need to find their niche. For Gemini, nobody needs the third-, fourth-, or fifth-best coding model. So they need to find a niche that will drive Gemini token sales if they want DeepMind to continue competing at the forefront of AI research. Right now, even with this Gemini 4 Argon release, it doesn't make it into the top 2 or 3 in coding.
The risk is what happens next year when OpenAI and Anthropic try to add 5–10 gigawatts of additional computing power to their labs, and Google adds only a few gigawatts. Now their computational deficit is growing. Previously, Google was way ahead in computing, and the others were somewhere behind or on par—mostly a little behind. Now they're even further behind, even though their computing power has remained the same.
So what happens when they have less computing power? Do they somehow magically catch up? It's possible. They have the talent and the people; they need to reorganize and really focus. But more and more of the best people are being lured to Anthropic and OpenAI, and that's kind of the goal.
A loop of discoveries.
A loop of discovery, yes.
A loop of discovery, for sure. Okay, I want to ask one more thing before we finish, if I may.
Dylan, I think you said before that the people who get the most value from the AI models that are created are the end users—people who use Claude, for example, at a basic level. Anthropic makes this model available via API for all these various purposes. They're going to have to do more licensing, and who knows if they'll be able to create products for all of this.
I'm wondering if you see a scenario, given the big numbers we've talked about in the past, where they keep their advanced models in-house for longer—much longer—and use them to build their own products. We know they have a biology lab, Claude Code, and Cowork. They don't make them available because they want to make more profit from the infrastructure they bought and built.
I think we've seen this before. Anthropic had the Mythos model ready back in February. Their model safety cards state that they released 2 more models after Mythos, but publicly, we still don't have the full version of Mythos. We have a limited-edition version of it called Fable, and it was many, many months before Fable even came out.
They had Glasswing, which was similar to Mythos, but only for some subsectors, and it was limited in some areas—not in all areas. So it's clear that Anthropic isn't releasing any new models.
OpenAI is the same. The Astra model was ready a few months ago, and it took several months before they released it. With their new Bell model, they even showed a blog post where they showed off its efficiency in math and how it solved all these problems, and they didn't release it. So obviously, they only use it internally, not externally.
Labs are already moving in that direction. It's like, “Oh, yeah, the gap between them and Chinese open source is still 6 months.” They're keeping it, but in reality, they've left months of progress for internal use only, because the moment you make it external, people can distill the model and do all that other stuff.
I think Anthropic will likely release a new model in the coming weeks, before its IPO. I think OpenAI will probably have to release a new model in the coming weeks, just because of this competitive environment. But at the same time, they keep a lot of things for internal use only.
Yes, we've discussed internal capitalist incentives. Will these incentives motivate people to release more models and provide them to consumers because of the incredible gross profits from API pricing per token? Or is the incentive to keep them internal and engage in end-user applications that have an even higher value than selling access to the model for a token?
It will be interesting to see how it all plays out. In one of these scenarios, the capitalist approach would be to set the pace of development yourself. How do you feel about where all this is heading? Which of these scenarios do you think will come true?
I think Anthropic has publicly stated that they are slowing down by limiting themselves, and that is the only way their blog posts are becoming a reality.
No one slows down anymore.
They only slow themselves down.
I think that's true.
I think that's exactly what's happening. And I think incentives for improving the state of AI by solving problems—well, let's say cancer, which is clearly beneficial and something that everyone can agree on—would be a positive incentive. I also think you can make a lot of money doing something like that.
So I'm betting that we'll continue to see minimally viable models from OpenAI and Anthropic to maintain an edge over Chinese developments and the rest of the open-source ecosystem.
The website semi-analysis.com, subscribe to the blog, like Dylan says, you know, subscribe to the newsletter, like I did. It's a source of deep knowledge and analysis, and I think you'll enjoy it if you're not there yet. And this conversation was just, you know, a manifestation of that.
Dylan, it's good to see you again. Jordan, it's very nice to meet you. You are always welcome, and we appreciate your time today.
Thanks for the invite, Alex.
Thank you, Alex. Okay, great material. Thanks to everyone who listened and watched, see you next time on the Big Technology podcast. Mhm.