Most of my Sunday weekend video focused on compute. It was a big week of compute-related news, and I reiterated my view that there is, and will continue to be, insatiable demand for compute as AI capabilities expand at an exponential pace, while supply, constrained by physics and human execution friction, moves at a far more linear pace.
I recorded my video on Friday and Saturday morning, I listened to the All-In Podcast, and the episode covered this same point from multiple angles. It is worth listening to as signal in a world of constant AI noise. I seldom write papers around a single podcast episode, but this one was Gavin Baker and David Sacks-heavy, and they attacked the compute debate in a way I rarely hear despite spending hours each week listening to AI podcasts. It was broad and surgical.
That is why I decided this week’s paper should focus on how one podcast can frame the debate so effectively, and why podcasts have become such an important source of information in today’s rapidly accelerating AI world. There is simply too much happening each week for any one person to absorb individually, let alone connect all the dots in real time.
Here is my summary. Go listen for yourself.
The Argument in One Sentence
David Sacks put both sides of the compute market into a single thought I say regularly. Demand for tokens, he said, “is just going to keep growing exponentially,” and the live question is “whether they can physically meet that demand.”
That is the entire investment problem I have written about through the framework of long scarcity, short abundance. Demand is exponential driven by AI agents, built on bits, software, code, and relentless competition. They surprised everyone with their sudden rise this year. Supply is built by humans and from atoms, scarce physical resources constrained by real-world friction.
Demand is now agent driven and compounds at the speed of software. Supply compounds at the speed of turbine manufacturing, power generation, data-center construction, financing, regulation and human execution. The two curves run on different clocks.
The mistake in simply calling AI a bubble is that it begins with the size of the capital spending rather than with the economic relationship between demand and supply.
The dot-com analogy only becomes useful if AI infrastructure begins to resemble dark fiber: capacity is built faster than economically valuable demand can absorb it. Gavin Baker explicitly identifies that as the risk. But the evidence discussed in this podcast points in the opposite direction today.
Demand is accelerating. Existing compute is being economically consumed. Older hardware remains profitable. The majority of tokens are profitable across the chain. And physical constraints are limiting how quickly new supply can arrive.
An AI overbuild is possible, and I think highly likely at some point in the future. However, the facts today suggest that future is still far away. Ironically, human friction is one of the reasons why.
Historical overbuilds occurred when both demand and supply were driven by humans. Acknowledging that line is important when comparing this time to past supply bubbles. This time is different. Demand growth is driven by agents and software, while supply is still constrained by humans, physics, power, permitting, turbines, data centers, and execution bottlenecks.
Agent-driven demand versus human-driven supply is not a fair fight.
Start With the Bubble Case
A useful investment process should begin by steel-manning the other side.
Baker does exactly that:
“The biggest risk is to me is not on the demand side. The biggest risk is that you get a glut of compute and you get an overbuild.”
He even uses the historical analogy directly:
“In the same way that we had dark fiber after the dot-com crash, if you had dark GPUs, that’d be a disaster for everyone.”
That gives us the proper test.
The question is not whether AI capital spending looks enormous. It obviously does.
The question is whether supply is beginning to outrun economically profitable demand.
If this were already a dark-fiber-style overbuild, we should expect some combination of idle capacity, falling utilization, uneconomic token pricing, rapidly obsolete hardware and infrastructure owners struggling to earn returns.
The transcript instead gives us evidence of nearly the reverse.
That is where the learning journey should begin.
Demand Is Compounding, Not Cycling
Anthropic provides the clearest real-time demand signal.
The company is discussed as potentially ending 2026 at a $100 billion to $120 billion annualized revenue run rate, after growing approximately 10x year over year for three consecutive years.
Sacks runs the extrapolation himself:
“If that rate of growth were to continue, they’d hit a trillion dollars of ARR by the end of next year.”
The remarkable point is that Anthropic may already be losing market share.
Baker says:
“Probably on the margin, Anthropic is losing share to OpenAI, to open source, and to Grok. And they’re still growing so fast.”
That observation matters far more than who currently has the best model.
Anthropic does not need to preserve market share for aggregate compute demand to rise. If Anthropic loses share while still growing rapidly, and OpenAI, Grok and open source are simultaneously expanding, then the real phenomenon is category expansion.
That is Jevons Paradox.
As Baker says:
“The pie is getting ginormous.”
OpenAI reinforces the point. Sacks says its growth rate has even reaccelerated to more than 20% month over month, which if sustained for twelve months would approximate another 10x annual growth rate.
Three competing ecosystems can gain simultaneously if the total market demand is expanding much faster than share is shifting between them.
This is the first major break from the bubble framing.
The bear case implicitly treats AI demand as a relatively fixed pool that everyone is fighting over.
The transcript instead describes a pool that is still expanding rapidly.
The End Market Is Much Larger Than Today’s AI Revenue
Baker then provides the broad economic context.
“Depending on how you count it, there’s, you know, 25 to, you know, 65 trillion in knowledge work.”
He adds that the head of the AI institute at one of the three largest investment banks told him that his $25 trillion estimate was “way low.”
And this is before robotics.
The more important question is what AI does to economic activity.
If AI simply replaces existing labor, the addressable market eventually becomes constrained by the value of the labor displaced.
But Baker says the early evidence points somewhere more interesting:
“Thus far, it really does look like it’s accelerating growth.”
He notes that there are more software-coding jobs open than there were a year earlier despite coding sitting near the center of the AI impact zone.
That matters because an accelerating-growth world consumes much more compute than a simple labor-substitution world.
Jason Calacanis reaches a similar conclusion from the bottom up. With roughly 150 to 160 million U.S. workers and approximately $10 trillion to $12 trillion of aggregate salaries, he argues that corporations could plausibly spend 5% to 10% of compensation on AI tools and tokens that increase worker productivity.
His framing:
“There is no world in which I don’t see corporations and people in the economy spending five or 10% of the salaries of their employees on the equivalent in tokens.”
At 5%, the number is already hundreds of billions of dollars.
At 10%, it approaches $1 trillion in the United States alone.
The precise estimate matters less than the implication.
Today’s AI revenue does not appear obviously large relative to the pool of economic activity AI is attempting to augment.
Cheap Intelligence May Increase Compute Demand
One of the more sophisticated bear arguments is that open source and model efficiency eventually solve the compute shortage.
The podcast acknowledges that open-source models can be dramatically cheaper than frontier models. Calacanis cites GLM-class models at roughly a 90% discount to Claude Opus and argues that corporate America will increasingly adopt open-source systems.
This is one of the loudest bear AI arguments today.
Baker argues the opposite.
His key point is that frontier intelligence and cheaper intelligence can become complements rather than substitutes.
He offers the Manhattan Project analogy: Oppenheimer and a small group of exceptional physicists still required thousands of highly capable people beneath them to execute the project.
The same architecture could emerge in AI.
A frontier model capable of orchestrating many cheaper models becomes more useful because cheaper intelligence exists underneath it.
Baker describes a plausible equilibrium in which:
Frontier tokens represent “65 to 85% of the economic value”
while:
open-source tokens represent roughly 80% of the volume.
Both consume compute.
The implication is important.
If the cost of intelligence falls dramatically but the quantity consumed rises even faster, lower unit costs expand the number of tasks worth computing.
The podcast also discusses inference becoming roughly five to ten times more efficient every twelve to eighteen months.
That should lower the cost of producing intelligence.
But in an elastic market, lower unit costs can stimulate much greater consumption.
Efficiency can become the mechanism through which AI spreads into vastly more use cases.
The Demand Is Paid For
The strongest rebuttal to the idea that this is merely speculative infrastructure comes from the fact around broad ecosystem economics.
Baker directly challenges investors who assume AI tokens are being subsidized.
He says:
“Anthropic is generating cash.”
He continues:
“Open source tokens are profitable. Anthropic is profitable. OpenAI, if they’re not generating cash, they will be imminently.”
Then he makes the broader claim:
“The overwhelming majority of tokens are profitable for everyone in the chain. Everyone.”
That distinction is fundamental.
Subsidized demand can disappear when financing conditions tighten.
Profitable demand is much more durable.
Sacks then explains how those economics travel through the infrastructure stack.
Anthropic monetizes the compute it purchases. That enables the compute provider to buy Nvidia GPUs. Nvidia can then pay TSMC, Micron, SK Hynix and the rest of its suppliers.
As Sacks puts it:
“It’s the entire food chain.”
This is important because one of the common bubble arguments is that AI spending is circular: capital is effectively moving between related technology companies without sufficient end demand underneath it.
The argument made here is different.
There is an end customer paying for intelligence.
That revenue supports compute rental economics.
Those economics support GPU purchases.
Those GPU purchases support the semiconductor supply chain.
The chain can still become overextended.
But their claim is that today’s AI buildout is supported by cash generation rather than expectations. That is why it is so important to monitor every earnings transcript across my agentic infrastructure thematic portfolio.
AI factories are converting atoms into intelligence. Power, land, chips, cooling, labor, permitting, and capital are being transformed into tokens, agents, software, productivity, and new economic output.
That makes AI larger than a normal innovation cycle. It is intelligence becoming an input into everything, and once intelligence becomes an economic input, it touches the entire economy.
Supply Is Atoms
Now move to the other side of the equation.
Baker’s description of building AI infrastructure is deliberately physical.
Thousands of workers must be coordinated in remote locations, sometimes in extreme heat. Turbines have to be manufactured. Power generation has to be built. Sites need to be developed and connected.
Calacanis summarizes the distinction:
“It’s atoms, not bits.”
That may be the simplest way to understand the current mismatch.
Demand can appear because a new model becomes useful overnight.
Supply cannot.
The turbine discussion illustrates the problem. Baker describes key turbine components being manufactured in only a handful of facilities and notes that those facilities are moving to 24-hour shifts.
Demand has become strong enough that old aircraft engines are reportedly being removed from private jets and repurposed to power data centers.
Caterpillar, Cummins, GE Vernova and Siemens Energy are all described as expanding capacity.
That is evidence of an industrial system racing to respond before supply has caught demand.
As Baker says:
“It’s really, really hard.”
Sacks makes the near-term energy constraint even clearer:
“The green stuff can’t scale in the short term. It’s all gas powered for the new data centers. That’s the only way to kind of get the amount of scale that you need in the next year or two.”
Later Baker describes the situation more starkly:
“We’re running out of energy.”
And:
“We’re running into physics and, like, planetary scale limitations.”
This is the core asymmetry.
Software demand can compound exponentially.
Physical capacity cannot.
Human Friction May Prevent the Overbuild
The infrastructure constraints are not purely mechanical.
They are political and social as well.
Texas is discussed as implementing energy audits around large new data-center loads. The speakers describe public concerns over electricity, water, pollution and local infrastructure.
Whether every concern is economically justified is secondary to the investment framework.
Friction still delays supply.
And Baker makes an important inversion.
The same obstacles that make individual projects harder may reduce the probability of industry-wide overcapacity.
Discussing dark-GPU risk, he says:
“All the political headwinds I think insure against that outcome because it is so hard to build data centers.”
And then:
“Those political headwinds I think will almost guarantee that there’s not an oversupply relative to the exponentially growing demand.”
Normally regulation is simply modeled as a cost.
Inside a supply shortage it also functions as a governor.
If every dollar of capital could instantly become a gigawatt of compute, the probability of oversupply would be much higher.
But it cannot.
That delay gives demand time to continue compounding while the physical world catches up.
Capital Was the Soft Constraint and It Is Being Solved
One constraint is becoming less binding: financing.
The podcast discusses Nvidia partnering with Goldman Sachs, BlackRock and other major financial institutions around a structure intended to facilitate up to $500 billion of AI compute financing.
Baker explains why:
“The numbers are getting so big that the TAM is getting constrained by the ability to finance this buildout.”
That sentence deserves attention.
He is not saying the addressable market is being constrained by a shortage of customers.
He is saying the buildout is becoming too large to finance through traditional corporate balance sheets alone.
His interpretation of Jensen Huang’s move is straightforward:
“What he’s doing is alleviating that finance constraint so that he can grow as big as the TAM actually is.”
The podcast compares the emerging structure to aircraft financing.
An aircraft can be financed not solely against the airline’s credit but against the residual economic value of the aircraft itself.
GPUs may increasingly be treated similarly: productive assets capable of generating rental income and retaining residual value.
That matters because it lowers the cost of capital and expands the pool of investors able to fund infrastructure.
The scale makes the need obvious.
Elon Musk is described as wanting to add roughly 6 to 8 gigawatts in the following year, which the speakers estimate could require $300 billion to $400 billion of capex.
Capital markets can adapt.
Turbine manufacturing, grid connections and physical construction cannot adapt as quickly.
Money is becoming solvable.
Atoms remain hard.
Old GPUs Are Not Going Dark
One of the most useful pieces of evidence in the entire discussion comes from CoreWeave.
Baker says:
“CoreWeave said that they are renting Ampères at economically profitable rates today for 2029. So an Ampere that came out in 2020 is going to have a nine-year life.”
This matters for two reasons.
First, it challenges the idea that rapid Nvidia product cycles automatically make older GPUs economically obsolete.
Second, it provides a direct test of the overbuild thesis.
In a true overbuild, old capacity should be the first thing to go dark. Instead, six-year-old Nvidia silicon is still being contracted profitably years into the future.
That is not what excess capacity normally looks like.
The nine-year economic life also changes the supply equation.
New GPU generations do not necessarily replace old generations one-for-one. Older vintages can remain economically productive while newer capacity is added on top.
That makes the installed compute base more durable but it also means today’s apparent demand is strong enough to absorb both old and new silicon simultaneously.
The Bulls Take the Under and Supply Is Why
This may be the most revealing exchange in the podcast because it shows stated conviction colliding with physical reality.
Anthropic could exit 2026 near $100 billion to $120 billion of ARR.
A simple continuation of its historical growth rate could produce something approaching $1 trillion the following year.
Neither Baker nor Sacks forecasts that.
After being pressed, both settle nearer $400 billion to $500 billion.
Why?
The reason is physical supply.
Sacks says:
“I do think that somehow you get into physical limitations at that point.”
Elsewhere he says:
“You start to get into physical constraints.”
This is important revealed preference.
Two of the strongest AI bulls in the discussion take the under on the mathematical demand extrapolation because they do not believe the physical infrastructure can support it. That says everything you need to hear about the exponential demand vs linear supply mismatch.
They are effectively haircutting potential revenue because supply cannot scale fast enough.
That is one of the cleanest definitions of an undersupplied market implied through the Anthropic ARR forecast.
Anthropic Becomes the Pace Car
A good thesis needs a falsification mechanism.
The transcript provides one.
Sacks argues that Anthropic’s eventual public quarterly reporting may become:
“Probably the most important signal that the entire industry has.”
Anthropic becomes the pace car.
If Anthropic continues converting compute into rapidly growing, profitable revenue, the infrastructure buildout retains economic justification.
If its demand abruptly weakens, the impact propagates through the chain.
Baker is explicit:
“If he’s slamming on the brakes because there’s not demand, there will be a pileup.”
That distinction matters.
If Anthropic slows because Grok, OpenAI or open source takes its market share, the compute thesis can remain intact. Demand has moved elsewhere.
The dangerous signal is aggregate demand destruction.
That is what they say investors should monitor and I agree.
How Long Does the Imbalance Last?
Four forces determine the duration.
Lead Times
The supply system cannot instantaneously respond to price signals. Turbine manufacturing, power infrastructure and data-center construction operate on industrial timelines.
The transcript’s explicit discussion focuses on a shortage extending through the next year and potentially the year after.
Political Friction
Regulation, grid concerns and local resistance limit the rate at which projects can be built.
Baker’s inversion matters: those headwinds may actually reduce the probability of oversupply because they prevent the industry from responding too aggressively to demand.
Long-Lived Hardware
The CoreWeave example suggests 2020-era Ampere GPUs can remain economically productive into 2029.
If older hardware stays utilized, new hardware is being added into a market that continues to absorb prior generations rather than merely replacing them.
Demand Is the Only Fast Breaker
The most important downside risk is aggregate demand disappointment rather than one model company losing leadership.
That is the scenario in which Baker’s dark-GPU analogy becomes relevant.
Probabilistic Assessment
The transcript gives us evidence rather than certainty.
It allows us to assign probabilities based on the evidence presented so I asked the AI models to give me probabilities based on the arguments in the podcast.
Outcome | Probability |
|---|---|
Compute remains supply-constrained through year-end 2027 | 75% |
Compute remains meaningfully supply-constrained through 2028 | 45–50% |
Supply catches demand without a major bust | 20–25% |
A genuine overbuild emerges as demand weakens and idle GPUs appear | 10–15% |
Thinking probabilistically about insatiable compute demand will be critical to investing success over the next few years. Upload this transcript into your favorite LLM, then upload almost any bear podcast alongside it, and the contrast is obvious. This discussion is balanced, fact-based, and grounded in what is happening today. Most bear arguments are telling a story about the future focusing on one argument while ignoring the broad facts on the ground.
The mistake is treating AI as a binary story.
It is not.
AI will change our lives. The real question is how fast demand grows, how quickly supply can respond, where the bottlenecks emerge, and when the market eventually moves from scarcity to overbuild.
The important point is the asymmetry.
For the bubble case to become dominant, we need to see new evidence:
- Anthropic and its peers slowing because aggregate token demand weakens rather than because market share shifts.
- GPU utilization deteriorating.
- Compute pricing collapsing.
- Older hardware failing to clear economically.
- Token economics becoming persistently unprofitable.
- The physical supply system adding capacity faster than demand can absorb it.
Today, the transcript describes the opposite conditions.
Signposts to Monitor
The most useful investment signals from the podcast are therefore straightforward.
Anthropic’s S-1 and quarterly financials.
Sacks believes these could become the pace car for the industry. Revenue growth, profitability and compute intensity will tell us whether demand continues justifying infrastructure spending.
Compute pricing.
The discussion references spot economics as high as roughly $30 to $50 per watt. A sustained collapse would be a meaningful warning that scarcity is easing.
Utilization of older GPUs.
The CoreWeave Ampere example is particularly powerful. The first evidence that older GPUs can no longer be rented economically would tell us supply is beginning to outrun demand.
The downstream profit chain.
Baker’s assertion that nearly everyone in the chain is profitable should continue to be tested. If profitability begins disappearing at multiple levels simultaneously, the probability of an overbuild rises sharply.
Conclusion
The dot-com analogy is attractive because the capital spending is enormous and the future being discounted is equally enormous.
But capital intensity by itself does not define a bubble.
The relevant question is whether supply is outrunning economically valuable demand.
The evidence in this discussion suggests that supply is not outrunning demand today.
Anthropic is growing rapidly even while losing share.
OpenAI is reaccelerating.
Open source lowers the price of intelligence and potentially expands the quantity consumed.
The majority of tokens are described as profitable.
The economic value flows through the entire AI supply chain.
Six-year-old Nvidia GPUs continue to clear at profitable rates years into the future.
And when Sacks and Baker are asked to extrapolate demand, they cut their own numbers because physical constraints limit revenue even when customers are likely to exist.
That is the central distinction.
Demand for intelligence is governed increasingly by software economics, model capability and falling unit costs.
Supply is governed by turbines, natural gas, grid interconnections, construction crews, permitting processes and human institutions.
One curve can compound exponentially.
The other cannot.
Sacks summarizes the entire investment framework:
“The demand for tokens is just going to keep growing exponentially.”
The real question is:
“Whether they can physically meet that demand.”
Until the pace car says otherwise, the higher-probability conclusion is not that AI has already overbuilt compute.
It is that compute remains short because demand is moving at AI speed while supply is still moving at human speed.
And that difference in clock speed is where the investment opportunity remains.
