● AI Bubble Fears Ignite as Demand Risk Grows
Anthropic IPO and the AI Investment Cycle: The Real Meaning of “AI turns $1 into as much as $180”
The core point of this issue is not simply whether Anthropic’s listing will go well.
The really important point is whether the AI industry is actually making money, whether AI infrastructure investment is a bubble, and how long data center and GPU demand can keep going.
In particular, this All-In podcast broke down the AI value chain from Nvidia to cloud infrastructure, Anthropic, and end customers in numbers, directly addressing whether “AI is really creating productivity.”
To start with the conclusion, based on the original source, Anthropic is not a company with an apparent $80 billion loss; rather, it is interpreted as an AI company with very low cash burn and abnormally strong revenue growth and margins.
However, all of these calculations rest on the assumption that AI demand will keep exploding.
If that assumption falters, AI data centers, cloud companies, and firms that raised debt to build infrastructure could come under pressure similar to the 2008 financial crisis.
1. Why the All-In podcast matters: the perspective of key Silicon Valley investors
This discussion came from the All-In podcast, which is highly influential in Silicon Valley.
The panel included investor Chamath Palihapitiya, a former early Facebook executive; Jason Calacanis, an early Uber investor; David Friedberg, who sold a climate data company to Monsanto; and David Sacks, who was involved in AI policy in the Trump administration.
They are not just tech commentators; they are people who simultaneously look at venture investing, policy, corporate operations, and global capital market flows.
So their view of the Anthropic IPO is not “AI is cool,” but is organized around revenue, margins, cash flow, debt, compute demand, and customer ROI.
2. The earlier debate: Ken Griffin’s warning that “AI is almost certainly a bubble”
One week earlier on All-In, they covered Citadel founder Ken Griffin’s warning about an AI bubble.
He pointed out that companies were starting to look at Anthropic-like AI service bills and ask, “So where is the productivity?”
This claim connects to a very important question in the current global economic outlook.
Whether companies are actually getting enough productivity improvement to keep paying for AI costs will determine the direction of the next AI investment cycle.
If AI only raises costs without creating productivity, then the current data center investment and GPU demand could end as overinvestment.
Conversely, if AI is truly driving revenue growth, labor efficiency, workflow automation, and better customer support, then today’s AI infrastructure investment may be an early growth phase rather than a bubble.
3. David Friedberg’s core argument: the entire AI value chain is making money
Friedberg viewed the AI industry as one value chain and explained with numbers where the money is being made.
The core structure runs from Nvidia → data centers and cloud infrastructure → AI model companies like Anthropic → end customers.
His argument is that in this chain, it is not just one segment making money; high margins are appearing at each stage.
- Nvidia: It is described as turning $1 of cost into about $4 of revenue.
- GPU-based infrastructure providers: They buy chips, put them into data centers, and collect high usage fees over the depreciation period.
- Anthropic: It is said to show an incremental gross margin of about 85% on each additional $1 of revenue.
- End customers: They are interpreted as potentially getting 2x to 10x the economic effect versus what they spend on AI.
This is where the structure behind the phrase “AI turns $1 into $180” comes from.
The number directly mentioned in the original source is a calculation in which Nvidia’s $1 of cost expands to roughly $18 of revenue at the Anthropic revenue level.
If you add the idea that end customers can get up to 10x productivity effects relative to AI usage costs, then the logic can extend to as much as $180 in total economic effect per $1.
In other words, it does not mean $1 directly becomes $180 of Anthropic net profit.
More precisely, it means that economic value is greatly amplified across the entire value chain, from AI infrastructure to customer productivity.
4. Nvidia’s role: the first clearly profitable stage in the AI investment cycle
The company that first and most clearly makes money in the AI investment cycle is still Nvidia.
Friedberg explained that $1 of cost used to make Nvidia chips turns into about $4 of revenue.
This implies a margin structure of about 75%.
Of course, actual accounting margins can vary depending on product mix, supply chain costs, HBM memory prices, and packaging bottlenecks.
But the broad direction is clear.
As AI model competition intensifies, GPU demand rises, and Nvidia maintains pricing power at the front end of that demand.
5. Data centers and cloud: the stage that buys GPUs and rents them out at a higher price
The second profitable stage is data centers and cloud infrastructure.
The original source referenced large AI compute infrastructure examples such as Colossus.
For example, if a $10 billion investment builds a GPU cluster and it is depreciated over five years, the annual cost works out to about $2 billion.
But if that infrastructure generates around $1.25 billion in monthly revenue, annual revenue reaches $15 billion.
In that case, profitability is extremely high relative to depreciation expense.
That is why the podcast suggested that infrastructure operators could generate margins above 80% relative to chip investment.
The key point is not just buying GPUs once, but creating recurring revenue by renting them out in cloud form.
6. Anthropic’s key number: an 85% incremental gross margin
The strongest number mentioned for Anthropic is an incremental gross margin of about 85%.
This means the existing computing cost base is already in place, and when additional customer usage rises, the remaining margin is very high.
AI model companies initially face huge training and infrastructure costs.
But once model performance stabilizes and inference costs fall, a large portion of incremental revenue can remain as profit.
The important point here is the speed of cost efficiency improvement.
The podcast said AI cost efficiency is improving by 10x per year for the same task or capability.
If the cost of handling a specific task falls to one-tenth in a year, and Anthropic only lowers customer prices by 50%, customers experience major savings.
At the same time, Anthropic can absorb the rest of the efficiency gains as margin.
If this structure continues, AI model companies can be valued like high-margin software businesses.
7. Are customers actually making money? The real validation point for AI demand
The most important question in the AI bubble debate is whether customers are actually making money.
Friedberg mentioned that customers may be getting roughly 2x to 10x in economic value relative to what they spend on AI.
If not, the logic goes, retention should fall and usage growth should weaken.
But right now, AI usage and spending are still rising.
That can be interpreted as a sign that at least many enterprise customers are seeing a certain level of productivity improvement from AI.
Especially in areas like customer support automation, code generation, document analysis, research automation, sales support, and security monitoring, AI is reducing labor time and increasing throughput.
This could affect future corporate earnings, the labor market, wage growth, and interest rate outlook as well.
8. The Microsoft example: revenue surged, but headcount barely grew
One of the most striking examples mentioned on the podcast was Microsoft.
Microsoft’s revenue reportedly rose from about $198 billion to $332 billion over the last four years.
Meanwhile, headcount increased by only about 2,000 people.
Under the old corporate growth formula, revenue growth of that scale would usually require much larger hiring.
But Microsoft essentially achieved massive revenue growth with almost the same workforce size.
This can be seen as a snapshot of the productivity gains created by AI and cloud software.
Of course, Microsoft’s growth is driven by many factors including Azure, Office 365, LinkedIn, security, and AI Copilot.
Even so, the structure of “revenue rising rapidly while headcount is nearly flat” is very important data in the debate over corporate productivity in the AI era.
9. The $80 billion Anthropic loss controversy: the claim that it is not an actual cash loss
The biggest twist in the original source is the controversy over Anthropic’s $80 billion loss.
Some media outlets used that number to argue that Anthropic is a massively loss-making company and a representative example of an AI bubble.
However, the All-In participants explained that this loss is closer to an accounting loss than an actual cash loss.
The key is the convertible debt structure.
The explanation is that Anthropic issued convertible debt to an infrastructure partner like Amazon, and as Anthropic’s valuation rose, the mark-to-market loss on that convertible debt was recorded as an accounting loss.
If the same funds had been raised as common equity, that loss might not have appeared on the income statement.
In other words, the interpretation is that $80 billion of cash did not actually leave the company; it was an accounting loss tied to rising enterprise value.
This is a core point that is easy to miss in general news coverage.
When looking at an AI company’s income statement, you should not focus only on net loss; you need to distinguish whether the loss is a cash loss, a valuation loss, stock-based compensation, or convertible debt accounting.
10. Anthropic’s capital efficiency: it raised a lot, but actually burned little
The original source says Anthropic may have secured, or be able to secure, around $110 billion to $125 billion in funding.
Of that, actual cumulative cash burn is described as only around $9 billion to $10 billion.
This is quite different from the common perception of AI model companies.
Many people thought Anthropic was a company rapidly burning cash because of massive compute costs.
But the view presented on the podcast is the opposite.
Anthropic has raised large amounts of capital, but its actual net burn is relatively low, and it is still interpreted as a company with roughly $100 billion in cash and cash equivalents.
There is also an argument that if it adds about $100 billion more through an IPO and roughly $80 billion in debt facilities, it could have around $300 billion in funding capacity.
This funding power is very important for future data centers, GPUs, cloud contracts, and long-term compute commitments.
11. Why capital strength matters: the real competitiveness of an AI company is not just the model
In the AI industry, model quality matters, but now compute access is becoming an even more important advantage.
To build a strong model, you need a large GPU cluster.
To serve more users, you need inference compute as well.
As services with heavy token usage such as AI agents, multimodal models, real-time voice, video generation, and cybersecurity automation increase, compute demand grows further.
So if Anthropic has enough cash and debt capacity, it is more likely to handle future compute orders and CAPEX commitments.
AI competitiveness is expanding from “who built the smarter model” to “who can secure more compute reliably.”
12. The CoreWeave plunge debate: volatility or the beginning of a bubble collapse?
The original source also mentioned CoreWeave’s stock decline and bond price drop.
It noted that CoreWeave bonds were valued around 80% and that the stock fell 15% to 20% in a few days.
This is a very important question.
We need to judge whether this is simply market volatility or the beginning of an AI infrastructure bubble collapse.
David Sacks leaned toward seeing it mainly as volatility.
The reason is that he believes AI compute demand has not yet weakened.
AI infrastructure companies are being valued now based on a future market worth trillions of dollars.
Therefore, even small changes in interest rates, debt costs, demand outlook, customer contracts, and GPU prices can move stock and bond prices sharply.
In short, in a market with high growth expectations, excessive volatility is natural.
13. Is compute demand really strong? Signs of supply shortage
One part repeatedly emphasized in the podcast is that compute demand remains extremely strong.
Large AI model companies are still saying they lack compute.
Startups are also asking venture capital firms to help them obtain compute resources.
Some VC firms are even trying to secure compute directly to strengthen their investment competitiveness.
There was even a case mentioned where one accelerator secured spare compute resources and used them as an incentive to recruit startup teams.
This means GPU access is becoming as important as cash in the AI startup ecosystem.
14. Old GPUs are still rented at high prices: the abnormal case of A100s and H100s
Another sign of strong AI infrastructure demand is the price of older GPUs.
The original source said previous-generation GPUs like A100s and H100s are still being rented at premium prices.
Normally, such equipment should depreciate over time and rental prices should fall quickly.
But in the actual market, GPUs from several generations ago are still fetching high prices.
This is a signal that supply is not keeping up with demand.
During the dot-com bubble, there were plenty of fiber optic cables laid down but not enough users to absorb them.
Today’s AI infrastructure market is almost the opposite: there are too many people who want to use the equipment if it exists.
15. Jevons Paradox: the lower AI prices fall, the more usage rises
A concept that must be understood in this discussion is Jevons Paradox.
Jevons Paradox refers to the phenomenon where, as technological efficiency improves and unit costs fall, total usage actually rises.
In AI, the lower token prices go, the more often companies and individuals will use AI.
For example, if people once only asked AI to summarize documents, now they will keep assigning it research, coding, meeting notes, customer service, contract review, security monitoring, and sales automation.
When AI usage becomes cheaper, companies do not say, “Let’s try it a little”; they say, “Let’s shift more work to AI.”
So even if token prices fall, total AI revenue and data center demand can grow instead.
16. Future AI apps will use more tokens
Early AI usage often involved entering a prompt and receiving an answer.
But future AI applications will use far more tokens.
AI agents do not just answer one question; they plan, execute, and verify multiple steps.
Cybersecurity AI will not just perform penetration testing once a year; it must continuously monitor attack surfaces and anomalies 24/7.
Multimodal AI processes text, images, voice, and video data at the same time.
All of these services require much more compute and many more tokens than a traditional chatbot.
So if you judge AI demand only by current chatbot usage, you may underestimate future compute demand.
17. Mentions of AWS and Nebius being sold out: supply shortages may last for a long time
The original source also mentioned that companies like Nebius and Amazon Web Services are sold out through 2028 and beyond.
If that is true, the AI data center supply shortage may be a medium- to long-term phenomenon rather than a short-term one.
Cloud capacity being reserved years ahead means companies view AI demand not as a short-lived trend, but as long-term business infrastructure.
Of course, such contracts need to be viewed alongside cancellation terms, pricing adjustments, minimum usage commitments, and customer credit risk.
Still, the demand signal itself is interpreted as very strong.
18. Key points not often covered in other YouTube videos or news
First, Anthropic’s $80 billion loss is likely not a cash loss, but an accounting issue.
Most news stories emphasize the loss number, but if you do not distinguish between a convertible debt valuation loss and actual operating cash outflow, you can reach a completely different conclusion.
Second, AI profitability should not be viewed by looking at just one model company; it should be viewed across the whole value chain.
Nvidia, data centers, cloud, AI model companies, and end customers each capture value in different ways.
Third, lower AI costs may not reduce revenue; they may cause usage to surge.
When token prices fall, customers assign more work to AI, and total usage is likely to grow further.
Fourth, AI competitiveness is shifting from model quality to compute acquisition capability.
Access to capital markets, long-term GPU contracts, power supply for data centers, and cloud partnerships become core advantages.
Fifth, the real risk is not “AI cannot make money,” but “the debt structure breaks if AI demand forecasts are wrong.”
If demand holds, high CAPEX is growth investment; if demand weakens, the same CAPEX becomes distressed debt.
19. The biggest risk: if the assumption that AI demand will continue turns out to be wrong
The most dangerous part of this analysis is that all the numbers assume continuing AI demand.
If AI users keep growing, companies keep expanding AI spending, and token usage surges, then current data center investment can be justified.
But if in reality only a small number of advanced users keep using AI and most companies do not get the productivity gains they expected, the situation changes.
In that case, data centers become oversupplied, GPU rental prices fall, and firms that built infrastructure with long-term debt come under pressure.
Especially in a high-interest-rate environment, debt repayment burdens become heavier.
Then AI infrastructure companies’ bond prices could fall, and some firms may fail to refinance.
This scenario could create a stress structure similar to the way real estate and structured finance collapsed during the 2008 financial crisis.
20. The difference between the dot-com bubble and the AI bubble
During the dot-com bubble, internet infrastructure was laid down too quickly, but actual usage demand could not keep up.
There were fiber cables, but not enough traffic and revenue models.
By contrast, today’s AI market has such strong compute demand that even older GPUs are rented at high prices.
This difference is very important.
Whether current AI infrastructure investment is a bubble cannot be judged simply by saying “there is too much infrastructure.”
The key is whether there are enough customers actually paying to use it.
So far, the data leans toward strong demand.
But oversupply always appears with a lag.
The real test is likely to begin when large data centers come online at the same time in 2026, 2027, and 2028.
21. Checkpoints investors should watch
- AI company revenue growth: You need to check whether revenue growth at Anthropic, OpenAI, and Microsoft AI services remains strong.
- Customer retention: You need to see whether enterprise customers reduce AI spending or actually increase it.
- GPU rental prices: A sharp drop in A100, H100, and latest GPU rental prices could be a sign of oversupply.
- Data center vacancy rates: You need to confirm whether long-term contracted volume turns into actual usage.
- Debt markets: Bond prices for AI infrastructure firms like CoreWeave can be a leading indicator.
- Interest rates and liquidity: In a high-rate environment, the CAPEX burden of AI data centers can become heavier.
- Power infrastructure: For AI data centers, securing electricity may become a bigger bottleneck than GPUs.
22. My view: the AI bubble debate is not “yes or no” but should be viewed by segment
It is too simplistic to lump the entire AI market together and call it a bubble.
There are segments like Nvidia that are already making money, and segments like Anthropic that show the potential for both high growth and high margins.
On the other hand, some data center companies that have taken on a lot of debt to bet on future demand could be much riskier.
In other words, the AI bubble is more likely to show up first not across the whole industry, but in companies with weak capital structures, low-quality customer contracts, and excessive power and GPU costs.
From a global economic outlook perspective, if AI truly raises productivity, it could help ease inflation and improve corporate profit margins.
Conversely, if excess AI infrastructure investment turns into a debt problem, it could increase volatility in credit markets and slow the economy.
So going forward, rather than focusing on AI model performance alone, we should pay closer attention to AI investment returns, data center utilization, cloud revenue growth, and corporate productivity indicators.
< Summary >
The core point of the Anthropic IPO debate is whether AI is actually making money.
On the All-In podcast, the analysis was that high margins and productivity effects are appearing across the entire AI value chain, from Nvidia and data centers to Anthropic and end customers.
Anthropic’s $80 billion loss is explained as likely not an actual cash loss, but an accounting loss related to convertible debt valuation.
Anthropic is instead interpreted as a capital-efficient company with low cash burn, an incremental gross margin around 85%, and enormous funding capacity.
As AI prices fall, Jevons Paradox is appearing, where usage rises even more, and new services like AI agents and cybersecurity require more tokens and compute.
However, all of the optimism depends on the assumption that AI demand will continue.
If demand weakens, data center oversupply, falling GPU rental prices, debt burdens, and credit risk could hit at once.
So the AI industry should be viewed not as simply bubble or not, but as a market where some segments are already profitable while others carry significant debt risk.
[Related Articles…]
- AI infrastructure investment cycle and global economic outlook: key takeaways
- Data center wars and changes in the next-generation cloud market
*Source: [ 월텍남 – 월스트리트 테크남 ]
– “AI는 1달러로 180달러를 번다” / 세계적 투자자 4인의 앤트로픽 IPO를 보는 관점… 꼭 보셔야 합니다.
