
Oct 1, 2026 · 1h 22m
Listen from 1:17:52
Listen at 1:17:52
Anthropic’s IPO exposes the debt beneath AI’s growth story
Ed Zitron on Anthropic's IPO (S-1), AI Debt, and Counterparty Risk
The episode connects Anthropic’s weak economics and infrastructure commitments to broader risks for lenders, hyperscalers, and markets.
- 1Anthropic’s spending, losses, and customer concentration complicate claims that it is financially safer than OpenAI.
- 2Run-rate revenue and adjusted profitability can make AI businesses appear healthier while obscuring infrastructure and compute costs.
- 3AI data-center debt is spreading risk across private-credit lenders, hyperscalers, banks, and public markets.
Don't miss
Ed Zitron uses CoreWeave’s rising financing costs to show how the AI boom’s infrastructure commitments can become repayment risks.
The brief
Ed Zitron and Jack Farley use Anthropic’s leaked IPO materials to challenge the idea that the company has superior economics or less concentrated revenue than OpenAI.
The discussion focuses on Anthropic’s losses, vast infrastructure commitments, and adjusted metrics, arguing that excluding compute and related costs obscures the business’s underlying economics.
The conversation then widens to consumer agents, where products such as Meta’s Muse raise questions about usefulness, security, and the cost of serving AI at scale.
CoreWeave becomes a case study in the debt behind the buildout, as rising financing costs and uncertain demand test whether infrastructure spending can earn durable returns.
The episode’s central warning is that AI enthusiasm sits above a fragile financing system whose losses could reach private credit, banks, hyperscalers, and markets.
What was said on this episode
66 statements · 4 positive · 58 negative · 3 mixed · 1 neutral
Anthropic was a worse business than OpenAI in 2025
“anthropic at least in 2025 was a worse business than open ai”
Listen at 0:50
“they lost eight billion dollars on 4.6 billion dollars of revenue”
Listen at 1:03
Google and Amazon reselling generated about 47% of Anthropic revenue
“about 47 percent of anthropics revenue came through sales from google and amazon reselling their models”
Listen at 1:11
Anthropic’s largest proxy customers are also competitors, creating counterparty risk.
“there is a real risk there they have that basically their largest... customers by proxy are also their competition”
Listen at 1:39
Anthropic’s largest proxy customers are also competitors
“their largest... customers by proxy are also their competition”
Listen at 1:43
A profitable AI lab may not be possible; Anthropic was worse than OpenAI in 2025
“i don't think it's possible and at the very least we know that at least in 2025 anthropic was in a worse condition than open ai”
Listen at 2:17
Anthropic was in worse financial condition than OpenAI in 2025.
“at least in 2025 anthropic was in a worse condition than open ai”
Listen at 2:20
Google and Amazon can discount Anthropic’s tokens
“Google and Amazon can discount Anthropics tokens all they want”
Listen at 3:08
One percent of customers generate 80% of OpenAI and Anthropic revenue
“80% of OpenAI and Anthropics revenue comes from 1% of their customers”
Listen at 3:29
Anthropic’s success depends on Google and Amazon selling its models
“their success is contingent on Google and Amazon's ability to sell their stuff”
Listen at 4:21
“if they're not, this is a company in distress”
Listen at 5:06
Anthropic is not a good business
“This company's a dog. This is not a good business.”
Listen at 8:32
Anthropic has over $400 billion in non-cancellable compute agreements
“Anthropic has over $400 billion worth of non-cancellable compute agreements”
Listen at 10:17
Anthropic’s operating obligations could require $100–200 billion annually despite profitability.
“if they're profitable but they have to pay 100 200 billion dollars a year in opex how do you get the money in”
Listen at 10:59
Anthropic’s total commitments are $518 billion, including $413 billion non-cancellable.
“They're 518 billion. There's 413 billion of them that are non-cancellable.”
Listen at 11:35
Anthropic’s compute agreements create large-scale counterparty risk
“there is counterparty risk here at scale”
Listen at 12:24
Anthropic’s obligations are hundreds of billions versus $11.5 billion quarterly revenue
“this is hundreds of billions of dollars for a company that i think its latest quarter had 11.5 billion dollars of revenue”
Listen at 13:38
Annualized run rate is a dangerous metric for AI companies
“annualized run rate is a dangerous metric”
Listen at 18:39
OpenAI revenue will be about $36 billion in 2026
“it's gonna be 36 billion for this year for open ai i think”
Listen at 19:22
AI-company run-rate figures are currently useless.
“at this point, these run rate figures are useless”
Listen at 22:17
AI companies’ run-rate figures are useless
“these run rate figures are useless”
Listen at 22:18
Anthropic’s annualized revenue figures poorly reflect actual revenue.
“these numbers don't interact with reality very much at all”
Listen at 25:26
SEC regulation should require private companies to disclose revenue transparently.
“we need at the end of this is SEC regulation around private companies and disclosures of revenue because these numbers are actively deceptive”
Listen at 25:41
Private AI-company revenue disclosures are actively deceptive
“these numbers are actively deceptive”
Listen at 25:46
Anthropic’s IPO has been delayed repeatedly despite its importance.
“It's peculiar for the most important IPO of the year to get delayed twice.”
Listen at 26:32
Anthropic’s reported numbers are worrying
“the numbers are quite worrying”
Listen at 27:08
A separate entity collected 47% of Anthropic’s cash
“47% of their cash was collected by a completely separate entity”
Listen at 31:02
Anthropic may struggle to fund its $400 billion obligations
“how the hell are you paying those $400 billion worth of obligations”
Listen at 31:47
Muse requires an eight-gigabyte virtual machine per user, creating infrastructure burden
“You've got a little virtual machine with eight gigabytes of RAM for each person using this. So you've just got this massive infrastructural burden.”
Listen at 33:21
Transaction-based agents would be high-volume businesses with heavy operating costs
“a very high volume business on your extremely opex heavy business”
Listen at 35:14
AI agents mishandling personal information can impose large real-world costs
“the real actual cost of having these things being willy-nilly about your personal information is huge”
Listen at 36:11
Banks should not allow language models to access or manage customer accounts.
“they're not going to allow they shouldn't allow an lm to do that for financial regulatory reasons”
Listen at 37:58
Muse generates negative revenue despite boosting Meta’s stock
“Meta's stock skyrocketing off of Muse is insane. However you may feel about the product, it's generating negative revenue.”
Listen at 38:45
Vague AI narratives could enable intentional market manipulation
“someone with really ill intentions could intentionally manipulate the market as a means of making money off of this”
Listen at 39:54
Trump’s rebranding of AI as superintelligence will not improve public acceptance.
“No, I don't think this will work.”
Listen at 41:09
Netflix’s data centers differ substantially from AI data centers.
“the kinds of data centers that Netflix uses and pretty much every major web service uses are much smaller and very different to the AI data centers”
Listen at 42:20
AI is not offering meaningful value to ordinary people
“AI is not offering anything.”
Listen at 43:37
CoreWeave’s repriced bond debt would carry roughly 13% financing cost
“If you reprice them, all of that debt, which is still around nine points at the time, you're looking at 13%.”
Listen at 53:30
Oracle debt is already priced like junk
“Oracle is already pricing like junk.”
Listen at 55:18
Future data-center projects will face financing costs of at least 8–13%.
“every data center project that is raising from here for the next forever will be raising minimum at eight, maybe nine, maybe 12, maybe 13”
Listen at 56:13
Data centers’ first five revenue years may not cover operating costs
“your first five years, which are the only really guaranteed revenue years, are not even going to cover your costs”
Listen at 57:18
A future AI infrastructure bond offering will fail or require higher interest.
“There's just going to be one that cannot raise, or they have to raise the interest at the end to get people in”
Listen at 58:24
Most private-credit debt tied to AI infrastructure may not be repaid
“In fact, I think most of the private credit debt doesn't get paid here.”
Listen at 1:03:22
OpenAI and Anthropic account for 80% of data-center demand.
“80% of all data center demand is OpenAI and Anthropic”
Listen at 1:04:08
Private credit exposure to AI infrastructure is already a brewing crisis
“this private credit is already a brewing crisis”
Listen at 1:05:07
OpenAI is projected to spend $750 billion on compute through 2030.
“OpenAI is projected to spend $750 billion on compute through 2030.”
Listen at 1:06:11
OpenAI and Anthropic must achieve historically unprecedented cash-flow growth
“we are either going to have to watch two companies pull off the biggest cash flow increase in history”
Listen at 1:07:12
Oracle will be downgraded
“I think we see Oracle downgraded.”
Listen at 1:07:58
Nvidia’s revenues will decline because of AI infrastructure stress.
“I think we see NVIDIA's revenues have dropped by them.”
Listen at 1:08:05
NVIDIA will not reach its projected fiscal-2028 revenue requirement
“I don't see that happening.”
Listen at 1:08:19
AI has not materially improved productivity so far.
“What has AI actually improved? Has it improved productivity? No, productivity is basically the same.”
Listen at 1:10:44
AI has not materially improved productivity or economic growth
“Has it improved productivity? No, productivity is basically the same. Growth is the same.”
Listen at 1:10:46
AI has created a complex financial crisis
“So what has it actually done? Created a financial crisis.”
Listen at 1:10:58
AI investment has created a financial crisis.
“Created a financial crisis.”
Listen at 1:11:00
GPU inference has a brittle business model
“So it's already quite a brittle business model.”
Listen at 1:14:48
Demand may be insufficient for even one-quarter of planned data centers
“there is no proof that there is enough demand for even a quarter of the data centers being built”
Listen at 1:15:01
AI infrastructure excess could cause a tech-sector depression
“I don't see how that doesn't work out to some form of minor or major depression in the tech sector.”
Listen at 1:15:23
An AI downturn could eliminate 80% of venture capital
“I think there's a scenario here where 80% of venture capital gets washed out with this.”
Listen at 1:17:52
Higher Japanese rates will prompt investors to repatriate funds from American equities.
“Japanese investors are going to start repatriating their money instead of investing in them in American equities.”
Listen at 1:18:16
Ratings agencies may classify Anthropic as investment grade.
“I think that we could see the ratings agencies say, yeah, this is investment grade”
Listen at 1:18:48
Anthropic’s IPO will proceed despite concerns.
“I think Anthropic will be pressed through.”
Listen at 1:19:16
Anthropic may reach a $2 trillion valuation with volatile shares.
“I think they get 2 trillion. I think they have a volatile ass stock.”
Listen at 1:19:18
Anthropic could reach a $2 trillion valuation at IPO
“I think they get 2 trillion.”
Listen at 1:19:18
Anthropic’s stock may trade with extreme volatility after IPO
“I think they have a volatile ass stock. I think maybe they pop and they drop and they pop and they drop.”
Listen at 1:19:20
Investors will become seriously concerned after Anthropic reports its first quarter
“the first quarter they share, people are going to get real worried”
Listen at 1:19:26
AI products reflect marketing and hype more than stable businesses
“the products reflect marketing and they reflect hype far more than they reflect stable businesses”
Listen at 1:20:43
Statements are attributed to the speaker as said on the episode and reflect their view at the time, not PodLume's. They are not advice.
Weekly email
State of the Pod, every Monday
The week's most-discussed names and stories from podcasts, with the moments behind them. Free, five minutes, unsubscribe in one click.