Podcast Summary: Better Offline — "Exclusive: Here's How Much Anthropic Spends on AWS"

Host: Ed Zitron
Network: Cool Zone Media / iHeartPodcasts
Date: October 20, 2025
Episode Theme: A revealing, data-driven exposé into the staggering cloud computing costs behind Anthropic, one of the leading AI firms, and the business and ethical implications for the entire generative AI ecosystem.

Episode Overview

In this special, investigative episode, Ed Zitron presents exclusive data on Anthropic’s actual spending on Amazon Web Services (AWS)—pulling back the curtain on the real economics underlying the most-watched generative AI companies. Zitron dissects confidential cost figures, explains why they matter for the future (or demise) of AI startups, and interrogates the murky, rent-seeking business tactics employed by tech power players like Anthropic, Amazon, and Google.

Key Discussion Points & Insights

1. Anthropic’s Cloud Spending Exposed ([02:26])

• Ed Zitron exclusive: For the first time, breaks down Anthropic's AWS bills:
  • 2024: $1.35 billion spent on AWS.
  • 2025 (Q1-Q3): $2.66 billion spent on AWS, versus $2.55 billion in revenue.
  • Q3 2025 alone: $1.225 billion on AWS.
• Anthropic spent essentially every dollar earned on AWS in 2025, sometimes even exceeding their revenue.
  • "Through the end of September 2025, Anthropic has spent around 100% of their revenue in 2025 on Amazon Web Services, spending $2.66 billion on comp, $2.55 billion in revenue." (Ed Zitron, [02:26])
• Zitron emphasizes these are net cash numbers, after credits: “They're after credits, though.” ([03:50])

2. Cost Spikes and Volatility ([05:00], [07:20])

• AWS bills have shown wild jumps within single months:
  • August 2025: $183.7 million
  • September 2025: $518.9 million—a $335 million month-over-month spike.
• Suggests root problem: users of products like Claude Code are very expensive to serve, costing potentially thousands per user even if subscriptions are $100–$200/month.
  • “My hunch is it's because they have a massive problem where Claude code users are each costing them thousands of dollars despite only paying 100 or $200 a month.” ([07:37])

3. Cursor: Anthropic’s Biggest Customer Caught in the Middle ([04:30], [08:40])

• Cursor, a major AI coding company, saw AWS bills double from $6.2M (May) to $12.6M (June 2025) right after Anthropic implemented new “Priority Service Tiers”—interpreted as an aggressive rent-seeking maneuver.
  • “Cursor's Amazon Web Services bills doubled from $6.2 million in May 2025 to $12.6 million in June 2025...since Anthropic increased the costs with the launch of Priority Service Tiers, an aggressive rent-seeking measure.” ([04:56])
• Cursor’s costs remained high and volatile, directly correlating with Anthropic’s price tier changes and new product launches.

4. Cloud Ecosystem and Conflicts of Interest ([06:52], [14:00])

• Anthropic’s investors and cloud providers: Google owns ~14%, Amazon ~15–18%. Both are “the main or primary compute partner” with little clarity on Google Cloud usage.
• Anthropic became Cursor’s competitor with the launch of Claude Code, just after introducing punitive service tiers:
  • “Anthropic supplied access to their models to a company Cursor and then released a product Claude Code that did exactly the same thing as that company Cursor, turning it both into a customer and a competitor...” ([13:26])
• This move created a massive conflict of interest, as Cursor was simultaneously customer, competitor, and a target for higher costs.

5. Business Model Critique: Unsustainable Economics ([15:00])

• Anthropic and companies like Cursor have no path to profitability, as compute costs scale linearly with revenue:
  • “Anthropic should have always charged sustainable rates, even if it meant that it wasn't possible to build a big company based on their models. Sadly, we don't live in that universe.” ([14:52])
• Zitron draws analogies with Uber but emphasizes Anthropic’s problem is not customer acquisition, but unsolvable compute costs (“fuel problem”):
  • “What Anthropic has is a fuel problem. They have a Compute problem for the amount that they're paying to run their goddamn services.” ([16:05])
• Projected attempts to scale (e.g., “$20 billion in annualized revenue targets for next year”) are lambasted as “an absolute fucking joke” ([17:47]), with the implication that costs would only rise faster.

6. Skepticism of AI Industry Narratives and Solutions ([18:34], [20:00], [25:39])

• Zitron challenges listeners and boosters on what truly warrants continued investment in companies like Anthropic:
  • “Why are we rewarding a company business practices for making a product that loses more money the more money it makes?” ([18:59])
• Debunks common solutions: advertising is not the answer (Perplexity’s failed ad revenue is used as a case study: “They made $20,000 in 2024 in advertising revenue...” ([25:39]))
• Technological optimism around specialized chips like Trainium and theoretical improvements is dismissed (“Chips aren't fixing this...you are being conned.” ([26:28]))

7. Environmental and Social Costs ([17:00], [29:00])

• Calls out the environmental and social recklessness of burning billions in compute, often subsidized by venture capital, for outputs that are “kind of cool, but a hundred times more often mediocre, unreliable and outright ridiculous.”
  • “Claude Opus and Claude Sonnet are not worth burning billions of dollars a year in cloud costs fueling an environmentally destructive, plagiarism charged pseudo company that would roll over and die within months if it didn't constantly get fed billions of dollars a year.” ([16:59])
• Urges public discussion on how much inefficiency and waste society should tolerate in pursuit of generative AI.

8. Call to Action: Honest Public Conversation ([29:00])

• Zitron urges everyone—AI boosters and skeptics alike—to join a frank conversation about the mismatch between costs and value in generative AI:
  • “What I am advocating for is for everybody to openly discuss that the amount of money it costs to run these companies is at odds with what they have built, are building and will build in the future.” ([29:52])

Memorable Quotes & Moments

• On the absurdity of costs:

  “These digital Mr. Beans very likely spend comparable sums on Google Cloud and likely another billion or two on salaries … This company absolutely fucking sucks. I don't care if you like Claude Sonnet or Claude Opus, I don't give a fuck.”
  —Ed Zitron, [17:00]

• On Cursor’s predicament:

  “Cursor hit $500 million in annualized revenue in the same month that they then saw their costs double, dramatically reducing the value of their subscription product at the apex of their success.”
  —Ed Zitron, [15:12]

• On business model futility:

  “The only thing funnier than that joke is that it will likely cost them $25 billion to make that fictional money. And where, pray tell, is that coming from?”
  —Ed Zitron, [17:52]

• On AI ad revenue hype:

  “Are we supposed to be impressed that Perplexity made enough revenue to buy a second hand Toyota Corolla? There are people making more money than that slinging fucking Herbalife.”
  —Ed Zitron, [25:39]

• On the AI industry’s supposedly heroic narratives:

  “No combination of Amazon, Trainium or Google TPUs is going to usher in the birth of the machine God. The products they make are at best and in inconsistent moments kind of cool, but a hundred times more often mediocre, unreliable and outright ridiculous.”
  —Ed Zitron, [30:16]

Timestamps for Key Segments

| Timestamp | Segment Summary | |---|---| | 02:26 | Ed Zitron introduces the exclusive AWS spend revelations. | | 04:30–08:40 | Details and analysis of Anthropic’s costs; deep dive into Cursor’s surging AWS bills and service tier impact. | | 13:26 | Explains Anthropic’s product/partner conflict with Cursor. | | 15:00–17:47 | The business model critique: lack of profitability, compute cost scaling. | | 18:34 | Asks what scientific or technological achievement justifies the ongoing investment. | | 25:39 | Explains why ad-based model is a non-starter for generative AI. | | 29:00 | Final call to action for honest public conversation on value vs. cost in AI. |

Tone & Language

Zitron is direct, unsparing, and often caustic—combining sharp economic analysis with dark humor and explicit exasperation at tech industry excess. He acknowledges hedging certain inferences with “I can’t say for sure,” but ultimately delivers a scathing indictment of business models predicated on endless subsidized growth.

Takeaways

• Anthropic’s economics are dire: Its cloud spend matches (or exceeds) revenue, making its growth model deeply unsustainable.
• AI startups are stuck on a burn treadmill: Compute costs scale linearly with usage; neither margins nor economy of scale are materializing.
• Hyperscalers hold the power: Amazon and Google profit regardless, deepening their role as industry gatekeepers.
• Conflict, not partnership: Anthropic’s move to undercut its top cloud customer, Cursor, illustrates the toxic competitive dynamics fueling the space.
• Industry PR is divorced from reality: Claims of “impending profitability” or “future AGI windfalls” are, in Zitron’s words, “total bullshit.”
• Urgent need for a public reckoning: Audiences and investors must confront the actual costs and consequences of propping up companies whose products don’t deliver commensurate value.

If you want the raw numbers and detailed breakdowns, Ed Zitron recommends reading his newsletter, freely available—he’s happier giving away this scoop than seeing the tech establishment paper over the structural rot.