Anthropic's $80 Billion 'Loss' Is Just an Accounting Illusion
The money is a paper loss from Amazon's convertible loan being repriced as valuation changes—no cash outflow; what's truly concerning is that training costs are growing faster than service costs and remain completely outside the company's control.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Training costs exceed service revenue growth and escape management control
The S1 filing shows Anthropic's H1 revenue surged while R&D—specifically training—expenditure grew faster than service costs. The reason: maintaining model leadership requires continuous model training, and how much to spend and when to spend it are entirely beyond the company's control. This is ongoing operational cost, not a one-time capital investment that can be managed once a problem is solved. Thompson points out that this collapses the distance between "limiting frontier speed for safety" and the reality of "burning money to survive".
— Ben ThompsonRevenue returned to cloud platforms should count as cost of sales
About half of Anthropic's revenue flows through AWS, Azure, and Google Cloud, but roughly 37 percent is paid back to these platforms. The company counts the full amount as revenue and records the payments back as a sales and marketing expense rather than cost of goods sold. Thompson argues this portion should properly count as cost of revenue, which would significantly compress the company's reported gross margins.
— Ben ThompsonThe $80 billion loss is mark-to-market repricing, not cash outflow
Early reports claimed Anthropic lost $80 billion in H1, which sounds catastrophic. Thompson explains this is not cash out the door. Amazon's investment took the form of a convertible loan, and every time Anthropic raises at a new valuation, the loan is repriced and the difference is recorded as a paper loss. The same money surfaces on Amazon's financial statements as gain on investment—it's an accounting identity on opposite sides of a ledger, not cash destruction.
— Ben ThompsonShareholder dilution is written into the equity structure itself
A cited tweet traces how large companies finance: first free cash flow, then debt, now equity dilution. Anthropic's situation is extreme: Amazon, Google, and Broadcom hold agreements to convert stakes into equity over time. Every valuation increase forces a new dilution round, and the agreements actually require issuing additional shares if the stock price crosses certain thresholds—dilution is baked into the company structure itself.
— Ben ThompsonAn AI bubble exists only if revenue cannot cover capex
Craig Moffett offers a simplified test: ignore the noise of companies trading revenue with each other along the value chain, and look only at the total revenue flowing into the system—is it enough to cover the capex required to build out that chain? Thompson agrees with this framework and notes that even Anthropic's rapid revenue growth remains far short of covering industry-wide capex, which is being deployed for years ahead. Real victory depends on whether revenue growth can accelerate further.
— Ben Thompson / Craig MoffettCloud vendors alone have the cash to bridge the AI capex gap
Thompson argues that whoever has the thickest cash flow, strongest borrowing capacity, and strongest ability to issue equity becomes the bridge across the AI capex chasm: Google, Amazon, and Microsoft. Nvidia, in reverse, provides financial backing to emerging cloud competitors—essentially hedging against the risk that "the entire industry runs out of money before revenue catches up to costs." This mirrors the path history took with railroad booms and collapses.
— Ben ThompsonTraining costs force Meta to invest ruthlessly in Muse platform
Thompson wrote the S1 analysis showing training costs as the primary, uncontrollable cost center, then analyzed Meta's earnings report and realized it illuminated Meta's massive Muse spending. To become the "aggregator of aggregators" in the AI era—the ultimate user-facing layer—you must control your own models, which means you must spend heavily on training. This logic chain is indivisible; opposing Muse funding is really opposing the pursuit of Muse itself.
— Ben ThompsonMeta's push for engagement indirectly drives Netflix viewers away
Thompson surfaces a rare angle: Meta's massive AI investment pressure creates enormous investor expectations for returns, so Meta maximizes engagement in Reels and the feed—flooding users with short video and ads to keep them addicted. This directly competes for time users would otherwise spend on Netflix, making Netflix an indirect victim of the capital demands imposed by AI spending.
— Ben ThompsonAI companies will ultimately answer to state control, not self-regulation
Around the dispute over Starlink's apparent limits on Ukraine use at U.S. government direction, Thompson reiterates his earlier argument about Anthropic and government relationships: this is not a moral question of "should it be used for surveillance," but a power question. If a company claims to wield power comparable to a nation-state, the government holding a monopoly on legitimate violence will ultimately decide how that power is used. A company cannot unilaterally say no to its government.
— Ben ThompsonIn their own words · checked verbatim
the largest portion of their costs is not in fact inference the largest portion of their costs is this training cost and the speed with which they need to spend money on that is out of their control
Ben Thompson8:22
So that $80 billion is fake. It's not money out the door.
Ben Thompson16:31
First hyperscalers used free cashflow. Next hyperscalers use debt. Now hyperscalers are turning to share dilution.
Vincent Jansen19:33
into the system which means coming into anthropic coming into chat gpt are those greater than the cost of building the whole value chain ignore all the revenues going on between the layers of the value chain that's just noise all that really matters is enough money coming into the system to support the system
Craig Moffett43:25
There is an extent to, if we cross the gap, they're like the girders of the bridge that get us across.
Ben Thompson45:30
Remember that deal we made, 2023, 2024? How about an even worse one?
Ben Thompson48:37
netflix is actually a casualty of ai because ai is pushing meta to ramp up engagement to 100 engagement
Ben Thompson1:08:19
if anthropic keeps saying and insisting perhaps accurately we are accruing power to an extent that rivals nation states the nation states are going to have a say in it
Ben Thompson1:21:04
Figures
| Anthropic revenue from three major cloud platforms | ~50% | 12:26 |
| Percentage of that revenue returned to cloud platforms | ~37% | 12:26 |
| WSJ-reported Anthropic H1 loss | $80 billion | 9:25 |
| Meta planned staged share issuance scale | up to $95 billion | 56:00 |
| Meta share price decline one week after earnings | ~10% | 57:03 |
| S1 financial data cutoff date | June 30, 2026 | 22:39 |
Glossary
- S1
- The detailed disclosure document a company files with the SEC before going public, covering financials, risks, and business details.
- neoclouds
- Startups focused on renting AI compute infrastructure, distinct from traditional cloud providers like AWS.
- Starshield
- SpaceX's encrypted satellite service for the U.S. military, separated from consumer Starlink.
- benefit corporation
- A corporate structure that legally permits pursuing a mission alongside profit, not profit alone.
- aggregator of aggregators
- Thompson's term for an AI model positioned as the ultimate user-facing layer above all platforms.
- circular financing
- Investment where capital flows back to the investor through purchase of their products or services.
How to listen
Investors tracking AI industry funding and valuing Nvidia's supply chain and cloud vendors; founders trying to understand Anthropic's true financial position.
The final segment about the Brewers' championship celebration and watching Everything Everywhere All at Once can be skipped.