Steve Eisman says Anthropic's future capital commitments total roughly $800bn to $900bn, yet they appear only in a footnote rather than the income statement.
Eisman says the S-1 omits third-quarter figures and shows revenue-to-compute ratio falling to 1.7 times as token maxing ended and sequential growth slowed from 147% to 45%.
Eisman argues Anthropic has no moats, so cutting prices by about 50% while committed to huge spending is dangerous, and Elson expects losses to rise.
Anthropic's IPO filing hides slowing growth and vast spending
Anthropic's S-1 looks strong on revenue growth, but Steve Eisman argues it hides the risk ahead. Growth is slowing after the end of token maxing, and more than $800bn of compute commitments have yet to reach the income statement. Ed Elson concludes that even at a possible $2tn valuation, it remains a highly risky venture bet rather than a proven safe investment.
Takeaways
Huge compute commitments have not hit earnings. Anthropic's future capital commitments total roughly $800bn to $900bn depending on what is counted, including a $110bn Amazon contract. They appear only in a footnote now but will eventually flow through the income statement and push losses up.
The S-1 is backward looking and stops at Q2. The filing gives numbers only through the second quarter, and Eisman has heard third-quarter results may be required before the IPO. Charts omit that quarter, and some speculate this was deliberate. He expects sentiment to turn sharply if the numbers disappoint.
The end of token maxing slowed growth. Token maxing was the early-2026 push to use AI regardless of cost. Revenue to compute expense rose from 0.69 times to 2.08 times, then fell to 1.7 times in the third quarter as sequential growth dropped from 147% to 45%.
Price cuts are risky without moats. An unconfirmed report says Anthropic cut prices about 50% in October to regain share as customers grew cost conscious and Chinese open-weight models gained. Eisman calls it the data point to watch, since competing on price while committed to huge spending is dangerous.
Governance leaves shareholders without rights. A class T share held by the Long-Term Benefit Trust chooses the board, and one class F share held by seven cofounders carries 50.1% of the vote. Eisman says objectors should not invest, and notes Meta draws the same criticism yet performed well.
Waymo leads robotaxis while Tesla wins on cost. Waymo runs in 15 US cities with 500,000 paid rides weekly. Tesla operates in six cities and has about 1.7m paying FSD customers, but its Cybercabs cost about $25,000 to build against more than $100,000 for Waymo vehicles.
Public support and regulation hold autonomy back. US rules vary state by state, and Waymo has sharply raised federal lobbying. Americans agreeing that autonomous vehicles improve safety was 52% in 2018 and fell 12% in a poll last year, though 94% of crashes stem from human error.
Autonomous vehicle investment angles
LiDAR makers, since 95% of passenger car LiDAR sensors are made in China.
Car rental firms such as Hertz, which is building an autonomous vehicle servicing business.
Tap a name to go there · tap a line to see why they are linked
On this episode 3
Show 1
Trucking, where driver pay is about 44% of marginal costs and Aurora is testing self-driving trucks.
The Middle East and Europe, where the real contest is expected and countries have so far mostly chosen Chinese companies.