“It’s now very clear that the whole “inference is profitable” and “most compute is being used for training” myths are dead, because if they weren’t, Anthropic would either need way more compute or way higher-quality compute.”
Our reading:
The claims that inference is profitable and that most compute goes to training have been disproven, as shown by Anthropic's compute purchases.
Assessment: Unresolved
What happened
Was it true for the period described?
Outcome: Unresolved
We checked, and the evidence so far doesn’t settle it.
Judged on a plain reading. They named no measure, so we used the reading a typical informed reader would take. It is spelled out under “How we judged it”.
Evidence up to . Checked .
Why
Anthropic's filing provides first-half 2026 financial figures and expense classifications, but does not establish inference profit for January–March 2026 [e1]. Crusoe's 2026 commentary says inference exceeds training across clouds, but supplies no measured training share for Anthropic in that quarter [e2]. Neither component is settled by these records [e1, e2].
Caveats
- First-half consolidated financial figures do not establish first-quarter inference revenue and serving costs on a consistent basis.
- Industry workload commentary does not measure Anthropic's training share in comparable computational units.
- The broader company scope and profitability cost boundary remain unspecified.
Evidence (2 items)
Context only: Anthropic / U.S. Securities and Exchange Commission,
personnel-related costs and allocated overhead costs
Why it matters: The filing's first-half revenue exceeds its reported cost of revenue, which cautions against a blanket inference-loss conclusion. However, free access and other costs appear in other expense categories, and these first-half figures do not establish the specified quarter's inference profit or physical training share.
- Publisher
- Anthropic / U.S. Securities and Exchange Commission
- Kind of source
- Regulatory filing, unaudited
- Measures
- Anthropic, Consolidated cost of revenue, USD millions, gaap, 2026-01-01 to 2026-06-30
- About events on
- Published
Context only: Crusoe,
across all clouds consolidated.
Why it matters: This supplier statement is consistent with the training-share component at an industry level. It supplies neither a measured Anthropic allocation nor a first-quarter result; the page displays both April 8 and September 30 dates, so the later date is used conservatively.
- Publisher
- Crusoe
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- About events on
- Published
How we judged it
- Kind of claim
- Observation
- Deadline
- None given. We don’t add one.
Where we had to interpret
- The ordinary reading rejects two propositions: inference earns more than it costs to serve, and training consumes more than half of compute.
- The immediate example is Anthropic, while the wording generalizes to AI labs. Anthropic is therefore the minimum relevant company scope; industry-wide conclusions require broader evidence.
- For this ongoing observation without a named reporting period, use January through March 2026, the latest completed calendar quarter before publication.
- Profitability requires matched inference receipts and serving costs. Company-wide net losses do not establish inference losses.
- Compute means computational resources used, not automatically expenditure. Research expense also need not consist entirely of training.
Right if: Matched evidence establishes that inference was not profitable and that training used no more than half of compute in the relevant company scope and target quarter.
Wrong if: Matched evidence establishes positive inference profit or a training share above half. A mixed result requires direct evidence settling the two components in opposite directions.
We wrote this test after the outcome was known.
What would settle it
Matched January–March 2026 inference revenue and serving costs with an explicit cost boundary, plus training and inference usage in comparable computational units. Anthropic-specific records are required; broader conclusions require equivalent records for a defined set of labs.
We found 36 claims in this post. See them all.
Full record
Source and dates
Where's Your Ed At, Where Are All The Data Centers?, paragraph 197.
- Said
- We read the source
- live page, full text, public
- Evidence up to
- Checked
We read this wording on 2026-10-28. The post may have been edited after 2026-05-12.
Where it was said (1)
Said once in the posts we read.
: Where's Your Ed At, Where Are All The Data Centers?
It’s now very clear that the whole “inference is profitable” and “most compute is being used for training” myths are dead, because if they weren’t, Anthropic would either need way more compute or way higher-quality compute.
- Where
- paragraph 197
- Published
- 2026-05-12
- We read it
- , live page
- Access
- full text, public
History
Versions
- Current: Unresolved. Evidence up to 2026-10-28.
What happened to this record
PublishedBy us
Published after both AI reviews and the final check passed.
Held backBy us
The final check failed: a cited source didn’t say what the write-up said.