The roadshow
On November 10, 2026, on the roadshow, Anthropic’s Tom Brown told investors that Claude Mythos Preview had brought Claude up on AMD’s MI355X in about three days — “almost 100% of the kernels,” “basically no human intervention.” The same work, he told the room, previously took dozens of engineers six to twelve months.
Sit with the shape of that claim before the paperwork arrives. Porting a frontier model to new silicon — writing and validating the kernels that make the model run on a different chip — was one of the labour moats around the frontier: slow, scarce, human work that any challenger had to redo before it could compete on new hardware. The roadshow claim is that the model crossed its own moat. A model ported itself, in days, with basically no one’s hands on it.
Compression
The port, two tellings · scale: 12 monthsDrawn to a 12-month scale: the solid bar runs to six months, the hatched extension to twelve. Three days is 0.8% of the scale — the orange sliver at the left edge is the claim.The filing gives no figure for either duration, so there is no scale to draw. The bars above are what “a fraction of the time this once required” looks like plotted honestly: a comparison with both numbers withheld.
The context is a partnership. Anthropic’s agreement with AMD — up to 2 GW of capacity — dates to July 22, and the chip in the roadshow claim is the partner’s silicon: the MI355X is the hardware the relationship runs on. None of that makes the claim false. It makes the claim placed — told to investors, on a roadshow, about a partner’s chip, in the season the partnership is being sold.
The footnote
The slide carried its own correction, in small print. The footnote limits the claim to prefill, for one model, at “a useful level of performance.” That is the claim’s actual border. Inside it: one phase of inference, one model, a standard of performance the slide does not define. Outside it: everything else a port involves.
The scope of the claim
Per the slide’s own footnoteCovered by the claim
- Prefill In scope
- One model In scope
- “A useful level of performance” The standard, undefined
Not addressed
- Decode Unaddressed
- Training Not covered
- Production tuning Not covered
- An Nvidia comparison None published
- Independent verification None
Decode is unaddressed; training and production tuning are not covered; no Nvidia comparison has been published; the claim is not independently verified.
None of this is hidden. It is printed on the slide — underneath the claim, in the type size reserved for the part of the document the room is not meant to read aloud. The roadshow number and the footnote are the same document telling two stories at two volumes, and which one a listener carries out of the room depends on which volume they were listening at.
The filing says “a fraction”
Then there is the other document: S-1/A No. 2, the amendment written by people who can be sued for what it says. It tells the same story in one careful phrase. The work, the filing says, took “a fraction of the time this once required.” That is all. No figure. No MI355X. No kernel percentage. The chip that headlined the roadshow is not named. The three days are not in it. The almost-hundred-percent is not in it.
Two documents
The same port, as told to two audiencesThe roadshow
November 10, 2026 · Addressed to buyers
TimeAbout 3 days
The chipAMD MI355X
The kernels“Almost 100% of the kernels”
The labour“Basically no human intervention”
The work beforeDozens of engineers, six to twelve months
Small print: prefill, one model, at “a useful level of performance.”
The filing
S-1/A No. 2 · Addressed to liability
“A fraction of the time this once required.”
A figure for the time — not given
The chip — not named
A kernel percentage — not given
The human role — not described
Roadshow lines as reported from the November 10 presentation; filing language quoted from S-1/A No. 2. The absences are marked ∅ — the items the roadshow specifies and the filing does not.
A filing is not a less honest roadshow. It is a document with a different reader: not a buyer pricing a future, but a plaintiff’s lawyer pricing a sentence. Every number in it is a commitment someone can be held to. So the number that was safe to say to a room of investors — three days, a named chip, almost a hundred percent — becomes, under liability, a fraction. Not a smaller number. No number.
The distance
If the claim is true at scope, the labour moat around porting models is gone. What took dozens of engineers most of a year becomes days of machine time, and the scarce thing is no longer the porting. It is verification — trusting the number. SemiAnalysis saw the shape of this in July: “generating a kernel is the easy part; trusting the number is not.” A kernel that exists is not a kernel that is right, and a port that runs is not a port whose outputs you can sign. When generation is cheap, the gate is the product.
If the claim is true only at the footnote’s scope — prefill, one model, a useful level of performance — then the roadshow sold the unfootnoted version, and the footnote is where the claim actually lives. The three days would be real and the impression false, which is the more durable arrangement: nothing said in the room was untrue, and nothing the room believed was the claim.
Either way, the two documents disagree in precision exactly where precision is priced. The one addressed to buyers rounds up: three days, a named chip, almost a hundred percent. The one addressed to liability rounds down to “a fraction.” The moat was labour, or the moat was the footnote — but the distance between the roadshow and the filing is measurable either way, and it is measured in the units each audience is owed: buyers get figures, liability gets care, and the same event is three days long in one room and a fraction long in the other.
Sourcing — graded
The porting claim itself is Anthropic’s, and it is graded unverified: no independent check of the three days, the kernel percentage, or the performance level has been published. Everything else in this piece is graded by the document it came from.