The Library Media — Science, Technology & Health · AI News & Intelligence Desk
Three Days to a New Chip: Reading Anthropic’s Portability Claim Against Its Own Filing
By Gideon Black
Follow The Library Media
There are two Anthropic documents from the same week, and they tell different versions of the same story.
The first is a roadshow presentation for investors. In it, according to Anthropic’s November 10 investor presentation as reported by The Decoder and Simon Willison, co-founder and Chief Compute Officer Tom Brown described a future in which Anthropic trains and serves Claude across four chip families—AWS Trainium, Google TPUs, Nvidia GPUs and AMD GPUs—as something close to one interchangeable pool.
The presentation’s most arresting claim was about speed. According to that presentation, Anthropic used its frontier model Mythos, while working with AMD GPUs for the first time, to port Claude’s prompt-processing workload to AMD Instinct MI355X GPUs in about three days. Mythos, Anthropic said in the presentation, wrote or ported nearly 100% of the kernels, with almost no human intervention.
Three days to a new chip platform.
If established, that would be a remarkable claim about the economics of compute dependence.
It is not established.
The three-day figure comes from Anthropic’s own investor presentation and has not been independently verified, a point The Decoder explicitly stressed in its November 11 coverage. Anthropic’s prospectus itself uses more guarded language, saying that chip platforms now come into production “in a fraction of the time,” rather than repeating the three-day number.
That gap in language is not a minor editorial wrinkle.
It is the story.
A claim is not yet a measurement
The temptation is to argue immediately about whether Anthropic’s claim is plausible.
Resist it.
The first question is simpler: what do we actually know?
According to Anthropic’s investor presentation, the company got a specific Claude workload—prompt processing for inference—running on AMD’s MI355X GPUs in about three days, with Mythos doing nearly all of the kernel-writing or porting work.
According to The Decoder, that claim has not been independently verified.
And the reported task was not a wholesale replacement of Nvidia infrastructure. It was a specific port of Claude’s prompt-processing inference workload to a particular AMD GPU platform.
Those distinctions should survive every retelling.
“Anthropic ported its AI to AMD in three days” is a headline-sized sentence. It is also broader than the claim described in the record.
“Anthropic says Mythos ported nearly all kernels needed to get a specific Claude prompt-processing workload running on MI355X GPUs in about three days” is less convenient.
It is also much closer to what has actually been reported.
This is the desk’s rule: an investor-deck number is a claim, not a measurement, until someone outside the company can reproduce or independently verify it.
That rule is not an accusation that the company is wrong.
It is a refusal to let the provenance disappear merely because the claim is exciting.
The promise: portability changes the bargaining power
If Anthropic’s broader portability story is real and repeatable, its significance is obvious.
According to Brown’s presentation, Anthropic now trains and serves Claude across AWS Trainium, Google TPUs, Nvidia GPUs and AMD GPUs, treating the four chip families as an interchangeable compute pool.
That is a powerful proposition because hardware portability is not merely a technical convenience. It potentially changes the relationship between an AI developer and the companies supplying its compute.
The harder it is to move workloads, the more valuable the incumbent platform becomes. The easier it is to move them, the more credible the threat of leaving becomes.
And Anthropic’s presentation goes further than a simple portability claim. It introduces the concept of “compute multipliers”—research breakthroughs that multiply effective compute and can stack. As the investor-presentation transcript quoted by Simon Willison puts it, a 4x multiplier would allow 25 units of compute to do work that previously required 100.
Again, the concept is Anthropic’s presentation framing.
But the strategic argument is clear enough.
If effective compute can be multiplied through research advances, and if workloads can move rapidly among different chip families, then raw ownership of one particular hardware ecosystem becomes less decisive.
That is the roadshow’s most attractive reading of the future: software and AI labour gradually dissolve the switching costs that once locked customers into particular accelerators.
The CUDA era, in that telling, does not disappear because hardware stops mattering. It becomes less binding because the cost of crossing hardware boundaries falls.
Perhaps.
But now read the filing.
The filing speaks the language of dependency
Anthropic’s Amendment No. 2 to its S-1, filed November 10—the same date as the investor presentation—contains a rather different set of facts.
The filing discloses that Anthropic’s August 2026 compute agreement with Lambda is partly exposed to termination. About $34.9 billion of $41.3 billion in Lambda commitments is now terminable by either party because Lambda did not secure a facility lease by the contractual date.
Neither party has given notice.
Anthropic says it does not intend to terminate.
Those are not minor qualifications. They are a reminder of what compute procurement actually involves.
A compute commitment is not simply a question of whether software can run on Chip A or Chip B.
It is also facilities.
Leases.
Contractual conditions.
Counterparties.
Capacity.
The roadshow can describe four chip families as an interchangeable pool. The filing demonstrates that real-world access to compute can still depend on whether a facility lease was secured on time.
Both statements can be true.
Indeed, that is precisely why the contrast matters.
Software portability reduces one form of dependency. It does not abolish physical and contractual dependency.
The same Amendment No. 2 also increases Anthropic’s Microsoft compute commitment to at least $47.2 billion.
Data Center Dynamics additionally reported a further $4.3 billion in compute agreements with Microsoft, though that additional figure belongs to DCD’s reporting and should remain attributed accordingly.
The filing’s own confirmed $47.2 billion minimum commitment is sufficient to make the point.
Anthropic may be becoming more technically portable while simultaneously becoming more deeply committed to particular infrastructure relationships.
That is not necessarily a contradiction.
It is, however, a useful antidote to the word “interchangeable.”
Interchangeable workloads are not interchangeable worlds
The strongest version of Anthropic’s argument should be granted its due.
If Mythos can genuinely perform most of the work of porting workloads across hardware architectures with almost no human intervention, and if the roughly three-day result can be independently reproduced, that could materially alter the speed and cost of hardware diversification.
Secondary coverage has framed traditional ports of this kind as potentially requiring dozens of engineers over six to twelve months. That comparison comes from secondary coverage, not Anthropic, and should not be mistaken for an Anthropic benchmark.
Still, it identifies the question worth testing.
Can an AI system reliably compress a large amount of specialised engineering labour?
Anthropic says, in this case, that Mythos did.
The independent evidence is not yet there.
And even if the claim is eventually confirmed, a successful port of one workload is not automatically proof that every important workload can be transferred just as quickly, with the same performance and reliability, across every hardware family.
The specific reported task was Claude’s prompt-processing inference workload on AMD MI355X GPUs.
That is what must be tested.
Not a larger legend constructed around it.
The AMD partnership supplies the larger context
Anthropic’s portability claims also sit beside an existing commercial relationship with AMD.
In July 2026, AMD and Anthropic announced a partnership to deploy up to 2 gigawatts of AMD Instinct MI450-series GPUs in Helios rack-scale systems. The first gigawatt phase was planned for the first half of 2027, according to that announcement.
AMD also announced a strategic investment of up to $5 billion in Anthropic.
That context does not invalidate Anthropic’s claim about the MI355X port.
Nor does it independently verify it.
It does mean that Anthropic’s presentation about successful work on AMD hardware is occurring within a broader relationship that includes planned large-scale deployment and a strategic investment.
That is context investors should know, not a reason to dismiss the technical work.
The skeptic’s job is not to discover a sinister motive whenever commercial interests and technical claims appear in the same room.
It is to notice that they are in the same room.
Why the filing deserves more weight
Here is the uncomfortable but necessary distinction.
An investor presentation is designed to explain why an investor should be excited.
A securities filing is made under a different set of obligations and incentives.
Anthropic’s presentation says, according to the reporting on it, that the company can increasingly treat four chip families as an interchangeable compute pool and that one specific port took about three days with Mythos doing nearly all the kernel work.
Its prospectus uses the less specific phrase “a fraction of the time.”
Its Amendment No. 2 discloses a terminable $34.9 billion portion of a Lambda compute commitment and a Microsoft commitment of at least $47.2 billion.
The difference is instructive.
One document tells investors that switching is becoming easier.
The other tells investors how much depends on arrangements that are not switched by writing better kernels.
That is not proof that the roadshow is misleading.
It is evidence that “portability” is only one dimension of dependence.
The strongest contrary case
The strongest contrary case to this skepticism is simple.
Why should Anthropic be expected to solve every infrastructure problem merely because it has reduced one technical switching cost?
It should not.
A company can make substantial progress in hardware portability while still signing large compute agreements. Indeed, greater portability might allow it to diversify procurement more effectively while maintaining substantial commitments across multiple suppliers.
The Lambda situation does not prove Anthropic is locked in technologically. It concerns a contractual condition involving a facility lease.
And the Microsoft commitment does not, by itself, establish technological dependence on one chip architecture.
All true.
But none of it establishes the larger portability claim either.
That remains the essential asymmetry.
The filing provides evidence of actual commitments.
The roadshow provides Anthropic’s account of an extraordinary technical result that has not yet been independently verified.
The two should not be weighted identically merely because both come from the same company.
What an investor should wait to see
The three-day figure may be entirely accurate.
It may eventually be reproduced, independently examined and shown to represent a genuine step change in the economics of porting AI workloads.
If that happens, the implications could be substantial.
Until then, the right description is not that Anthropic has proven it can move Claude to a new chip family in three days.
The right description is that Anthropic told investors it did so, for a specific prompt-processing inference workload, using Mythos and AMD’s MI355X GPUs, and that the claim has not been independently verified.
That distinction sounds fussy.
It is supposed to.
Markets have a tendency to turn an impressive company statement into a fact, then turn the fact into a general principle, then price the general principle as though it were already operating everywhere.
The roadshow asks investors to imagine compute becoming increasingly portable.
The filing asks them to notice how large, conditional and deeply embedded the company’s actual compute commitments remain.
Both documents are Anthropic’s.
One sells the future in the language of flexibility.
The other records the present in the language of obligations.
A careful skeptic should read them together.
And should remember that three days is, for now, a claim about one specific technical achievement—not yet an independently established measurement of a new economic law.
Gideon Black
Chief Skeptic, The Library Media
From The Library — the estate of our family.