Brian Reed: Hey — good to be here. Busy morning.
Eliza Ward: Yeah, understatement. Anthropic dropped something this morning.
Brian Reed: Right — so the confirmation came through August 5th. Anthropic officially said they are building an in-house silicon team to design custom chips for Claude. First time they've said it out loud, though Reuters had the outline earlier.
Eliza Ward: The thing that caught me — no timeline. No specs. They're announcing a hiring push and basically nothing else is confirmed.
Brian Reed: Which is a little — I mean, that's unusual, right? You announce the team exists before you've apparently figured out what the team is building.
Eliza Ward: Wait — and Samsung is still just exploratory? The Information had that piece about manufacturing talks, but Samsung declined to comment. So there's no foundry locked in either.
Brian Reed: No foundry, no timeline, no specs — but they are hiring engineers with experience moving semiconductor designs from concept to production. That part is confirmed.
Eliza Ward: So what's confirmed is: the intention and the job postings. Everything else is still open.
Brian Reed: But here's what breaks the clean 'no timeline, no specs' read — the revenue number does a lot of work. Anthropic's annualized revenue is past $30 billion this year. That's up from something like $9 billion just months ago. At that scale, the Nvidia margin hit isn't theoretical anymore.
Eliza Ward: Right — that's the 'why now.' Not strategy, just math.
Brian Reed: Exactly. So — let me try to say this plainly, because I think the jargon buries it. The core idea of co-design is: you build the chip and the model at the same time, so the chip's architecture matches exactly how Claude thinks. Instead of squeezing Claude onto a chip that was built for everything.
Eliza Ward: It's like — okay, bad analogy incoming — it's like having shoes custom-made for your feet instead of buying the most popular size and hoping it fits.
Brian Reed: That's actually it. Every inference dollar goes further when the chip isn't carrying overhead for workloads that aren't Claude.
Eliza Ward: And the competitive pressure — wait, this is the part that I think the headline undersells. OpenAI unveiled Jalapeño on June 24th. That was six weeks before Anthropic's August 5th announcement. Jalapeño, developed with Broadcom, nine months start to finish. Meta was already moving before that. Anthropic isn't ahead of this curve.
Brian Reed: No, they're catching up. And building in-house rather than partnering with Broadcom the way OpenAI did — that's either a longer runway or a much bigger bill. Probably both.
Eliza Ward: The Reuters estimate is $500 million to $750 million just for design. At a 2nm or 3nm process node. And Anthropic has not — I mean, they've disclosed nothing about their own budget. That number is Reuters, not Anthropic.
Brian Reed: So the honest version of this announcement is: we have the revenue to justify this, staying on the sidelines was becoming a liability, and everything else — foundry, timeline, specs — still open.
Eliza Ward: And that honest version is exactly the one I don't see circulating. The take I keep running into is — Anthropic has a chip coming. As if today's announcement is a product announcement.
Brian Reed: That's the one I wanted to press on. Because — hang on — the Samsung piece from The Information specifically says those talks are exploratory and pre-blueprint. There's no foundry commitment. You can't have a chip coming if you haven't told a foundry what to build.
Eliza Ward: And the tape-out question is where it actually falls apart. Tape-out — sending the finalized design to manufacturing — that's the real milestone. Anthropic hasn't named one. Not a target date, nothing.
Brian Reed: Which means we're not even at design yet, not really. We're at — I mean, they're still hiring the people who will do the designing.
Eliza Ward: Right. So the critics on August 5th calling this a '2029 story' — that's not unfair.
Brian Reed: No, and Google's TPU ramp is the honest precedent here. That took approximately seven years to become a genuine production workhorse. And Google started with way more infrastructure than Anthropic has right now.
Eliza Ward: Wait — and the $500 million to $750 million Reuters estimate is design only. That's before manufacturing, before iteration, before any foundry deal with Samsung or anyone else actually closes.
Brian Reed: So the people saying 'chip coming' are reading an announcement about a hiring program as — actually, no, they're reading it as a delivery schedule. Those are completely different things.
Eliza Ward: And the piece I want to get into — whether the multi-chip language around AWS, Google, Nvidia, AMD is genuine hardware pluralism or something else entirely — it gets more complicated from there.
Brian Reed: And that's the thing — the multi-chip language isn't just reassurance, it's load-bearing. Anthropic cemented a long-term agreement with Google and Broadcom that's part of a broader $50 billion infrastructure commitment. That language about AWS Trainium, Google TPUs, Nvidia, AMD staying central — those partners can read the subtext. They have to be told it isn't a threat.
Eliza Ward: Right — and the $50 billion figure makes the 'we're not abandoning partners' framing actively necessary. That's not diplomatic courtesy, that's contract management.
Brian Reed: So picture this — let me try to make it concrete. An Anthropic infrastructure engineer is routing ten million Claude inference calls. Today that person looks at Nvidia pricing versus Trainium pricing and picks whichever is cheaper at that moment. The choice is real, it's external, it's competitive. Once Anthropic's chip exists, that decision — I mean, it doesn't go away, it just moves entirely in-house. That's not pluralism anymore.
Eliza Ward: That's consolidation with a delay baked in.
Brian Reed: Which is exactly what the multi-chip framing obscures. Because they're building in-house specifically so Claude's workloads run best on Anthropic silicon. If it works, why would that engineer keep routing to Trainium?
Eliza Ward: Wait — and that's the co-design logic biting back. You optimize the chip for Claude's specific computational patterns, and suddenly the general-purpose argument for Nvidia or TPUs weakens every year the custom chip matures. The multi-chip strategy is real now. It's genuinely uncertain in 2029.
Brian Reed: So what actually tells us which version this is — genuine pluralism or diplomatic cover? Because I don't think we can know yet from the announcement.
Eliza Ward: Three things. A tape-out announcement with a target date — that's the first real signal. Second, a named foundry commitment, not the Samsung exploratory talks, an actual signed partner. And third — actually, this one's probably the most telling — any disclosed shift in the ratio of external chip spend. If Anthropic's Nvidia and Trainium orders start shrinking relative to revenue, the diplomatic language stops holding.
Brian Reed: None of those exist right now. Which means the multi-chip framing gets to stay untested — and load-bearing — for however long the design phase runs.
Eliza Ward: And that's — I mean, none of those signals exist, and the program is premised entirely on compute scarcity staying painful. That's the assumption doing all the work here. If the scarcity eases before Anthropic has anything to tape out, the whole logic shifts.
Brian Reed: Which is actually the question I can't get past. The chip — if it ships, call it 2029, which the skeptics on August 5th said wasn't unfair — by then, does the problem it's solving still exist in the same shape? New foundry capacity comes online, Nvidia keeps iterating, market pressure does what market pressure does. Anthropic is locking in architectural control over Claude's specific workloads right now, betting that control is still worth something in three years. And I genuinely don't know if that bet lands.
Eliza Ward: We don't. That's not evasion — the sourced evidence just doesn't reach it.
Brian Reed: So what Anthropic actually bought on August 5th — I mean, assuming this goes well — is the option to own their stack. Whether that option is worth $500 million to $750 million in design costs alone depends entirely on whether Claude's revenue trajectory holds, and whether the supply gap that made this feel urgent is still a gap when the chip is real.
Eliza Ward: Neither of those is settled. Story's still open.
