Adam: There are two statements about Anthropic's hardware strategy. One of them is incomplete.
Adam: Anthropic's on-record comment to TechCrunch: compute strategy centered on Google, Amazon, Nvidia. That's the public position.
Adam: The Information reported it — Bloomberg confirmed it — Anthropic is in early-stage talks with Samsung Electronics to manufacture a custom AI accelerator chip.
Adam: Anthropic was asked directly. Didn't confirm. Didn't deny.
Adam: Here's what I think actually happened — and this is the argument — Anthropic is moving on custom silicon while holding the public posture of a company that isn't. Because the moment you say you're building your own chip, you're making a claim about your roadmap, your foundry relationship, your workload targets. None of those are locked in. The Samsung talks are early-stage. Nothing is finalized.
Adam: But the context makes the move legible, even if Anthropic won't confirm it.
Adam: June 24, 2026 — OpenAI and Broadcom unveil a chip. Jalapeño. Custom inference ASIC. Built in roughly nine months. Claiming 50% lower cost per token compared to Nvidia GPUs. Production target: Q4 2026.
Adam: Inference is what this is about. Not training — inference. Running a deployed model at scale, generating outputs for real users. That's where the costs compound. Estimated 70% of Nvidia's data center GPU revenue comes from inference workloads.
Adam: An ASIC is a chip designed for exactly one category of task. Not flexible. Not general-purpose. But for that task — faster, and cheaper than a GPU. The risk is that AI workloads evolve and the chip doesn't. That's a REAL risk. But look at what Jalapeño is claiming and tell me the calculus hasn't shifted.
Adam: Vertical integration — owning your chip design rather than buying from Nvidia — is now the dominant strategic move at the frontier. OpenAI did it. Google built TPUs in 2016. Amazon has Trainium and Inferentia. Anthropic is — quietly, apparently — next.
Adam: Here's what that late June 2026 moment actually was. OpenAI, Etched, Amazon, SambaNova — all crossing from prototype to shipping product at the same time. That's not a trend line. That's a break.
Adam: And the number every one of them is pointing at is the same number. Inference — estimated 70% of Nvidia's data center GPU revenue. That's the target. Not training, not research workloads. The part of the business where a deployed model generates outputs for real users, at scale, every second it's running.
Adam: Every custom ASIC announced this year is aimed at that specific number.
Adam: Jalapeño claims 50% lower cost per token than Nvidia GPUs. Nine months from concept to product. That's the number OpenAI and Broadcom want you to hold onto. And look — it might be right. But that figure comes from sources who need it to be true. There is no independent verification. None.
Adam: Worth sitting with that.
Adam: And then there's the other thing cost-per-token doesn't capture. CUDA. Nvidia's software ecosystem is a decade of tooling, libraries, workflows — it's embedded in every serious AI research stack. A better cost number does not dissolve that overnight. You don't migrate off CUDA the way you switch cloud providers. The switching cost is real, and it's not on any spec sheet.
Adam: Google has been running TPUs since 2016. A decade. And Nvidia still owns the market.
Adam: Amazon's Trainium and Inferentia are now in talks to sell to external data centers — not just internal use. Meta is moving on custom silicon. The structural shift is real. But the question is not whether custom chips are cheaper. The question is whether cheaper is ENOUGH.
Adam: Cheaper than Nvidia, locked to one workload, with a migration cost nobody's fully accounting for — and AI workloads are still evolving fast enough that the chip you design today may not fit the model architecture you're running in two years.
Adam: That's the real risk. Not that the ASICs fail. That they succeed — and then the workload moves. Nvidia's moat isn't just the hardware. It's the optionality. And that's what a custom chip, by definition, gives up.
Adam: Here's what the Samsung talks don't have yet. No finalized workload targets. No performance spec. No server integration plan. The chip is described as nascent — and that word is doing a lot of work. Nascent means it may never reach tape-out.
Adam: That's not a minor gap.
Adam: The workload target is the engineering brief. If Anthropic hasn't decided what Claude inference profile the chip is optimized for, Samsung is being asked to manufacture something that doesn't have a definition yet.
Adam: What Samsung brings to the table is real — 2-nanometer process node, advanced packaging, high-density chip stacking, high-bandwidth memory integration. Those are serious manufacturing options. But manufacturing capability without a locked target spec is just potential.
Adam: And TSMC is sitting right there.
Adam: Samsung's yield rates at advanced nodes are not established at the same level as TSMC's. That's a real consideration. Anthropic would be accepting more manufacturing risk to avoid — what, exactly? A supplier relationship with the dominant foundry? That choice, if they make it, will tell you something about what's actually driving the Samsung talks.
Adam: Then there's the obsolescence trap. Chip design cycles run two to three years. AI model architectures are moving faster than that. A chip locked to today's Claude inference profile — today's attention patterns, today's context window assumptions — could be partially obsolete before it ships. That's not a hypothetical. That's the structural problem with every ASIC in a fast-moving stack.
Adam: Reuters flagged earlier intent — April 2026, Anthropic was already exploring its own chip production as a response to chip shortages. So this isn't a sudden pivot. But intent and execution are separated by exactly the decisions that aren't finalized.
Adam: The nearest hard date in this space is Jalapeño's Q4 2026 production deployment. That lands before Anthropic's Samsung talks have likely moved past early-stage. That matters. Because Q4 2026 is the first real-world data point — actual cost, actual performance, actual scale — on whether custom inference ASICs deliver what they claim. If Jalapeño stumbles, every lab rethinks the calculus. If it holds, the pressure on Anthropic accelerates.
Adam: Either way, Anthropic is watching it happen from outside.
Adam: The tell — the thing worth watching if any Anthropic chip eventually ships — is one decision. Claude-only, or open ecosystem. A chip optimized solely for Claude workloads is a cost-reduction tool. An open-ecosystem chip is an attempt to compete with Nvidia at the infrastructure layer. Those are different companies. Different risk profiles. Different claims on the market.
Adam: That single design choice will reveal whether Anthropic is building a genuine Nvidia alternative — or building a toll road on top of Nvidia's foundation. Watch for it.
Adam: Anthropic told TechCrunch that Google, Amazon, and Nvidia remain central. That's the on-record position. And then The Information reports Samsung talks — early-stage, no locked spec, no workload target — and Anthropic doesn't confirm, doesn't deny. Both of those things are true at the same time.
Adam: What that tells you — and honestly, sit with this — is that these chip programs aren't really about replacing Nvidia. Not yet. Maybe not ever. They're about having the conversation. Having Samsung in the room. Having a chip in early development that you can point to when the next supply crunch comes, or when Nvidia's pricing moves, or when OpenAI's Jalapeño actually ships in Q4 and the board wants to know what Anthropic's answer is.
Adam: That's leverage. Not infrastructure.
Adam: Anthropic's public statement and its private Samsung talks aren't a contradiction — they're the same strategy. The chip is the negotiating position.
