Anthropic Moves to Design Its Own Claude Chips, With Samsung's 2nm in Its Sights
Anthropic confirmed this week it is standing up an in-house silicon team to design custom AI chips built around Claude, making the model maker the last of the big frontier labs to commit to its own hardware. The announcement, first reported by Business Insider and picked up by TechCrunch, is still thin on specifics — no chip specifications, no production date, no named foundry partner. The broad shape of the plan is clear enough, though, and it has the semiconductor world watching Samsung Foundry as the likely beneficiary.
The company is hiring for a hardware team that will work alongside its model group rather than underneath it. That is the core of the co-design approach: the silicon is tuned to how Claude actually runs, instead of forcing Claude to fit what general-purpose hardware already does. According to the reports, senior chip engineers on the effort are being offered annual salaries near $485,000 — a clear sign Anthropic is ready to pay for scarce talent in a part of the industry that has grown crowded.
The hiring itself is the firmest evidence that this is a real program and not a rumor. Anthropic had already been quietly adding people who understand what it takes to build silicon from scratch, and this week's confirmation restates that commitment in public. The company has not said how large the team will be, where it sits inside the organization, or whether it reports directly to the engineering leadership that runs Claude's training and serving. Those details will matter to suppliers that are trying to gauge how serious the effort is, because a hardware program that lacks internal backing tends to stall once design work gets hard.
Samsung's Foundry Could Be the Manufacturing Partner
The chip-design announcement is the sharper echo of something that surfaced months ago. In early July, reporting from The Information and Bloomberg said Anthropic had held early talks with Samsung Electronics about producing a custom AI chip. The specific pieces under discussion were Samsung's SF2P 2-nanometer foundry process and its advanced packaging lines — both of which would represent a meaningful order for a contract-manufacturing business that has spent years trying to close the gap with TSMC.
Those discussions were, and remain, preliminary. Sources said Anthropic was still deciding what the chip should do, what its power limits should be, and how it would fit into the company's server infrastructure. Nothing had moved into detailed design, testing, or manufacturing when the earlier reports ran. Even so, the direction of travel is unmistakable, and it lines up with an existing tie: Samsung, along with SK Hynix and Micron, joined Anthropic's $65 billion Series H round in May as a strategic infrastructure partner. Of those three memory makers, only Samsung runs a foundry business, so the link between the funding and the manufacturing chatter was easy to see.
Anthropic has also been pulling in people who have done this before. The Information reported that Clive Chan, an early member of OpenAI's own custom-chip team, has joined the effort, part of a deliberate engineering build-out that predated this week's formal announcement.
A Crowded Field That's Been at It for Years
Anthropic is arriving a bit late. Google runs its own TPU infrastructure at scale. Amazon has Trainium3. Meta fields its MTIA chips, and Microsoft deployed its Maia 200 silicon for internal workloads. OpenAI moved the earliest and the farthest in public, teaming with Broadcom in 2024 and later unveiling "Jalapeño," an inference chip that the company says turns in better performance per watt than general-purpose parts. That chip reportedly went from design to announcement in nine months, with AI models used inside the design flow itself.
The milestone on the clock for Anthropic is therefore somewhere around 2028 or later. Between hiring a team, defining the chip, tapping it out, qualifying manufacturing yields, and wiring it into Claude's serving stack, custom silicon does not arrive quickly. The company has said a multi-chip strategy stays in force through it all: cloud-provider Trainium chips, Google TPUs, Nvidia GPUs, and AMD hardware all remain part of the picture.
Why the Push Matters Beyond the Chip
The business reason for the shift is straightforward. Nvidia still holds an estimated 74% share of the AI chip market, so every lab that buys compute is effectively renting from one supplier. Custom silicon gives a lab control of its own cost structure, and that cost structure decides which models are affordable to serve and at what price. Midjourney's public numbers offer a sense of the scale: when the image tool moved its inference off Nvidia GPUs and onto Google TPUs, its monthly compute spend fell from $2.1 million to $700,000, a drop of about 65%. At that ratio, a hardware story starts to look like an API-pricing story, because lower inference costs can flow through as cheaper token prices or let a lab run bigger models for the same money.
Latency matters just as much. A general-purpose GPU carries overhead built for jobs that Claude never performs. A chip built around the transformer attention mechanism, with custom memory bandwidth and custom interconnects, can shed that overhead and hold more consistent utilization. The result is a steadier time-to-first-token, which agents and real-time tool calls notice quickly.
The suppliers feel it first. A marquee AI client like Anthropic landing on Samsung's 2nm would be a welcome win for a foundry that has struggled to match TSMC's yields on leading-edge ramp-ups. The Information has also suggested Google is separately weighing Samsung for a future tensor processing unit, which would cement any contract-manufacturing momentum.
There is a constraint that keeps the timeline honest, and it is physics as much as business. A design that has not even settled on its transistor budget cannot be taped out, and a tape-out that has not happened cannot be qualified through manufacturing yields. Each of those steps runs on a calendar of months, not weeks, and any change of course along the way pushes the whole schedule out. That is why observers put the realistic first production window at 2028 at the earliest, and why nobody in the industry treats this week's announcement as the start of a countdown. It is the start of a decision process.
What Changes Now
In the near term, the announcement changes very little for developers. The Claude API prices and rate limits in force today are not about to move because of a hiring announcement. The realistic window for the first Anthropic chips to reach production is 2028 at the earliest. What has changed is the strategic direction: vertical hardware integration is now a standard move for frontier labs. The one that controls its own silicon controls its own economics, and that edge now separates the leading players from those that stay dependent on a single vendor's prices and allocations.
Anthropic is betting it can close that gap. The chip the team eventually designs will run where the money already matters, and the semiconductors world is watching which foundry finally wins the business.
TrendForce laid out the Samsung talks in detail, and the specifics on the 2nm process and packaging come from Yonhap News and The Information's original report.