Anthropic has started laying the groundwork for its own Anthropic custom chip, and the first concrete sign is a set of early talks with Samsung Electronics about manufacturing it. The report came from The Information on July 2, 2026. Outlets including TechCrunch and The Decoder quickly followed. The discussions are preliminary. Anthropic has not settled on what the chip would do or how much power it would draw, let alone how it would slot into a server rack. Even so, the mere fact that the maker of Claude is shopping for a foundry partner tells you how the economics of frontier AI are shifting.
- Anthropic is in early talks with Samsung to build a custom AI chip, per reporting from The Information, though no design or power target has been locked.
- The company recently hired Clive Chan, an early member of OpenAI's custom-silicon team, signaling that the effort is more than idle curiosity.
- Anthropic still calls its rented mix of cloud and Nvidia hardware central to its strategy, so any in-house chip would add to that mix rather than replace it.
- The move follows OpenAI's Broadcom-built "Jalapeño" inference chip and mirrors a broader push by AI labs to control the cost of running their models.
What The Information reported about the Anthropic custom chip
The core of the story is short. Anthropic has held conversations with Samsung about fabricating a bespoke processor, and the project sits at an exploratory stage. TechCrunch reporter Lucas Ropek wrote that the company has not yet determined the chip's intended use or how it would integrate into a server, and no performance target has been set. The Decoder's Matthias Bastian framed it similarly, noting that the effort has no detailed design and no finalized power requirements.
When reporters asked for comment, Anthropic did not confirm a roadmap. Instead it pointed back to the hardware it already leans on. In a statement relayed by TechCrunch, the company said a diversified hardware stack, including chips from its cloud partners and from Nvidia, will continue to be pivotal to its compute strategy. That phrasing matters. It positions any future silicon as an addition to an existing portfolio, not a break from Nvidia or the cloud providers that host Claude today.
Several outlets, including reporting aggregated by Yahoo Finance and Bloomberg, added a technical wrinkle: Anthropic is said to be weighing Samsung's 2-nanometer manufacturing process and its advanced packaging facilities. If accurate, that points toward a high-performance accelerator rather than a modest, low-cost part. A 2nm node is leading edge, the kind of process reserved for chips meant to run demanding workloads at scale. Anthropic has not publicly confirmed the node, so treat the detail as sourced reporting rather than a settled spec.
The Clive Chan hire that turned talk into intent
A chip program is only as real as the people staffing it, and Anthropic's most telling recent move happened in the hiring pipeline. The company brought on Clive Chan, described by The Decoder as an early member of the custom-chip teams at both Tesla and OpenAI. At OpenAI, Chan spent roughly two and a half years working on the Broadcom-designed inference accelerator that the company later unveiled as "Jalapeño." Reporting suggests he is expected to build out a dedicated chip group inside Anthropic.
That detail reframes the Samsung talks. Foundry conversations can start and stall for many reasons, but a senior silicon hire is a durable commitment. You do not recruit someone who helped ship a frontier lab's first inference chip unless you intend to ship one of your own. As one analysis of the move put it, personnel tends to precede product, and roadmaps follow headcount. The Chan hire is the strongest evidence that Anthropic's chip ambitions have moved past the whiteboard.
Chan's background also hints at the likely target. His prior work centered on inference, the phase where a trained model actually answers user requests, rather than training, the far more compute-hungry phase of building the model in the first place. Inference is where a company like Anthropic feels cost most directly, because every Claude API call and every chat message runs through it. A chip tuned for that job would attack the part of the bill that scales with usage.
Inference and training place very different demands on hardware. Training rewards raw floating-point throughput and enormous memory bandwidth across thousands of chips. Inference rewards low latency and predictable, energy-efficient unit costs. A custom inference accelerator lets a lab optimize for the workload it runs most often once a model is deployed.
Why Samsung, and why now
Samsung is an interesting choice of partner because it already sits at several points in the AI supply chain. TechCrunch noted that Samsung manufactures chips for Nvidia and is collaborating with Nvidia on an AI chip factory in South Korea. The company has also been discussing chip work with Google. A foundry that already fabricates leading accelerators for the market leader is a credible place to build a competing part.
There is a financial thread too. According to reporting compiled after the news broke, Samsung took part in Anthropic's large 2026 funding round as a strategic infrastructure backer, alongside memory suppliers such as SK Hynix and Micron. When a manufacturer is also an investor, the incentives to build together line up. Anthropic gets fabrication capacity and packaging expertise, and Samsung gets a marquee AI customer at a moment when it is trying to catch TSMC at the most advanced nodes. Anthropic has not detailed the terms of any such arrangement, so the investment link should be read as context rather than a confirmed manufacturing contract.
The competitive angle for Samsung is worth dwelling on. TSMC has spent years as the default foundry for the most advanced AI accelerators, fabricating Nvidia's Blackwell GPUs and Google's TPUs among other leading accelerators. Samsung has trailed on yield and on customer confidence at the newest nodes, which is why landing a flagship AI lab as a 2nm customer would be a meaningful validation. For Anthropic, going with a foundry that is hungry for the business could translate into more favorable terms and closer collaboration than a slot in TSMC's crowded queue. It also carries the flip side of that trade: a first-generation chip on a less-proven manufacturing line is a harder engineering bet than one built on the industry's most battle-tested process.
The timing is not accidental either. Back in April 2026, Reuters reported that Anthropic was already considering developing its own AI chips to soften the impact of chip shortages. The Samsung talks look like the next step in a line of thinking that predates them. Supply has been the constraint that shapes almost every strategic decision at the frontier, and owning a slice of the silicon roadmap is one way to loosen it.
The real driver: inference margins
Strip away the hardware romance and the logic is financial. Running a frontier model is expensive, and the cost lands squarely on inference. For a company whose revenue increasingly comes from API usage and subscriptions, every point of gross margin depends on how cheaply it can serve a token. Custom silicon is the most direct lever a lab can pull on that number, because it lets the company stop paying a third party's markup on the chips that do the serving.
This is the same calculus that pushed the cloud giants into their own silicon years ago. Amazon built Trainium and Inferentia, Google built the TPU line, then Microsoft and Meta followed with Maia and MTIA. In each case the pitch was similar: a company that designs and runs its own AI infrastructure more economically keeps more of the money the infrastructure generates. The Decoder made the point plainly, describing custom chips as a cost-efficiency strategy in which the labs that run AI infrastructure more cheaply retain more revenue.
For Anthropic, the stakes are rising with scale. As Claude adoption grows across coding tools, enterprise deployments, and consumer chat, the inference bill grows with it. At that size, compute cost stops being an operating footnote and becomes a variable that shapes the whole business. A chip that shaves even a modest percentage off the cost of each served token compounds into real money across billions of calls. That is the prize, and it explains why a company that has never made hardware is suddenly hiring people who have.
Anthropic's existing compute web
A custom chip would not arrive in a vacuum. Anthropic has spent the past year assembling one of the largest and most diversified compute footprints in the industry, and any in-house silicon would plug into that web rather than replace it.
The Google relationship is the headline. In an expansion announced by Anthropic, the company committed to using up to one million Google TPUs, a deal that DataCenterDynamics reported would bring well over a gigawatt of capacity online during 2026. On the Amazon side, Anthropic continues to treat AWS as its primary training partner through Project Rainier, a cluster that started with roughly half a million Trainium2 chips spread across multiple US data centers and is set to scale toward several gigawatts over the life of the agreement.
Nvidia sits in the mix as well. Anthropic trains and serves Claude across Amazon Trainium and Google TPUs as well as Nvidia GPUs, matching workloads to whichever silicon suits them best. That multi-vendor posture is exactly why the company can insist Nvidia still matters while quietly exploring its own part. A custom inference chip would become one more option in a stack designed for flexibility, useful for specific workloads without forcing an all-or-nothing bet on any single supplier.
A diversified hardware stack that includes chips from Google, Amazon, and Nvidia will continue to be pivotal to its compute strategy.
The wider race for custom silicon
Anthropic is joining a movement, not starting one. The clearest recent precedent is OpenAI, which on June 24, 2026 unveiled its first in-house inference chip, code-named "Jalapeño" and built with Broadcom. OpenAI pitched the part on energy efficiency, the same lever Anthropic would likely pull. The two announcements landing about a week apart underline how fast the frontier labs are converging on the same conclusion: to control their destiny, they need to control their chips.
The hyperscalers got there first. Google has iterated on the TPU for close to a decade and now runs much of its own AI on that silicon. Amazon has pushed Trainium into a genuine training alternative and offers it to customers through AWS. Microsoft and Meta both operate custom accelerators tuned for their internal workloads. Broadcom has quietly become the connective tissue of this trend, co-designing accelerators for several of these players, and now Samsung is angling to add Anthropic to the list of labs it fabricates for.
What is notable is how the design work has spread beyond the traditional chip vendors. A handful of AI labs and cloud providers now employ silicon teams that would have been unthinkable for a software company a few years ago. The talent market reflects it. Engineers who cut their teeth on the TPU, on Trainium, or on OpenAI's Broadcom project are among the most sought-after people in technology, and hires like Clive Chan move the needle on which companies can credibly attempt their own designs.
The stakes for Nvidia are the subtext of the whole trend. Nvidia still supplies the overwhelming majority of AI accelerators, and its CUDA software ecosystem remains the default that most developers build on. Yet some analysts covering the custom-silicon wave have begun to model a future where Nvidia's slice of inference compute erodes as hyperscalers and labs route more of their internal workloads onto homegrown parts. Those projections apply mainly to the captive workloads that big players run for themselves, not to the broad market of teams that rent GPUs from a cloud. Even so, they capture why every additional lab building its own chip is a small pressure on the incumbent. Anthropic's statement that Nvidia still matters reads, in that light, as both sincere and strategic.
Energy efficiency is the other thread tying these efforts together. OpenAI leaned on it when introducing Jalapeño, and it is the metric that most directly governs how many tokens a data center can serve for a given power budget. Grid capacity has become one of the binding constraints on AI expansion, with new clusters measured in gigawatts rather than server counts. A chip that does more inference per watt does not just cut the electricity bill, it stretches the finite power a company can actually secure. For a lab that has reserved capacity measured in gigawatts across multiple partners, efficiency is close to a strategic weapon.
How a homegrown chip could reach Claude users
It is worth being clear-eyed about the distance between a foundry conversation and anything a customer would notice. If Anthropic did ship an inference accelerator, the effects would show up indirectly. Cheaper inference gives a company more room to hold prices steady while covering rising demand, or to run larger models at the same cost point. Lower latency can make interactive products feel snappier, which matters for coding assistants and agentic workflows where a model may be called many times in a single task.
None of that is promised by the current reporting. The chip does not exist in a shippable form, and Anthropic has been careful not to attach product claims to it. The reasonable read is that the company is trying to protect its long-term unit economics so that the pricing and capability of Claude are shaped by its own roadmap rather than dictated entirely by what its chip suppliers choose to charge. Whether users ever feel a direct benefit depends on a chain of execution that is only just beginning.
What is still unknown
For all the momentum, the honest summary is that very little about the Anthropic custom chip is settled. The company has not confirmed the project's existence in detail, let alone a foundry contract or a launch window, and certainly not a performance target. Custom silicon takes years to move from concept to production, and plenty of ambitious chip programs stall in the gap between a promising hire and a shipping part. Samsung, for its part, has spent years trying to close the manufacturing lead TSMC holds at the most advanced nodes, so betting a first-generation accelerator on a 2nm process would carry real execution risk.
There is also the question of what a custom chip actually buys Anthropic given how much capacity it has already reserved on TPUs and Trainium. Google and Amazon silicon are mature, well-supported and available at enormous scale today. An in-house part would need to beat those options on cost or performance for specific workloads to justify the investment, and it would have to do so while competing for the same scarce fabrication capacity everyone else wants. The strategic logic is sound, but the payoff depends on execution that has not happened yet.
What the Samsung talks confirm is direction, not arrival. Anthropic is telling the market, through its hiring and its foundry conversations, that it intends to own more of its compute future rather than rent all of it. Whether that intent becomes a working chip is the story to watch over the next couple of years. For now, the company is doing what every frontier lab under margin pressure eventually does: it is looking at the biggest line item in its budget and asking whether it can build a cheaper version itself.