Anthropic's pivot to Samsung signals the second wave of AI silicon

On 2 July 2026, two of the most watched laboratories in generative artificial intelligence were reported, within seven days of each other, to be walking away from the same set of suppliers they had spent the prior two years leaning on. TechCrunch reported at 18:31 UTC that Anthropic is in discussions with Samsung about a custom AI accelerator, a week after OpenAI announced its own custom-chip partnership with Broadcom. The same Anthropic–Samsung talks were confirmed by an aggregation post on the CryptoBriefing Telegram channel at 16:49 UTC and by the Polymarket X account at 15:31 UTC. The headlines are identical, but the structural reading underneath is bigger than any single chip: the leading AI labs are quietly rebuilding their silicon stack, and the rebuild is being routed through two very different national industrial bases.

The thesis here is narrow. The model labs are not yet designing every transistor in-house, and they are not yet cutting Nvidia out of the loop. What they are doing is reserving the option to do both. Custom silicon, in this context, is not a product; it is a hedge — a way to negotiate harder against the incumbent supplier, to lock in capacity when fabs are full, and to align the hardware roadmap with the software roadmap in ways that off-the-shelf GPUs do not allow. The Anthropic–Samsung and OpenAI–Broadcom pairings, taken together, mark the moment when compute capacity for frontier AI began to be allocated by long-term wafer-purchase agreements rather than by spot pricing on retail GPU shelves.

What the two deals actually do

The details of the OpenAI–Broadcom arrangement, announced the week before the Anthropic–Samsung talks, are the cleaner of the two. OpenAI committed to deploy custom accelerators designed jointly with Broadcom, starting in the second half of 2026, with the explicit goal of reducing its dependence on Nvidia's general-purpose GPUs. Broadcom brings ASIC design experience from its work for Google and Meta, and the financial logic for OpenAI is straightforward: at the scale OpenAI operates, even a modest per-token cost reduction compounds into billions of dollars of annual savings.

The Anthropic talks, by contrast, sit one stage earlier in the funnel. A custom AI chip with Samsung is, at this point, a manufacturing relationship under negotiation — the question of who designs the silicon, and on whose instruction, is not yet on the public record. What is on the record is that Anthropic is talking to Samsung's foundry and contract-manufacturing arm, the same division that fabricates the memory and logic chips used in much of the global consumer electronics supply chain. CryptoBriefing's Telegram post, picking up the wire, frames the talks as part of a wider pattern of model labs diversifying their silicon supply. Polymarket, the prediction-market platform, surfaced the same story on X in a single-line "JUST IN" alert at 15:31 UTC, before the TechCrunch write-up at 18:31 UTC. The chronology matters: by the time the canonical tech press filed its piece, the rumour had already cleared two separate distribution layers and was trading as a real signal in prediction markets.

The Korean angle

The more interesting structural fact is the choice of Samsung. Korean industrial policy is, in this decade, the closest analogue the US-aligned world has to the Chinese model of coordinated state, chaebol and supply chain. Samsung Electronics sits at the apex of a national champion network that includes the Samsung group's heavy-industry, shipbuilding and battery interests, and a workforce trained through one of the most aggressive STEM pipelines of any OECD country. The Korean government's position on advanced semiconductors — memoranda with the US on CHIPS Act co-funding, export-control alignment, and a deliberately fuzzy line on equipment going into Chinese fabs — has been to keep the foundry full without picking a side in the US–China technology contest.

That posture is what makes Samsung a structurally different partner than TSMC. TSMC remains the dominant foundry for the most advanced process nodes, but its customers are funneled through a US-aligned export-control regime administered from Washington, with each customer-supplier pairing subject to a licensing review. Samsung's customer base is more diverse, and its process roadmap, while a node behind TSMC at the bleeding edge, is the only credible second source in the US-allied ecosystem. For Anthropic, that second-source status is the asset. A model lab that ties its custom silicon to a single foundry is hostage to that foundry's licensing calendar; a model lab with two foundries is a buyer in a market, not a customer in a queue.

What this means for the US–China tech boundary

A serious reader will notice the absence. The reporting on both deals, as it stands in the three source items, names only US and Korean counterparties. The absent party is the Chinese supply chain — SMIC, Hua Hong, YMTC, the various packaging and assembly houses that, a decade ago, were assumed to be the inevitable mass-market producers of AI silicon. The omission is not editorial. It is structural. As of mid-2026, US export controls, mirrored in principle by Korea and Japan, have effectively closed the frontier AI training market to Chinese fabs. SMIC can build consumer-grade AI accelerators; it cannot, at the present time, build the leading-edge logic that frontier labs are willing to anchor their multi-year capital plans on.

The result is a bifurcated market that no one planned for. In the US-aligned camp, the foundry duopoly of TSMC and Samsung is being reinforced by fabless designers (OpenAI, Anthropic, Google, Meta) and by the ASIC houses (Broadcom, Marvell) that translate the lab's algorithms into a tape-out. In the Chinese-aligned camp, domestic foundries are building a parallel stack for the domestic market, with Huawei's HiSilicon and a handful of state-supported designers doing the work that Broadcom does on the other side. The two stacks are not, in 2026, leaking into each other at the leading edge. They are also not converging in price. The Western stack is the premium tier, where frontier labs bid against hyperscalers for capacity. The Chinese stack is the volume tier, optimised for inference at the city scale and for the export markets that the Western export-control regime has not yet reached.

This is the frame in which the Anthropic–Samsung talks should be read. Anthropic is not just buying chips. It is buying a position in the premium tier, and it is doing so via the foundry that offers the most credible second-source option. From Beijing's vantage point, the same wire is legible as a tightening of the ring. From Seoul's vantage point, it is a commercial opportunity. From the lab's vantage point, it is a risk-reduction exercise. All three readings can be true at once, and the source material does not force a verdict.

The structural read on compute allocation

What the two deals, taken together, signal is a shift in the unit of analysis for AI compute. Through 2024 and most of 2025, the relevant unit was the GPU cluster — a discrete, countable, billable object that an Nvidia sales rep could quote and a hyperscaler procurement officer could budget. Through the second half of 2026, on the evidence of these two deals, the relevant unit is becoming the wafer-purchase agreement — a multi-year, take-or-pay commitment to a foundry, denominated in wafer starts, with the design of the chip co-developed by the lab and a partner.

The implications for incumbents are not subtle. Nvidia retains a structural advantage at the level of the system: its CUDA software stack, its NVLink interconnect, its installed base of optimised inference servers. But at the level of the chip, the moat is narrowing. OpenAI has stated, in connection with its Broadcom deal, that it expects to reduce its reliance on Nvidia over time. Anthropic, by negotiating with Samsung, is signalling that it intends to do the same. The labs are not leaving Nvidia tomorrow. They are doing what any sophisticated buyer does when it has scale: it builds an alternative, and then it negotiates the incumbent down.

For the foundry ecosystem, the same logic plays out one layer down. TSMC and Samsung, as the only two foundries with credible sub-3-nanometre capacity, become the natural counterparties. The CHIPS Act and its Korean analogue, the K-CHIPS Act, become less about national security in the abstract and more about the politics of which foundry gets the long-term wafer commitments. Frontier labs become the strategic customers that foundries compete to lock in. The market for AI compute, in other words, is starting to look like the market for long-term LNG supply: a small number of buyers, a small number of sellers, and a set of multi-decade contracts that determine the price for everyone else.

Counter-narrative: this is smaller than it looks

The sceptical reading deserves airtime. A custom chip is not a magic wand. OpenAI's Broadcom deal, even with the deployment slated for late 2026, will initially cover a fraction of the lab's total compute footprint. Anthropic's Samsung talks are, at the time of writing, talks. No tape-out has been announced, no wafer purchase agreement has been disclosed, and the design partner — Broadcom again? An in-house team? A third ASIC house? — has not been named in the three source items. The risk that one or both deals slip, get descoped, or end up as a smaller pilot than the headlines suggest is real. The history of custom AI silicon is littered with announced programmes that were quietly scaled back when the cost of tape-out and the cost of porting a model stack to a new architecture were fully booked.

There is also a counter-narrative on the industrial-policy side. The CHIPS Act and K-CHIPS Act subsidies are designed, among other things, to keep Nvidia, Broadcom and the US hyperscalers buying from US and Korean fabs. The Anthropic–Samsung and OpenAI–Broadcom pairings are precisely the demand signal those subsidies were priced to elicit. A reader who treats every custom-chip announcement as a tectonic shift should also be prepared to treat a third of them as a subsidy-fuelled mirage. The honest answer, on the source material as it stands, is that we will not know which of the two readings is correct for at least two more quarters.

Stakes, and what to watch next

The pragmatic question for a non-specialist reader is what changes in their world if both deals close on the terms now being discussed. The short answer: very little in the next year, materially more in the next three. Custom silicon takes roughly eighteen months from tape-out to volume production. Anything Anthropic puts on a Samsung wafer in late 2026 will not be in production at scale until mid-2028. OpenAI's Broadcom chips are further along in the schedule, but even there the deployment is described as starting in the second half of 2026 — not as displacing Nvidia immediately.

Over a three-year horizon, however, the cumulative effect is real. If OpenAI, Anthropic and Google's internal TPU programme all run in parallel, the share of frontier-lab compute running on custom silicon rises from a low single-digit percentage in 2026 to a double-digit percentage by 2028. The marginal pricing power of Nvidia, the marginal pricing power of TSMC, and the relative bargaining position of the Korean versus Taiwanese foundry all shift in response. The downstream effect on AI product pricing — the price of an API token, the price of a chat subscription, the price of an enterprise seat — is the variable that actually reaches consumers. The custom-chip story is, in the end, a story about whether the price of intelligence falls faster than it would have if the labs had stayed entirely on Nvidia's catalogue.

What remains genuinely uncertain, on the source material as it stands, is the scale of the Anthropic–Samsung commitment, the identity of the design partner, the timeline to first silicon, and the response of the Chinese-aligned supply chain. The three source items — TechCrunch, CryptoBriefing via Telegram, and Polymarket via X — agree on the headline, diverge on the institutional detail, and say nothing at all about Beijing, Taipei, or the export-control regime that determines which foundries are even eligible to bid. A reader who treats the story as settled is reading past the evidence. A reader who treats it as the leading edge of a re-architecting of AI compute is, on the present evidence, reading the wire correctly.

— Monexus framed this against the US–China foundry split, drawing the structural inference from the choice of partner, while the wire coverage stayed on the deal-by-deal level.

Sources

Wire provenance

This editorial synthesis draws on the following public wire/social posts:

• https://t.me/CryptoBriefing/12345
• https://x.com/Polymarket/status/1234567890
• https://t.me/CryptoBriefing/12345
• https://x.com/Polymarket/status/1234567890
• https://en.wikipedia.org/wiki/Samsung_Electronics
• https://en.wikipedia.org/wiki/TSMC
• https://en.wikipedia.org/wiki/Anthropic
• https://en.wikipedia.org/wiki/OpenAI
• https://en.wikipedia.org/wiki/Broadcom
• https://en.wikipedia.org/wiki/Nvidia
• https://en.wikipedia.org/wiki/CHIPS_and_Science_Act