Anthropic negotiates its custom chip with Samsung and prepares its IPO for October: AI's most disciplined lab
🔎 Why Anthropic is building its own silicon — and why it's the strongest signal in the industry
Anthropic just filed a confidential S-1 with the SEC and has begun discussions with Samsung to manufacture a custom 2nm AI chip. Two announcements separated by a few days, but they tell the same story: the San Francisco lab no longer wants to be just a cloud compute customer. It wants to control its value chain end-to-end, exactly as Google, Amazon, and Microsoft did before it.
The timing is not coincidental. OpenAI announced its Jalapeño chip with Broadcom a week before Anthropic released its own news. The two frontier AI leaders are converging on the same conclusion: renting Nvidia GPUs at $30,000 a pop is not a sustainable strategy at the scale of tens of billions of dollars in revenue.
Except Anthropic is playing this game with a radically different financial position than OpenAI.
The essentials
- Anthropic confidentially filed an S-1 registration statement with the SEC, paving the way for an IPO expected in October 2026.
- The lab is in talks with Samsung Foundry to manufacture a custom AI chip, specifically considering the 2nm process and advanced packaging.
- Anthropic is showing a revenue run rate of $47 billion (May 2026), with a trajectory toward $93 billion by May 2027.
- Samsung participated in Anthropic's $65 billion Series H in May 2026, transforming an investment relationship into a supply discussion.
- The custom chip aims to reduce compute costs — Anthropic's largest expense item — while decreasing dependence on Nvidia.
- Samsung's 2nm foundry is lagging behind TSMC in yield, making this bet risky in the medium term.
Recommended tools
| Hostinger | Web hosting for AI startups | starting at €2.99 (July 2026, check on hostinger.com) | Landing pages and product docs |
|---|---|---|---|
$47 billion run rate: Anthropic's revenue machine
Anthropic went from a $9 billion run rate in late 2025 to $47 billion annualized in May 2026. That is growth of over 400% in six months, a pace that has almost no equivalent in the history of enterprise software.
Two engines explain this acceleration. Claude Code has become the reference tool for coding assistants in large enterprises. And the adoption of Claude for agentic workflows — autonomous tasks, tool chains, decision-making — is exploding in the finance, healthcare, and legal sectors. Anthropic has thus surpassed OpenAI in revenue in the frontier AI race, a situation that seemed unthinkable just a year ago.
FutureSearch forecasts place Anthropic's ARR at nearly $93 billion by May 2027. If these projections materialize, the company would not only be the AI revenue leader, but also potentially profitable as early as 2026. This is a crucial detail for the IPO: public investors are no longer paying for promises of growth without profit. They want to see unit economics that hold up.
With a 20x multiple on revenue, the Series H valuation reaches $965 billion. A figure that places Anthropic among the very top tier, a hair's breadth from OpenAI's $730 billion valuation — except that Anthropic generates significantly more revenue with a higher valuation.
The lead banks on the deal are Goldman Sachs, Morgan Stanley, and JPMorgan. A trio signaling to the markets that Anthropic is aiming for an institutional-sized offering, not a niche tech IPO.
The confidential S-1: an IPO calibrated for October 2026
Filing a confidential S-1 is a standard yet strategic step. It allows Anthropic to negotiate with the SEC privately, adjust the figures, and avoid exposing its finances publicly until it is ready. According to Luminix, the listing is expected for October 2026.
The choice of October is not neutral. It falls after the half-year earnings season, when the markets have digested the first-half figures. It is also before the traditional end-of-year volatility. Anthropic wants a window of maximum attention.
The contrast with OpenAI's path to its IPO is striking. OpenAI is navigating lawsuits with Elon Musk, debates over government equity participation, and a governance structure that already exploded once in November 2023. Anthropic, on the other hand, has maintained a more classic control structure with a Benefit Corporation framework that reassures institutional investors.
The long-term compute deals already signed by Anthropic also play a key role in this IPO preparation. Public investors love revenue predictability. When Anthropic signs a three-year contract with a Fortune 500 company to deploy Claude at scale, it's not just revenue: it's an indicator of retention and growth that is visible in financial models.
The custom chip with Samsung: reducing compute cost, the real battle
Compute represents the majority of operational costs for a frontier AI lab. Each Claude Opus 4.7 request consumes GPU resources that amount to fractions of a cent, but multiplied by billions of calls, the bill becomes astronomical. Building a custom chip is therefore a defensive move as much as an offensive one.
Anthropic has started preliminary work on its own AI chip and is holding discussions with Samsung Electronics as a manufacturing partner, according to The Information. The company is specifically considering Samsung Foundry's 2-nanometer process, coupled with advanced packaging. TechCrunch points out that Anthropic has not yet decided exactly what the chip will be used for or how it will integrate into its existing infrastructure.
This is an important point: we are at the stage of preliminary discussions, not an imminent tape-out. Anthropic is exploring. But the strategic signal is clear. Following OpenAI's Jalapeño a week later shows that market pressure is pushing all players toward the same axis of optimization.
The objective is twofold. First, reduce the cost per token by eliminating Nvidia's margins on H100/B200 GPUs and the margins of cloud providers that rent out these GPUs. Second, specialize the silicon for Claude's architectures — optimized inference, memory management tailored to long contexts, agentic routing. An Nvidia GPU is a general-purpose product. An Anthropic-Samsung chip would be tailored for a specific workload.
Samsung Foundry in 2nm: the risky bet on a foundry that isn't TSMC
This is the weak point of the whole story. Samsung Foundry has a documented yield lag on leading-edge nodes compared to TSMC. Yield — the percentage of functional chips per wafer — directly determines the unit cost. A yield of 60% versus 90% at TSMC radically changes the economics of a custom chip.
TrendForce reports that Anthropic is considering Samsung's 2nm as well as advanced packaging, two areas where Samsung is investing heavily but where TSMC remains dominant. SamMobile adds that Samsung Foundry would simultaneously be in talks with Anthropic and Meta for 2nm chips, which suggests that Samsung is aggressively positioning its foundry as the alternative to TSMC for AI clients who want capacity and a less monopolistic partner.
The problem: if Samsung's 2nm yield does not improve sufficiently, the effective cost of the Anthropic chip could be higher than that of an off-the-shelf Nvidia GPU. This is the worst-case scenario that several custom chip attempts have encountered in the past.
Anthropic knows this. This is why the discussions remain preliminary and why the company has not yet defined the exact use case of the chip. The caution is justified. But not engaging in the discussion would have been worse: in 18 months, if TSMC is saturated by Apple, Nvidia and Google, Anthropic would have no backup option.
From Series H to the foundry: how an investment becomes a strategic partnership
Samsung participated in Anthropic's Series H, which closed at $65 billion in May 2026, alongside SK Hynix and Micron. According to AI Weekly, this participation positions chip discussions as an investment relationship that is deepening into a supply relationship.
This is a classic pattern in the silicon industry. When a foundry invests in a client, it gets two things: visibility into future demand and a seat at the table where architecture decisions are made. For Anthropic, it's the reverse: an investor that also becomes a supplier creates a powerful alignment of interests. Samsung has every interest in Anthropic's chip succeeding, because its investment depends on it.
The simultaneous presence of SK Hynix and Micron in the Series H suggests that Anthropic is building a complete memory ecosystem around its future chip. HBM (High Bandwidth Memory) is the most critical component for an AI chip — often more limiting than the compute itself. Having the big three of memory as investors guarantees priority supply in a market where demand far exceeds supply.
Anthropic vs OpenAI : two silicon strategies for two opposing philosophies
The contrast between the two labs' approaches is revealing. OpenAI announced Jalapeño, its custom inference chip with Broadcom, with the explicit goal of reducing costs by 50%. Broadcom is a fabless company specializing in design and packaging — not a foundry. OpenAI will still need to find someone to manufacture it, most likely TSMC.
Anthropic, on the other hand, is going straight to the foundry: Samsung. The advantage is a more integrated relationship. The risk is committing to a foundry whose 2nm yield is uncertain. Technology.org summarizes the dynamic well: both labs are following the same playbook — controlling their silicon — but via different paths.
Financially, Anthropic surpasses OpenAI in revenue with $47 billion against a valuation exceeding $965 billion. OpenAI, valued at $730 billion, is preparing its own IPO in a more chaotic context. The question investors are asking is simple: which of the two business models is more sustainable?
Anthropic is betting on operational discipline: higher revenue, stable governance, and costs optimized by custom silicon. OpenAI is betting on the network effect and brand. In the short term, both work. In the long term, unit economics will decide.
What the Anthropic-Samsung chip would change for Claude users
For developers and companies using Claude Opus 4.7 or Claude Sonnet 4.6 on a daily basis, a custom chip won't change anything tomorrow. The AI chip development cycle — from design to scaled deployment — typically takes 2 to 4 years.
But in the medium term, the impact would be significant. A chip optimized for Claude inference could reduce latency on agentic tasks, where every second counts. It could also enable wider context windows without a linear increase in cost, opening up new use cases in document analysis and code review.
The table below illustrates the current hierarchy of Anthropic models and their competitive positioning, as established by the June 2025 benchmarks:
| Model | Type | Overall Score | Agentic Score |
|---|---|---|---|
| Claude Opus 4.7 (Adaptive) | General / Agentic | 90 | 94.3 |
| Claude Opus 4.6 | General / Agentic | 87 | 84.7 |
| Claude Sonnet 4.6 | General / Agentic | 83 | 81.4 |
A custom chip could notably bridge the gap between Opus 4.7 at 94.3 in agentic and the leaders GPT-5.5 at 98.2 or Gemini 3 Pro Deep Think at 95.4. Not by improving the model itself, but by enabling it to be served more efficiently and at a lower cost.
The industry context: why every lab wants its own chip in 2026
Anthropic is not an isolated case. The trend is clear: any AI player that reaches a certain revenue scale starts designing its own silicon. Google has had its TPUs since 2016. Amazon has Trainium and Inferentia. Microsoft has Maia. Meta has MTIA.
The reason is structural. The AI GPU market is dominated by Nvidia with gross margins exceeding 75%. When your compute cost represents 60 to 70% of your total server cost, every margin point ceded to Nvidia is a point that does not go into your profit.
UPI points out that AI developers are explicitly looking to reduce costs and their dependence on Nvidia. The problem is that compute providers — AWS, Azure, GCP — also serve your rivals. When Anthropic rents GPUs on Azure, Microsoft rents the same GPUs to OpenAI, its exclusive partner. There is no competitive advantage in rented compute. Only proprietary silicon creates one.
Yahoo Finance recalls that Anthropic's annualized revenue recently exceeded $47 billion, fueled by the explosive adoption of Claude in the enterprise for coding and agentic workflows. At this scale, custom compute is no longer a luxury. It is an economic necessity.
The specific risks of the Samsung bet
Despite all the arguments in favor of a custom chip, the risks are real and Anthropic is probably well aware of them.
The yield risk
This is the number one risk. If Samsung fails to achieve a competitive yield at 2nm, the unit cost of each Anthropic chip will be too high to justify the design investment. Defective chips are literally thrown away — and the cost of 2nm wafers is considerable.
The timing risk
Anthropic is preparing an IPO for October 2026. A custom chip will not be ready by that date. Investors will evaluate the company based on its current revenue and trajectory, not on silicon that currently only exists in discussions. If the market turns before the chip is ready, the strategic argument loses its strength.
The capital tie-up risk
Designing a custom AI chip costs hundreds of millions of dollars in R&D, with no guarantee of success. Anthropic has $65 billion in Series H funding, but money burns fast when you're doing frontier AI. Every dollar invested in silicon is a dollar not invested in model research or commercial expansion.
The reverse dependency risk
By engaging with Samsung, Anthropic is replacing a dependency on Nvidia with a dependency on Samsung. If the foundry encounters production issues, Anthropic has no immediate Plan B. Supplier diversification is a fundamental principle of the supply chain, and a custom chip makes it more difficult, not less.
❌ Common mistakes
Mistake 1: Confusing preliminary discussions and production
Several commentators have presented the Anthropic-Samsung chip as a done deal. This is not the case. Anthropic has not yet defined the chip's use case, the final process node, or the production schedule. The discussions with Samsung are real and significant, but they remain exploratory. The difference between a tape-out and a foundry conversation is immense.
Mistake 2: Directly comparing Jalapeño and the Anthropic chip
OpenAI Jalapeño is an inference chip designed with Broadcom, a fabless company. Anthropic is discussing directly with a foundry (Samsung) for a 2nm process. Both projects aim for the same goal — reducing dependence on Nvidia — but through different partnership architectures and stages of maturity. Comparing them as competing products is premature.
Mistake 3: Underestimating Samsung's yield lag
TSMC dominates the leading-edge foundry market with yields that remain unrivaled. Samsung has made progress, but each node transition (3nm, then 2nm) has been marked by yield difficulties. An AI lab betting on Samsung at 2nm is taking a real technological risk that should not be minimized in the analysis.
❓ Frequently Asked Questions
Will Anthropic stop using Nvidia GPUs?
No. Even if a custom chip comes to fruition, Anthropic will continue to use Nvidia GPUs for training and part of inference for years to come. Custom silicon would serve as a complement, not a total replacement. The transition is always gradual in AI infrastructure.
Why Samsung and not TSMC?
TSMC is the dominant foundry, but its capacity is saturated by Apple, Nvidia, and legacy clients. Samsung offers available capacity and a deeper strategic partnership through the Series H investment. The yield risk is higher, but availability is a real argument.
Is Anthropic's IPO guaranteed for October 2026?
No. A confidential S-1 filing does not guarantee an IPO. Anthropic could withdraw the filing if market conditions deteriorate. But with a $47 billion run rate, top-tier lead banks, and a $965 billion valuation, all signals point to an effective listing.
What is the difference between the Anthropic chip and Google's TPUs?
Google's TPUs are accelerators designed for Google workloads (Search, YouTube, Gemini). The Anthropic chip would be optimized specifically for Claude architectures — inference, long contexts, agentic. The principle is similar, but the specialization would be different.
Is Anthropic really profitable?
Sources indicate that Anthropic would be profitable in 2026 thanks to Claude Code and enterprise adoption. But "profitable" in the context of a frontier AI lab often means "positive gross margin on inference," not necessarily positive net income after R&D and capital expenditure. The exact details will be in the public S-1.
✅ Conclusion
Anthropic is executing the plan that every AI lab of this scale must follow: control its silicon to control its costs, prepare for an IPO with revenues that justify the valuation, and differentiate itself from OpenAI through operational discipline rather than media hype. The Samsung 2nm chip is a risky but necessary multi-year bet. The October 2026 IPO will be the first real test of whether the market buys this narrative.