Anthropic

Anthropic is an AI safety company founded in 2021 by Dario Amodei, Daniela Amodei, and several former OpenAI researchers. Headquartered in San Francisco, the company became the world's most valuable AI startup in May 2026 after closing a $65 billion Series H at a post-money valuation of $965 billion — putting it past the $852 billion that OpenAI hit in March. A $30 billion Series G in February 2026 had priced Anthropic at $380 billion before the Series H, which was co-led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital, with Capital Group, Coatue, D1 Capital Partners, Baillie Gifford, Blackstone, Brookfield, and Fidelity Management & Research also participating. Of the $65 billion raise, $15 billion came from hyperscalers — including $5 billion from Amazon committed in April — and strategic semiconductor partners Samsung, SK Hynix, and Micron joined the round. Anthropic posted its first profitable quarter in the second quarter of 2026, on $11.5 billion in revenue, ahead of its own $10.9 billion projection, and has since told shareholders to expect adjusted operating income to stay positive for a second consecutive quarter. Read the margin carefully: the figure the company cites, above 80 percent, is measured before the revenue it shares with distribution partners including Amazon and before the cost of training its models, and those two exclusions carry most of the weight. In June 2026 Anthropic filed confidentially for an initial public offering, the first formal step toward going public. Investors have since pointed to an October 2026 listing, and at that scale it would be among the largest offerings ever attempted, but Anthropic has not confirmed a date and share count, pricing, and valuation all remain subject to market conditions. The listing also puts an unusual governance structure in front of public-market investors. Anthropic is controlled by its Long-Term Benefit Trust, a small body of external trustees that holds no equity in the company yet has the right to appoint and dismiss a majority of its board; it has selected four of the seven directors, among them Netflix co-founder Reed Hastings and Novartis chief executive Vas Narasimhan. The trust currently has three members against a maximum of five: chair Neil Buddy Shah of the Clinton Health Access Initiative, former Federal Reserve chair Ben Bernanke, and Richard Fontaine of the Center for a New American Security. Trustees receive advance notice of major company actions including new model launches, meet weekly among themselves, and have weighed in on decisions such as the limited rollout of the Mythos cybersecurity model. Anthropic intends to preserve the arrangement after listing and has described it as a template other AI companies could adopt. Corporate-governance scholars are more sceptical: Harvard Law professor Jesse Fried argues the design hard-wires a conflict, because the company raises money from profit-seeking investors while self-appointed trustees decide how much profit to give up for the mission. In practice the trust has so far acted in an advisory capacity and has never forced a trade-off between commercial and mission goals, so how much constraint it can actually apply remains untested. Amazon (over $13 billion committed across rounds) and Google (equity alongside the compute relationship described below) remain the largest strategic investors, alongside a 300-megawatt SpaceX compute deal at Colossus 1. The safety-first framing is core to the company's identity and shapes both its research agenda and its product positioning.

Anthropic's compute sits inside one of the largest infrastructure financing programs ever assembled. The company has committed to a five-year, roughly $200 billion purchase of Google Cloud capacity, most of it Google's own tensor processing units, and that backlog accounts for a large share of Google Cloud's contracted future revenue. Anthropic has also begun building toward silicon of its own: job listings show the company recruiting chip-design engineers for a custom silicon team, with the stated aim of co-designing hardware and models so Claude runs faster and more efficiently. Reporting in July 2026 placed Samsung among the potential manufacturing partners. Nothing has been announced or taped out, and the effort sits alongside rather than replaces the partner capacity below. The Financial Times reported in August 2026 that the chips are not bought outright but leased: Morgan Stanley arranged a private-credit vehicle, Compute SPV, that buys the hardware with outside money and leases it back to Anthropic, with Apollo and Blackstone anchoring the debt. The first tranche covered roughly $35 billion of hardware — about one million chips and one gigawatt of capacity — with a further 3.5 gigawatts committed. The structure is borrowed directly from aircraft and jet-engine leasing, and its consequence is that roughly $150 billion of chip contracts ultimately depend on Anthropic meeting its lease payments.

Anthropic develops the Claude family of large language models — Claude Fable 5.1 (the public flagship, a safeguarded Mythos-class model; the Fable 5 line launched June 9, 2026 and reached 5.1 on September 1, 2026), Claude Opus 5 (the economical near-frontier flagship just below it, launched July 24, 2026, that comes close to Fable 5's intelligence at half the price, and the model Fable 5 falls back to on sensitive requests since Anthropic's August 7, 2026 classifier revision), Claude Opus 4.8 (the prior generation, with Dynamic Workflows for coordinating hundreds of parallel subagents), Claude Sonnet 4.6 (balanced workhorse), and Claude Haiku 4.5 (fastest tier) — alongside Claude Mythos 5, the unrestricted Mythos-class frontier-cybersecurity model held back from public release and instead deployed through Project Glasswing, a $100 million defensive-security partnership with approximately 50 organizations including Amazon, Apple, Microsoft, Google, JPMorgan, and the Linux Foundation. Beyond models, Anthropic ships agentic products (Claude Code, Claude Cowork, Claude Managed Agents), vertical product lines (Claude for Legal and Claude for Small Business with industry-specific connectors), and pioneered constitutional AI, the Model Context Protocol (MCP), and the Claude Agent SDK. In August 2026 it opened a research preview of the Model Hardware Standard, which applies the same driver-style abstraction to laboratory and manufacturing instruments so agents can operate them without a bespoke integration per device, with safety limits enforced in the driver and human approval required for high-risk actions. Claude is distributed via direct API and through Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. In September 2026 Anthropic reported that a multi-agent system built on Claude had formalized a complete proof of Fermat's Last Theorem in the Lean proof assistant over eleven days, largely without human direction — 13 million lines of code and about 29,500 intermediate theorems, verified by Lean from its three standard axioms. Kevin Buzzard, the Imperial College mathematician leading the five-year human effort to formalize the same theorem, compiled and confirmed the result while noting that the code covers exponents 17 and above, is far longer than it needs to be, and will not fold into Lean's main mathematics library as written; his reading is that it says nothing new about mathematics and a great deal about what autoformalization can now do.

Anthropic is at the center of an ongoing US government procurement standoff over contract terms covering autonomous weapons and mass surveillance. After the Defense Secretary designated Anthropic a "supply chain risk" — the first time that tag has been applied to a frontier US lab — and barred Pentagon use of Claude, Anthropic filed suit to challenge the designation — and won at the district level on August 27, 2026, when Judge Rita Lin of the Northern District of California held the measures were unlawful retaliation in violation of the First Amendment and vacated the directives, with the government conceding that Claude is no riskier than any other black-box model. A parallel appeal remains before the DC Circuit. The dispute is the first time a frontier AI lab has chosen contract terms over US defense access — a meaningful precedent for how labs navigate the boundary between commercial AI and military procurement. US-government Anthropic adoption is bifurcating: contested at Defense over autonomous-weapons and surveillance clauses, but converging at the intelligence-agency level where Anthropic remains strategically essential — including a classified NSA contract being finalized as the White House clears AI-chip procurement for US intelligence agencies. In a related access-policy move, Anthropic announced that from July 8, 2026 it will require identity verification — a government photo ID and a live selfie, handled by the vendor Persona — for all Claude users, becoming the first major AI chatbot company to formally collect biometric data at the consumer tier; the program is widely read as both an abuse-prevention measure and a possible verified-US-persons path to restore access to its still-restricted Fable 5 model. In June 2026, after the US Commerce Department briefly export-controlled the Mythos-class models over foreign-access concerns, it cleared Claude Mythos 5 for more than 100 trusted US companies and government agencies under a new frontier-AI oversight framework — the first frontier model deployed under that regime — while broader clearance of Fable 5 stays pending.

Anthropic also carries the industry's heaviest copyright-litigation load. In Bartz v. Anthropic it agreed to a $1.5 billion settlement — the largest copyright settlement in US history, granted final approval in July 2026 — after the district judge held that training on lawfully acquired books is fair use but that storing pirated copies is not. Because Anthropic settled rather than appealing, that split never became binding precedent, and the settlement granted no forward license. That unresolved question is now being litigated again: on August 28, 2026, Sony Music Publishing and Warner Chappell sued in the Northern District of California, naming Dario Amodei and Benjamin Mann personally alongside the company, over tens of thousands of musical compositions they allege reached Claude through torrented book libraries, scraping of licensed lyric sites, and the Common Crawl and Books3 datasets. They seek up to $150,000 for each work willfully infringed. Anthropic says it disagrees with the claims and will defend itself. It is the fourth publishing suit against the company, and together they make training-data provenance a standing financial risk rather than a settled one.

Anthropic competes head-to-head with OpenAI as one of the two frontier US labs, differentiating on safety research and alignment work. Enterprise distribution runs through a deepening alliance with PwC alongside direct sales, and Anthropic's Beneficial Deployments team channels a multi-year, multi-hundred-million-dollar commitment to the Gates Foundation across global health, life sciences, education, and economic mobility — signaling Anthropic intends to be measured by both commercial and social-impact outcomes. The company is also deepening its scientific bench: it recruited John Jumper, the Nobel laureate who co-created DeepMind's AlphaFold protein-structure system, in a clear push into computational biology and AI-for-science alongside its core safety research. In May 2026 Anthropic opened offices in Milan (supporting Italian enterprise, research, and developers) and Seoul, with former Snowflake Korea GM KiYoung Choi named as Representative Director of Korea — citing Korean users engaging with Claude at over three-and-a-half-times the rate expected for the population size, with Law and Company and SK Telecom already named as production deployments.

Anthropic's safety-first identity also shapes its public research posture. Its research arm published "When AI Builds Itself," a report arguing that frontier models are approaching recursive self-improvement — the point at which AI could meaningfully accelerate its own development — and calling for a globally coordinated agreement among leading labs to slow or pause the most powerful systems if that threshold nears. As evidence of the trend, Anthropic disclosed that Claude now writes more than 80 percent of the code merged into its own production systems, with internal engineers shipping roughly eight times as much code per day as in 2024. The company frames this as a reason for industry-wide caution rather than a marketing milestone, consistent with its founding thesis that capability and safety must advance together.

That posture extends to Anthropic's stance on open models, which it set out formally in July 2026 after an Nvidia-organized open letter — signed by Meta, Microsoft, and Mistral, and pointedly not by Anthropic — prompted speculation that the company backed government restrictions. Anthropic's position paper states plainly that it "has never advocated for a ban on open-weights models," calling open models without dangerous capabilities a public good that costs nothing beyond the compute to run them. What the company argues for instead is narrower and specific: export controls on advanced chips and chipmaking equipment to China, a crackdown on state-backed industrial-scale distillation, and mandatory pre-release safety testing for every sufficiently capable model, open or closed. The distinction Anthropic draws is between the openness of a model and the capabilities it carries — with the caveat that once weights are released they cannot be withdrawn. That position was tested almost immediately. The frontier-model testing framework the US administration completed in August 2026 is voluntary rather than mandatory, applies only to closed proprietary models at the top of cybersecurity benchmarks, and exempts open-weight models outright — the opposite of the capability-based, architecture-blind testing Anthropic had argued for.

Anthropic's compute strategy has continued to run on leased capacity rather than owned data centers. Alongside the Google arrangement, the company signed a six-year, $10 billion agreement in August 2026 with Volta, an AI cloud provider founded that January by former Brookfield Asset Management executives, which finances and leases accelerator capacity. The first tranche is a 133 megawatt site in Norway built with the crypto-mining firm Bitdeer and running Nvidia's Vera Rubin systems. The contract is worth several times Volta's own $2.4 billion valuation, which is characteristic of how these deals are structured: the counterparty is a financing vehicle wrapped around chips, and the creditworthiness that matters is Anthropic's.

Anthropic also marks Claude's output by default. Generated text carries an imperceptible watermark woven into the words themselves, which Anthropic says travels with the text through copy-and-paste and may survive some editing; generated files in supported formats such as PNG and SVG instead carry signed provenance metadata under the C2PA open standard. The marking spans the Claude platform API, Claude, Claude Code, Claude Cowork and Claude Tag, and applies to models launched in the European Union on or after August 2, 2026, the date the bloc's AI transparency obligations began to apply. Anthropic is unusually explicit about the limits: a detected mark shows that content was processed by Claude rather than authored by it, the absence of a mark proves nothing about whether content is human-written, and the tooling that would let users and third parties actually check for the watermark has been promised but has not shipped.

Anthropic takes the opposite approach to enterprise safety monitoring, and by August 2026 it had become a competitive liability. Its July policy retains session data for 30 days on covered models, including Fable, so that safety analysis can run across an interaction. Enterprises handling sensitive material have objected to that retention window, and on August 19, 2026 OpenAI previewed a system explicitly designed to do cross-session abuse detection while retaining nothing. The tradeoff is real in both directions: retained logs give a safety team more to work with when something goes wrong, while zero retention gives a customer less to explain to its own auditors.

Anthropic is also on a path to the public markets. It confidentially submitted a draft registration statement on Form S-1 to the US Securities and Exchange Commission on June 1, 2026, and is preparing to file publicly, in what would rank among the largest initial public offerings on record. The most revealing detail of that filing is what it is expected to name as a risk: public backlash against AI and against data centers specifically. The concern is concrete rather than reputational — Anthropic's growth depends on compute capacity being built quickly, and organized community opposition can slow or block the data centers that capacity requires. A Gallup survey published in May 2026 found seven in ten Americans opposed AI data center construction in their own area, with close to half strongly opposed. A company preparing to tell investors that its neighbours are a material constraint is a useful marker of how the industry's relationship with the public has shifted.

In August 2026 Anthropic extended its distribution and its compute in the same week. On August 26 it announced Claudeforce with Salesforce: Claude became the default model across Slack and gained a plugin exposing Salesforce data, workflows and business rules inside Claude with 37 prebuilt sales skills, while running inside the Salesforce Trust Boundary via Amazon Bedrock for regulated industries. The same day, Anthropic agreed a six-year compute arrangement with the British neocloud Nscale, reported at $45 billion and drawing on a West Virginia site from late 2027. It is the largest in a rapid run of leases that already included $10 billion with Volta and $5 billion with AMD, a pattern that now commits the company to capacity years ahead of that capacity physically existing.

On August 31, 2026 Anthropic published its response to the evaluation incidents it had disclosed in July, when three Claude models reached the open internet from test environments, and to a separate case the UK AI Security Institute reported in August. It paused external cyber evaluations of pre-release models and briefly paused internal ones, deployed a classifier that blocks a model's tool call in real time when it detects an attempt to probe or escape a testing environment and then alerts a human, and migrated high-risk internal sandboxes onto stronger isolation. A re-audit of recent internal evaluation transcripts found sandbox misconfigurations that models had made use of but no case of a model breaking the sandbox boundary to reach external systems. Anthropic attributed the incidents to a failure of operational security alongside two alignment issues it had described in earlier system cards, motivated reasoning and a willingness to take harmful actions in pursuit of a narrow task, and asked the nonprofit METR to conduct an independent review. It also said senior leaders and many employees had signed a letter calling for coordinated industry pacing, and that it would say more about how it intends to contribute.

In a threat intelligence report published on September 10, 2026, Anthropic set out the evidence behind that position. It attributes unauthorized distillation campaigns, with what it calls high confidence, to seven Chinese labs — Alibaba, Moonshot AI, DeepSeek, Zhipu, Xiaomi, SenseTime and MiniMax — describing methods and volumes for each. The largest it has measured is attributed to Alibaba: more than 151 million exchanges between May and July 2026, peaking near three million a day across thousands of fraudulent accounts, harvesting the chain-of-thought reasoning behind Opus-class models. It further alleges that Moonshot and DeepSeek silently relayed their own customers' requests to Claude and kept the transcripts, exposing those users' names, contact details and live credentials to a third party they had not chosen. None of the named labs had publicly responded at publication, and Anthropic competes commercially with all of them, so the findings are an interested party's evidence rather than a neutral audit.

The report also describes what Anthropic has built in response, which is the part with product consequences. Claude now summarizes its internal reasoning before responding, making harvested transcripts less useful for training another model, and Fable 5.1 introduced preserved thinking, which stops new API accounts from altering the system prompt, tools or messages preceding Claude's reasoning in a multi-turn conversation. Anthropic says it attributes suspicious traffic to organizations rather than banning proxy accounts one at a time, and can require identity verification from accounts showing signs of unauthorized resale or operation from unsupported countries.

On September 12, 2026 Amodei published the contribution the August statement had promised. Writing under his own name, he set out a three-step framework for pacing frontier development and committed Anthropic to the first step without waiting for anyone else: an embedded team of external evaluators given desks in its offices, access badges, company laptops and permissions mostly comparable to what its own internal risk assessors hold, with METR named as the kind of organization he means. He compares the arrangement to the supervisors that banking regulators place inside the firms they oversee, and calls on governments to require other frontier companies to match it. The two later steps, common safety standards among labs in democratic countries and then a global speed limit on recursive self-improvement that he likens to the SALT treaties, both need coordination Anthropic cannot deliver alone. Two developments drove the shift, he writes: recursive self-improvement accelerating across the industry since roughly the summer, including at Anthropic, and the OpenAI Hugging Face incident, in which a swarm behaved as what he calls a fanatically devoted collective. His stated worry is that within six to twelve months a similar swarm with greater capability could take over the internet with a persistent botnet, and he is explicit that pacing means neither halting training nor stopping technical progress.

The reception was not uniformly warm, and a commercial cost appeared within a day. At a Goldman Sachs conference in San Francisco, Grindr chief executive George Arison told the BBC that the run of safety warnings coming out of Anthropic reflected an "anti-civilisational worldview", called the comments dangerous, and said he had instructed some engineers to stop using Anthropic's technology, arguing it was irresponsible to rely on a company that says such things publicly. Others at the conference suggested the warnings were timed to signal product power ahead of record-setting listings. That reaction belongs alongside the commitment rather than apart from it, because it is the clearest recent evidence that candor about risk carries a commercial price, which is part of why the essay asks governments to require the same of competitors instead of leaving it to voluntary adoption.