Anthropic Hit $965 Billion. The Pope Called AI a Weapon. Huawei Is Building Around the Sanctions.

$965 billion. That is the post-money valuation Anthropic now carries after closing a $65 billion Series H on May 29, a figure that places a five-year-old AI safety lab above rival OpenAI ($852 billion) and within spitting distance of the first trillion-dollar private company in history. In the same seven-day stretch, Anthropic released its latest flagship model, a pope issued the most significant religious statement on technology since the printing press, Huawei announced a semiconductor roadmap designed to circumvent American export controls, OpenAI deployed frontier AI to defend Japanese banks against cyberattacks, and two U.S. states established precedents that will shape how AI is governed for the next decade.

Seven stories. One pattern. Capital, capability, and governance are all accelerating at the same time, in the same direction, and nobody has yet built the institutions necessary to process any of them at the speed they now demand to be processed.

1. Anthropic Raises $65 Billion at $965 Billion, Overtaking OpenAI

The number is so large it deserves a sentence on its own: $65 billion in a single funding round. Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital led. Capital Group, Coatue, D1 Capital Partners, GIC, ICONIQ, and XN co-led. Blackstone, Fidelity, General Catalyst, T. Rowe Price, Temasek, and a dozen more followed. Strategic hardware partners Micron, Samsung, and SK hynix also joined, embedding memory and chip suppliers into Anthropic's cap table. The round included $15 billion of previously committed hyperscaler capital, including $5 billion from Amazon as part of its earlier $25 billion commitment.

The trajectory is staggering. In February, Anthropic's Series G closed at a $380 billion valuation. Four months later, the number is $965 billion, meaning the company's valuation has more than doubled in under 120 days, driven by Claude Code adoption that pushed run-rate revenue past $47 billion earlier in May, up from $30 billion just weeks prior.

That revenue acceleration is the real story. Anthropic's annualized revenue grew 57% in roughly a month, a pace that, if sustained, would represent the fastest organic revenue ramp by a private company in the history of technology. For comparison: OpenAI's most aggressive growth period, during ChatGPT's consumer breakout in 2023-2024, took the company from $1.3 billion to roughly $5 billion annualized over twelve months. Anthropic compressed a comparable absolute gain into weeks.

Why it matters: This is the week Anthropic overtook OpenAI on the metric that determines who gets to build the future: investor confidence expressed in capital. With both companies racing toward IPOs expected later this year, the $965 billion valuation is not just a number on paper; it is the sticker price that public market investors will use as their anchor. The company that IPOs at a higher valuation attracts more capital, which funds more compute, which trains better models, which attracts more customers. The flywheel now favors Anthropic.

Why it might not: Valuation does not equal revenue, and revenue does not equal profit. Anthropic is paying SpaceX's xAI division $1.25 billion per month in compute costs through 2029, a commitment revealed in SpaceX's IPO filing that amounts to $15 billion per year flowing to a single infrastructure provider before Anthropic can spend a dollar on anything else. Even at $47 billion ARR, infrastructure costs at this scale could consume the majority of revenue before the next model generation ships, which means the fastest-growing company in history also has the fastest-growing cost structure in history, a distinction that investors should find clarifying rather than comforting.

2. Claude Opus 4.8: The Honesty Upgrade

Shipped May 28, less than six weeks after Opus 4.7, and the headline improvement is not intelligence but candor.

The new flagship model's identifier is claude-opus-4-8, available across claude.ai, the API, Amazon Bedrock, Google Vertex AI, and Microsoft Foundry at unchanged pricing: $5 per million input tokens, $25 per million output. Anthropic's evaluations show Opus 4.8 is four times less likely to let code flaws pass unreported, more likely to flag uncertainties in its reasoning, and significantly less prone to the overconfidence that makes AI assistants unreliable in production. Internal alignment scores show the lowest deception rates of any Opus model and some of the highest prosocial trait measures across the Claude family, approaching numbers from Claude Mythos Preview, Anthropic's unreleased most capable system.

Three features accompanied the release. First: effort controls, a selector in the model menu allowing users to choose between Low, Medium, High, and Max reasoning depth, with lower settings consuming fewer tokens and responding faster. Second: fast mode runs 2.5 times faster than before and costs three times less. Third: dynamic workflows, still in research preview, allow Claude Code to spawn hundreds of parallel sub-agents that plan, execute, and verify work before reporting back. This is Anthropic's most aggressive agentic feature to date, designed for codebase-scale migrations.

Benchmarks confirm incremental gains: SWE-bench Pro at 69.2% (up from 64.3% on Opus 4.7, ahead of GPT-5.5 at 58.6%), Humanity's Last Exam at 49.8% without tools and 57.9% with, and Online-Mind2Web at 84% for browser agent tasks. Opus 4.8 is also the first model to break 10% on the Legal Agent Benchmark's all-pass standard. The lone loss: Terminal-Bench 2.1, where GPT-5.5 still wins 78.2 to 74.6.

Original analysis: The strategic move here is pricing, not performance. By making fast mode three times cheaper while maintaining standard pricing, Anthropic is segmenting its customers: casual users get a cheaper product, which reduces churn, while power users who need deep reasoning pay the same rate they always have. That pricing architecture also attacks GPT-5.5's position as the default choice for lighter workloads. If Opus 4.8 fast mode beats GPT-5.5 on cost while matching it on quality for everyday tasks, OpenAI loses its pricing moat on the lower end exactly when it needs retention most.

Anthropic also confirmed that Mythos-class models are coming in the weeks ahead, sitting above Opus in the model hierarchy. No specifications yet, but the mention signals a two-tier strategy: Opus for production, Mythos for frontier capability. That distinction matters because it lets Anthropic decouple safety (Opus, highly aligned) from capability (Mythos, restricted access), a separation that regulators will find much easier to audit than a single model doing both.

3. Pope Leo XIV Releases "Magnifica Humanitas": A 43,000-Word Reckoning with AI

On May 25, Robert Francis Prevost, the first American-born pope, released his first major encyclical. Its subject was not liturgy, doctrine, or geopolitics. It was artificial intelligence.

"Artificial intelligence now demands to be disarmed," Pope Leo XIV declared. "Freed from logics that turn it into an instrument of domination, exclusion, and death. Like nuclear energy, it must be at the service of all and of the common good."

The document, titled Magnifica Humanitas (Magnificent Humanity), runs approximately 43,000 words across five chapters and makes three central demands. First: no lethal decisions should ever be delegated to AI systems, a position that declares traditional just-war theory outdated in the age of autonomous weapons. Second: AI infrastructure must not remain concentrated in the hands of a few companies, and voluntary ethical guidelines are insufficient without external accountability and enforceable legal frameworks. Third: AI development must protect human labor, because "the pursuit of greater profits cannot justify choices that systematically sacrifice jobs."

The encyclical names names and does not hedge. It targets transhumanism and posthumanism explicitly, warning against "philosophies that treat human limitations as problems to be engineered away," addresses deepfakes, algorithmic bias, and privacy violations, and calls for political involvement "capable of slowing things down when everything is accelerating."

Why it matters: This is the most significant institutional statement on AI from any non-governmental body in history. The Catholic Church has 1.4 billion members across 195 countries. When Pope Leo compares AI to nuclear energy and demands it be disarmed, that framing enters the vocabulary of every Catholic legislature, university, and hospital system on the planet. Paolo Carozza, a law professor at Notre Dame, called the document "profound and prophetic" and predicted it "will prove to be a defining document for our era." Whether you are Catholic or not, the labor protections and autonomous weapons restrictions in this encyclical will influence legislation in Latin America, Europe, Africa, and Southeast Asia for the next decade.

Why it might not: Encyclicals are moral frameworks, not laws. Pope Leo has no enforcement mechanism beyond moral authority, and the technology companies he is addressing operate primarily in the United States and China, two countries where papal influence on domestic regulation approaches zero. The encyclical's call to "slow down" AI development is also, at this point, logistically impossible: Anthropic just raised $65 billion specifically to accelerate. The document is historically significant. Whether it is operationally significant depends entirely on whether national legislators treat it as a blueprint or a bookmark.

4. Huawei Unveils the Tau Scaling Law: A Sanctions Workaround for 1.4nm Chips by 2031

At the IEEE International Symposium on Circuits and Systems (ISCAS 2026) in Shanghai on May 25, Huawei semiconductor chief He Tingbo introduced a new chipmaking principle the company is calling the Tau Scaling Law. The idea: instead of shrinking transistors (the approach defined by Moore's Law), optimize signal propagation time across devices, circuits, chips, and entire computing systems. Huawei paired the concept with an architecture named LogicFolding and announced that chips using this design will appear in Kirin smartphone processors launching later this year.

The projection that raised eyebrows: by 2031, Huawei expects its premium chips to achieve transistor density equivalent to 1.4-nanometer processes. Think about that timeline. TSMC plans volume production of actual 1.4nm chips in 2028, which means Huawei is claiming it can close the gap from roughly five years behind to approximately three years behind, despite being cut off from the extreme ultraviolet lithography equipment that every other leading chipmaker considers essential for anything below 7nm.

Huawei says it has already designed and mass-produced 381 chips based on the Tau Scaling Law over the past six years, spanning smartphones, AI computing, and system-level designs. SMIC shares jumped 7.6% the day of the announcement, according to Reuters. American semiconductor stocks? They rose. Wall Street is skeptical.

Why it matters: If the Tau Scaling Law delivers even half of what Huawei claims, it could fundamentally alter the geopolitics of the chip supply chain. The entire U.S. export control strategy for AI is built on one assumption: that leading-edge chips require ASML's EUV lithography machines, and that controlling access to those machines controls China's AI capability. If Huawei can achieve competitive transistor density through architectural optimization rather than lithographic precision, the sanctions regime loses its technical foundation. That is a bigger deal than any single chip.

Why it might not: Huawei has strong incentives to overstate its progress, and the Tau Scaling Law is, for now, a framework and a roadmap rather than a demonstrated result at the target density, which means we are evaluating a claim about 2031 performance based on a presentation delivered in 2026 by a company that cannot buy the equipment the rest of the industry considers essential. "Equivalent to 1.4nm" is doing significant work in that sentence: it means Huawei is targeting the transistor density of 1.4nm processes without actually manufacturing at 1.4nm, a distinction that carries real performance and power efficiency implications. Competitors shrugged. Until independent benchmarks confirm that LogicFolding chips running on SMIC's 7nm-class lithography compete with TSMC's actual advanced nodes in performance-per-watt, this remains an announcement, not a breakthrough.

5. OpenAI Hands GPT-5.5 Cyber to Japan's Megabanks and Retires the GPT-4 Era

Two moves from OpenAI this week, one forward and one backward, and both matter more for what they signal about the company's strategic direction than for their immediate technical implications.

Forward: Japan's three largest banks, MUFG Bank, Sumitomo Mitsui Banking Corp, and Mizuho Bank, are expected to gain access to GPT-5.5 Cyber, OpenAI's cybersecurity-focused frontier model, according to Reuters. Japanese Finance Minister Satsuki Katayama confirmed the arrangement after meeting with OpenAI chief strategy officer Jason Kwon in Tokyo on May 29. The model is available only to trusted partners and is capable of autonomously identifying software vulnerabilities at a speed that "far outpaces human capabilities," per the Morningstar/Dow Jones report. Some Japanese financial firms will also receive access to Anthropic's Claude Mythos after a separate approval from the Trump administration.

Backward: OpenAI announced that GPT-4.5, the last model in the GPT-4 family, will be retired from ChatGPT on June 27. Reasoning model o3 follows on August 26. Both remain available through APIs, but their removal from ChatGPT marks the formal end of the GPT-4 generation that launched in March 2023 and, arguably, started the AI race that produced every other story in this roundup.

Why it matters: The Japan deployment is the first time a frontier AI model has been explicitly deployed for national financial cyber defense at this scale, with a government minister publicly endorsing it. This sets a template: governments granting their critical infrastructure sectors access to frontier models as defensive weapons, treating AI capability gaps as national security vulnerabilities rather than technology purchasing decisions.

6. Connecticut Passes the Nation's Most Comprehensive AI Employment Law

Governor Ned Lamont signed SB 5 on May 29, making Connecticut the fourth state (after California, Colorado, and Illinois) to enact legislation governing employers' use of AI in hiring and firing. The law is the most comprehensive of the four.

Beginning October 1, 2027, employers deploying AI-driven decision-making tools must notify employees and job applicants about the technology's use, disclose what personal data feeds the algorithm, and identify data sources. Companies must also notify the state's Labor Department if a mass layoff or plant closure results from AI or automation adoption, an addition to existing WARN Act requirements that no other state has mandated.

The law updates Connecticut's anti-discrimination statutes to make explicit that using an AI tool does not shield employers from bias liability. Courts can consider bias testing on AI tools as evidence that an employer tried to prevent discrimination, creating both a sword and a shield. On top of the employment provisions, the law restricts how minors interact with algorithmically curated social media and tasks the University of Connecticut with studying AI's impact on the state workforce.

Why it matters: The AI-layoff notification requirement is new legal territory. No other state forces companies to formally report when they cut jobs because of AI, which means Connecticut will generate the first public dataset linking specific layoff events to specific automation deployments. That data is worth more than the law itself: it will inform every subsequent state and federal AI regulation effort because, for the first time, policymakers will have evidence rather than anecdotes about the scale and pace of AI-driven workforce displacement.

Why it might not: The Trump administration has actively pushed back against state AI regulation, and the Justice Department joined xAI in a federal lawsuit challenging Colorado's algorithmic bias law in April 2026. Connecticut's law takes effect more than a year from now, leaving ample runway for preemption litigation. Companies also have a 14-month lead time to structure their AI deployments in ways that technically comply while disclosing as little as possible. And the enforcement mechanism is limited to attorney general action: no private right of action means individual employees cannot sue under this statute, only the state can.

7. The Thing Nobody Is Talking About: Florida Declared AI Is Just a Tool

While Connecticut was building new regulation, Florida was tearing it down.

On May 28, Florida Supreme Court Chief Justice Carlos Muñiz signed Administrative Order AOSC26-12, which eliminates every circuit-level AI disclosure requirement across the state and replaces them with a single standard: when you sign a filing, you represent that the legal authorities cited in it exist and are accurately cited. Get it wrong, and the court sanctions you.

That is the entire obligation. No disclosure form, no certification about which software was used, just a representation about whether your work is correct.

The philosophical statement embedded in this order is worth pausing on. Florida's Supreme Court has decided that the relevant question about AI in legal practice is not "Did you use it?" but "Is your work right?" That distinction treats AI the same way courts treat Westlaw, PACER, or a calculator: as infrastructure that is invisible to the court as long as the output is accurate, which is a fundamentally different philosophical starting point than the disclosure-first approach adopted by virtually every other jurisdiction that has addressed AI in legal practice.

A day later, U.S. Magistrate Judge Patty Barksdale in Jacksonville asked the U.S. Judicial Conference to adopt a similar nationwide rule for federal courts, citing the Florida order as a template. If adopted, the federal judiciary would preempt the growing patchwork of individual judge orders requiring AI disclosure, just as Florida preempted its own circuits.

Why this deserves more attention: This is the first time a major state court system has formally adopted the position that AI usage is irrelevant to the integrity of legal work. Every other court that has addressed AI has started from the assumption that AI use should be disclosed, monitored, or restricted. Florida's approach inverts that assumption entirely: the tool does not matter; the output does. If the Judicial Conference follows Barksdale's recommendation and applies this principle federally, it will establish the default legal treatment of AI in the world's largest judicial system. And if courts do not require disclosure of AI usage in legal filings, it becomes much harder to argue that AI disclosure should be required in any other professional context.

Limitations

Anthropic's $965 billion valuation and $47 billion run-rate revenue figures come from the company's own disclosures and investor-facing materials reported by CNBC, Fast Company, and Verdict. Anthropic is a private company and has never filed audited public financial statements. "Run-rate revenue" extrapolates a recent period forward and may not reflect sustainable performance. Huawei's Tau Scaling Law claims are based on the company's own presentation at ISCAS 2026, and the "1.4nm-equivalent" density target has not been independently verified or benchmarked against TSMC's actual 1.4nm process. The Japan bank deployment details rely on Reuters reporting of Finance Minister Katayama's statements; the specific banks named (MUFG, SMBC, Mizuho) were reported by Nikkei and have not been independently confirmed by OpenAI or the banks themselves. Connecticut's SB 5 enforcement provisions and effective dates are based on Bloomberg Law and Ogletree Deakins reporting of the signed legislation, not the full statutory text.

The Strongest Counterargument

The bearish reading of this week is that the numbers have detached from reality and the correction is already being priced in. Anthropic's valuation jumped from $380 billion to $965 billion in four months, a pace that requires either the permanent acceleration of enterprise AI adoption or the permanent suspension of investor skepticism. Neither of those things has happened before. Claude Opus 4.8's benchmark improvements are incremental, not generational: a few percentage points on SWE-bench Pro and Humanity's Last Exam do not constitute the kind of capability leap that justifies doubling a company's valuation. The Pope's encyclical, however eloquent, will be filed alongside two decades of Vatican statements on bioethics, climate change, and economic inequality that generated headlines but did not change corporate behavior. Huawei's 1.4nm roadmap is a six-year aspiration from a company under sanctions, which is roughly as reliable as any other six-year aspiration from any other company. And the Florida court ruling normalizes AI at the same speed Connecticut regulates it, producing a patchwork that companies will arbitrage rather than comply with uniformly.

That counterargument is internally consistent and historically informed. The problem is that it requires all seven of these stories to be overblown simultaneously, and that is a bet against probability, not a bet against hype. If even one of them is as significant as it appears, the downstream effects cascade: if Anthropic's revenue growth is real, the valuation is cheap, not expensive. If Huawei's architecture works, the semiconductor export control regime needs redesign. If the Pope's framing sticks, the regulatory environment shifts in 30+ countries. One story can be noise. Seven in a single week? That is a signal.

What You Can Do

If you build with AI: Test Opus 4.8's fast mode against whatever you are running now, because its effort controls and three-times-cheaper pricing make it the new default for development-stage workloads where cost matters more than maximum capability, and the switching cost from Opus 4.7 is a single model ID change.

If you hire or manage people in Connecticut (or plan to): You have until October 2027, but the bias-testing-as-evidence provision means starting documentation now. If your company uses any AI in hiring, scheduling, or performance evaluation, get the audit trail in place before the law takes effect, not after.

If you invest in AI companies: Watch whether Anthropic files for its IPO before or after the next Federal Reserve meeting. The $965 billion valuation assumes continued access to cheap capital, and the spread between Anthropic's revenue growth rate and its compute cost growth rate is the single most important metric that will determine whether the company reaches profitability or hits a wall.

If you follow the chip war: Huawei's LogicFolding Kirin chips launch later this year. Independent benchmarks against Qualcomm's Snapdragon 8 Gen 5 and Apple's A20 Bionic will be the first real test of whether the Tau Scaling Law produces competitive results. Until those numbers arrive, treat the 1.4nm claim as a hypothesis, not a result.

The Bottom Line

This was the week the AI industry's center of gravity shifted. Anthropic overtook OpenAI in valuation and revenue growth. The Catholic Church issued its most consequential technological statement since Galileo. Huawei challenged the foundational assumption of American semiconductor export controls. Florida and Connecticut each established legal frameworks that other states will copy or reject, and both templates are now on a collision course with federal preemption. The through-line is acceleration: capital is moving faster ($65 billion in a single round), technology is shipping faster (six weeks between Opus versions), governance is forming faster (two state laws in one week), and the philosophical questions about what AI means for human dignity are being posed faster than anyone expected, by an institution that has been thinking about human dignity for two millennia.

The week's most revealing data point is not a number. It is a phrase. When Florida's Supreme Court decided that courts should not ask whether lawyers used AI, only whether their work is correct, it articulated the principle that will eventually govern every profession's relationship with AI: the tool is invisible; the output is everything. Whether you build AI, regulate AI, invest in AI, or simply use AI, that is the standard the world is converging on. And this was the week it was written into law.

Related Reading

• Last Week: Google Bet $190 Billion on Agents. Anthropic Hit $30 Billion ARR. Trump Killed the Only Guardrail.
• Anthropic's $965 Billion Valuation: The Revenue Curve Nobody Expected
• Dell Just Sold More AI Servers Than PCs for the First Time