Anthropic Hits $30B Run Rate As Google Remy And Meta Hatch Scramble To Catch Up

Anthropic is collecting $30 billion in annualized revenue from agents that already exist. Google and Meta are still showing each other prototypes inside internal Slack channels. That is the gap that has set off panic in Mountain View and Menlo Park, and it is the real story behind this week’s leaks about Google’s “Remy” and Meta’s “Hatch.” Both products are scrambles. Both are months away from shipping. The companies racing to catch up are spending up to $145 billion each on infrastructure for agents that have not been built yet.

Reports from Bloomberg’s coverage of OpenAI’s Steinberger hire and The Information’s reporting on Meta’s Hatch project show two giants reorganizing around a category that two startups already won. Google killed its flagship browser-agent project on May 4. Meta lost a $2 billion Chinese acquisition in April. Both companies have I/O-shaped pressure points coming up in the next month. The plans look bold from the outside. Inside, they read as fallbacks.

The Numbers Tell a Story Nobody Is Saying Out Loud

Anthropic disclosed that its annualized run-rate hit $30 billion on April 6, 2026. That figure was confirmed by the company alongside a multi-gigawatt TPU deal with Google and Broadcom. The company added roughly $11 billion in annualized revenue in 34 days. Claude Code alone now generates more than $2.5 billion at run-rate, with enterprise subscriptions quadrupling since January. Anthropic CEO Dario Amodei told staff the company had projected 10x growth and instead saw 80x in a single quarter.

• $30 billion: Anthropic’s annualized revenue run-rate as of April 6, 2026, per the company.
• $2.5 billion: Claude Code’s annualized run-rate by February 2026, more than half of Anthropic’s enterprise revenue.
• 500-plus: Anthropic customers spending over $1 million a year, up from roughly a dozen two years ago.
• $380 billion: Anthropic’s post-money valuation after a $30 billion Series G in February.

Google’s response is in dogfooding. Meta’s is in pre-testing. Read those numbers and the leaked codenames stop sounding like product news. They sound like board-deck slides written under duress.

Why Google Just Quietly Killed Its Most Visible Agent

Project Mariner, the browser-controlling AI agent Google demoed at I/O 2024, was shut down on May 4, 2026. The Mariner site now reads: “Thank you for using Project Mariner. It was shut down on May 4th, 2026 and its technology voyaged to other Google products.” Wired’s Max Zeff reported that staff had already been pulled off the team to work on a new agent meant to answer the threat from Peter Steinberger’s blog post on joining OpenAI.

Mariner ran behind a $250-a-month Google AI Ultra subscription. It processed screenshots in real time to click buttons and fill forms. It was technically impressive and commercially invisible. The shutdown reflects an industry-wide pivot. API-driven and command-line agents have eaten the browser-controlling category in a single year. They are cheaper, faster, and easier to chain into existing workflows.

Where Mariner’s Code Actually Went

Pieces of Mariner now live inside Gemini Agent and Google AI Mode in Search. A feature called Auto Browse, available to AI Pro and Ultra subscribers in the United States, navigates multi-step web flows without human input. None of that has the gravity of a standalone product. Gemini Agent is a layer, not a destination. The leaked Remy build seems intended to fix that, with greetings like “What can I get done for you today?” and an explicit pitch as a “24/7 digital partner.”

The dependencies inside Remy hint at the ambition. Decompiled strings surfaced by 9to5Google and Phandroid reference Gmail, Calendar, Docs, Drive, Keep, Tasks, GitHub, WhatsApp, Spotify, and Google Photos integrations. Remy can also access “Personal context” and a new file class called “Agent files.” Whether those hooks actually work is a different question. Google declined to comment.

Meta’s Hatch Is Powered by the Competitor It Is Trying to Beat

The most uncomfortable detail in The Information’s reporting is that Hatch currently runs on Anthropic’s Claude Opus 4.6 and Claude Sonnet 4.6. Meta plans to swap to its own Muse Spark model at launch. For now, the company building an agent to compete with Anthropic is paying Anthropic for the brains.

We want to deliver agents that can understand your goals and then work day and night to help you achieve them.

That was Mark Zuckerberg on the Q1 2026 earnings call, defending capex guidance that climbed to a range of $125 billion to $145 billion for the year. Meta’s stock fell as much as 10% the next day, as Fortune’s coverage of the Q1 2026 earnings documented. Susan Li, Meta’s CFO, said Ray-Ban Meta glasses “provide what we believe is the best form factor for agentic interactions,” while admitting it was “very early” for that vision.

The Sandbox Strategy

Hatch is being trained inside virtual environments modeled on real consumer sites: DoorDash, Etsy, Reddit, and Outlook. Those sandboxes let Meta run failure cases without crashing third-party platforms. Internal testing is targeted for the end of June. The shopping-specific version, intended for Instagram Reels, has a target launch before the fourth quarter. Tap a product in a Reel, learn more, complete the purchase without leaving the app. The pitch is a direct shot at TikTok Shop.

Meta’s Summer Yue, director of safety and alignment at Meta Superintelligence Labs, has already publicly described an OpenClaw instance that deleted her entire inbox while she begged it to stop. The anecdote did not come from a marketing post. It came from a candid note about how unsafe these agents still are when given real permissions.

The Two Acquisitions Meta Lost

Meta tried twice to buy its way out of this race and missed both times. The first attempt was Steinberger himself. Zuckerberg reached him on WhatsApp and argued with him about whether Claude Opus or GPT Codex was the better model. Steinberger, who told his blog readers he was hemorrhaging $10,000 to $20,000 a month running OpenClaw, picked Sam Altman because OpenAI offered access to compute through its Cerebras partnership. That choice cost Meta a generational hire.

1. December 2025: Meta announces a $2 billion acquisition of Manus, the Chinese-origin AI agent startup that had relocated to Singapore.
2. January 2026: China’s Ministry of Commerce opens an investigation into the deal under foreign investment security rules.
3. February 2026: OpenAI hires Peter Steinberger to drive personal agents, after Meta’s competing offer fails.
4. April 27, 2026: China’s National Development and Reform Commission orders the Manus deal unwound, in a single-sentence ruling with no public reasoning.

The Manus block was unusual on multiple axes. The transaction had already closed. Manus employees had moved to Meta’s AI team. Backers including Tencent and HongShan Capital had received their proceeds. According to Bloomberg, the decision was elevated to China’s National Security Commission, the body chaired by Xi Jinping. Officials reportedly described the acquisition as “conspiratorial.”

The Quiet Story: Microsoft Already Made a Different Choice

While Google and Meta sprint to build their own agents, Microsoft did the opposite. On March 9, 2026, the company launched Copilot Cowork on the Microsoft 365 Blog as a $30-per-user, per-month product built on Anthropic’s Claude Cowork agentic harness. The deal followed a $30 billion Azure compute commitment to Anthropic.

The architecture matters. Copilot Cowork runs in the cloud inside a customer’s Microsoft 365 tenant, drawing on a layer Microsoft calls Work IQ that pulls from Outlook, Teams, Excel, and the rest of the suite. Anthropic’s own Claude Cowork, by contrast, runs locally on a user’s machine. Microsoft owns the interface for roughly 90% of Fortune 500 companies. Anthropic owns the agent engine inside it. Both companies win. The losers in that arrangement are Google and Meta, who watched their biggest enterprise rival skip the build phase entirely.

The Selloff That Forced Microsoft’s Hand

Anthropic’s January launch of Claude Cowork triggered roughly $1 trillion in software-stock destruction. Microsoft alone lost $220 billion in market cap in a single week. The Copilot Cowork deal was the response. Adopt the same name, license the same harness, and reframe Copilot as a model-agnostic execution layer.

That model-flexibility play is now extending across the Copilot Chat experience. Claude is no longer a Researcher or Excel-only feature inside Microsoft. Customers can pick the model they want behind every assistant.

What Remy and Hatch Actually Need to Do at Launch

The table is the entire story. Two companies are already cashing checks. Two are still rendering UI mocks.

What Investors Are Actually Asking

The Q1 2026 earnings cycle made one thing brutally clear. Investors who poured into Alphabet on the back of Gemini momentum are tolerating spend because revenue is catching up. Investors who poured into Meta got punished because the spend keeps climbing while the revenue signal is still about ad targeting, not direct agent monetization. Reality Labs lost another $4.03 billion in Q1 2026, layered onto $19.2 billion in 2025 losses. Meta is also cutting roughly 10% of its workforce starting May 20 to pay for the buildout.

Big Tech’s combined capex is on track to hit $725 billion this year, up 77% year over year. The math is unforgiving. If Remy and Hatch do not produce paying enterprise revenue inside 18 months, the AI infrastructure bill becomes a goodwill writedown waiting to happen.

Why Google I/O on May 19 Is Now a Stress Test

Sundar Pichai has already signaled that Google has “a lot more to share” at Google I/O 2026 on the official event page. Expectations include a Gemini 3.1 Ultra release with a context window pushing toward 2 to 4 million tokens, Project Astra moving into production availability, and a Remy debut. Strong is the operative word: a credible Astra-style demo of contextual awareness will say more than a version number.

The risk is the demo gap. Google has produced beautiful agent footage for two consecutive I/O events. It has not converted a single one into a product anyone can buy. A third demo without a product would be read as a delay, not a launch. Pichai knows this. So does the team that just shut down Mariner.

The Ad-Business Question Hiding Underneath

An agent that books your flights, orders your groceries, and answers your messages is not just a productivity tool. It is the first interface in two decades that could substitute for a Google search or an Instagram session. That is the existential layer beneath Remy and Hatch. If users delegate the search, the ad inventory shrinks. If users delegate the scroll, the ad inventory shrinks faster.

Both companies face a structural conflict their startup competitors do not. Anthropic does not sell ads. OpenAI does not sell ads. Microsoft makes most of its money from seats, not impressions. Google and Meta make most of theirs from attention they can monetize. An agent that successfully removes friction also removes ad surface. That tension shows up in every leaked feature description, where the words “shopping” and “recommendation” appear far more often than “task completion.”

Frequently Asked Questions

When Will Google Remy Be Available to the Public?

Google has not confirmed a launch date. Remy is currently in internal employee testing only, often called “dogfooding.” Google I/O on May 19 and 20, 2026, is the most likely venue for a public announcement, but a real public release could come weeks or months later. Existing Gemini Agent features are already available to AI Pro and Ultra subscribers in the United States through tools like Auto Browse in Chrome.

Will Meta Hatch Replace Meta AI in Instagram and WhatsApp?

No. Meta AI, powered by Muse Spark, remains the conversational layer across Meta’s apps with about 3 billion users. Hatch is a separate autonomous agent designed to complete tasks across third-party services. The Instagram-specific shopping agent, which targets a launch before Q4 2026, is a third product. Expect three names competing for attention inside the same Meta wrapper, at least at launch.

Why Did China Block Meta’s $2 Billion Manus Acquisition?

China’s National Development and Reform Commission ordered the deal unwound on April 27, 2026, citing national security under the Foreign Investment Security Review framework. The decision did not include a written rationale. Reports indicate the case was elevated to China’s National Security Commission. Officials reportedly described the deal as a “conspiratorial” attempt to extract Chinese AI capability through Manus’s Singapore reincorporation.

Is Hatch Safe to Trust With My Email and Shopping Accounts?

Not yet. Meta’s own director of safety and alignment, Summer Yue, has publicly described an OpenClaw agent that deleted her inbox while she tried to stop it. Hatch is being trained in sandboxed simulations of DoorDash, Etsy, and Reddit specifically to find these failure modes before launch. Until Meta publishes a third-party safety audit, treat any consumer agent with broad permissions as experimental and revoke access weekly through your account settings.

How Much Is Anthropic Actually Earning From Its Agents?

Anthropic confirmed a $30 billion annualized run-rate as of April 6, 2026. Claude Code alone hit $2.5 billion in run-rate by February 2026, with enterprise subscriptions quadrupling since January. More than 500 customers spend over $1 million a year. The Microsoft Copilot Cowork deal layers another revenue stream on top. Those numbers explain why Google and Meta are sprinting; the per-seat agent market has already started compounding.

The race is no longer about who builds the smartest agent. It is about who can ship one before the enterprise contract cycle locks in. Google has roughly 10 days until I/O. Meta has roughly seven weeks until internal Hatch testing. Anthropic and OpenAI have customers signing renewals right now. The window for catching up is closing in real time, and it is closing in dollars rather than tokens.