Google Bet $190 Billion on Agents. Anthropic Hit $30 Billion ARR. Trump Killed the Only Guardrail.

$190 billion. That is how much Google committed to spending on AI infrastructure this year, announced onstage at I/O 2026 on May 19 in front of 7,000 developers and roughly 900 million monthly Gemini users watching from home. In the same 120-hour stretch, Anthropic told investors its quarterly revenue had doubled to $10.9 billion and it would record its first-ever operating profit, the White House canceled the only executive order that would have imposed any federal review of frontier AI models, SpaceX filed an IPO that values a rocket-and-AI conglomerate at $1.75 trillion, and the week's human toll came into focus: 11,000 jobs eliminated at Meta and Intuit alone, pushing 2026's tech layoff count past 114,000.

Seven stories. One throughline. The AI industry has entered a phase where the amounts of money moving in, the speed of structural change, and the absence of oversight are all accelerating simultaneously, and nobody with the authority to slow any of it down appears interested in trying.

1. Google I/O 2026: $190 Billion, 100 Announcements, One Message

Google's two-day developer conference in Mountain View was, by the company's own count, a 100-announcement event. The sheer volume was the strategy: bury the audience in product launches so no single one can be isolated and criticized, force reporters to write listicles instead of analysis, and bury the capex number inside a keynote where it competes with a dozen product demos for attention. Three of those announcements matter more than the other 97 combined.

First: Gemini 3.5 Flash, the new default model powering AI Mode across Google Search, replaces the prior generation with faster inference and what Google claims is a 40% reduction in hallucination rates on medical and legal queries. AI Mode now serves over one billion monthly users globally, according to Engadget's recap. At the new $100/month AI Ultra tier (five times the usage limits of the $20 Pro plan), Google is pricing AI access like a SaaS product, not a search engine.

Second: Gemini Omni, the multimodal generation model that accepts text, images, audio, and video as input and produces video output with what Google describes as physics-aware rendering. Omni reportedly simulates gravity, kinetic energy, and fluid dynamics in generated scenes. It is rolling out to the Gemini app, Google Flow, and YouTube Shorts. This is Google's answer to OpenAI's Sora and Runway's Gen-3, but integrated directly into a platform with 2.3 billion monthly active users across YouTube alone.

Third, and most consequential for the platform war: Android XR smart glasses, co-developed with Samsung and Qualcomm, with two models teased alongside partnerships with Gentle Monster and Warby Parker. The glasses support real-time Gemini conversation, audio translation in the speaker's voice, text translation overlaid on your visual field, and photo capture. Geeky Gadgets reported that Xreal's Project Aura, the maximalist variant, packs a 70-degree field of view and three cameras into the form factor. Google is not building a single device. It is building a platform that third-party manufacturers will populate, exactly the way Android colonized the smartphone market between 2010 and 2014.

Why it matters: The $190 billion capex commitment represents roughly 10% of Alphabet's market capitalization being plowed into AI infrastructure in a single year, a concentration of capital expenditure that has no precedent among publicly traded technology companies. Google is processing 3.2 quadrillion tokens per month across its ecosystem. The I/O announcements were not individual products competing with individual rivals; they were the declaration that Google intends to make AI the operating layer between users and everything else, from search and shopping via Universal Cart to scheduling via Gemini Spark to video creation and wearable computing. Read that list again. If this works, Google does not sell AI. Google becomes AI.

Why it might not: Everything announced at I/O depends on sustained user engagement with agentic features that require trust. Gemini Spark, which autonomously monitors your credit card statements and your kids' school emails, is exactly the kind of ambient agent that provokes privacy backlash. Google's history with privacy scandals (Google+, Street View Wi-Fi collection, Location History disclosures) makes it the least trusted major company to operate a 24/7 personal agent. Distribution advantage means nothing if the first Spark privacy incident goes viral. And $190 billion buys a lot of servers, but it does not buy a reason for consumers to pay $100 per month for AI features they can get from Claude, ChatGPT, or Perplexity for $20.

2. Anthropic: $10.9 Billion Q2, First Profit, and the Fastest Revenue Ramp in Tech History

The number that should keep every AI competitor awake at night is not the $10.9 billion quarterly revenue figure. It is the trajectory.

Anthropic's annualized revenue run rate sat at $9 billion at the end of 2025. By February it was $14 billion. March: $19 billion. By April, according to The Wall Street Journal and investor materials reviewed by multiple outlets, the annualized run rate exceeded $30 billion, surpassing OpenAI's $24-25 billion. The company told investors it expects $10.9 billion in Q2 revenue, a 130% increase over Q1's $4.8 billion, producing an operating profit of $559 million.

Nobody expected this, least of all Anthropic itself, whose own financial projections shared with investors as recently as late 2025 did not forecast profitability until 2028. The driver, by every account, is Claude Code: the developer tool that generated $2.5 billion in annualized recurring revenue by February 2026 alone, embedding itself so deeply into engineering workflows that Ramp's April spending index showed Anthropic overtaking OpenAI in U.S. enterprise AI adoption for the first time, 34.4% to 32.3%.

The profitability, however, may be temporary. Anthropic is paying SpaceX's xAI division $1.25 billion per month to rent data center compute through 2029, a commitment revealed in SpaceX's IPO filing this week. That is $15 billion per year in compute costs from a single provider. Add Google Cloud, AWS, and the Microsoft Maia 200 chips Anthropic is now in early talks to access, and total infrastructure spend could easily exceed quarterly revenue by late 2026 as training runs scale for Claude 5.

Original analysis: The financial story of 2026 in AI is the speed at which the enterprise revenue pool is growing versus the speed at which compute costs are escalating, a race between two exponential curves where the winner determines whether the AI industry becomes self-sustaining or collapses under the weight of its own infrastructure bills. Anthropic's 130% quarter-over-quarter revenue growth would be, if sustained, the fastest revenue ramp in the history of the technology industry. For context, the previous record holder in percentage terms at comparable scale was ChatGPT's consumer growth in 2023, which took OpenAI from roughly $1.3 billion annualized to $3.4 billion in a year. Four months. That is how long it took Anthropic to go from $9 billion to $30 billion annualized. But the $15 billion annual xAI compute bill alone represents 50% of the current annualized revenue, and training frontier models requires spending money before revenue arrives to pay for it. Anthropic's profitable quarter may be a snapshot of a fleeting window between enterprise revenue catching up and infrastructure costs leaping ahead again.

Why it might not matter: At $30 billion ARR and accelerating, Anthropic is raising a round that could value it at $900 billion, surpassing OpenAI's $340 billion valuation. If Anthropic IPOs at that number (discussions are reportedly happening as early as October), profitability becomes a secondary concern because the capital markets will fund the infrastructure buildout regardless, the same way Amazon ran unprofitable for two decades while investors funded warehouse expansion. The real question is not whether Anthropic is profitable today. It is whether Claude Code's developer adoption creates switching costs durable enough to justify the valuation once compute costs inevitably rise.

3. Karpathy Joins Anthropic: The Talent War Gets Personal

"I've joined Anthropic," Andrej Karpathy posted on X on Tuesday, May 19. "I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D."

That post carried weight. Karpathy co-founded OpenAI, led Tesla's Autopilot AI division, founded Eureka Labs, and is one of perhaps twenty people on the planet who have shipped production-scale neural network systems at both a research lab and a manufacturing company, which means he understands not just how to train models but how to deploy them in physical systems where failure has real consequences. His specific mandate at Anthropic: start a new team focused on using Claude itself to accelerate pre-training research, working under pre-training lead Nick Joseph, according to TechCrunch.

The recursive element is the headline. Anthropic is not just hiring Karpathy to build better models; it is hiring him to build systems where Claude builds better versions of Claude, a recursive loop that sounds circular because it is circular, and that circularity is the entire strategy. Pre-training is the most expensive and compute-intensive phase of building a frontier model, sometimes consuming hundreds of millions of dollars in compute for a single run. Automating portions of it with the existing model reduces the cost-per-capability-gain of each subsequent training run, and Karpathy's specific expertise in data pipelines, scaling laws, and training infrastructure makes him uniquely qualified to execute it.

Why it matters: OpenAI has now lost its co-founder to its primary rival. Combined with the Ramp enterprise adoption data and the revenue numbers, the Karpathy hire is the third data point in a trend that should concern OpenAI's leadership: Anthropic is winning on product (Claude Code), winning on revenue, and now winning on talent. The last time an AI company simultaneously led on all three dimensions was OpenAI itself in early 2024.

Why it might not: Talent moves are symbolically powerful but operationally slow. Pre-training research operates on six-to-twelve-month cycles. Karpathy will not ship anything visible for at least two quarters. OpenAI still has the largest consumer user base (ChatGPT's 300+ million weekly active users), the deepest Microsoft relationship, and a model lineup (GPT-5.5, Codex) that competes credibly at every tier. Losing one person, even an important one, does not lose the war.

4. Trump Kills AI Oversight Hours Before Signing

On Thursday, May 21, a signing ceremony was scheduled at the White House. An executive order would have established the first federal framework for reviewing frontier AI models before public release: a 90-day pre-launch evaluation process, with the NSA likely leading the security assessment. The order had already been watered down from mandatory compliance to voluntary participation after pressure from technology companies. Hours before the ceremony, Trump called it off, saying he "didn't like certain aspects" and needed to ensure the measures would not "hamper the U.S. in its competition with China."

The timing was remarkable. The cancellation happened the same week Anthropic confirmed that Claude Mythos, a model it described as too powerful to release publicly, had autonomously discovered hundreds of previously unknown software vulnerabilities in critical systems, prompting the company to restrict access and deploy real-time detection systems to prevent misuse. This is the model the executive order was specifically designed to address.

Why it matters: The United States now has zero federal oversight mechanisms for frontier AI models. None. The Biden-era executive order on AI was revoked in January 2025, and this replacement was killed before it was born, which means the voluntary safety commitments that major tech companies signed in 2023, commitments with no enforcement mechanism and no penalties for noncompliance, are the only governance framework that exists for an industry spending $190 billion per year on capability development at Google alone. The European Union's AI Act, for all its flaws, at least exists. The U.S. has nothing, and the political dynamics that killed this order, tech industry lobbying combined with the China competition framing, ensure that nothing will replace it before the 2028 election.

Why it might not: Voluntary frameworks have worked tolerably well in some industries. The nuclear power industry's Institute of Nuclear Power Operations (INPO) is a self-regulatory body that emerged after Three Mile Island and has maintained a credible safety record for 47 years. If the major AI labs genuinely fear an uncontrolled release more than they fear regulation, self-regulation could work. Anthropic's decision to restrict Mythos is evidence that at least one lab takes the risk seriously enough to eat the cost. The question is whether all labs will, or whether competitive pressure forces the least safety-conscious player to release first and let the others suffer the consequences.

5. SpaceX Files for the Largest IPO in History

SpaceX filed its S-1 prospectus on Wednesday, May 21, targeting a valuation of approximately $1.75 trillion with plans to raise up to $80 billion. If completed, the June 12 listing would be the largest initial public offering in history by a factor of roughly eight. The filing revealed a company that is no longer just a rocket manufacturer. It is an AI infrastructure conglomerate.

The numbers: $18.7 billion in 2025 revenue (Starlink contributed $11.4 billion), a $4.9 billion net loss for the year, 10.3 million Starlink subscribers, and an xAI division generating $818 million in Q1 2026 revenue with a $2.5 billion quarterly operating loss. The filing also disclosed $2.9 billion in gas turbine investments to power AI data centers, a detail that surfaced alongside ongoing litigation over pollution from those generators.

But the buried bombshell was this: Anthropic has committed to paying SpaceX $1.25 billion per month to rent compute from xAI's Colossus data center through 2029. That is a single customer providing roughly $15 billion in annual revenue to a division that currently loses $10 billion per year. SpaceX is not just launching rockets. It is becoming a cloud provider, and its biggest customer is the company that just overtook OpenAI.

Why it matters: The SpaceX IPO reframes the AI infrastructure supply chain. The three largest AI companies by revenue (Google, Anthropic, OpenAI) are all dependent on compute infrastructure they do not fully control. Google builds its own TPUs. OpenAI depends on Microsoft Azure. Anthropic is now dependent on SpaceX/xAI, Google Cloud, AWS, and potentially Microsoft Maia chips. The $1.25 billion monthly compute bill means Anthropic's profitability is structurally dependent on Elon Musk's willingness to provide competitive pricing to a company that directly competes with his other AI venture. That is an extraordinary business risk hiding inside a revenue number.

Why it might not: Multi-year compute contracts are standard in the industry and typically include pricing protections. Anthropic's diversification across four compute providers (SpaceX/xAI, Google, AWS, Microsoft) limits any single vendor's leverage. And SpaceX has strong financial incentives to keep Anthropic as a customer, because $15 billion per year turns xAI from a money-losing side project into a revenue engine that justifies the conglomerate's valuation to IPO investors.

6. 11,000 Jobs in One Week: The AI Restructuring Reaches Escape Velocity

On Tuesday, May 20, Meta eliminated approximately 8,000 positions, roughly 10% of its 78,000-person workforce. The layoff notices rolled across time zones starting at 4 a.m. Singapore time. An additional 7,000 employees were reassigned to AI-focused teams, and 6,000 open roles were closed. Combined, the changes touch nearly 20% of Meta's people.

The same week, Intuit announced it would cut 3,000 positions, 17% of its global workforce. The TurboTax and QuickBooks maker framed the restructuring as simplifying management layers. CEO Sasan Goodarzi insisted "none of it had to do with AI," though the company simultaneously disclosed multi-year deals with both Anthropic and OpenAI and is shuttering its Reno and Woodland Hills offices. Intuit's stock fell 20%, its worst single-day decline since 2003.

Across the industry, TechSpot reports that 2026 tech layoffs have surpassed 114,000 across 150 companies, a pace that projects to 370,000 by year's end, exceeding 2023's total.

Original analysis: The math tells a story the corporate memos do not. Meta's $145 billion capex guidance for 2026 divided by 8,000 eliminated positions produces a figure of $18.1 million in infrastructure spending per job cut, which means one human salary funds roughly 24 hours of operation at a 100-megawatt data center, and that exchange rate is the clearest signal of what this restructuring actually represents. The capital is not disappearing; it is being redirected from labor to silicon with a speed and scale that has no precedent in corporate restructuring. Intuit's denial is particularly transparent given the timing: you do not sign multi-year AI platform deals, cut 17% of your workforce, and close two offices in the same announcement cycle without the AI deals being the reason for the cuts. The workforce is being replaced, and the only question is how many companies will admit it before 2027.

Why it might not be as bad as it looks: Meta is simultaneously hiring for AI roles and reassigning 7,000 employees to AI teams, which means the net headcount reduction is smaller than the headline number suggests, and some displaced workers will find roles at the companies building the AI infrastructure that displaced them in the first place. And Intuit's 17% cut may reflect genuine operational bloat: the company grew headcount 40% between 2020 and 2024 during the pandemic hiring frenzy, so these layoffs could be a correction to that overcorrection, with AI providing the strategic narrative rather than the causal mechanism.

7. The Thing Nobody Is Talking About: Anthropic's Maia 200 Chip Talks Signal the Post-Nvidia Inflection

Tucked inside a Thursday Reuters report: Anthropic is in early-stage talks with Microsoft to run Claude on Microsoft's custom Maia 200 AI chips. The conversations are preliminary, no deal has been announced, and the market barely reacted.

It should have.

Maia 200 is Microsoft's second-generation custom silicon, designed specifically for AI inference workloads in Azure data centers, and it has never been offered to an external customer, which means Anthropic would be the first company outside Microsoft to run production workloads on it. If the deal closes, it means the fastest-growing AI company in the world has decided that Nvidia GPUs are not the only viable path forward, and Microsoft has decided that selling custom chips to Nvidia's competitors is worth the diplomatic cost.

This matters because the AI industry's single largest bottleneck is GPU supply, and Nvidia controls roughly 80% of the AI training chip market. Every AI company's growth plan depends on securing more GPUs than the competition. If Anthropic can run inference workloads on Maia 200 at competitive cost, it reduces its Nvidia dependency and creates a second source of leverage against every compute provider in its stack: Google (TPUs), SpaceX/xAI (Nvidia clusters), AWS (Trainium), and now Microsoft (Maia). That is a company building optionality into its infrastructure at the exact moment when infrastructure constraints determine who wins the AI race.

The broader signal: two years into the AI infrastructure buildout, the biggest customers are starting to push back against GPU monopoly pricing. Cerebras IPO'd at $67 billion last week, Anthropic is testing Maia, and Google has been running its own TPUs for a decade, so the post-Nvidia era has not arrived, but the conditions for it are being assembled, one chip deal at a time, by companies whose compute bills have grown large enough to justify building alternatives.

Limitations

This roundup relies on revenue projections that Anthropic shared with investors and reported by The Wall Street Journal and Financial Times, which are forward-looking estimates rather than audited financials, since Anthropic has never filed public financial statements and is not required to as a private company. The $30 billion annualized run rate figure, cited by multiple outlets, could reflect a single exceptional month extrapolated forward rather than a sustainable trend. SpaceX's IPO prospectus is a legal document designed to attract investors, and the $1.25 billion monthly Anthropic compute commitment has not been independently verified outside the S-1 filing. Google's $190 billion capex figure is a commitment, not a spend, and actual expenditure could differ materially if market conditions change. Layoff figures from Meta and Intuit are based on company disclosures and Reuters reporting, but the distinction between eliminated roles and unfilled positions varies across sources.

The Strongest Counterargument

The bearish reading of this week is that none of these numbers are sustainable. Anthropic's 130% quarter-over-quarter growth is the kind of inflection that happens exactly once in a company's life, during the transition from early enterprise adoption to mainstream deployment, and the growth rate will inevitably decelerate as the addressable market saturates. Google's $190 billion capex is an act of competitive panic, not strategic conviction, since Alphabet spent that much because Anthropic and OpenAI forced it to, not because users demanded agents that monitor their credit card statements. The executive order's cancellation reflects the fact that frontier AI is not yet dangerous enough to override the political power of the technology industry, and voluntary safety measures have so far prevented catastrophic incidents. The layoffs are a cyclical correction, not a structural displacement wave, and 114,000 jobs in a U.S. economy that creates 180,000 jobs per month is a rounding error.

Each of these counterpoints is individually plausible. Together, they amount to the argument that the AI industry is in a garden-variety bubble, and the correction will be painful but survivable. That may be right. The problem is that bubbles generate real infrastructure, and the infrastructure being built this week will exist regardless of whether the valuations that funded it turn out to be justified. The internet bubble produced the fiber-optic backbone that made modern cloud computing possible. Whatever happens to Anthropic's stock price, the data centers, the chip architectures, and the 114,000 displaced workers will still be here when the hype cycle fades.

The Bottom Line

This was the week the AI industry's scale became impossible to process at human speed. A $190 billion infrastructure commitment, an $80 billion IPO raise, a $10.9 billion quarter, 114,000 jobs lost, and no federal oversight framework. If you work in technology, your employer is almost certainly evaluating which of its functions can be automated by the tools announced this week. If you manage people, the Meta and Intuit playbook is now the template: sign the AI deals, flatten the org chart, reassign whoever survives. If you invest, the question is not whether AI is overhyped (it might be) but whether you can afford to be wrong about it being underhyped (you can't). If you build with AI, Anthropic's Claude Code and Google's Antigravity 2.0 are the two platforms to bet on right now, with OpenAI's Codex as the open-source hedge.

And if you are a policymaker: the window for establishing governance before the technology outruns the institutions is closing. This week, it got smaller.

Related Reading

• Last Week: Cerebras Raised $5.5 Billion. Anthropic Dethroned OpenAI.
• Google I/O 2026: The Token Price Paradox
• The AI Layoffs Paradox: Productivity Up, People Out