AI in 15 — May 22, 2026

Anthropic is about to post its first profitable quarter. Five hundred fifty-nine million dollars in operating profit on ten-point-nine billion in revenue. Two years ahead of internal projections. The narrative that AI companies are bottomless money pits just took a serious hit.

Welcome to AI in 15 for Friday, May twenty-second, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Heavy slate to close the week, Marcus. Anthropic discloses its first profitable quarter at ten-point-nine billion in revenue. Trump postpones his AI executive order hours before signing. Nvidia posts an eighty-one-point-six billion dollar quarter and tells the world demand has gone parabolic. Reuters finds Grok deployed at just three federal agencies out of more than four hundred. The Five Eyes intelligence alliance issues its first joint guidance on agentic AI security. The DRAM crisis starts hitting smartphone prices. And Samsung memory workers are taking home an average three hundred forty thousand dollar bonus.

An AI lab finally turns a profit.

Trump blinks on the AI executive order.

And your next phone is getting more expensive because of a chip you'll never own.

Lead story, Marcus. Anthropic's profitability disclosure. Walk me through it.

Genuinely consequential, Kate. Anthropic disclosed yesterday that Q2 2026 revenue will land at approximately ten-point-nine billion dollars. That's more than doubling Q1's four-point-eight billion. Estimated operating profit — five hundred fifty-nine million. That would be the company's first profitable quarter, arriving roughly two years ahead of internal projections. Compute cost per revenue dollar dropped from seventy-one cents to fifty-six cents. Enterprise accounts spending over a million dollars annually doubled to more than one thousand. CNBC, TechCrunch, and WinBuzzer all confirmed the numbers.

How does this square with the SpaceX deal we covered yesterday, Marcus? Fifteen billion a year in compute spend.

That's the right question, Kate. Anthropic warned in the same disclosure that profitability may not hold for the full year as new compute commitments hit. The SpaceX one-point-two-five billion per month obligation kicks in fully later this year. Enterprise revenue is growing faster than compute costs right now, but the curves cross and uncross depending on which quarter you measure. Still — a half-billion-dollar operating profit at a frontier lab is a milestone. Even one quarter of it changes the conversation.

And the IPO timing implications.

Direct, Kate. Anthropic is targeting an October listing at roughly nine hundred billion. OpenAI is filing confidential paperwork this week — possibly today — targeting an eight-fifty-billion-to-one-trillion-dollar valuation in September. Whichever files first sets the comparable metrics for the second. The Anthropic profit milestone counters the dominant AI-companies-are-bottomless-money-pits narrative — a narrative that has been politically convenient for skeptics but is increasingly hard to sustain when an AI lab posts a five-hundred-million-dollar operating profit. The pro-Western libertarian read here, Kate — competition and capital markets are doing exactly what they should. The frontier labs are now being asked to show their books, and at least one of them can.

Quick hits. Marcus, Trump was supposed to sign an AI executive order this week. What happened?

It got scrapped, Kate, hours before the ceremony. The order would have established a voluntary ninety-day pre-launch review framework for frontier AI models, with NSA involvement. The White House had summoned major lab CEOs to Washington — Dario Amodei among them. Then it all came apart. Trump reportedly said he didn't like certain aspects. Specifically, he was unwilling to slow what he sees as American AI dominance. AI adviser David Sacks reportedly opposed the order outright. Some labs had pushed for a much shorter window — fourteen days, not ninety. And not enough CEOs could clear their calendars on short notice.

Trump's public framing.

Quote, we're leading China, we're leading everybody, and I didn't want to do anything to get in the way of that lead, end quote. Kate, the postponement reveals the live fault line inside this administration. Security-focused factions — DoD, intelligence community — want guardrails before things go wrong. The pro-acceleration camp — Sacks, Musk-adjacent advisers — wants nothing slowing the labs. For companies about to go public, regulatory clarity matters. But Trump's instinct to keep the field unconstrained is arguably more valuable to their valuations than any voluntary framework would be. Expect this back on the agenda within weeks, probably in a softer form.

Nvidia earnings, Marcus. Eighty-one-point-six billion.

Record quarter, Kate, by a wide margin. Revenue up twenty percent sequentially, eighty-five percent year over year. Data center alone — seventy-five-point-two billion, up ninety-two percent annually. Gross margins at seventy-five percent. Q2 guidance is ninety-one billion plus or minus two percent. The company authorized an additional eighty billion in share buybacks and raised the quarterly dividend twenty-five-x — from a penny to twenty-five cents. Jensen Huang called it the largest infrastructure expansion in human history and said demand has gone parabolic.

What does this tell us about the cycle.

Nvidia is the single best leading indicator for AI capex, Kate. A ninety-one-billion-dollar forward quarter means hyperscaler and sovereign-AI spending is still accelerating, not normalizing. The dividend hike is the louder signal — Nvidia would not be returning capital at that scale if they thought this was a cyclical spike. They think the AI buildout is a multi-year secular trend. Quick perspective check — Anthropic's one-point-two-five billion monthly compute bill, which sounded enormous yesterday, is a rounding error on Nvidia's order book. The constraint at the frontier is no longer algorithms. It's wafers, power, and the willingness of hyperscalers to keep writing checks.

Grok in Washington, Marcus. Reuters had an exclusive.

Brutal one, Kate. Reuters reviewed the 2025 federal AI inventory. Of more than four hundred publicly identified vendor-specific AI deployments across the U.S. government, only three involve Grok — small pilots at Lawrence Livermore National Laboratory and the Election Assistance Commission. By contrast, two hundred thirty-four use OpenAI's models, thirty-three use Gemini or other Alphabet products, twenty-six use Anthropic's Claude — and Claude has since been blacklisted by the Trump administration over a contractual dispute. At DARPA, sources say staff prefer Gemini for engineering analysis and Claude for coding and writing. One Pentagon source told Reuters Grok was passed over because — quote — it's just not the best model out there.

And the IPO angle.

Sharp, Kate. SpaceX's S-1 leans heavily on the xAI segment being a credible growth story. The Reuters piece lands at the worst possible moment for that narrative. xAI has a two-hundred-million-dollar Pentagon classified-systems deal, but actual adoption is thin. The more important point — government technical evaluators, even in an administration that is enormously friendly to Musk personally, are picking on capability rather than politics. Career civil servants at DARPA chose Gemini and Claude over Grok despite the political environment pointing the other way. That's the system working at the procurement level. And it makes the Reuters story doubly painful for the SpaceX prospectus — buyers aren't political, even when sellers are.

Five Eyes guidance, Marcus. First joint statement on agentic AI.

Significant move, Kate. Earlier this month CISA and NSA in the U.S., plus their counterparts in the UK, Canada, Australia, and New Zealand, published a thirty-page joint document titled Careful Adoption of Agentic AI Services. First time the alliance has issued unified policy on a single AI attack surface. Five risk categories — privilege, design and configuration, behavioral, structural, and accountability risks. Core warning — AI agents capable of taking real-world actions are already inside critical infrastructure with broader access than organizations can safely monitor. Core recommendations — zero trust, defense-in-depth, least-privilege. Require human approval for high-impact actions. Fold agents into existing cybersecurity frameworks rather than treating them as a new vertical.

Practical impact.

Procurement, Kate. This is the kind of intergovernmental signal that drives enterprise procurement requirements. Expect Five Eyes alignment to start appearing in RFPs within months. Combine it with the CISA GovCloud leak we covered Wednesday and you can see the doctrine forming in real time. Agentic AI is moving from demo to deployment fast, and the security community is publicly saying the deployment is outrunning the controls. For frontier labs that want enterprise revenue — and that's all of them — Five Eyes alignment becomes table stakes. For agent orchestration startups, this is either a major tailwind or a serious compliance burden, depending on whether they engineered for it from day one.

DRAM crisis, Marcus. The cheap smartphone is dying.

Genuinely a story to pay attention to, Kate. A widely-shared piece on Hacker News this week walked through the mechanics of the worst memory pricing cycle in fifteen years. Building a single state-of-the-art DRAM fab costs fifteen to twenty billion dollars. Hyperscalers — Microsoft, Google, Meta, Amazon — are buying every wafer of high-bandwidth memory that Samsung, SK Hynix, and Micron can produce, because that's what stacks on Nvidia GPUs. Every wafer routed to HBM is a wafer not made into LPDDR5X for a phone or DDR5 for a laptop. Data centers now consume an estimated seventy percent of memory production. Q2 2026 DRAM contract prices rose fifty-eight to sixty-three percent sequentially. NAND surged seventy to seventy-five percent.

What it means downstream.

Real money, Kate. Samsung Galaxy S26 reflects a sixteen percent LPDDR5 cost increase. Xiaomi has warned of higher prices. Base-tier smartphones may regress to four gigabytes of RAM in 2026 — which is a meaningful step backward for any user who runs more than three apps at a time. This is one of the first concrete, mass-consumer ways the AI boom is hitting wallets. It's also a useful corrective to the AI-is-just-software framing. At the bottom of the stack, AI is wafers, fabs, and physical scarcity. Apple's upgrade cycle, the entry-level smartphone market, and PC gaming all sit downstream of choices being made by Microsoft Azure procurement.

Samsung bonuses, Marcus. The flip side.

Roughly three hundred forty thousand dollars per employee on average in the memory chip division, Kate. Quartz reported it. Hacker News flagged it with one hundred ninety-nine points. Korean cost of living means several workers commented that the bonuses are effectively early retirement money. The average likely masks heavy skew toward senior staff, but the scale is unprecedented. The sharper edge to the story — at the same moment Samsung memory workers are getting three hundred forty thousand dollar bonuses, American tech workers continue to be laid off as companies cite AI optimization. We've been tracking that pattern all week — Intuit, Cloudflare, Standard Chartered, Meta, Amazon, Cisco, Microsoft. The AI value chain is allocating its surplus right now to the physical layer — wafers, fabs, and the people who run them. South Korea, Taiwan, and to a lesser extent Japan are quietly becoming the biggest non-U.S. winners of the AI boom. Pro-Western libertarian read, Kate — these are allied democracies. The supply chain still concentrates in friendly hands. That's a good outcome from a strategic competition standpoint, even if it's a painful one for Western consumer electronics buyers.

One more, Marcus. The viral essay arguing AI is just unauthorized plagiarism at a bigger scale.

Seven hundred sixty-seven points on Hacker News, Kate. Six hundred seventy-two comments. The essay argues that the AI industry's fair-use defense is a category error — that learning from data at scale, by entities optimizing the output for commercial sale, is meaningfully different from a human learning to draw by studying art. The comment thread split. Some defend the math-not-memorization view — you're estimating token distributions, not reproducing originals. Others argue that even if it's legal, it's hollowing out the underlying creator economy. Website owners pay to host content that gets crawled and then summarized away from their traffic. A companion piece, Shunning AI Is the Human Choice, hit three hundred fifty-three points the same day.

Why it matters.

This is the cultural temperature on AI from a technical audience that mostly uses these tools daily, Kate. And it's getting colder, not warmer. The backlash matters because it shapes the political environment for the trillion-dollar IPOs about to land. Connect it to the booed commencement speakers we covered Monday, to the Standard Chartered apology, to the Trump executive order falling apart. Public sentiment is the variable that politicians actually respond to. Right now the people who use AI most fluently are the ones turning most skeptical about how the industry behaves around copyright, labor, and provenance. The labs that figure out how to share more value with creators win the next political cycle. The ones that don't, don't.

Big picture, Marcus.

Three through-lines closing the week, Kate. First — AI is moving from exciting technology to load-bearing infrastructure for the global economy. OpenAI is producing original mathematical discoveries we covered Thursday. Anthropic is posting real profits today. SpaceX is selling compute at one-and-a-quarter billion dollars a month. Nvidia is shipping data center revenue at a three-hundred-billion-dollar annual run rate. Smartphone prices are rising because too many wafers are headed to AI servers. The U.S. government is writing security doctrine for AI agents. And three companies that together may be worth three-and-a-half trillion dollars are racing to file IPO paperwork inside the next six months. Second — the countervailing forces are real and they keep showing up. Public sentiment turning skeptical. The plagiarism essay hitting the front page. Trump postponing an AI executive order because the politics weren't lining up. Grok's flop in Washington showing that not every well-funded lab wins. The viral discourse is harder on AI than it was six months ago. Third — the strategic competition picture is genuinely good for the West. Anthropic posts a profit. Nvidia's order book accelerates. Samsung memory workers — South Korean, allied — get unprecedented bonuses. Five Eyes coordinates on doctrine. The pro-Western libertarian read here, Kate, is that markets and alliances are doing what they should. Competition is forcing real numbers into daylight. Allied supply chains are capturing the surplus. The government is writing rules through coalitions of consenting democracies rather than imposing them top-down on industry. The risk is straightforward — the gap between what AI can do and what the public is being asked to swallow keeps widening, and the political response to that gap is unpredictable. Watch next week. Both IPO filings should be public by then, and every story we covered this week gets repriced against the actual numbers.

That's your AI in 15 for today. See you tomorrow.