AI in the News

AI in the news — week of May 17, 2026

Meta confirms 8,000 layoffs against a $135B capex print. OpenAI hands the EU GPT-5.5-Cyber while Anthropic holds Mythos. Anthropic overtakes OpenAI for business customers, eyes $950B. Google I/O ships Tuesday. The concentration story keeps writing itself.

Sid Smith

17 May 2026 • 8 min read

Week ending Sunday May 17. The story I was going to lead with — Anthropic eyeing a $950 billion valuation and quietly overtaking OpenAI among business customers — got displaced by the Meta layoff print. Then displaced again by OpenAI handing the EU access to GPT-5.5-Cyber while Anthropic still holds Mythos behind Glasswing. Google I/O ships Tuesday and is about to add another data point to the agent race. The four positions I keep returning to are all in this week's news: concentration is getting worse, sovereignty is getting harder, the labor frame is accelerating, and the case for owning your own stack keeps getting stronger.

Plus a personal one: this is the week the vibe-coding series kicks off on the schedule I'll be posting it on, and the news this week makes the case for itself.

Meta confirms the 8,000-job cut. May 20.

Meta confirmed this week that the 8,000-job cut goes live on May 20. That's about 10% of the company, plus an additional 6,000 open requisitions cancelled, for an effective headcount reduction in the neighborhood of 14,000. Severance is sixteen weeks plus two weeks per year of service. Health coverage runs eighteen months. The internal framing is "structural, not performance," and the org is being reshaped into AI-focused pods. Additional cuts are planned for the second half.

Zuckerberg's town-hall quote is the part worth reading slowly:

We basically have two major cost centres in the company: compute infrastructure and people-oriented things. If we're investing more in one area to serve our community, then that means we have less capital to allocate to the other.

That's an executive saying out loud what the rest of the cohort has been saying through proxies. The Meta capex print for 2026 sits at $125–145B. Across Amazon, Alphabet, Microsoft, and Meta the total crosses $725B for the year, up 77% from 2025's already-record $410B. Layoffs.fyi has already tracked 62,000+ tech-sector cuts in the first 17 weeks of 2026 — 71% of the full 2025 total before the calendar's even half done.

The Zuckerberg quote is more honest than most of what's been said in this category. He's not pretending the AI is taking the jobs. He's saying the capital is. Which is structurally the same outcome for the people losing their jobs, and a different argument about what the company is doing. The labor frame keeps moving faster than the realistic view said it would. I'd rather be wrong about the pace than be caught off guard.

OpenAI hands the EU GPT-5.5-Cyber. Anthropic still holds Mythos.

May 11. OpenAI announced it would grant the EU access to GPT-5.5-Cyber, a variant of its latest frontier model tuned for security research. Anthropic, by contrast, still has no European government on the Glasswing partner list — and Euro-area finance ministers convened to talk about it. The framing from Brussels: equivalent treatment, or escalation.

This is the disclosure-shape question I flagged last week coming due faster than I thought it would. Anthropic's Glasswing structure was the right call on day one — give defenders the model first, hold broad release. But the geometry of "defenders first" only works if "defenders" includes every government that has to defend critical infrastructure. The current Glasswing partner list is heavily US-skewed (AWS, Apple, Microsoft, Google, JPMorgan, Linux Foundation, the usual). European banks aren't on it. Euro regulators aren't on it. That's a sovereignty problem and the EU is right to push on it.

The other side of this week's Mythos coverage: TechTimes reported that a private Discord group got unauthorized access to Mythos Preview through a third-party vendor environment on the same day Glasswing was announced. So the choice isn't "Mythos contained inside the partner list" versus "Mythos leaked to attackers." It's "Mythos contained inside most of the partner list, plus a Discord that wasn't supposed to have it." That's a worse fact pattern for the contained-disclosure argument.

Watch the EU response. If Anthropic ships Mythos access to European supervisors in the next two weeks, the model holds together. If it doesn't, expect Brussels to use the Omnibus deal leverage they just bought.

Anthropic eyes $950B, overtakes OpenAI among business customers

May 14. AI Insider reported Anthropic overtook OpenAI as the leading provider to business customers, with active enterprise customer count crossing OpenAI's. Same window, the company is in talks for a funding round at a $950B valuation. Three weeks ago Google committed up to $40B. Anthropic also closed a $30B Series G in late March.

This is the part of the concentration story I keep landing on. The lab+PE consulting vehicle (TDC from OpenAI, the $1.5B Blackstone joint from Anthropic — both announced May 4) is the integration-and-deployment side. Add the capital allocator (Google's $40B, Blackstone, Hellman & Friedman, Goldman) and the customer-base distribution (PwC, the Wall Street push into Moody's and Microsoft 365). And then the small-business push (Claude for Small Business, shipped May 13) at the bottom of the funnel.

That's not the three-party transaction I was writing about last week. That's the same party showing up on all three sides of the table, plus the cohort sponsoring the lawyers. If you sit on a PE-backed company that's about to get the deployment-services pitch from TDC or its Anthropic-Blackstone equivalent, the data-flow terms are the part to push back on. The on-prem case gets stronger every week.

The PwC partnership specifically is worth watching. PwC has the relationship with every Fortune 500 board. Anthropic now has a delivery channel into all of them. That is exactly the kind of one-handshake distribution that built the original consulting+software stacks (the SAP-Accenture pattern, the Salesforce-Deloitte pattern). It works for the providers. It tends to be a long-run net loss for the customers — measured in lock-in, in pricing power, in the dilution of internal capability.

Google I/O 2026 — Tuesday

Google I/O kicks off Tuesday May 19 with the main keynote at 1 p.m. ET. The expected drops:

Gemini 3.5 (some sources naming it 3.2 instead). Roughly the OpenAI GPT-5.5 class. Incremental rather than breakthrough.
Gemini Spark / Gemini Agents — Google's answer to the OpenAI Operator and Anthropic computer-use stack.
"Remy" — described internally as a 24/7 personal agent that "elevates the Gemini app into a true personal assistant that can take actions on your behalf." Cross-app, screen-aware, proactive.
Veo upgrades and Omni video generation.
A reported Google-built laptop nicknamed the "Googlebook" running Gemini on-device.

The Remy framing is the one I'll be reading for. "Takes actions on your behalf" is the agentic frame that I've been writing about for two years, and Google's track record on shipping agent surfaces in a way that's actually trustworthy (not the Mariner project that just got shut down) is mixed. The capability is probably there. The discipline around it is the question.

If you've been reading the vibe-coding series I'm starting this week, the Remy announcement is going to land in a particular light. Persistent agents with cross-app access and the ability to "take actions on your behalf" are exactly the category where AI's current defaults compound badly. The 35 services on my laptop were small-time compared to a Remy that's also writing to your filesystem and your calendar and your email all at once. I'll write a real reaction post once I see the keynote.

Smaller items

Gates Foundation + Anthropic announced a $200M partnership on May 14 covering global health, agricultural development, and education. Focus is on AI applied to underserved populations rather than enterprise customers. The right kind of move and an unusual one in this market.
PwC + Anthropic expanded their strategic alliance on May 14 (Anthropic announcement). The two firms have been working together since 2023; this is the formal "Claude is now the primary AI substrate inside PwC's enterprise practice" rev.
OpenAI shipped "Helping ChatGPT better recognize context in sensitive conversations" on May 14 and "Work with Codex from anywhere" on May 15. The sensitive-conversations work is overdue. The Codex-from-anywhere flag is the agent push by a different name.
The Pentagon has now signed AI deals with 8 of the major frontier labs and pointedly excluded Anthropic. Reporting from early May; the exclusion is sticking. Anthropic's red-team posture and refusal to bid on certain DoD profiles is the proximate cause. The longer-run effect is that Anthropic gets to keep its non-DoD positioning and the other labs absorb the political surface that comes with the contract.

What this week tells me

Three threads.

Concentration is the story of 2026. Anthropic-Blackstone, OpenAI-TPG, Anthropic-PwC, Anthropic-Google's $40B. The lab + PE + integration-partner triangle is now the dominant shape, and the small-business push at the bottom of the funnel means the same triangle is being run at every tier of customer. Hyperscaler concentration is what made the cloud era what it was, for better and for worse. The AI version is consolidating faster.

Sovereignty is the next pressure point. The EU vs. Mythos question this week is the test case. If Brussels can't force Anthropic to share capability with European supervisors on equivalent terms, expect a sovereign-AI investment surge in the back half of 2026 — France's Mistral push, Germany's Aleph Alpha rebuild, EU public infrastructure projects. The Pentagon-Anthropic split is the US-side analog. Governments are going to build or buy their own frontier models when the commercial labs won't sell them.

The labor frame is faster than the realistic view said. Meta is the third major cohort cut of 2026 and the calendar isn't half done. The Zuckerberg quote about "two cost centres" is the most honest thing any executive has said about it. The right answer isn't to slow AI down. It's to be honest about what's actually getting paid for with the layoffs, and to design the human+AI work patterns that take the pace of the cuts back down to something the labor market can absorb.

Plus the personal angle, since I started writing this on the Saturday before the vibe-coding series ships: every news item this week makes a quiet case for owning the stack you can own, watching the data flows you can watch, and being honest about what AI's defaults compound into when nobody's looking. That's most of what the next several months of writing here is going to be about.

Next Sunday: Google I/O fallout, EU response on Mythos, and whatever the post-Meta layoff dominoes look like across the cohort.