Move-in day. Last week the walls between the AI labs came down. Microsoft and OpenAI ended exclusivity, the Pentagon took eight vendors, Musk admitted distillation under oath. We called it the end of the team-sport era.

This week, the labs moved in furniture.

Apple turned the iPhone into a multi-AI app store. Anthropic ran a four-day Wall Street blitz that ended with a reported $200 billion check to Google. The Commerce Department got pre-release login credentials at Google DeepMind, Microsoft, and xAI. Cloudflare cut 20% of its workforce on record-breaking quarterly revenue and called it an “agentic AI-first operating model.” DeepMind workers in London filed for union recognition — the first such request at any frontier AI lab. And in Beijing, a four-billion-product catalog quietly slid behind a chat box.

The model isn’t the bottleneck anymore. The storefront is.

📱 Apple Just Made Itself an AI App Store

And ChatGPT’s exclusive ride is over

On May 5, Bloomberg’s Mark Gurman reported that Apple is building “Extensions” into iOS 27, iPadOS 27, and macOS 27 — a system-level framework that lets users pick which third-party AI powers Siri, Writing Tools, Image Playground, and the rest of Apple Intelligence. Test-build text reportedly reads: “access generative AI capabilities from installed apps on demand.” Models from Google and Anthropic are already in internal testing. Expected unveil at WWDC on June 8.

Translation: ChatGPT’s de facto exclusivity since iOS 18.2 is over. Apple is doing what it has done at scale exactly once before — opening a tightly-controlled platform to outside choice — and turning iOS into a neutral host for whichever frontier model the user prefers.

The architectural move matters more than the announcement. Apple has spent a year stitching together two separate AI strategies. The January Gemini deal puts a custom 1.2-trillion-parameter Gemini model behind the next-generation Siri at the OS layer (~$1B/year, eight times the parameter count of Apple’s current cloud models). Extensions sits on top of that, letting any user opt out of Apple’s stack entirely and route work to Claude, Gemini, ChatGPT, or whoever else ships an Extensions-compatible app. Both layers coexist by design.

Why it matters: if you’re shipping a consumer AI product, the iPhone just became the single most important distribution surface arriving in 2026. The Apple App Store-style gallery for compatible AI apps is the new Top of Funnel. And if you’re an iOS developer who’s been building chat features around the OpenAI SDK because that’s what’s available — you have until WWDC to think about which model you’d actually want your users to default to, because they’re about to get the choice.

Hype vs. Reality: 8/10 — Bloomberg’s Gurman has the strongest Apple-leak track record in the industry, and the Extensions branding has now appeared in at least three separate Bloomberg pieces over two months. WWDC will confirm the rest.

💰 Anthropic Spent the Week Becoming Goldman Sachs

Then leaked the bill

Across four days, Anthropic launched a $1.5 billion enterprise services JV with Wall Street’s largest private equity firms, deployed Claude across the Microsoft Office stack, took over SpaceX’s entire Memphis data center, disclosed 80x annualized Q1 growth against a planned 10x, and watched The Information drop the reported size of its Google Cloud commitment — $200 billion.

Monday, May 4 — the joint venture. Anthropic, Blackstone ($300M), Hellman & Friedman ($300M), and Goldman Sachs ($150M) launched a new enterprise AI services firm backed by Apollo, GIC, General Atlantic, Leonard Green, and Sequoia. Total committed capital: ~$1.5B. The structure is explicitly modeled on Palantir’s forward-deployed-engineer playbook — Anthropic engineers embedded inside mid-sized portfolio companies, redesigning workflows around Claude. Hours earlier, Bloomberg reported OpenAI was raising $4 billion for “The Deployment Company” at a $10B valuation, with TPG, Brookfield, Advent, and Bain Capital. In a single news cycle: Anthropic launched a ~$1.5B enterprise-services JV; OpenAI was reportedly raising $4B for its own deployment company at a $10B valuation. Accenture’s $2.7B GenAI revenue line just acquired serious competition.

Tuesday, May 5 — Wall Street briefing, and the receipt. Anthropic CEO Dario Amodei shared a New York stage with JPMorgan CEO Jamie Dimon for the first time. Anthropic released ten ready-to-run finance agent templates (pitchbook builder, KYC screener, earnings reviewer, model builder, valuation reviewer, GL reconciler, month-end closer, statement auditor) as plugins for Claude Cowork and Claude Code. The big news: Excel, Word, and PowerPoint add-ins were already in early access; Outlook would land in public beta two days later (more on that below). Moody’s launched an MCP app embedding credit data on 600 million companies directly into Claude. Eight new connectors (Dun & Bradstreet, Verisk, Third Bridge, Guidepoint, IBISWorld, SS&C IntraLinks, Fiscal AI, Financial Modeling Prep). FIS announced a co-built AML agent compressing anti-money-laundering investigations from days to minutes. Citadel and Walleye Capital named as customers — Walleye said 100% of employees use Claude Code.

The same day, The Information reported (and Reuters picked up) that Anthropic has committed $200 billion to Google Cloud over five years as part of the April TPU deal we covered in Issue #011. Reuters could not independently verify the figure, but Alphabet stock moved on the report. The math, per The Information: contracts with Anthropic and OpenAI now account for more than half of the $2 trillion in revenue backlog at AWS, Azure, Google Cloud, and Oracle combined. Anthropic alone is reportedly ~40% of Google’s $460B backlog.

Wednesday, May 6 — compute reveal. At Anthropic’s Code with Claude developer conference, Amodei told the room: “We tried to plan very well for a world of 10x growth per year. And yet we saw 80x. And so that is the reason we have had difficulties with compute.” Annualized revenue run-rate: $30 billion in April 2026, up from $9 billion at the end of 2025 — a 3.3x climb in five months. Same day: the SpaceX Colossus 1 deal. Anthropic took the entire Memphis facility — 220,000+ Nvidia GPUs, 300 MW within a month. Claude Code rate limits doubled across Pro, Max, Team, and seat-based Enterprise. Peak-hour reductions removed. Buried in the press release, a sentence most outlets skipped: “we have also expressed interest in partnering with SpaceX to develop multiple gigawatts of orbital AI compute capacity.”

Thursday, May 7 — Microsoft’s backyard. Claude for Microsoft 365 hit GA for Excel, Word, and PowerPoint, with Outlook in public beta. Context carries automatically across all four apps. Two AppSource listings, deployable from the Microsoft admin center. Microsoft 365 Copilot customers can also work with Claude AI models directly within Excel and PowerPoint. Microsoft’s own Copilot just got Claude piped into the apps it runs — a structural concession that Claude is now distribution-grade competition inside Microsoft’s own product surface.

Friday, May 8 — the second compute deal. Anthropic inked a $1.8B deal with Akamai. Geographic redundancy is now compute strategy.

Two AI labs now control more than half of the cloud industry’s reported forward revenue.

That isn’t a customer-vendor relationship. That’s a structural dependency that runs in both directions, and it cuts the same way for Microsoft, Amazon, Google, and Oracle.

Why it matters: the consulting market the labs are coming for is roughly $375 billion. Accenture booked $5.9B in GenAI bookings in fiscal 2025 and is the #1 player. The labs aren’t competing with that on price — they’re competing on architecture: we own the model, we ship our engineers, we capture the implementation revenue too. If your product strategy includes “we’ll use Anthropic or OpenAI to build vertical workflows for our customers,” you have a new competitor with a tighter cost structure and a more ambitious cap table.

Hype vs. Reality: 9/10 — Every product link is primary-source. The $200B figure is reported, not independently verified by Reuters. Everything else is on Anthropic’s own blog, in Blackstone’s press release, or on Wall Street stages.

🚨 The Government Just Got a Test-Lab Login

And the Pentagon answered the Anthropic question

Tuesday, May 5: Google DeepMind, Microsoft, and xAI signed renegotiated agreements with the Center for AI Standards and Innovation (CAISI), the Commerce Department’s NIST-housed safety arm. Pre-deployment evaluations and post-deployment research, on the most capable models, with safety guardrails reportedly stripped back so federal cyber experts can probe latent capabilities for biosecurity, chemical-weapons synthesis, cyber-offensive operations, and autonomous exploitation. CAISI Director Chris Fall confirmed the agency has already completed 40+ evaluations including unreleased state-of-the-art models. OpenAI and Anthropic signed similar pacts in 2024 under Biden — these renegotiated agreements are explicitly aligned with the Trump AI Action Plan.

OpenAI’s Chris Lehane confirmed via LinkedIn that the company handed over GPT-5.5 and a separate GPT-5.5-Cyber model before public release. The aviation-style certification regime for frontier AI just stopped being a hypothetical.

Then on Thursday, May 7, the Pentagon answered the question we asked in #009 about whether the Anthropic blacklist meant the DoD was abandoning safety-aligned vendors. At the AI+ Expo, Pentagon Defense Under Secretary Emil Michael reframed Mythos — the Anthropic model that drew the supply-chain-risk designation — as opportunity rather than threat: “The Mythos moment is really a cyber moment, and it’s how is the U.S. government going to deal with cyber, how do we operationalize fixing things that need to be fixed?” Defense One headlined Michael’s broader stance: the Pentagon would “never again” rely on a single AI provider.

Pentagon CDIO Cameron Stanley confirmed the operational footprint at the same event. “Operation Epic Fury leveraged Palantir’s Maven Smart System in order to conduct strike missions across the entire battlespace — 13,000 targets in 38 days,” Stanley said, alongside “894 million tokens per day in agentic workflows.”

Why it matters: if you’re selling AI into anything defense-adjacent, regulated, or security-sensitive, “we passed our internal evals” is no longer the answer. The CAISI evaluation regime is rapidly becoming the de facto floor — and developers handing over jailbroken versions of their own models is becoming the de facto cost of doing federal business. Two months ago that would have read as compliance theater. After Mythos, after the Glasswing leak, after the Bedrock-OpenAI deal, after the May 5 announcement — it reads as procurement.

Hype vs. Reality: 9/10 — The CAISI agreements are confirmed across Reuters, CNN, and Nextgov, with Microsoft, OpenAI, and Commerce all on the record. The Pentagon reframe is one event covered well by DefenseScoop and Breaking Defense. What’s still TBD: whether voluntary pre-release vetting becomes statutory.

🛠️ “Agentic AI-First” Is the New “Lean”

Two layoffs, one earnings beat, one new corporate vocabulary

Tuesday, May 5, 6:55 AM — Coinbase CEO Brian Armstrong emailed the company that 14% of staff (~700 people) were being cut. Layered on top: a structural reorg eliminating “pure managers,” capping the org at 5 layers, and creating “AI-native pods” including one-person teams directing agents that handle the work of engineers, designers, and product managers. The line that’s going to live in HR-deck history: “rebuilding Coinbase as an intelligence, with humans around the edge aligning it.” Armstrong, separately: mass layoffs are “coming to every company.” $50–60M restructuring charge. System access revoked the same day.

Thursday, May 7 — Cloudflare cut 1,100 employees (~20% of a 5,000+ person workforce). Same morning: Q1 2026 revenue of $639.8M, up 34% YoY, the highest quarterly print in company history. Cloudflare has never done a mass layoff in its 16-year history. CEO Matthew Prince’s framing in the memo and earnings call: “agentic AI-first operating model.” Internal AI usage is up 600% in the last three months. Sales-quota carriers were exempt; cuts hit eng, HR, finance, marketing, and support across all geographies. Severance was unusually generous — full base pay through end of 2026, equity vesting through August 15, cliff waivers. The market wasn’t impressed: Cloudflare stock sank 24% after earnings.

Friday, May 8 — the BLS jobs report put the macro on the layoffs: information-sector unemployment ticked from 3.6% in March to 3.8% in April. The information sector lost 13,000 jobs in a single month. Total economy added 115K, beat expectations, unemployment held at 4.3%. Tech jobs as a share of all jobs are at record lows even as tech stocks set records. The Challenger Gray & Christmas April report attributed 26% of all April layoffs to AI — 21,490 of 83,387 cuts. Tech-sector cuts YTD: 85,411, up 33% vs. the same period in 2025.

Then OpenAI handed us the company-side answer. On May 6, OpenAI published “How Frontier Firms Are Pulling Ahead” — a study of de-identified enterprise usage data. Companies in the 95th percentile of AI usage now use 3.5x more intelligence per worker than typical firms, up from 2x a year ago. Message volume only explains 36% of the gap. The other 64% comes from depth — richer context, more complex tasks, more substantive outputs. The single largest delta: frontier firms send 16x more Codex messages per worker. Delegated work, not chat assistance.

Issue #011 gave you the worker-side data: Anthropic’s 81,000-user economic survey found that the workers reporting the highest AI productivity gains also reported the highest displacement anxiety. OpenAI just gave you the company-side data on the same vector. Same arrow. Different end of it.

Why it matters: “agentic AI-first” is the new “lean” — the corporate vocabulary that rebrands a margin-improvement initiative into an operating-model story. Cloudflare’s 600% internal AI usage is the metric of the moment, and you can expect every public-company CFO to start citing it next quarter. Two practical takeaways for builders:

  • If you’re an employee anywhere AI productivity is being measured, document your gains in writing this quarter. Not in your head, not in a 1:1, in a system the comp committee can read. The frontier-firm data isn’t going away.
  • If you’re a founder, the talent floor just dropped. The same Cloudflare cut that made the news this week put 1,100 people who survived the agentic-AI-first internal reorg on the open market. That’s a hiring signal, not just a labor story.

Hype vs. Reality: 9/10 — Coinbase’s filing is public, Cloudflare’s earnings are public, the BLS data is public, Challenger’s report is public, OpenAI’s frontier-firms study is OpenAI’s own data. The receipts are clean.

✊ The 600 Didn’t Kill the Deal. So They Filed for the Union.

DeepMind UK files for the first union recognition at a frontier AI lab

A week and a half ago, 600 Google employees signed an open letter urging Sundar Pichai not to let Google’s Pentagon deal allow autonomous weapons or domestic surveillance. The deal closed anyway — Google joined the eight vendors covered in Issue #012. On May 4, ~1,000 DeepMind staff in London escalated.

DeepMind workers formally requested recognition of the Communication Workers Union and Unite the Union as joint representatives. 98% of CWU members at DeepMind voted yes. The letter gave Google management 10 working days to voluntarily recognize the unions or face a legal compulsion process. Google told Fortune it had “received” the letter but disputes that a formal unionization vote has occurred. The demand list includes commitments not to develop AI for weapons or surveillance that violates international norms, stronger whistleblower protections, and a worker right to abstain from projects that violate ethical or moral beliefs.

If recognized, this would be the first union at a frontier AI laboratory. Anywhere.

The Guardian quoted an anonymous DeepMind worker on motivation: the U.S. war in Iran combined with the Trump administration’s feud with Anthropic showed the Pentagon was “not a responsible partner.” CWU’s national tech officer John Chadfield told Truthout: “This is a really important moment for workers in the tech sector.” Workers are also reportedly considering “research strikes” — refusing to work on specific projects — alongside the union recognition push.

In #012 we noted that the threshold for collective tech-worker resistance had moved up dramatically — Project Maven took 4,000 signatures and a dozen senior resignations to kill in 2018, while last week’s 600 signatures didn’t move Google an inch. The escalation ladder went from open letter to union recognition request in three weeks. That’s not a slow burn. That’s a step change in how AI engineers organize against the products they ship.

Pair this with the cultural pushback texture from the same week — the Academy of Motion Picture Arts and Sciences ruled May 1 that AI-generated actors and AI-written screenplays are ineligible for Oscars, full stop (“only roles demonstrably performed by humans with their consent”; “screenplays must be human-authored”). The Academy and the union don’t share members, but they share an instinct.

Why it matters: if you run an AI-adjacent team, the labor stack just got more complicated. “Agentic AI-first” reads differently to engineers who can now choose between working on the agent or organizing against it. Coinbase’s “no pure managers” memo and Cloudflare’s “support roles aren’t the roles that drive companies forward” memo are corporate ideologies — and DeepMind workers are showing you the counter-ideology that’s now legally viable to organize around. Founders who build AI tooling for engineers should pay close attention to how this plays in the next two quarters.

Hype vs. Reality: 9/10 — The union request is verified across Fortune, Truthout, Gizmodo, the Guardian, Business Insider, and Breitbart. Google has confirmed receiving the letter but hasn’t recognized the unions. The 10-working-day clock is real and started May 5.

📡 Nvidia Bought Its Own Customers Again, And It’s Working

$40 billion in equity bets in five months. The circular thesis just got priced.

Wednesday, May 6: Nvidia announced up to $3.2 billion in Corning — warrants for 15 million shares at $180 (+$500M pre-funded warrant for 3M more shares) tied to three new US fiber-optic factories, 3,000 jobs, and a 10x expansion of US optical manufacturing capacity for AI data centers. Corning stock +12%, NVDA +6%.

Thursday, May 7: Nvidia committed up to $2.1 billion to IREN — a 5-year right to buy 30 million shares at $70 each, alongside an agreement for IREN to deploy up to 5 GW of Nvidia DSX-aligned AI infrastructure. IREN’s Q3 contracted ARR jumped to $3.7B against a Microsoft contract worth $1.9B/year and the new Nvidia contract worth $0.7B/year.

That’s two multibillion-dollar deals in 24 hours, with the same playbook: Nvidia takes equity, partner agrees to deploy Nvidia silicon, both stocks go up, both companies announce commercial growth. Per CNBC’s tally backed by FactSet data: Nvidia has now committed over $40 billion to AI equity investments in the first five months of 2026 — anchored by the $30B OpenAI stake, plus seven multi-billion deals in public companies and ~24 private rounds. Wedbush analyst Matthew Bryson called the pattern “squarely into the circular investment theme” — money moving between the same companies — but added it could create “a competitive moat” if it works.

Then The Information closed the loop. Anthropic’s reported $200B Google Cloud commitment isn’t just a compute deal — it’s the spending side of the same circular structure that runs through every major lab right now. Google invests up to $40B in Anthropic. Anthropic spends a reported $200B back on Google’s TPUs. Google’s TPUs use Broadcom-designed silicon. The pattern repeats with Nvidia and OpenAI, with AWS and Anthropic, with SpaceX and Anthropic now too. Engadget called these “expensive circular deals.” Engadget is being polite.

The “is this sustainable” question is now less interesting than the “what does it look like when one of these counterparties stops paying” question. Anthropic projects $20B in 2026 compute costs. OpenAI projects $45B. The chips have to be ordered before the revenue exists to pay for them.

$2 trillion in revenue backlog at AWS, Azure, Google Cloud, and Oracle combined. Two AI labs account for more than half of it.

That’s not a customer relationship. That’s mutual hostage-taking with quarterly accounting.

Why it matters: if you’re building on top of any frontier API, the counterparty you actually depend on isn’t your model vendor. It’s whoever is funding their compute. Run the chain on your dependencies. Whose money is paying for the GPUs my agent runs on, and what do they need to be true for me to keep getting the latency I expect? That used to be a question for venture analysts. Now it’s a stack-architecture question.

Hype vs. Reality: 9/10 — The Corning, IREN, and FactSet numbers are public-company filings. The $200B figure is reported by The Information, picked up by Reuters, and Reuters explicitly couldn’t verify. Treat that one as directionally true, magnitude-uncertain.

🛒 The Agent-Commerce Layer Just Moved East

4 billion products, one chat box, and Alipay underneath

Sunday, May 10: Reuters reported Alibaba is integrating its Qwen AI app with Taobao and Tmall — a single conversational shopping interface that browses, compares, recommends, and purchases against Alibaba’s 4-billion-product catalog. Backed by an Alibaba-built “skills library” handling logistics, customer service, and after-sales. Alipay native checkout. Virtual try-ons. 30-day price tracking. Inside Taobao itself, a separate Qwen-powered shopping assistant launches alongside.

Per Qwen’s own metrics: 300 million monthly active users across Alibaba properties, with ~140 million first-time AI shopping experiences logged during the Chinese New Year campaign. This isn’t a US-style “agent recommendation pilot” — it’s the most ambitious test of agentic shopping at any scale, anywhere.

The contrast with the US side of the same week is the story. Amazon has been cautious about full agent autonomy in shopping. Shopify punts to external agents instead of running its own. The closest US analog this week was OpenAI’s ad expansion on May 7 — ChatGPT ads now testing in UK, Mexico, Brazil, Japan, and South Korea on logged-in Free/Go users, alongside a self-serve Ads Manager pilot for SMBs. ChatGPT is still the discovery layer. Alibaba just collapsed discovery, comparison, payment, and post-purchase into a single agent surface.

The US protocol stack is being built — the Cloudflare/Stripe Projects spec from Issue #012, Google’s AP2 (Agent Payments Protocol) at FIDO, Mastercard’s Verifiable Intent. The Chinese stack is shipping the consumer surface that uses it.

Underneath both: the compute landlords. Hut 8 signed a 15-year, 352 MW data center lease for a base value of $9.8 billion at its Beacon Point campus in Texas (Nueces County, ~525 acres). Three 5-year renewal options could lift the total to $25.1 billion. Tenant: confidential, “high-investment-grade.” Hut 8 stock jumped ~30%. The site was originally configured for 224 MW Bitcoin mining and got redesigned for 352 MW once Nvidia’s DSX rack-density spec evolved. Bitcoin miners pivoting to AI compute landlords is the #012 MARA→Long Ridge story scaling up — Hut 8 now has 597 MW under contract worth ~$16.8B in base-term value and a 7,500 MW pipeline. Whoever your storefront is, somebody else owns the parking lot.

Hype vs. Reality: 8/10 — Alibaba is “Reuters citing a source familiar with the plan,” not an official launch event. Hut 8 is a public-company filing. OpenAI ads is OpenAI’s own announcement.

📊 Quick Signals

OpenAI’s full week of shipping. GPT-5.5 Instant becomes the new default ChatGPT model (May 5). GPT-5.5-Cyber launches as a security-focused variant for first-line cyber defense partners (May 7). New real-time voice and translation models for the API (May 7). Trusted Contact safety feature in ChatGPT (May 7). Running Codex Safely (May 8) — sandboxing, network policy, OS-keyring credentials, agent-native telemetry, auto-approval subagent. Read the Codex post if you’re shipping anything with agent permissions; it’s the cleanest published spec for production-grade agent governance.

Cursor 3.3 shipped “Build in Parallel” with async subagents (May 7). Cursor identifies independent parts of a plan and runs them simultaneously instead of in queue. New PR review experience with inline threads, commits view, and split-PR actions. The agent stack from the Issue #012 OS war is rapidly maturing into production tooling.

Skills repos exploded. addyosmani/agent-skills added ~4,000 stars in a single week to land at 39K (as of publication) — Addy Osmani productizing engineering skills for Claude Code, Cursor, Gemini CLI, Windsurf, OpenCode, Copilot, Kiro, and Codex. anthropics/financial-services launched at 13.3K stars with the 10 finance agent templates. Skills as portable middleware is now the dominant agent-tooling pattern.

Publishers v. Meta gets teeth. The May 5 Manhattan federal court filing by Hachette, Macmillan, McGraw Hill, Elsevier, Cengage, and Scott Turow against Meta over Llama training alleges Meta engineers actively stripped Copyright Management Information from ingested works — that’s a separate DMCA §1202 claim with statutory damages, not just a fair-use argument. Mark Zuckerberg is named personally. Class-action sought.

Musk v. Altman, Week 2. Brockman’s August 2017 journal entry — “Financially, what will take me to 1B?” — became courtroom evidence; his current OpenAI stake is ~$30B. Musk reportedly demanded $80B for Mars City colonization as the price of supporting OpenAI’s for-profit conversion. Former board member Shivon Zilis described the 2018 split as a “weird halfway breakup.” OpenAI nonprofit reportedly holds a stake in the for-profit estimated at $200B.

Oregon’s $110K AI hallucinations sanctions. Federal Magistrate Judge Mark D. Clarke’s December opinion got a May 10 OPB wrap-up that puts the nationwide tally at ~900 cases involving AI-fabricated citations in legal filings. Oregon State Bar is now actively evaluating disciplinary frameworks. The HEC Paris AI hallucination database catalogues 1,174+ documented incidents globally; JD Supra reports Q1 2026 alone saw $145K+ in US sanctions.

Oscars: AI ineligible. AMPAS announced May 1 that AI-generated actors and AI-written screenplays are ineligible. Per TechCrunch’s read of the rule changes: “only roles demonstrably performed by humans with their consent”; screenplays must be “human-authored.” Effective for the 99th Academy Awards. Tilly Norwood, the AI-Val-Kilmer trailer, and Disney/Marvel’s AI pivot speculation made this inevitable.

AWS Q Developer EOL. Amazon set Q Developer’s IDE plugins and paid subs to sunset on April 30, 2027 — 12-month migration window to Kiro. New signups blocked May 15. Opus 4.6 leaves Q Developer Pro May 29; Opus 4.7 ships only on Kiro. If you’re on Q, the migration calendar is already running.

Five Eyes agentic-AI guidance landed in week one of May. Six national cyber agencies (CISA, NSA, ASD ACSC, CCCS, NZ NCSC, UK NCSC) published Careful Adoption of Agentic AI Services on May 1. First-ever joint Five Eyes policy on a single AI attack surface. Read as a likely procurement-floor document for defense-relevant agent deployments going forward.

🎯 The Playbook

Six moves for builders this week

  1. Build for iOS Extensions, not iOS-as-it-is. If you ship a consumer AI product, your TAM just changed. WWDC is June 8. The Apple App Store-style gallery for AI apps will be the highest-leverage distribution surface arriving in 2026. Audit your iOS app and identify the surface area you’d want to expose: text generation, summarization, image, voice. If you don’t have a partnership-track conversation with Apple by Memorial Day, you’re behind.
  2. Run the funding chain on your model dependencies. Pull up your top three frontier API vendors. Trace whose money funds their compute. Identify the counterparty whose default would cost you latency or capacity. That is your real single point of failure, not the API itself. The reported $200B Google/Anthropic loop is the bellwether — it’s not unique, it’s the template.
  3. Document AI productivity wins in writing this quarter. Cloudflare just announced 600% internal AI usage as the precedent for cutting 20% of staff. OpenAI’s frontier-firms data is going to be on every CFO’s desk by Q3. Get your name on the productivity gains that already happened, in a system the comp committee can actually read.
  4. Build the harness, not the magic. OpenAI’s “Running Codex Safely” is the closest thing to a published spec for production-grade agent governance. Sandboxing, network policy, OS-keyring credential storage, agent-native telemetry, auto-approval subagents. Read it. Bake it into anything you ship that takes agent actions on user systems. The Cursor RCE patches earlier this year were the warning shot.
  5. Treat CAISI evaluation baselines as enterprise procurement table stakes. If you’re selling AI into anything defense-adjacent, regulated, healthcare, or finance, “we passed our internal evals” stops being the answer this quarter. Start designing the eval-artifact bundle you’ll hand to a buyer’s CISO: model cards, capability evaluations, jailbreak resistance, audit logs, role-based access. The federal floor is coming — whoever ships it first as a procurement deliverable wins enterprise pipeline.
  6. If you manage engineers, stop saying “agentic AI-first” without saying what you actually mean. The Coinbase memo, the Cloudflare memo, and the DeepMind union letter are reading the same vocabulary differently. Tell your team specifically: which workflows you expect AI to take over, what the headcount implications are, what severance and equity treatment look like, and what kinds of work are off-limits to the agent. The companies treating “AI-native operating model” as a black-box buzzword are the ones generating union drives. The ones being specific are generating retention.

🔥 What’s Viral Right Now

“Rebuilding Coinbase as an intelligence, with humans around the edge aligning it.” — Brian Armstrong, Tuesday May 5, 6:55 AM email. Going to be a HR-deck quote for the next decade. Either prophetic or appalling depending on which side of the layoff you’re on.

Anthropic’s 80x. “We tried to plan very well for a world of 10x growth per year. And yet we saw 80x. And so that is the reason we have had difficulties with compute.” — Dario Amodei at the May 6 dev conf. Annualized Q1 growth, run rate at $30B in April from $9B at year-end 2025. Also explains why they’re now writing reported $200B compute checks.

Cloudflare cuts 20% on a record quarter and the stock drops 24%. First mass layoff in 16-year history, agentic AI-first operating model, $639.8M in Q1 revenue, +34% YoY. Investors voted with the sell button. The mismatch is the story.

addyosmani/agent-skills adds 4,000 stars in a week. Addy Osmani (Google Chrome) productized engineering skills for every coding agent that matters. The “skills as middleware” thesis we flagged in Issue #011 is the dominant agent-tooling pattern now. If you’re not writing skills for your team’s workflows, you’re behind your peers who are.

DeepMind workers vote 98% to file for union recognition. First union recognition request at a frontier AI lab. Worker quote of the week: “the Pentagon is not a responsible partner.”

The walls came down. The labs moved in. The receipts came fast.

Build the harness. Document the wins. Pick your storefront.

Stay building. 🛠️

— Matt