The walls came down
Two divorces, eight Pentagon contracts, one sworn confession — and the open-source layer just drew even.

The walls came down. In five business days, Microsoft and OpenAI ended their exclusivity, AWS brought OpenAI’s latest models and Codex into Bedrock the next morning, the Pentagon cleared eight frontier AI vendors for classified networks and pointedly excluded Anthropic, and Elon Musk acknowledged under oath that xAI partly used OpenAI’s models or outputs in training Grok — the same kind of practice the State Department wrote diplomatic cables warning foreign governments about last week.

The team-sport era ended this week. The big labs aren’t pretending to be partners anymore. They’re competing for the same agentic stack, and the moats are publicly drying up. Here’s what actually matters.

---

⚡ Microsoft and OpenAI Got a Divorce

Cordial co-parenting now. Multi-cloud kids.

Apr 27: Microsoft’s exclusive license to OpenAI’s IP became non-exclusive through 2032. Microsoft no longer pays revenue share to OpenAI; OpenAI keeps paying Microsoft 20% through 2030, now subject to an undisclosed cap. OpenAI is converting to a public benefit corporation, with the nonprofit reportedly retaining a major equity stake; Microsoft’s restructured equity reportedly locks in around the high-20s percent range. The legal peril Microsoft had been threatening over OpenAI’s $50B February Amazon deal — gone. Resolved by the restructure.

Apr 28, the next morning: AWS Bedrock launched OpenAI’s latest models, Codex, and Bedrock Managed Agents in limited preview. Codex available through CLI, desktop, and VS Code extension. OpenAI usage applies against AWS commitments. Andy Jassy posted the announcement on X the same day from an AWS event in San Francisco.

Translation: the single-cloud-dependency story is over. Codex and ChatGPT are about to start showing up everywhere in weeks. If your enterprise architecture is bet on Azure-OpenAI lock-in, you’re now holding a 2024 deal in a 2026 market. Anthropic is already on Bedrock and Vertex. DeepSeek runs locally, on cloud, on the China stack. The cloud providers are becoming AI app stores with compliance wrappers — the model isn’t the product anymore. The procurement path is the product.

Hype vs. Reality: 9/10 — Both companies’ own blogs documented the restructure. Limited preview isn’t GA, but the strategic shift is already done.

---

🚨 The Pentagon Cut Anthropic, Picked a Startup With No Public Model

One of the eight chosen vendors hasn’t shipped a product yet

May 1: the DoD signed eight AI vendors to deploy on classified IL6/IL7 networks — AWS, Google, Microsoft, OpenAI, Nvidia, SpaceX, Reflection AI, and Oracle (added later that same Friday, per Breaking Defense). Anthropic is the only frontier lab missing from the list. Anthropic’s absence reads like the procurement consequence of the supply-chain-risk fight covered in Issue #007 — even if the official DoD release doesn’t say that out loud. The court win didn’t repeal the buyer’s preference.

The wild inclusion is Reflection AI. Founded March 2024 by ex-Google DeepMind researchers Misha Laskin and Ioannis Antonoglou. Has not released a public model. Their code-research agent “Asimov” is still on a waitlist. Reportedly raising at very high private valuations despite limited public product surface. NVIDIA-backed. Their pitch: open-weight as the American answer to DeepSeek. The Pentagon just made them an operational-tier defense contractor before their first public product.

The 600-vs-4,000 contrast deserves the receipt. Roughly 600 Google DeepMind and Cloud employees signed an open letter to Sundar Pichai opposing the deal. Google did the deal anyway. In 2018, Project Maven took ~4,000 signatures and a dozen high-profile resignations to kill. The threshold for collective tech-worker resistance just went way up. Per Defense One reporting, senior officials around Defense Secretary Hegseth have framed the Anthropic dispute by comparing it to “an airplane manufacturer trying to tell the military who they are allowed to shoot at.”

The contrast that completes the picture: Anthropic was simultaneously fielding reported preemptive equity offers around the $850–900B range that would surpass OpenAI’s February valuation. Cut from the Pentagon Friday morning. Reportedly worth nearly a trillion by Wednesday afternoon. The contrast IS the story.

The real story: refusing to permit autonomous-weapons or domestic-surveillance use cases is now formally a procurement liability. If you’re an enterprise buyer in any defense-adjacent vertical, your AI vendor’s stance on use-case restrictions is part of due diligence now. It wasn’t last week.

Hype vs. Reality: 10/10 — Confirmed by official DoD release. Reflection AI’s inclusion is the part nobody’s quite digesting yet.

---

⚖️ Musk Acknowledged xAI Used OpenAI’s Models

Same week the State Department was warning foreign governments about Chinese distillation

The Musk v. Altman trial opened Apr 28 in Oakland. Liability phase. Advisory jury, Judge Yvonne Gonzalez Rogers makes the final call. Microsoft is co-defendant. Two surviving claims: unjust enrichment, breach of charitable trust.

Apr 30, on cross-examination by OpenAI lawyer William Savitt: Musk acknowledged that xAI had partly used OpenAI’s models or outputs in training Grok, while defending model-to-model training as common industry practice. “It is standard practice to use other AIs to validate your AI,” Musk said from the stand.

This is the same kind of practice OpenAI accused DeepSeek of in February. And the same kind of practice the State Department issued diplomatic cables about last week — covered in Issue #011 — warning foreign governments that Chinese firms including DeepSeek were extracting and distilling US AI models. The Frontier Model Forum’s anti-distillation initiative just got a sworn admission from a US frontier-lab founder, on the federal court record, while the same US government argues equivalent foreign practice is industrial espionage.

Bonus that says everything: under questioning, Musk ranked the labs Anthropic > OpenAI > Google on capability — placing his own xAI behind all three despite using their work as training data. The trial that’s supposed to resolve a 2018 ownership dispute is busy mapping the credibility ranking of the labs that will define the next decade.

Hype vs. Reality: 9/10 — Sworn federal court testimony. Multiple primary outlets in the courtroom.

---

🛠️ The Agent Stack Got Built This Week

And GitHub admitted the load broke them

While Anthropic and the Pentagon were generating the headlines, the agent OS war from Issue #010 productized in fourteen days. Pieces shipped this week:

Cloudflare + Stripe Projects (Apr 30) — Open protocol where agents can create accounts, buy domains, start subscriptions, and deploy code from a single Stripe login. Default $100/mo per-provider cap. Initial partners: Cloudflare, Vercel, Supabase, Clerk, PostHog, Sentry, PlanetScale, Inngest, Hugging Face, Twilio. Plus $100K Cloudflare credits for new Stripe Atlas startups.

Cursor TypeScript SDK (Apr 29) — @cursor/sdk public beta. Same agent runtime as the desktop app, exposed for CI/CD. Three deployment modes (local/cloud/self-hosted), sandboxed VMs per run, codebase indexing, MCP, hooks, subagents, skills. Default model Composer 2 at $0.50/M in / $2.50/M out. Production users: Faire, Rippling, Notion, C3 AI.

Mistral Medium 3.5 + Vibe Remote Agents (Apr 29) — 128B dense, 256K context, modified MIT open weights, self-hosts on 4 GPUs. $1.50/M in / $7.50/M out — half of Sonnet 4.6’s price. 77.6% on SWE-Bench Verified. Vibe Remote Agents adds cloud async coding sessions; Le Chat Work mode adds multi-step agentic workflows.

FIDO Alliance + AP2 (this week) — Google donated Agent Payments Protocol to FIDO and shipped AP2 v0.2 with “Human Not Present” payments — agents executing transactions on pre-authorized instructions. Mastercard contributed Verifiable Intent, co-developed with Google, designed to work with AP2. The cryptographic identity standard for agent commerce. Pair this with Cloudflare/Stripe and you have a complete agent-payments stack from rails to identity.

Plus: OpenAI Symphony (open-source coding-agent orchestration spec), Warp open-sourced its agentic IDE, Sentry Seer Agent for production debugging, IBM Granite 4.1, Nvidia Nemotron 3 Nano Omni.

Then the receipt landed. GitHub CTO Vlad Fedorov posted Apr 28 apologizing for recent incidents and admitting GitHub started a 10x capacity plan in October 2025 — and by February 2026 realized they needed to design for 30x current scale. The cause, in their own words: agentic development workflows accelerating sharply since late December. Repository creation, PR activity, API usage, automation, large-repo workloads — all growing fast enough that “small inefficiencies compound at high scale.”

The signal buried in the noise: the Agent OS war from #010 productized in two weeks, and GitHub just admitted the load model changed. Everyone wanted autonomous coding agents. Nobody wanted to pay the distributed-systems bill.

Builder action: add agent-load assumptions to your CI/CD now. Rate limits, queue design, idempotency, API quotas, PR automation rules. They’re not boring anymore — they’re the difference between “AI-native” and “we accidentally DDoS’d ourselves with interns made of software.”

Hype vs. Reality: 9/10 — Every product link above is primary-source. GitHub’s blog is GitHub’s own confession.

---

📡 Open Source Just Drew Even with Mythos

Anthropic’s $100M coalition got matched by a $50/run repo

The tagline writes the section: “Shannon’s job is simple: break your web app before anyone else does. The Red Team to your vibe-coding Blue team. Every Claude (coder) deserves their Shannon.”

This is the rebuttal to Anthropic’s “we kept Mythos to 50 hand-picked partners” stance from Issue #009. The democratized autonomous-pentester model just hit 21,000 stars on the same Anthropic stack the Glasswing coalition runs on. The capability that was supposed to be a $100M-coalition exclusive is on a laptop running on $50.

Concurrent on the defense side: Anthropic Claude Security public beta launched Apr 30 on Opus 4.7. Reads source like a security researcher rather than signature-matching — every finding ships with confidence rating, impact, repro steps, and a proposed patch. Slack/Jira webhook integration for Claude Enterprise customers.

The policy floor moved this week too. Reuters reported May 1 that US cybersecurity officials are weighing cutting federal patch deadlines on actively exploited vulnerabilities from the current 2–3 weeks down to three days — a proposal under consideration, not enacted policy. The cited reason: AI tools are collapsing exploitation timelines. Mandiant’s M-Trends 2026 report measured the handoff window from initial-access broker to follow-on threat actor collapsing to 22 seconds in 2025, and warned that AI abuse is showing up inside compromised environments at scale.

Builder action: point Shannon at systems you own or are explicitly authorized to test, before someone points comparable tooling at your prod. The Glasswing moat is publicly drying — and your AI-generated code shipping at 65% of your commit volume is the attack surface.

Hype vs. Reality: 9/10 — Reproducible benchmark, public agent logs, real CVEs in OWASP Juice Shop reports. The honest caveat is dual-use: same $50 autonomous pentester for defenders is a $50 autonomous attacker in unauthorized hands.

---

📊 Quick Signals

Big Tech earnings printed AI into the income statement. Apple Q2 FY26: $111.2B revenue (record March quarter), +17% YoY. Tim Cook becomes Executive Chairman Sept 1; John Ternus (SVP Hardware Engineering) takes over as CEO. WWDC AI announcements coming June. Microsoft Q3 FY26: AI run rate $37B, +123%. Azure +40%. Alphabet Q1: Google Cloud +63% — fastest growth ever. Capex guide raised to $180–190B. Meta Q1: $56.3B revenue, +33%. Stock dropped sharply on a $145B capex ceiling. The $690B bet from Issue #002 is paying — but investors rewarded everyone except the company spending the most.

China’s NDRC blocked Meta’s $2B+ Manus acquisition [Reuters / The Guardian, Apr 27]. A notably terse single-sentence order. Direct callback to Issue #001’s “Manus sells for $2B+ with no model” lede. Wrappers are now strategic assets nation-states will block at the border. The wrapper is the workflow. The workflow is the moat.

MARA Holdings → Long Ridge Energy & Power for $1.5B including debt. Acquires a 505 MW combined-cycle gas plant + 1,600 contiguous Ohio acres for AI data center expansion. Bitcoin miners pivoting to AI compute landlords.

KKR raising $10B+ to launch a new AI infrastructure company [Bloomberg, this week]. Private equity’s answer to the hyperscaler capex race.

EU regulators expanding Digital Markets Act scope to cloud and AI services [Reuters, Apr 28]. Includes whether some AI services should be treated as “virtual assistant” core platform services. The cloud-AI distribution war is becoming an antitrust story.

Federal judge sanctioned a managing attorney over a junior lawyer’s AI-assisted brief with hallucinated citations [Reuters, May 1]. AI legal liability climbing the supervisory chain. “The AI did it” is no longer a governance strategy.

Tobi Lütke is shipping public OSS. Shopify’s CEO trending on GitHub this week with qmd — a local CLI search engine for docs and notes with BM25 + vector + LLM reranking, all running locally via node-llama-cpp. Includes an MCP server out of the box. The CEO of a $200B-class company writing public code on the weekends. The old guard wouldn’t have done that. 👀

---

🎯 The Playbook

Your move this week

1. Build a model gateway, not a model marriage. OpenAI is now Azure / direct / AWS Bedrock. Anthropic is Bedrock / Vertex / Azure. DeepSeek is local / cloud / China stack. Hard-coding to one vendor SDK was a 2024 mistake — and now it’s a 2026 procurement risk.
2. Test OpenAI on Bedrock if you live in AWS. Don’t wait for GA. Compare latency, IAM fit, CloudTrail logging, PrivateLink, billing drawdown, and agent deployment ergonomics this week.
3. Add agent-load tests to your CI/CD. GitHub just told you the future failure mode. Simulate 10x PRs, retries, webhooks, automation runs. Your queue design and idempotency patterns aren’t boring anymore.
4. Run Shannon against systems you own or are authorized to test. ~$50, 1.5 hours, reproducible, on your existing Anthropic key. Pair vibe-coding offense with vibe-coding defense.
5. Treat agent actions like financial transactions. Bounded authorization, scoped identity, full audit logs, revocation paths. AP2 + Verifiable Intent is the standard you’ll be implementing in 18 months. Read the spec now.
6. Audit your defense-adjacent enterprise pipeline. If your AI vendor’s use-case stance could trigger a procurement risk designation, you have a buyer-side problem you didn’t have last week.

---

🔥 What’s Viral Right Now

South Africa pulled its national AI policy because the citations were AI-hallucinated. Reuters, Apr 27. The minister responsible called it proof that vigilant human oversight is critical. The most on-the-nose AI governance failure imaginable is a national AI policy face-planting on fake AI citations. If you sell anything in the AI policy/legal/research/compliance space, source verification needs to be a first-class product feature — not a disclaimer.

Musk’s distillation admission. Apr 30. Under cross-examination by William Savitt, Musk acknowledged xAI partly used OpenAI’s models or outputs in training Grok. Same week the State Department was warning foreign governments about Chinese distillation. The Frontier Model Forum’s anti-distillation initiative just got a US frontier-lab credibility problem on the federal record.

Shannon by KeygraphHQ. Past 21K GitHub stars, top of trending this week. 96.15% on the hint-free XBOW benchmark for $50 a run. “The Red Team to your vibe-coding Blue team.” If you’re shipping AI-generated code, you need this in your authorized-test workflow.

---

The walls came down. Build through them. 🛠️

— Matt