The Investor

The Revenue Inversion

Anthropic disclosed $30B+ in annualized revenue as of April 7, 2026 — up from $19B one month ago and roughly triple its ~$9-10B year-end 2025 run rate. This surpasses OpenAI's $25B+ reported at end of February. Fewer than 135 S&P 500 companies book $30B+ in annual sales. Anthropic is generating Fortune 100 revenue while still private.

The growth trajectory is historically unprecedented: $6B of ARR added in February 2026 alone — more monthly accretion than most SaaS companies achieve in their entire lifecycle. The vast majority comes from API access, confirming enterprise developers are choosing Claude at scale.

The Accounting Nuance That Will Dominate Both S-1s

A critical caveat: Anthropic books 100% of Claude sales through AWS, Azure, and GCP as revenue. OpenAI books only 20% of Azure OpenAI Service sales because Microsoft holds exclusive IP rights. Normalizing this closes the gap by "low billions" — bringing OpenAI to roughly $27-29B. This narrows but does not close the gap. For IPO valuation purposes, this accounting difference will be the most debated line item in either S-1.

The Profitability Trap

The number investors need to pair with $30B: Anthropic's 2025 gross margins came in 10 percentage points below expectations due to spiking inference costs. Both Anthropic and OpenAI remain deeply unprofitable. The central tension: is this a software business (70%+ margins, 20x+ revenue multiples) or an infrastructure business (40-50% margins, 8-12x)? The answer changes every AI valuation in your portfolio.

Anthropic targets cash flow positive by 2028 — two years ahead of OpenAI's 2030 target. Its multi-GW TPU commitment with Google/Broadcom (3.5 GW, up from 1 GW) secures compute through 2027 but deepens supplier dependency on its primary competitor.

OpenAI's Governance Crisis Escalates

While Anthropic accelerates, OpenAI's internal fractures widened dramatically this week:

• CFO Sarah Friar told colleagues OpenAI isn't ready to IPO in 2026 and questioned whether revenue can support compute commitments — then was excluded from financial planning conversations
• A New Yorker investigation (100+ interviews) corroborated by Sutskever internal memos and Amodei private notes alleges career-spanning deception by Altman
• A Microsoft executive compared Altman to "Madoff/SBF-level scammer"

When the CFO is sidelined for raising the exact questions investors will ask on an IPO roadshow, that's not a personality clash — it's a governance red flag with material valuation implications.

The IPO Sequencing Collision

Three mega-IPOs are converging on the same window while SpaceX's $75B raise at $1.75T (June roadshow) threatens to drain the institutional pool. Whoever IPOs second faces depleted allocations. Anthropic's Q4 2026 window looks increasingly likely to beat OpenAI — and the first AI model company to list captures the "platform" narrative premium.

The Capability Leap

Claude Mythos Preview scored 93.9% on SWE-bench Verified — a 13-point jump from Opus 4.6's 80.8%, the largest single capability leap on this benchmark. More importantly, it has discovered thousands of high-severity zero-day vulnerabilities across every major operating system and web browser, including:

• A 27-year-old OpenBSD vulnerability
• Linux kernel flaws enabling full machine takeover
• An FFmpeg bug that escaped 5 million prior automated tests

The critical insight: these capabilities emerged from general reasoning improvements, not specialized cybersecurity training. Every frontier lab pursuing general capability gains will cross this threshold. Per Alex Stamos (former Facebook/Yahoo security chief), open-weight models are ~6 months from matching this capability — at which point every ransomware actor gets frontier-grade exploit discovery for free.

Project Glasswing: The Defensive Response

Anthropic launched Project Glasswing: a coalition of 40+ companies — Apple, Google, Microsoft, Cisco, Broadcom — backed by $100M in Mythos usage credits and $4M in open-source donations. Cisco's chief security officer stated: "AI capabilities have crossed a threshold that fundamentally changes the urgency required to protect critical infrastructure."

This simultaneously validates the threat and creates Anthropic's deepest enterprise ecosystem yet. Glasswing builds switching costs through security-critical API integrations that are far stickier than general-purpose model access.

Converging Attack Vectors

The Mythos disclosure arrives alongside three other cybersecurity inflection signals this week:

For the first time, all five of SANS's most dangerous new attack techniques carry an AI dimension — the offensive/defensive AI arms race is no longer theoretical.

Where Capital Should Flow

The 6-month window creates extreme urgency. Companies with deployed vulnerability scanning products today — not research labs, not vapor — will see explosive pipeline growth. AI agent security is a separate but equally urgent category: OpenClaw has suffered six critical authorization CVEs in six weeks, with 63% of 135,000+ public instances running without authentication. OWASP split its GenAI guidance into separate LLM and agentic AI tracks, confirming category formation.

The Dark Factory Proof Point

An OpenAI internal team shipped 1 million lines of code with zero human-written or human-reviewed code — approximately 7 engineers managing Codex agents processing 1 billion tokens per day (~$2-3K/day). Ryan Lopopolo describes his role as equivalent to "group tech leading a 500-person organization" where all reports are AI agents. The product: a production Electron app with 500 NPM packages, 1,500 PRs, and full CI/CD.

The caveat matters: this was greenfield with no legacy constraints, and Lopopolo explicitly says it shouldn't be extrapolated. But as a directional signal, it's the strongest data point for where development is heading.

Codex itself is at 2M weekly active users with 25% week-over-week growth. Even with aggressive decay assumptions (halving growth monthly), that's 8-10M WAU by Q3. If 1% of power users consume $2-3K/day, implied Codex-only annualized revenue is $7-15B.

GitHub's Infrastructure Crisis

GitHub is on pace for 14 billion commits in 2026 — 14x its 2025 milestone. Claude Code's weekly public repo commits surged 25x in six months (100K → 2.5M). Agent PRs grew from 4M to 17M in six months. But GitHub's availability has dropped to 90% — roughly 2.4 hours of downtime daily — because its databases, Redis clusters, and failover mechanisms weren't built for machine-scale traffic.

The critical investment insight: GitHub's flat per-seat pricing doesn't capture third-party agent traffic. Anthropic and OpenAI capture the subscription/API revenue while Microsoft/GitHub bears the infrastructure cost. The former GitHub CEO has launched a competing startup — the ultimate insider signal that the architecture has structural gaps for the agent era.

Where Value Accrues in the New Stack

The sources agree on one critical point: models are commoditizing while orchestration, data, and infrastructure capture durable value. Anthropic's own Claude Code architecture relies on a "roughly constant" model underneath while a three-layer skeptical memory system drives production gains. Harness design alone shifts agent performance by 20+ ranks on benchmarks — the LLM underneath is table stakes.

A 7-person team with $90K/month in token spend demonstrated 500-person output — every developer-tools investment thesis written before April 2026 needs stress-testing against the agent-native paradigm.