Edition 12 · June 1–7, 2026
Public markets
S-1 before the roadshow
On June 1, Anthropic confidentially submitted a draft Form S-1 to the SEC — days after Series H closed at a $965 billion post-money mark. SpaceX is already in its IPO roadshow. OpenAI is preparing its own confidential filing. The cap-table story from last week became a liquidity clock.
Safety architecture
A pause you can verify
On June 4, the Anthropic Institute published “When AI builds itself,” arguing recursive self-improvement may arrive before institutions adapt — and proposing a verifiable global slowdown if frontier labs can prove others have actually stopped. Same week: a White House EO asks for voluntary 30-day pre-release model access. Two governance forks, one velocity.
June 1: Anthropic’s announcement states it confidentially submitted a draft registration statement on Form S-1 for a proposed IPO of common stock. Share count and price are unset; any offering depends on SEC review and market conditions. Rule 135 language applies — this is not yet an offer to sell. The move positions Anthropic ahead of OpenAI’s expected confidential filing and alongside SpaceX’s already-public prospectus process.
S-1
Confidential draft filed Jun 1
$965B
Post-money from May 28 Series H
$47B
Run-rate revenue cited in May releases
Follow-up from Edition 11: Series H and Opus 4.8 shipped last week; this edition covers what changed since — the filing itself, not a rehash of the round terms.
Anthropic — confidential S-1 CNBC — IPO prospectusShipped · Jun 4
recursive-self-improvement
Anthropic Institute essay + internal metrics on Claude-authored code
The June 4 Anthropic Institute post warns that models may soon contribute materially to their own successors — and that a unilateral lab pause “accomplishes much less” without verification. Anthropic says it would slow or pause if other frontier developers did so in a verifiable way, and will research systems to detect covert training. Separately, reporting on internal engineering metrics cites 80%+ of Anthropic production code now written by Claude, with engineers shipping roughly 8× more code per day versus 2024 — treat vendor metrics as directional, not audited.
“We believe it would be good for the world to have the option to slow or temporarily pause frontier AI development to enable societal structures and alignment research to keep up.”
— Anthropic Institute, “When AI builds itself”
June 2: Anthropic expanded Project Glasswing to approximately 150 additional organizations across 15+ countries — power, water, healthcare, communications, hardware, and critical open-source maintainers. Partners must meet security requirements before Mythos Preview access. The post explicitly shifts emphasis from discovery to verify → disclose → patch → deploy, acknowledging that Mythos-class models will likely proliferate from other labs within 6–12 months, potentially without equivalent safeguards.
Anthropic — Expanding Project Glasswing Anthropic — Claude Partner Network (Jun 3)June 1 GTC Taipei keynote: Jensen Huang declared Vera Rubin NVL72 in full production — a pod-scale, multi-rack system built for agentic AI, not inference-only clusters. NVIDIA’s blog cites 36 Vera CPUs and 72 Rubin GPUs per NVL72 rack, ConnectX-9 SuperNICs, Spectrum-X Ethernet, and BlueField-4 — with hundreds of Taiwan ecosystem partners ramping across 350+ factories. Edition 11 previewed this keynote; this is the shipped chapter.
Same keynote: NVIDIA unveiled RTX Spark (N1/N1X) — an Arm-based superchip with MediaTek, pairing a Blackwell RTX GPU (~RTX 5070 laptop class per press) with a Grace CPU and up to 128 GB unified memory. Eight OEM laptops announced (Asus, Dell, HP, Lenovo, Microsoft Surface Laptop Ultra, MSI); first systems expected autumn 2026. Microsoft Build keynotes were teased for Windows agent primitives and NVIDIA OpenShell.
Digital Foundry — RTX Spark specsJune 2–3 Build: Microsoft AI announced seven in-house models — the strategic break from sole OpenAI dependence Edition 11 previewed. Headliners: MAI-Thinking-1 (35B active parameters, first reasoning model, trained without third-party distillation), MAI-Code-1-Flash (5B active, Copilot-tuned), MAI-Image-2.5 (+ Flash), MAI-Transcribe-1.5 (43 languages), and MAI-Voice-2 (+ Flash). Microsoft claims MAI-Thinking-1 matches leading models on SWE tasks and is preferred to Sonnet 4.6 in blind human evals — verify against independent harnesses.
Build · Jun 2–3
Microsoft IQ
Work IQ · Web IQ · Fabric IQ · Foundry IQ
Build’s enterprise thesis: Microsoft IQ aggregates workplace signal (Work IQ from M365 graph), MCP-native web grounding (Web IQ), structured business data (Fabric IQ), and retrieval planning (Foundry IQ). Agent 365 (GA May 1) gained an SDK for local agents; Entra and Defender extensions cover agent identity. Project Solara teased chip-to-cloud reference designs for wearable and desk agent devices. Work IQ API GA slated June 16. Windows shipped Aion 1.0 Plan — a 14B in-box reasoning model with 32K context for tool calling.
Microsoft — Microsoft IQ Microsoft Build sessionsJune 2: President Trump signed Promoting Advanced Artificial Intelligence Innovation and Security. Section 3 directs Treasury, DHS/CISA, and War/NSA — within 60 days — to design a voluntary framework where developers may grant up to 30 days pre-release access to “covered frontier models” for classified cyber-capability benchmarking. The order explicitly prohibits mandatory licensing or preclearance. An earlier draft reportedly sought 90 days; industry pushback narrowed the window. OpenAI CEO Sam Altman met lawmakers June 3, opposing mandatory approval regimes while engaging on the voluntary framework.
EO signed — voluntary pre-release access framework; no mandatory licensing
Classified benchmarking process for “covered frontier model” designation — NSA leads threshold determination
Altman meets Johnson, Jeffries, Schumer — discusses light-touch vs mandatory approval (per press)
June 4: OpenAI began rolling out Dreaming V3 — a compute-efficient memory architecture that replaces the explicit “saved memories” list with asynchronous background synthesis across conversations. Plus and Pro users in the US first; Free tier follows after ~5× compute reduction for the dreaming process. Users get a readable memory summary with edit/delete controls and memory sources (introduced with GPT-5.5 Instant). Privacy shift: the system defaults to remembering and lets users audit afterward — not the inverse.
OpenAI — Dreaming memory Engadget — Dreaming V3 rolloutOpen weights · Jun 4
nemotron-3-ultra
550B total · 55B active MoE
June 4: NVIDIA released Nemotron 3 Ultra — 550B-parameter MoE with 55B active, hybrid Mamba-Transformer layers, LatentMoE routing, and NVFP4 quantization for cross-architecture deployment (Hopper, Blackwell, Ampere). Positioned for long-running agent workflows; NVIDIA cites up to 5× throughput vs comparable open models and up to 30% lower agentic task cost in internal experiments. RULER @1M context: 95% per NVIDIA table — compare setups carefully.
NVIDIA — Nemotron 3 UltraJune 6: Alibaba’s Qwen team shipped Qwen3.7-Plus — proprietary multimodal agent combining visual perception with coding and tool use. Alibaba tables show leadership on AndroidWorld and ScreenSpot Pro vs GPT-5.4, Opus 4.6 Max, and Gemini 3.1 Pro; mixed results on scientific multimodal reasoning. Pricing: $0.40 / $2.40 per M input/output tokens — roughly 6× cheaper input than Qwen3.7-Max. No open weights.
The Decoder — Qwen3.7-PlusJune 7: Perplexity rolled out Search as Code — models author Python search workflows in a sandbox instead of hitting fixed search APIs. Perplexity reports 85% fewer tokens on a cited research task vs its prior pipeline and claims wins on four of five benchmarks vs OpenAI Responses API and Anthropic Managed Agents — self-reported; WANDR benchmark forthcoming. Rolling out in Perplexity Computer and Agent API. Fits the week’s pattern: deterministic runtimes + model strategy.
The Decoder — Search as CodeAnthropic filing S-1 while publishing pause architecture is coherent: public scrutiny funds safety research. Vera Rubin production + MAI models prove agent economics are real revenue lines, not slideware. Voluntary EO avoids China-speed regulatory capture.
IPO roadshows and pause essays can coexist because neither binds compute purchases already locked in May. Voluntary 30-day review is opt-in theater if classified benchmarks lag deployment. Recursive-self-improvement warnings without verifiable pause infrastructure may read as liability management ahead of a prospectus.
Anchors immediately after this edition — confirm in primaries before booking travel or compliance work.
Gemini-powered Siri overhaul, iOS/iPadOS/macOS 27 betas, on-device and cloud AI split — verify only against Apple’s stream.
Enterprise agent context layer from Build — check Foundry docs for scope and residency.
Developer tools, Foundation Models framework updates, third-party AI default settings on Apple platforms.
Largest prospective public offering in history per press — sets comp multiples for Anthropic/OpenAI filings.
Classified frontier-model benchmarking framework due from NSA/CISA/Treasury coalition.
Aug 2 still legally scheduled; Digital Omnibus may delay Annex III to Dec 2027 if enacted.