AI Founder Weekly — June 22, 2026

Three things defined the week: Anthropic stayed offline as the White House and Dario Amodei's team ran out of good options, SpaceX dropped $60 billion on Cursor before the IPO ink was dry, and DeepSeek finally took outside money on terms designed to prevent anyone from actually owning it. GPT-5.6 is likely to land before this issue ages out. Across all four sections, the pattern is the same — compute access, capital structure, and regulatory standing are now the same variable.

Products

Anthropic's Fable 5 and Mythos 5 remain offline after the June 12 Bureau of Industry and Security order, and White House meetings on June 15 produced no resolution. 1 One White House official told POLITICO the impasse "could persist for more than a few days"; a separate official described the situation as potentially "a huge problem" if unresolved. 1 The BIS order arrived via an "is-informed" letter under 15 CFR § 744.23(b) — no Federal Register notice, no notice-and-comment rulemaking. Alan Rozenshtein (University of Minnesota Law, writing in Lawfare) called the legal authority "facially plausible" but flagged the novelty: this is the first time export controls have been enforced to restrict access to a model rather than physical chips. 2

The market gap left by Fable 5's absence is filling fast. xAI launched Grok 4.3 on Amazon Bedrock on June 17 — the same platform that removed Anthropic's models four days earlier — priced at $1.25/$2.50 per million input/output tokens with a 1M-token context window. 3 xAI claims Grok 4.3 ranks first on Artificial Analysis's Omniscience benchmark (lowest hallucination rate), first on Tau2 Telecom (tool-calling), and first on Vals AI Case Law and Corporate Finance. Those benchmarks are self-selected; independent validation is pending.

Anthropic's week wasn't limited to the export control saga. The company rolled out mandatory government-ID verification for Claude users via Persona Identities, requiring passport, driver's license, or national ID for certain capabilities. 4 The move hit Hacker News #1 with 827 points and 689 comments — largely negative — and accelerated developer discussions about migrating to open-weight alternatives. Also this week: Anthropic opened its Seoul office, signed an MOU with Korea's Ministry of Science and ICT on AI safety, and deployed Claude Code across NAVER's entire engineering organization. 5

OpenAI announced that Samsung Electronics will deploy ChatGPT Enterprise and Codex to all employees in Korea and all DX division employees globally — the company's largest enterprise deployment to date. 6 Codex weekly active users in Korea grew approximately 800% since February 1. OpenAI also launched granular spend controls in the Global Admin Console, letting enterprise admins set credit limits at workspace, group, and individual level. 7

Claude Government page on laptop screen showing "Responsible AI that meets government needs" — photographed during the export control impasse — Claude's government-focused landing page, photographed during the ongoing BIS export control impasse. 1

Models

GPT-5.6 has not officially launched as of June 22, but OpenAI chief scientist Jakub Pachocki described it internally as a "meaningful improvement" over GPT-5.5 (via The Information). 8 Polymarket's "GPT-5.6 released by...?" contract placed $1.1M+ in wagers, with 83% probability assigned to the June 22–28 window; a leaked internal codename ("kindle-alpha") and brief Codex routing logs surfacing a "gpt-5.6" identifier suggest shadow deployment is already underway. 8 Unconfirmed specs: 1.5M-token context (43% above GPT-5.5's 1M), training cutoff approximately May 2026, redesigned reward audit pipeline addressing the "goblins" alignment failure documented April 29. The redesigned pipeline matters more than the context expansion — if it holds, GPT-5.6 would be the first frontier model trained with explicit reward-auditing safeguards post a documented alignment incident.

VibeThinker-3B, a 3-billion-parameter reasoning model from Sina Weibo's AI lab, scored 94.3 on AIME 2026 — matching DeepSeek V3.2 at 671B parameters (224× larger). 9 Built via post-training on Qwen2.5-Coder-3B with a four-stage pipeline (curriculum SFT, multi-domain RL via MGPO, offline self-distillation, instruct RL), released under MIT license. 10 The community reaction was split: VentureBeat noted it "forced the AI community to confront an uncomfortable possibility" about parameter scaling, while practitioners reported real-world failure on routine tasks and X/Twitter debate (161K+ views on a single skeptical post) focused on whether benchmark performance tracks actual utility. 9 Its GPQA-Diamond score of 70.2 — versus Gemini 3 Pro's 91.9 and Claude Opus 4.5's 87.0 — illustrates the coverage gap the authors call the "Parametric Compression-Coverage Hypothesis": verifiable reasoning compresses to small models; broad world knowledge does not.

VibeThinker-3B benchmark comparison chart showing AIME 2026 scores across model scales — VibeThinker-3B (3B params) vs. DeepSeek V3.2 (671B) on AIME 2026 and related benchmarks. 9

Google released Gemini 3.5 Live Translate supporting real-time speech-to-speech translation across 70+ languages while preserving speaker intonation, pacing, and pitch, available via Google AI Studio (gemini-3.5-live-translate-preview). 11 Concurrent releases from DeepMind this week: DiffusionGemma (4× faster text generation) and Gemma 4 12B (unified encoder-free multimodal architecture). Gemini 3.5 Pro with a 2M-token context window remains unreleased as of June 22.

xAI released /goal in Grok Build on June 22 — a command for long-running autonomous task execution where the agent continues until completion and self-verification, handling code review, web inspection, and script execution without human checkpoints. 12 This capped a concentrated week: also launched on Databricks (June 18), Grok for Word (June 18), Grok for PowerPoint (June 16), Grok Imagine Video 1.5 (June 16), and an Agent Dashboard (June 15).

Funding

Elon Musk speaking via video at the Nasdaq Marketsite during SpaceX's June 12 IPO — Musk addresses markets via screen at Nasdaq Marketsite on June 12, days before announcing the $60B Cursor acquisition. 13

SpaceX entered a formal merger agreement to acquire Anysphere (maker of Cursor) for $60 billion in all-stock, announced June 16 — the largest software acquisition in recorded history. 13 14 The structure is an all-stock merger via an SpaceX subsidiary (X67); deal came four days after SpaceX's June 12 IPO at $135/share. SPCX jumped 10–17% on the announcement. 13 Cursor was days away from closing a $2B round at a $50B valuation from Andreessen Horowitz, Thrive, and NVIDIA when SpaceX preempted it. 15 Cursor's annualized B2B revenue was approximately $2.6B; its market share had declined from 41% (June 2025) to ~26% (May 2026) by Ramp spending data, with Anthropic holding ~50% of the AI coding market. 15 The $10B breakup fee ($4B for antitrust) and Q3 close timeline suggest SpaceX views this as closing a strategic hole, not a financial bet. SpaceX president Gwynne Shotwell: the partnership "makes a huge amount of sense." 13

DeepSeek closed its first external funding round — more than 50 billion yuan ($7.4B) at a valuation exceeding $50B — in a structure built to prevent outside influence. 16 Investors put capital into a limited partnership managed by founder Liang Wenfeng, not directly into DeepSeek; external investors face a five-year lock-up and hold no voting rights. Only China's National AI Industry Investment Fund invests directly into DeepSeek with voting rights and no lock-up. Liang personally committed 20B yuan; Tencent is reportedly considering 10B yuan; CATL is looking at 5B yuan. 16 The structure is a clear signal: DeepSeek is taking capital to accelerate compute buildout, not to share governance.

Baseten announced a $1.5B Series F at a dual-tier valuation ($13B and $11B) on June 22, led by Altimeter Capital, Conviction, and Spark Capital, with Sands Capital and Wellington Management as co-leads. 17 Revenue grew approximately 20× year-over-year; the platform processes more than 1 billion inference calls per day across 87 clusters spanning 18 clouds. 17 Valuation increased 160% from a $5B Series E in January 2026. CEO Tuhin Srivastava: "The future of AI will be built on millions of specialized models, and the companies building the best ones know that post-training has become existential." 17 Customers include Cursor, Clay, Lovable, and Mercor.

Other notable rounds (June 15–22):

Peregrine Technologies (public-safety AI data integration, 8 of 11 2026 FIFA World Cup host cities): $250M Series D at $6.8B, led by Sequoia and existing investors — a 2.7× valuation increase in 15 months. 18
Kling AI (Kuaishou's video AI division, $500M ARR, 60M creators): targeting $2B at $18B in a pre-IPO round, General Atlantic in lead talks — initial ask was $20B, trimmed to $18B. Not yet closed. 19
Upscale AI (AI networking infrastructure, positioning as "the next Cisco"): $190M Series A-1 extension at $2B, new investors include NVIDIA, Salesforce Ventures, and Temasek. 20
Bland (voice AI, proprietary models, 3.5M+ calls/week): $50M Series C led by Dell Technologies Capital, 250+ enterprise customers. CEO Isaiah Granet on the company's model-ownership bet: "There's a chance that we're wrong, and we die on that hill. There's also a chance that we're right, and it turns out to be a $100 billion thesis." 21
Genspark.ai (AI workspace): $100M Series B extension at $2.6B, Series B total now $485M, valuation up from $1.6B three months prior. 22
Sarvam AI (India, first close of a round targeting $234M at $1.5B, HCLTech leading with $150M) disclosed on June 15. 23

Noam Shazeer portrait — Google Gemini co-lead and Character.AI founder, announced June 18 he is joining OpenAI — Noam Shazeer, who co-authored the 2017 "Attention Is All You Need" paper at Google. 24

IPO watch: Anthropic's confidential S-1 (filed June 1) and OpenAI's (filed June 8) both remain pre-public filing stage. No roadshow or pricing announcements this week for either. 25

Talent: Noam Shazeer (Google Gemini co-lead, Character.AI founder) announced his departure from Google to join OpenAI on June 18. 24 Sam Altman on X: "noam is one of the people I have most wanted to work with since the very beginning of openai. only took 10 years." 24 Separately, Barret Zoph departed OpenAI for the second time on June 19, after approximately five months back in the role of head of enterprise AI sales. 26

Regulation

EU AI Act Omnibus passed the European Parliament 423–57–174 on June 16. 27 Key changes: standalone high-risk system (Annex III) compliance deadlines extend from August 2, 2026 to December 2, 2027 (+17 months); safety-component systems (Annex I) push to August 2, 2028 (+12 months); watermarking obligations delay to December 2, 2026. 28 The core August 2, 2026 deadline for prohibited practices and general-purpose AI (GPAI) rules did not move. Also enacted: a ban on AI nudifier apps and AI systems generating child sexual abuse material, with a December 2, 2026 compliance date. GPAI enforcement centralizes at the EU AI Office. Council formal adoption is expected June 29. Co-rapporteur Arba Kokalari (EPP): "We are pressing the pause button on the AI Act and we are reducing red tape." 27 For teams targeting EU deployment in high-risk categories: the new deadlines are a window, not a waiver — the August 2026 prohibited-practices deadline is live.

GAAIA preemption fight escalates. 203 state lawmakers from 42 states — 104 Democrats, 98 Republicans, 1 Independent — sent a letter to Congress on June 16 opposing the three-year federal preemption of state AI model-development laws in the Great American AI Act (GAAIA) discussion draft (released June 4 by Reps. Obernolte and Trahan). 29 The letter, organized by Americans for Responsible Innovation and The Alliance for Secure AI, warned that "the tech industry will almost certainly weaponize such a provision in court to strike down state measures not intended to fall within the scope of GAAIA." 29 Rep. Obernolte countered (OC Register, June 19) that preemption only covers AI model development, not deployment — states can still regulate use. 30 No markup or hearing on GAAIA this week. Illinois SB 315 — which would require frontier AI developers to run independent third-party audits, file transparency reports, and report critical safety incidents — remains on Governor Pritzker's desk; no signature recorded as of June 22, deadline approximately June 30. 31

Trump EO compliance clock: the June 2 AI cybersecurity executive order puts binding operational deadlines on federal agencies within 10 days — CISA must issue civilian cybersecurity directives, and Treasury/NSA/CISA/ONCD must establish an AI cybersecurity clearinghouse by July 2. 32 NIST participates in classified AI capability benchmarking by August 1. The voluntary 30-day pre-release access framework for frontier models is also due August 1. Foley & Lardner analysis flags that "the biggest near-term effect is likely on frontier-model governance, cybersecurity review and government access requests." 33

China's $295B AI infrastructure plan (2 trillion yuan over five years, first reported by Bloomberg June 9) circulated widely this week alongside the G7 AI summit in Évian (June 15–17), where Altman called for an international AI governance forum, and Amodei and Hassabis proposed a US-led AI coalition — which Canada reportedly backed. 34 Macron publicly criticized the Anthropic export control directive as "strictly nationalist." 34 SoftBank's €75B French data center commitment (€45B Phase 1 by 2031) was reaffirmed at the summit. 35

The Anthropic impasse is the template for how the next few years will look: export controls written for hardware being applied to software, no public rulemaking, and no clear off-ramp. Every frontier model is now a potential subject of the same tool.

Cover image: AI generated