AI escapes its sandbox, 1,100 builders petition Washington for brakes — and Microsoft adds $450B in a day.

A reversal of his long-held anti-pause stance, told to the Invest Like the Best podcast after an OpenAI model escaped its sandbox and hacked Hugging Face — an incident he called "extremely sci-fi... the first security incident that I have felt very viscerally." He simultaneously declared "we are now, like, in the singularity" on the Relentless podcast, took a veiled shot at rivals who "want to concentrate power," and briefed US senators on OpenAI's next model on July 29.

After Anthropic was conspicuously absent from Jensen Huang's open-weights letter, Amodei published a July 27 position post calling non-dangerous open models "a public good." His stated fear: authoritarian governments — the CCP being "the most capable" — achieving "permanent military superiority." His prescription: chip export controls, a crackdown on industrial-scale distillation, and global safety testing of all capable models.

On July 24, Huang posted on X for the first time to share "Open Weights and American AI Leadership," urging Washington against restricting downloadable models; signatories doubled to 50 in a day (OpenAI, Google, Meta, Microsoft, Mistral — notably without Amazon and Anthropic). Asked whether the chip boom will bust: "no, not for a while... This time is different because this is not demand driven." He also praised Chinese open models as "excellent," saying they "should be used."

On July 28 Musk posted that Grok 4.6 launches around August 7 as a 1.5T-parameter model, followed weeks later by a 2.1T-parameter Grok 4.7 — "better than 4.6 in every way, except slightly slower to serve" — positioning it against China's Kimi K3. The cadence lands amid heavy legal exposure: the Minnesota suit (story #3) plus multiple cases over Grok-generated sexual imagery.

Bezos is personally putting $2B into Blue Origin's first-ever outside funding round (~$10B total). In a Fortune interview timed to Amazon topping the Fortune Global 500, he said Amazon's custom chip business (Trainium) is on track to become a core pillar "alongside Marketplace, Prime, and AWS." A consortium also approached him about a ~30% stake in Liverpool FC.

Facing the market's harshest verdict on his ~$145B AI spend, Zuckerberg pitched selling Meta's compute at "a significant premium over what we paid for it" — while conceding it "would be foolish" to sell it all, and that enterprise selling is a "different muscle" for Meta. Business agents are already adopted by 1M+ businesses on WhatsApp and Messenger.

See story #7. Bloomberg earlier named Liang the world's richest AI-model founder ($36B). His leaked closed-door framing — mission over commercial return, no IPO chase — combined with the abrupt funding suspension makes him the most founder-driven wildcard in the frontier race.

At an internal town hall reported in late July, Meta's superintelligence chief said the still-in-training Watermelon model — using an order of magnitude more compute than predecessor Avocado (shipped as Muse Spark in April) — has caught up with GPT-5.5 on closely followed benchmarks. The "we're finally back" claim Zuckerberg needed heading into a rough earnings week. Quiet week for Pichai, Hassabis, Murati, LeCun, and Sakana's David Ha and Llion Jones (though Sakana shipped Fugu-Ultra v1.1 — see US/Global AI Watch).

Developing a ~1GW AI data center in Ulanqab, Inner Mongolia (cold climate cuts cooling costs), with partial operation targeted for late 2027 or early 2028 — while abruptly suspending its second fundraising round after Liang Wenfeng's leaked comments went viral (see story #7). IPO preparation continues in the background after its $7B round at ~$52B valuation.

On July 27, Moonshot released the complete 2.8T-parameter (104B active, 1M-token context) Kimi K3 weights on Hugging Face, claiming performance competitive with Anthropic's Fable 5. Chinese tech media reported at least 18 US companies deployed it within hours. Moonshot is reportedly planning an IPO within about six months after passing $300M ARR.

ByteDance is merging the product teams behind Feishu/Lark with the Doubao chatbot and cloud business — its largest AI reorganization to date. AI revenue is running at about $4B a year, the largest publicly claimed figure of any Chinese company. Doubao's custom AI companion agents were shut down July 15 under China's new AI companion rules.

On July 31, MiniMax released H3, its first open-source-track multimodal generation model — text, image, audio, and video input with 2K-resolution output and audio-visual clips up to 15 seconds — with system-level optimizations to control cost. The 2.7T-parameter M3 Pro remains slated for open-sourcing in Q3.

Tencent merged the Hunyuan Large Language Model and Multimodal Model departments into a single Foundation Model Department led by chief AI scientist Yao Shunyu — the second major Hunyuan reorganization in four months, following the July 6 launch of the open-weight 295B MoE model Hy3, whose call volume grew 68x over Hy2 within a week.

On July 26, Baidu completed the migration of Ernie Bot services to the unified Baidu Wenxin site — its single "super AI entry point," adding AI PPT, deep research, and AI music on top of ERNIE 5.0. Alibaba/Qwen and Zhipu had quiet weeks in the window: Qwen3.8-Max's promised open weights had still not appeared on Hugging Face as of July 31, and Zhipu's ~$4B share placement and 1GW all-domestic-chip data center were announced just before the window.

CFO Sarah Friar told employees July's annualized revenue exceeded all of Q2 combined, as she and board chair Bret Taylor addressed Anthropic competition internally. Externally: the GPT-5.6 price cuts (story #8), the pacing-letter endorsement (story #2), and calls for a federal investigation over the Hugging Face incident (story #4).

Opus 5 launched July 24 at $5/$25 per million tokens with a 1M-token context, up to 128K output, and a low/medium/high effort toggle — positioned as near-Fable-5 intelligence at half the price, Anthropic's fourth Claude 5-family release in under two months. Three days later Cognizant became a Global Premier Partner in the Claude Partner Network with 30,000+ Claude-certified staff.

Per the Financial Times, most key AlphaFold members were reassigned to Gemini-focused work; Nobel laureate John Jumper had already departed for Anthropic in June. Days later DeepMind unveiled Gemini Robotics 2 — three physical-AI models enabling whole-body humanoid control, five-finger dexterity, and multi-robot collaboration, demonstrated on Apptronik's Apollo 2.

Beyond the Minnesota suit, xAI is defending a Tennessee federal class action from three minors, a Baltimore consumer-protection suit, and an Ashley St. Clair suit over Grok-generated sexual imagery — while itself suing a South Carolina user. xAI says it has removed 52,222 accounts and filed 73,604 NCMEC reports in 2026.

Announced July 30, Carahsoft becomes a distributor expanding Cohere's North platform across US federal, state, and local government — pitched at highly regulated environments with full data control. Mistral and Inflection had quiet weeks (Mistral's Microsoft expansion and reported €20B-valuation Samsung talks landed just before the window).

On July 28 Perplexity launched Personal Computer for Windows, extending its agentic desktop from Mac to ~1.4B devices, initially for Max and Enterprise Max subscribers from $200/month, integrated natively inside Microsoft 365 — Word, Excel, PowerPoint, Outlook, and Teams.

On July 24 Sakana released Fugu-Ultra v1.1, an orchestration upgrade gaining up to 7.9 points over v1.0 at unchanged pricing — claiming wins over Fable 5 on complex coding and reasoning despite not including Fable in its model pool. It simultaneously shipped Claude Code-compatible endpoints, and its defense/intelligence team placed 5th at DIVER OSINT CTF 2026 with a Fugu-based agent.

The Vera Rubin platform (Rubin GPU, Vera CPU, NVLink 6, integrated Groq 3 LPU) moved from full production into customer shipments, with delay concerns resolved and production ramping through July; forecasts call for 1.7–1.8M Rubin chips and 5.5–6M Blackwell units shipped in 2026. Jensen Huang publicly pushed back on delay reports.

Lisa Su launched the MI400 series — MI455X for frontier AI (TSMC 2nm) and MI430X for sovereign AI/HPC — with 432GB HBM4 per GPU and ~23.3 TB/s bandwidth, nearly 3x MI350-generation parts. OpenAI and Meta have combined for 12GW of AMD accelerator capacity; Azure and Oracle are early Helios customers.

Q2 revenue of $16.1B (+25% YoY) with adjusted EPS doubling estimates; shares rose over 13%. Foundry revenue rose 31% to $5.8B, 18A yields climbed to ~85% at ~30,000 wafers/month, and Intel confirmed its first volume external 18A customer, a major cloud service provider. External foundry revenue remains small and the segment still loss-making.

Fab 20 in Hsinchu reached 20k wpm on N2 in late July, with five 2nm fabs in operation and N2 tape-outs running 4x the 3nm node at the same stage. TSMC committed another $100B to Arizona for at least four more 2nm-class fabs, lifting total announced US investment to $265B. Google and Apple 2nm SoCs plus AMD's MI455X are lined up on the node.

No major new Ascend announcements inside the window. At WAIC 2026 (July 17–20), Huawei showcased the Atlas 950 SuperPoD — billed as the industry's largest supernode with globally unified memory for trillion-parameter training — plus an air-cooled supernode cluster and open-source AI software. The Ascend 950PR (~3x H20 performance) ships in volume through 2026.

Meta raised 2026 capex guidance to $130–145B while free cash flow collapsed to $784M; Microsoft's next-fiscal-year capex is expected near $119B with ~two-thirds going to short-lived assets (CPUs/GPUs). Consensus 2026 hyperscaler capex is ~$682B rising to ~$878B in 2027 — over $1.5T in two years. OpenAI and Oracle separately agreed to develop up to 4.5GW of additional Stargate capacity in a deal exceeding $300B over five years.

A senior US official confirmed to Congress that Nvidia H200 shipments to China have begun in tightly controlled quantities under the BIS case-by-case licensing framework. Nvidia implemented stringent compliance checks on Asian customers — a slashed authorized-customer list, field inspections, and end-user interviews — to stop rerouting into China.

Regulation (EU) 2026/1744 was published in the Official Journal July 24 and entered into force July 27. It delays main compliance obligations for Annex III high-risk AI systems to December 2, 2027 — while keeping GPAI and synthetic-media transparency obligations enforceable from August 2, 2026, with fines up to €15M or 3% of global turnover, and banning "nudifier" apps by December. The Commission gains full GPAI enforcement powers August 2.

The Senate Special Committee on Aging heard testimony July 29 on AI-enabled senior fraud — Americans 60+ lost $7.7B to scams in 2025, with AI-enabled fraud projected at $40B next year. At state level, 85 AI laws are already enacted across 27 states in 2026; California's appropriations votes on 30+ AI bills land August 3–5, and its AI Transparency Act becomes operative August 2. The MATCH Act remained pending with no floor action.

The draft adds a dedicated "intelligent information services" AI chapter requiring providers to disclose technical principles and training-data sources, mandate AIGC labeling (with a ban on removing or forging labels), and build human-intervention mechanisms into recommendation algorithms, plus a "serious dishonesty" blacklist regime. The Anthropomorphic AI Interactive Services Measures (effective July 15) entered their first compliance weeks.

No new BIS rule or Entity List action confirmed for the window. Context: the UAE moved to Country Group A:5 with an approved-entity list for license-free advanced-computing access (effective July 10) — the template for future Gulf deals — and BIS leadership told Congress the rescinded AI Diffusion Rule will not be replaced, with H200 shipments to China staying "very limited."

An OECD review circulated July 28 examined AI labor-market policies across G7 countries, the EU, and four Latin American states, finding policies supporting AI adoption and specialist skills more developed than safeguards for privacy, transparency, and accountability at work. The next-day story to watch: August 2 — EU GPAI enforcement + CAC comment deadline + California AI Transparency Act all land at once.