AI Weekly Digest: June 2‑8 2026 Highlights

SpaceX IPO – On June 3, SpaceX listed on Nasdaq with a valuation of $1.75 trillion, setting a global IPO record. The offering raised $35 billion, with a first‑day price increase of 28% and a peak market cap of $2.24 trillion. Elon Musk retained 42% ownership, pushing his net worth above $940 billion.

OpenAI GPT‑5 – OpenAI announced GPT‑5 on June 5, a month ahead of schedule. Compared with GPT‑4o, GPT‑5 doubles the context window to 5 million tokens (+150%), expands multimodality to include video, audio, and real‑time interaction, and improves benchmark scores (reasoning 180 vs 100, +80%; code 96.5% vs 90.2%, +6.3%). Pricing also doubles: $10 per million tokens for input and $30 per million tokens for output.

Anthropic financing – On June 4, Anthropic secured $35 billion in funding, valuing the company at $1.2 trillion, surpassing OpenAI. Lead investors include Sequoia Capital, Dragoneer, Altimeter, and Greenoaks, with additional participation from Saudi PIF, UAE ADQ, and Singapore GIC. Funds will build a 1.5 million‑GPU compute cluster (launch Q1 2027), develop Claude 5 (target 2027), and expand data‑center presence in Europe, Asia, and the Middle East.

Google Gemini 3.0 – Released on June 6, Gemini 3.0 shifts focus from pure model capability to advanced agent functionality. Context window expands to 5 million tokens (+150%). Tool usage jumps from 50 to 500 calls (+900%). Multimodal support moves from image+video to full‑modal real‑time interaction. New features include Project Astra (real‑time multimodal agents), Deep Research 3.0 (autonomous research generating 100‑page reports), and Workspace Agent 2.0 (deep integration with Gmail, Docs, Sheets, Slides). Pricing is $0.0015 per k‑token input and $0.006 per k‑token output, a 50% reduction versus Gemini 2.5 Pro.

China AI compute – The Ministry of Industry and Information Technology reported that China’s AI compute capacity reached 2000 EFLOPS in Q2 2026, the world’s highest. The mix is 35% training (Huawei, Alibaba, Tencent), 50% inference (Huawei, Alibaba, Tencent, ByteDance), and 15% edge (Huawei, Horizon Robotics, Black Sesame). Growth drivers include domestic chip capacity (Ascend 920, Cambricon 590), exploding large‑model training demand, and the East‑to‑West compute initiative. Target: 4000 EFLOPS by end‑2027.

Huawei Ascend 920 – Huawei announced mass production of 1 million Ascend 920 chips. Benchmarks show 2000 TFLOPS FP16 performance (vs 1800 TFLOPS for Nvidia B200, +11%), 6 TB/s memory bandwidth (vs 4.8 TB/s, +25%), 500 W power (vs 700 W, 40% better efficiency), and a price of ¥150 k (vs ¥300 k for B200, 50% cheaper). Customers such as Baidu and Alibaba reported 35% faster training and 45% lower cost.

ByteDance Doubao – Doubao daily active users surpassed 200 million on June 7, with over 30 million paying subscribers (15% conversion). Token usage averages 2 million billion per day. Revenue composition: 50% subscription, 30% API calls, 15% enterprise customization, 5% other. Targets: 300 million daily users and 50 million paying users by end‑2026.

Tencent Hunyuan 5.0 – Launched on June 8, Hunyuan 5.0 reached 50 000 enterprise customers. Context window doubled to 20 million tokens (+100%). Multi‑agent capabilities grew from 10 to 50 collaborative agents (+400%). Code ability advanced from full‑stack to full‑stack + DevOps automation. Industry models increased from 20 to 50 (+150%). Pricing: $0.2 per million tokens input, $0.8 per million tokens output (30% cheaper than version 3.5).

US HBM export controls – On June 5, the US Department of Commerce restricted China’s access to high‑bandwidth memory (HBM). HBM3E exports to China are banned; HBM3 and HBM2E require licenses with limited quantities. The move impacts Chinese AI chip manufacturers (short‑term) and accelerates domestic HBM development (long‑term). Samsung and SK Hynix may see 15‑20% revenue declines.

EU AI Act fine – The EU issued its first AI‑related fine on June 4: Google was penalized €150 billion for systemic violations of AI transparency obligations under the AI Act. Violations included insufficient disclosure of Gemini training data, lack of labeling for AI‑generated content, missing algorithmic bias audits, and no effective appeal mechanism. Cumulative AI‑related fines now total €157.3 billion.

UN AI Governance Fund – Southeast Asia – The sixth batch of projects launched across ten Southeast Asian nations, allocating $5.35 million in total. Initiatives include AI‑enabled port logistics in Indonesia, smart manufacturing in Vietnam, AI tourism services in Thailand, disaster‑early‑warning in the Philippines, palm‑oil optimization in Malaysia, and AI‑driven fintech in Singapore. China contributed 300 technical experts, trained 5 000 local engineers, and donated compute equipment worth $200 million.

Overall, the week underscores rapid scaling of AI models, intensifying geopolitical competition over compute resources, and increasing regulatory scrutiny worldwide.