AI News Roundup — June 10, 2026
SpaceX's record-breaking IPO enters its final hours with $250 billion in demand and pricing set at $135/share for a $1.75 trillion valuation, while the global AI chip supply chain fractures further as Taiwan considers extending export controls to all Chinese buyers — not just Huawei — and Apple's new Siri AI is blocked from 450 million European users over a DMA standoff.
SpaceX IPO Eve — $135/Share, $250B Demand, Trading Opens Friday
SpaceX's landmark IPO is pricing Wednesday evening at $135/share, targeting $74.4 billion in net proceeds at a $1.75 trillion valuation — the largest IPO in recorded history, surpassing Saudi Aramco's $29.4 billion in 2019. Demand has reached $250 billion, roughly 3.5–4x the offering size, with several institutions placing single orders exceeding $10 billion. SpaceX perpetual futures on Binance were already trading around $162.50/share — about 20% above the IPO price — suggesting a meaningful first-day pop. Trading opens on Nasdaq under the ticker SPCX on Friday June 12. SpaceX is pitching investors on a $23 trillion AI opportunity and its unique ability to build space-based data centers beyond earthbound infrastructure constraints.
Yahoo Finance / Motley Fool · CNBC · Reuters
Claude Fable 5 Benchmark Deep-Dive — SWE-Bench Pro 80.3%, Beating GPT-5.5 by 21 Points
With a day of independent testing complete, Claude Fable 5's benchmark advantage over rivals has become clearer. On SWE-Bench Pro — the industry-standard agentic software engineering benchmark — Fable 5 scores 80.3% versus GPT-5.5's 58.6%, a gap of over 21 percentage points. GDPval-AA knowledge-work evaluation: 1,932 (up from Opus 4.8's 1,890). Tool use score: up 17.4% versus Opus 4.8. Legal benchmark: up 13.3%. OSWorld-Verified: 83.4%. Early enterprise testers reported that Fable 5 completed in 36 hours a physics research task that took GPT-5.5 four days. The model's stated principle: "the longer and more complex the task, the larger Fable 5's lead." Available immediately via API and at no extra cost to Pro/Max/Team subscribers until June 22, after which paid usage credits apply. Pricing: $10/$50 per million input/output tokens.
Build Fast With AI · Anthropic
Apple vs EU — Siri AI Blocked for 450 Million European iPhone Users Over DMA Standoff
One day after Apple's WWDC unveiled Gemini-powered Siri AI, the European Commission confirmed none of it will reach European users. Apple asked for an 18-month DMA interoperability exemption for Siri AI, arguing that the model's deep agentic capabilities — reading emails, editing files, taking cross-app actions — create privacy and security risks if rival AI assistants are granted equal platform access. The EU's response: "That's not an option." The Digital Markets Act requires equal access for competing AI assistants. Apple claims compliance would require a "Trusted OS Agent" sandboxed environment, which the EU has rejected as legally insufficient under the DMA's full-access requirement. Commercial impact: Apple's $1 billion/year Google Gemini licensing deal now excludes the EU's 450 million iPhone users. Google gets paid for global deployments only outside Europe, meaningfully reducing the effective value of the partnership.
Build Fast With AI · MacRumors
Taiwan Weighs Sweeping AI Chip Export Controls on All of China — Not Just Huawei
Taiwan is considering extending its AI chip export restrictions beyond blacklisted firms like Huawei to cover all Chinese customers for the first time, aligning with the US framework. Current rules require export licenses only for Huawei and SMIC, leaving the broader Chinese market unrestricted — a loophole allowing servers containing restricted chips to reach China through downstream channels. The proposed rules would target chips exceeding US-defined Total Processing Performance (TPP) thresholds, equivalent to roughly NVIDIA's H200 and AMD's MI325X and above. The impact would be substantial: Foxconn, Quanta, Wistron, Wiwynn, and Inventec all assemble NVIDIA and AMD server systems in Taiwan for Chinese data centers. Taiwan's Ministry of Economic Affairs confirmed it is "strengthening oversight of strategic high-tech exports in line with global export-control trends." Taiwan's ITA also added 265 entities to its Strategic High-Tech Commodities Entity List.
Google Releases DiffusionGemma — 4x Faster Text Generation Using Diffusion Instead of Autoregression
Google unveiled DiffusionGemma, an experimental model that generates text using diffusion rather than the token-by-token autoregressive approach used by virtually all major LLMs today. Instead of generating one token at a time, DiffusionGemma creates and refines 256-token blocks in parallel, enabling output speeds exceeding 1,000 tokens per second on an NVIDIA H100 GPU and over 700 tokens per second on a GeForce RTX 5090 — roughly 4x faster than comparable autoregressive models. The model uses a 26-billion-parameter mixture-of-experts architecture but activates only 3.8 billion parameters during inference, fitting within ~18 GB of VRAM. Bidirectional attention means every token can attend to every other token — better for code completion, in-line editing, and mathematical structures. Google notes it prioritizes speed over quality and recommends standard Gemma 4 for production. Released under Apache 2.0 on Hugging Face, with support for vLLM, MLX, Hugging Face Transformers, NVIDIA NeMo, and Unsloth.
Jeff Bezos Backs $500M Bet on Brain-Inspired AI — Startup Flourish Targets Neurons as Blueprint
Jeff Bezos led a $500 million funding round for Flourish, a startup valued at $2.5 billion that is attempting to reverse-engineer the brain's "core algorithm" to build a fundamentally different kind of AI. Unlike current deep learning approaches, Flourish is placing real neurons under the microscope — literally — to try to identify the computational principles that make biological intelligence so sample-efficient and energy-efficient compared to large language models. The bet is that current transformer-based architectures are hitting fundamental scaling limits, and that the next leap in AI will come from neuromorphic or bio-inspired computing architectures rather than more GPUs and more data.
CrowdStrike Warns of Escalating Chinese AI Cyberattacks on US Tech Infrastructure
CrowdStrike published a new threat report warning that Beijing has significantly ramped up AI-powered cyberattacks targeting US technology companies, accelerating efforts to copy American AI model architectures and close the capability gap. The timing is notable: it comes the day after China announced its $295 billion domestic AI data center plan explicitly designed to run without NVIDIA and AMD chips, and as Taiwan weighs comprehensive chip export restrictions that would close existing loopholes. CrowdStrike said the attacks are becoming increasingly automated and sophisticated due to AI augmentation, and are specifically targeting the underlying model weights and training pipelines of frontier AI labs.