hot events · 2026-06-03

Six reports converge on the same numbers: $75 billion sought, $135 a share, and a $1.8 trillion valuation. Bloomberg also frames Reuters as the source, so this looks like one financing document spilling into a broad media chase. I don’t read this as just a giant IPO. It is AI capex extending into orbital infrastructure. The body only gives headline-level detail; AI revenue, compute budget, and Starlink training workloads are not disclosed. Still, SpaceX putting “AI” beside “launch” in the use-of-proceeds pitch is a hard signal. OpenAI and Anthropic are still negotiating cloud, chips, and power; Musk is taking rockets, satellite networks, and data access to the public market as one package. If investors underwrite $1.8 trillion, the definition of an AI infrastructure stock gets stretched again.

Two sources picked this up, but the numbers don't match. The HN post links directly to Cloudflare Radar's live dashboard, which shows bots at 34.1% of HTML requests — not above human traffic as the headline claims. The 57.5% figure appears in aihot's headline and likely comes from a third-party report, not Cloudflare's own data. Cloudflare's methodology matters here: they only count HTML page requests, not all HTTP traffic. API calls, image loads, and other non-HTML requests are excluded. That 34.1% is actually a conservative number — if you included API traffic, the bot share would only go higher. The gap I can't close: where does the 57.5% number originate? Neither source cites the original report. Two outlets running with different figures and no cross-verification means I'd hold off on any "bots have taken over" narrative until the primary source surfaces.

HN and Reddit are both pointing to Google's official blog post — same source, same framing. That means we're working with Google's own narrative right now, no independent evals yet. The headline change is "encoder-free." Most multimodal models use a separate vision encoder to convert images into features before the language model touches them. Gemma 4 12B drops that step entirely — images and text go straight into the same transformer. Simpler architecture, lighter deployment, and Google says it runs on a laptop. The open question is whether this design holds up on complex visual reasoning against models with dedicated encoders. I'd take the "laptop-ready multimodal" pitch at face value for now, but hold off on performance claims. No MMLU or MMBench numbers in the post, no pricing, no inference speed data. Once someone runs real evals, we'll know if this is genuinely competitive or just conveniently small.

This is NVIDIA's own CVPR 2026 research roundup, covered by two blog posts that are essentially the same announcement. I'd discount the confidence a bit — everything here is NVIDIA highlighting their own wins, with no third-party benchmarks or peer reviews in the mix yet. The work spans three areas: robot grasping (a framework called DexGrasp that claims much higher success rates in cluttered environments), autonomous driving (end-to-end planning with large models, ditching HD maps), and vision agents (a new VLA architecture that trains visual understanding and action execution jointly). None of these directions are new — grasping, end-to-end driving, and vision-language-action models have been hot for two years. NVIDIA's edge is their full stack: Isaac Sim for simulation, their training frameworks, and their deployment chips, so the methods can run directly on their own hardware. Whether the paper numbers hold up in real-world settings is still an open question.

Trump’s AI order cuts the earlier 90-day pre-release ask to 30 voluntary days and rejects mandatory licensing. That leaves frontier labs with paperwork, review channels, and political exposure, but not a deployment gate. OpenAI, Anthropic, and Google DeepMind can live with that trade. The sharper part is the Anduril-Meta prototype: eye tracking plus voice commands for drone-strike orders. That is not consumer smart-glasses theater. It compresses sensing, command, and weapons interaction into one headset loop. Anduril has been selling the Lattice OS story for years; plugging Meta’s hardware stack into battlefield UX makes the governance fight concrete before the product is mature.

The UK's Competition and Markets Authority just handed Google a hard rule: website owners can block their content from appearing in AI Overviews and from being used to fine-tune Google's AI models. Both The Verge and TechCrunch reported this, and their accounts line up, which tells me they're working from the same regulatory document rather than spinning their own angles. What I'd hold off on: neither source explains how the opt-out actually works. Is it a new robots.txt directive? A toggle in Search Console? We've got the principle but zero technical detail. Also, this ruling is UK-only for now, and no one's saying whether Google will extend the same mechanism globally. If you're building on public web data for training or RAG, keep an eye on whether other regulators follow suit. But don't change anything yet—wait for Google's actual implementation spec.

Daxiao Robotics and NTU Singapore released PhysX-Omni, a framework that claims to unify rigid body, soft body, and articulated body physics in one 3D generation pipeline. Both Chinese tech outlets ran nearly identical angles — "unified physical 3D generation" and "completing the physical AI infrastructure" — which tells me they're working off the same press release, not independent reporting. I'd discount this for now. The WeChat article is behind a verification wall, so all we have are the headline and a one-line summary. No paper, no code, no benchmarks. "Unified framework" is a big claim in physics simulation — rigid bodies, soft bodies, and articulated bodies use fundamentally different math, and historically they've been handled by separate solvers. If PhysX-Omni actually pulls this off without sacrificing accuracy, that's real. But without numbers, this reads more like a direction statement than a verified result. What's missing: a paper link, a GitHub repo, and comparison data against existing simulators. Don't read this as a milestone yet — wait for the actual release.

DeepSeek’s rumored round reads less like startup financing and more like a cap table for China’s AI infrastructure stack. The numbers are huge: RMB 50B raised, RMB 350B–400B post-money, Liang Wenfeng putting in RMB 20B, Tencent RMB 10B, CATL RMB 5B. That mix does not price simple model revenue. Tencent buys distribution and cloud leverage; CATL buys exposure to power, storage, and data-center load. I’m skeptical of the framing. The body is only an RSS snippet, with no terms, board seats, compute purchase commitments, cloud tie-ins, or source of Liang’s RMB 20B disclosed. DeepSeek V3 and R1 earned real mindshare on cheap reasoning, but a RMB 400B valuation prices a national infrastructure seat, not a chatbot business.