
Google pays SpaceX $920M a month for GPU bridge capacity, citing unexpected Gemini Enterprise demand
Google will pay SpaceX $920M per month from October 2026 through June 2029 for roughly 110,000 NVIDIA GPUs — bridge capacity Google says is needed to absorb demand for its Gemini Enterprise agent platform that exceeded its own forecasts, making it the second major AI lab to rent Colossus capacity after Anthropic's $1.25B/month deal. The agreement landed one week before SpaceX's planned Nasdaq debut at a $1.75 trillion valuation, establishing Colossus as the de facto overflow infrastructure for frontier AI demand even the world's largest compute owner couldn't absorb fast enough on its own.
techcrunch.com →- 02
OpenAI's Dreaming V3 makes ChatGPT memory fully automatic, extending it to free-tier usersDreaming V3, rolling out to US Plus and Pro users since June 4, replaces ChatGPT's manual saved-memories list with a background process that synthesizes a user profile across years of past conversations and updates it over time — no user action required. The 5x compute reduction enabling the free-tier expansion comes with a structural consequence: synthesized memories live in a separate data layer, survive conversation deletion, and are injected into every new session's system prompt — the same architecture Tenable Research identified in 2025 as a prompt-injection exfiltration channel that OpenAI has not confirmed Dreaming V3 addresses. The rollout lands eight weeks before EU AI Act chatbot transparency obligations take effect on August 2, 2026.
techtimes.com → - 03
Apple rebuilds Siri on a custom 1.2T-parameter Gemini model at $1B per yearAt WWDC on June 8, Apple will unveil a rebuilt Siri whose cloud features run on a custom ~1.2-trillion-parameter Gemini model licensed from Google for roughly $1 billion per year — a structural acknowledgment that Apple's own LLM program couldn't reach production quality in time. The arrangement installs Gemini as the default cloud AI across roughly 1.4 billion active iPhones, handing Google the largest single distribution gain in the assistant market.
theverge.com → - 04
Meta built an unlabeled AI clickbait feed, then killed it after The Verge reportedMeta's standalone AI app ran a months-long personalized feed of entirely AI-generated articles — stereotyped localized prompts, images of real public figures riddled with errors, no attribution, no AI labels — until The Verge asked questions and Meta said it would deprecate the feature. The incident is a disclosed case of AI-generated content running at social-network scale without editorial safeguards, publisher licensing, or disclosure, killed only under press scrutiny.
theverge.com → - 05OpenAI Lockdown Mode expands to all personal and Business accounts, blocking exfiltration at the network layer
OpenAI's Lockdown Mode — previously enterprise-only — now covers Free, Go, Plus, Pro, and self-serve Business accounts; it disables outbound network channels — live web, Agent Mode, Canvas, and live connectors — to cut the exfiltration leg of prompt-injection attacks. The control is deterministic rather than AI-evaluated, which matters: a model-based filter can itself be manipulated by a devious injection, while a network block cannot. The rollout implicitly confirms that default ChatGPT offers no robust protection against determined exfiltration — a useful baseline for developers building on the API.
simonwillison.net → - 06
Bots now generate 57.5% of web traffic — a year ahead of Cloudflare's own forecastCloudflare Radar shows automated systems now account for 57.5% of HTTP requests globally — crossing 50% more than a year before CEO Matthew Prince's own 2027 prediction — driven by agentic AI programs that visit thousands of pages per user task where a human opens a handful. Cloudflare's HTTP 402 pay-per-crawl system is live today, with customers sending over one billion 402 responses per day, giving any site operator a concrete mechanism to charge crawlers rather than just block them.
techtimes.com → - 07
Japan removes consent requirement for AI training on medical and criminal recordsDigital Minister Hisashi Matsumoto invoked 'AI colony' risk to defend a bill now in upper-house debate that strips the individual-consent requirement from Japan's data-protection law, clearing the way for AI developers to train on medical and criminal records. Japan joins the EU — which this week unveiled its own tech-sovereignty package — in treating data-access rules as industrial policy, a pattern that compresses the timeline for domestic AI players like SoftBank and Sakura Internet to close the gap on US and Chinese models.
kfgo.com → - 08
Trump administration in active talks to take an equity stake in OpenAICNBC confirmed the Trump administration is in active discussions to acquire equity in OpenAI — a structure Altman has been pitching since early 2025, one OpenAI itself endorsed through its 'Public Wealth Fund' proposal — while Sanders separately proposed a 50% stock tax on OpenAI, Anthropic, and xAI payable in shares. With all three companies potentially going public this year, government co-ownership is now live on both sides of the aisle, adding a new layer of uncertainty to OpenAI's conversion from capped-profit to fully for-profit.
techcrunch.com → - 09
Waymo channels retired I-PACE battery packs into California and Texas grid storage via B2UWaymo signed a supply agreement with B2U Storage Solutions to divert retired Jaguar I-PACE battery packs — which retain capacity after falling below vehicle-use thresholds — into California and Texas grid storage, targeting hundreds of megawatts. Autonomous fleet operators generate a steady, predictable supply of partially depleted EV batteries, and this deal signals that second-life battery pipelines are becoming a standard layer of the AV fleet business model.
waymo.com →