The Trump administration and OpenAI are in early talks over the US government taking an equity stake in the company. The mechanism under discussion isn't a purchase: OpenAI would donate shares to the federal government, which would seed a "Public Wealth Fund" that invests in AI companies and pays returns directly to American households. Trump floated the concept aboard Air Force One on June 5: it "almost becomes a partnership with the American public." Altman first raised the idea in early 2025, and OpenAI codified it in an April policy paper that proposes the Fund explicitly, while calling its own ideas exploratory.

No deal is signed, no percentage is official. The 1–5% figures in circulation come from unnamed sources, not a term sheet. OpenAI sits around an $850B valuation with an IPO reportedly targeted as early as September. Anthropic, notably, says it isn't in these conversations.

The dissent is the interesting part. David Sacks, the administration's former AI czar, broke with the direction, warning it builds toward "corporate-government fusion," and called Senator Sanders' harder version, a one-time 50% tax paid in stock plus board seats at OpenAI, Anthropic, and xAI, a "stupidity tax." The structural novelty cuts both ways: it would be the first time the US government held a direct equity position in a major tech company by design rather than through a bailout, which means the same entity writing AI rules would profit from the largest AI company's success.

Per the Financial Times, a senior OpenAI employee put it bluntly: "Chat is dead." The plan is to collapse OpenAI's sprawl, ChatGPT, the Codex coding product, agents, image generation, and third-party integrations (Canva, Booking.com, Spotify, Dropbox), into a single agent-first desktop app. The reorg is real, not cosmetic: the ChatGPT and Codex product teams are being merged under one structure, with co-founder Greg Brockman taking product strategy in mid-May and Fidji Simo (currently on medical leave) involved in the restructuring. Rollout is described as "coming weeks," timed near a confidential IPO filing.

The thesis is that the next battleground is persistent, autonomous task execution, not single-turn answers, and that one app tells a cleaner story to public markets than a portfolio of experiments (Sora, OpenAI for Science, a standalone Codex subscription). The more granular details, an internal "Aria" codename, a built-in email-drafting widget, sit in secondary coverage and aren't confirmed on the record; the FT's account rests on more than a dozen mostly anonymous current and former staff. Treat the direction as solid and the specifics as soft.

A model-maker paying a rocket company for chips is the kind of sentence that needs an SEC filing behind it, and there is one. A Form FWP disclosure lays out the terms: Google pays SpaceX $920 million a month for access to roughly 110,000 Nvidia GPUs (plus CPUs, memory, and supporting hardware) at a Memphis facility, running October 2026 through June 2029. At full rate that's about $29 billion. Either side can walk after December 31, 2026 on 90 days' notice, and SpaceX must stand up the capacity by September 30 or face termination. Google frames it as a short-term bridge for Gemini Enterprise demand while it builds its own.

The "why SpaceX" resolves cleanly once you remember SpaceX acquired xAI in February, absorbing the Colossus data-center infrastructure, which is why some headlines tangled this up with "xAI compute." Google pays SpaceX; SpaceX owns the metal. Anthropic separately rents the same Memphis site at about $1.25 billion a month. The filing landed days before SpaceX's planned market debut, and on these numbers SpaceX's data-center revenue would dwarf its launch and Starlink businesses. Back-of-envelope: roughly $8,400 per GPU per month is what frontier capacity costs at scale right now.

After three years as OpenAI's host and best customer, Microsoft showed at Build 2026 that it can be a frontier lab in its own right. It launched seven in-house MAI models, spanning reasoning, code, image, transcription, and voice. The headliner, MAI-Thinking-1, is a sparse mixture-of-experts with 35B active parameters (~1T total) and a 256K context window; Microsoft reports 97.0% on AIME 2025 and parity with Claude Opus 4.6 on SWE-Bench Pro, trained on its own Maia 200 silicon at a claimed 1.4× perf-per-watt over GB200. The pointed design choice: it's trained without distillation from any third-party model, unusual for a reasoning model that would normally bootstrap off a stronger teacher.

AI chief Mustafa Suleyman tied it to the renegotiated OpenAI contract finalized last October, which "set us free" to pursue superintelligence independently. Two caveats temper the pitch. Microsoft's "clean, commercially licensed data" framing is misleading, the paper itself describes a proprietary web crawl of ~1.2 trillion pages (filtered to ~794B), which Simon Willison flagged after reading it; the no-distillation claim holds, the licensed-corpus claim doesn't. And every benchmark here is Microsoft grading its own work, with the human-preference eval run through a vendor it paid.

The market had priced the AI-chip complex as if guidance raises were a standing order. Broadcom broke the spell. Its results on June 3 actually beat ($22.19B revenue, $2.44 EPS), but Q3 AI-chip guidance of $16B came in ~7% under the $17.2B expected, and CEO Hock Tan declined to lift the full-year target. A flat reaffirmation read as deceleration. Layered on top: a May jobs report of 172,000 against ~80–88k expected, which flipped rate-cut bets to a roughly 70% chance of a hike by October.

The two shocks compounded on June 5:

• Nasdaq −4%, its worst day since April 2025, snapping the S&P's nine-week winning streak
• Philadelphia Semiconductor Index −6%+; Nvidia −6%, market cap back below $5T
• Asia followed: Kospi −5.1% (circuit breaker tripped), SK Hynix ~−8%, SoftBank ~−6%
• The Russell 2000 rose 1.45% the same day

That last line is the tell analysts are chewing on: small caps up while megacap tech fell looks like money rotating within equities, not fleeing them, which argues against a structural break. Dollar tallies vary by what you count, roughly $1.3 trillion off the chip complex, ~$1.7T across all equities. The open question Broadcom actually raised: whether hyperscaler AI capex keeps compounding, or just stopped surprising to the upside.

Anthropic put Claude Opus 4.7, a general model with no chemistry fine-tuning, against the dedicated tools chemists use to interpret NMR spectra, and it held its own on both directions of the problem. Forward prediction (structure → expected peaks) and the harder inverse (peaks + molecular formula → structure) were benchmarked against ChemDraw 25 and MestReNova 17.

The inverse result is the sharper one: from a peak list and formula alone, Opus 4.7 recovered all 8 simpler structures on every attempt, and 4 of 7 harder fused-ring targets when handed a starting-material hint. On the densest cases without the hint, it sometimes looped without committing.

The honest framing comes from Anthropic itself: this is self-reported, not peer-reviewed, and "indicative rather than precise", 20 compounds across four scaffold families and 15 inverse problems, where a real benchmark would need several hundred. It also skips 2D NMR (COSY, HSQC, HMBC) and stereochemistry entirely, which is where actual structure elucidation gets hard. Still, "general model matches the specialist tool from a plain-text paste" is a real result, and the inverse direction is the one that normally eats expert time.

Per a report in The Information, xAI spent months distilling its coding models on outputs from Anthropic's Claude. Anthropic revoked xAI's API access around January, pointing to the clause in its commercial terms that bars using Claude to train competing models, and xAI engineers allegedly kept pulling outputs afterward through personal accounts and the third-party service Blackbox AI. These are allegations sourced to unnamed people; neither company has confirmed the training claim on the record.

The geometry around it is almost funny. xAI, now under SpaceX, rents its Colossus GPUs to the very company it allegedly copied, Anthropic pays ~$1.25B a month for that compute, and to Google at ~$920M a month. The reporting also sketches an xAI under strain: a pre-training team reportedly down to fewer than five people, four Grok code leads gone within months, and a data-deletion mishap that cost weeks. Musk has previously conceded xAI "partially" used OpenAI models for early Grok, calling it industry standard, so the pattern isn't new; what's new is the claim it continued after an explicit cutoff.

Clive Chan, the second hardware hire on OpenAI's in-house accelerator effort, announced he's leaving for Anthropic as a Member of Technical Staff, summing up his focus as "perplexity per picojoule", performance per unit of energy. He came to OpenAI from Tesla's Dojo team in January 2024 and spent ~2.5 years on its Broadcom-partnered silicon, whose first chips are due in production in the second half of 2026.

The signal is in the timing. Anthropic has no chip team yet, Reuters reported in April it was only weighing whether to build custom silicon, so this reads less like a lateral move and more like a founding hire. With a run rate past $30 billion (up from ~$9B at the end of 2025) and an IPO on the horizon, the economics of owning your hardware, rather than renting Google TPUs and Amazon chips, start to pencil out. Poaching the person who stood up a rival's program from zero is the fastest way to compress the head start. What his actual charter is, Anthropic hasn't said.

Banking runs as an apprenticeship: hire large junior classes, work them hard, promote the survivors into senior bankers. AI is hollowing out the bottom rung. Standard Chartered has announced ~7,800 corporate-function cuts over four years, with CEO Bill Winters describing the shift as "replacing, in some cases, lower-value human capital with financial capital", a line he apologized for two days later. The sharper figure is a McKinsey QuantumBlack projection that junior analyst intakes could shrink by up to two-thirds industry-wide, even as roughly 62% of banks' new AI hires get drawn from those same graduate pools.

The C-suite chorus is unusually synchronized: Dimon ("eliminate jobs", "old jobs"), Citi's Fraser ("no longer required"), Goldman's Waldron (a "human assembly line ripe for automation"), with HSBC and Barclays leaning on reskilling language. Two things keep this from being a clean trend. The two-thirds number is McKinsey's projection, not any bank's stated policy, and Standard Chartered's 7,800 is corporate functions, framed partly as attrition. And Bank of America is cutting the other way, confirming 2,000 interns and 2,000 full-time hires. The bet underneath: that the next generation of senior bankers is either AI-native or simply smaller.

Most vaccines chase one strain. The Cambridge spinout DIOSynVax (short for "Digitally Immune Optimised Synthetic Vaccines," founded by Jonathan Heeney in 2017) does something stranger: it feeds every known sequence from the Sarbecovirus subgenus, human and animal-reservoir alike, into a machine-learning pipeline that designs a single synthetic antigen from the features those viruses share, then refines it against structural data. The product isn't any real spike protein; it's a computed composite of the whole family. Heeney's pitch is that it ends "the constant cycle of chasing the variants... like a dog chasing its tail." The construct, a DNA vaccine called pEVAC-PS, just cleared a Phase 1 trial, the first time a fully computationally designed antigen has been put into people.

It lands in a crowded field. The other serious universal-coronavirus bets are nanoparticle designs that physically stud a protein scaffold with real receptor-binding domains from many sarbecoviruses at once — Caltech's mosaic-8b (eight RBDs on a 60-mer), Walter Reed's SpFN, the GBP511 candidate now in Phase 1/2. They show the immune system many real examples and let it generalise; DIOSynVax computes the shared essence first and builds one synthetic antigen from it. The work isn't fringe: it carries up to £31m from CEPI under the "100 Days Mission" to stand up a pandemic vaccine within 100 days of a new pathogen, and the same platform is being aimed at pan-H5 influenza and Ebola.

The trial ran 39 healthy adults, 18–50, dosed needle-free by jet injector, with no significant adverse events and immune responses against SARS-CoV-2, SARS-CoV-1, and bat coronaviruses. Two headwinds sit behind the "modest" immunogenicity the paper reports, not one. The obvious one is pre-existing immunity, nearly everyone in the cohort had prior COVID exposure or shots, swamping the signal. The quieter one is the platform: DNA vaccines are notoriously weak in humans, where naked plasmid struggles to get inside enough cells. Only one has ever been approved for people (India's ZyCoV-D), and the field leans on delivery tricks like electroporation and needle-free jets to claw back potency. Phase 1 proves safety; whether a synthetic super-antigen on a DNA backbone can induce broad protection is the Phase 2 question, in a bigger and more varied group.

The scarcest component in an AI accelerator isn't the GPU, it's the high-bandwidth memory stacked beside it, and HBM supply has been the field's most persistent bottleneck. On June 7, Nvidia and SK Hynix signed a multi-year co-development and supply pact spanning four product lines: Vera Rubin systems (HBM4, LPDDR5X, 3D NAND), the standalone Vera CPU, RTX Spark PCs, and Jetson Thor robotics. The deal also pushes Nvidia's own software, CUDA-X, the CuLitho lithography tooling, Omniverse fab digital twins, into SK Hynix's factories.

The shift is from buying memory generation by generation to aligning roadmaps years out, which lowers the risk that memory slips behind GPU timelines. Embedding Nvidia's design tools in the fab also couples future memory to Nvidia's architecture from the drawing board. No volumes or pricing were disclosed; analyst estimates (not contract terms) put SK Hynix at ~60–70% of HBM4 for Vera Rubin, ahead of Samsung and Micron. Shipments are expected to start in Q3.

Prompt injection, hiding instructions in a web page or file so an AI with tools does an attacker's bidding, is the canonical unsolved problem for chatbots that browse and act. OpenAI's new Lockdown Mode for ChatGPT doesn't claim to solve it. It's an opt-in setting that disables the capabilities an attacker would use to get data out: live browsing (cached content only), web image retrieval, Deep Research, Agent Mode, Canvas networking, live connectors, and ChatGPT-initiated file downloads.

The design logic is refreshingly honest. You can't reliably detect a hidden injection, so instead of trying, you remove the exfiltration channel, cut the third leg of the attack (access to data, exposure to untrusted content, a way to send it out) rather than the first two. CISO Dane Stuckey frames it for "elevated risk profiles" and concedes the functionality tradeoff. OpenAI is explicit that the mode does not prevent injections from happening: a malicious instruction can still sit in cached content or an uploaded file and still skew what ChatGPT does. It's a surface-reduction control, not a fix, and it's rolling out across Free, Plus, Pro, Go, and Business self-serve accounts.

The American Federation of Musicians filed suit on June 5 in the Southern District of New York against Universal Music Group and Warner Music Group. The allegation: after the labels settled their own copyright fights with AI song generators Suno and Udio in late 2025, they licensed members' recordings into those settlements, pocketed the proceeds and ongoing revenue, and paid the session musicians whose performances trained the models nothing.

The legal hook is a clause the musicians already bargained for, the "new use" provision in their collective agreement, which requires payment whenever recorded performances are put to a new commercial purpose. The union wants damages and a court order forcing the labels to disclose exactly which recordings went into the AI training sets. Sony, the one major that hasn't settled with either AI company, isn't named. Warner calls the suit "unproductive" amid ongoing negotiations; Universal says it prefers to resolve this through collective bargaining. The settlement terms are confidential, so the live legal question, does feeding a recording to an AI count as a "new use", is exactly what's untested.

Telling a coding agent "make the button in the top-right a bit bigger" is lossy, it has to guess which element you mean. Cursor's Design Mode, expanded in version 3.7 on June 5, removes the translation step: you direct the agent on your live running app by clicking an element, multi-selecting several, drawing a region on the viewport, or just talking, with the mic staying live mid-run so you can queue the next change without waiting.

A click hands the agent two things at once: the element's identity (XPath, component name, computed styles, React fiber-tree props) and a frozen screenshot for spatial context, after which it edits the code directly. Multi-select and voice are the new bits in 3.7; single-element clicking shipped with Cursor 3 in April, and canvas support landed June 4. The distinction from Figma Dev Mode or Builder.io is that this operates on the actual app, real data, real state, not a static design file, though the fiber-tree introspection leans on React, and the reliability claims are Cursor's own.