On June 1 GitHub Copilot swaps premium-request units for token-metered "AI Credits" (1 credit = $0.01). Inline code completion and next-edit suggestions stay unlimited and free on every paid plan; everything agentic (Chat, the CLI, the cloud agent, Spaces, Spark, third-party coding agents) now draws from a monthly pool priced at published per-model token rates. The pools: Pro $10/mo gets 1,500 credits, Pro+ $39 gets 7,000, Max $100 gets 20,000, with Business ($19/user) and Enterprise ($39/user) carrying promotional bonuses through August. Sample rates tell you where the risk lives: Claude Opus runs $5/M input and $25/M output, GPT-4.1 runs $2/$8. Two structural changes matter more than the sticker. Cost now tracks actual token volume, not a request count, so "one request" can cost wildly different amounts depending on model and context depth. And the free-model fallback is gone: where Copilot used to silently downgrade you to a cheaper model at the ceiling, you now hit a hard wall. The community announcement thread logged roughly 901 downvotes. The viral "one agentic session zeroes out a $10 plan" claim is plausible arithmetic (Opus output at $25/M against a $10 budget is about 400K output tokens, which a multi-file refactor eats fast), but it's user extrapolation, not a GitHub-published session cost, and for anyone living in chat and completions the change is mostly cosmetic. Note also that Opus is pulled from Pro entirely and sits on Pro+ at a 27x multiplier.

Meta is laying off about 8,000 people, roughly 10% of its end-2025 headcount of 78,865, with 6,000 open reqs frozen on top. The cuts hit Reality Labs, the Facebook app division, recruiting, sales, and global operations; AI infrastructure, foundation models, and AI monetization were explicitly walled off. The tell is that none of this was forced by the numbers. Q1 2026 was a record: $56.31B revenue, up 33% year over year, the fastest growth since 2021. What changed is where the dollars go. Meta raised 2026 capex guidance to $125–145B, against $72.2B actually spent in 2025, and leadership framed the headcount cuts as offsetting that infrastructure surge. This is the clearest case yet of a hyperscaler openly trading payroll for compute. A few corrections to the loose framing: the "7,000 reassigned to AI" figure has no traceable Meta source (one disclosed number is roughly 1,000 transfers into Applied AI Engineering); "doubling the AI budget" is about +87% at the midpoint, not a clean 2x; the headline $26.8B net income leans on an ~$8B one-time tax benefit (adjusted is closer to $18.7B); daily actives slipped to 3.56B, down 5% sequentially, which raises the stakes on the bet; and more cuts are flagged for H2 with no number attached.

At the Choose France summit (Versailles, May 26; SoftBank's own press releases landed today), SoftBank committed up to €75B, roughly $87B, to build up to 5GW of AI data-center capacity in the Hauts-de-France region. The concrete part is Phase 1: €45B for 3.1GW across three named sites by 2031. Dunkirk pairs with Schneider Electric, which will run a robotized plant making data-center power modules next to the site, a vertically integrated supply-chain play rather than a plain campus. Bosquel is a 1GW joint venture with French operator Sesterce. Bouchain reuses a former EDF power-plant site, with EDF as partner. For scale, 5GW is roughly the entire installed UK data-center market, and the bet leans on France's cheap, low-carbon nuclear grid and existing heavy-industrial land. Hold the ceiling numbers loosely: every headline figure carries SoftBank's "up to," only Phase 1 is tied to sites and a timeline, and Fortune flags real financing questions, noting Son recently trimmed a planned $10B margin loan toward $6B. Choose France pledges also have a documented habit of overstating what actually gets built.

Within three weeks, the high-bandwidth-memory boom minted a trillion-dollar trio. Samsung crossed first, around May 6. Micron hit $1T for the first time on May 26, closing up 19.3% at $895.88 after UBS nearly tripled its target to $1,625. SK Hynix followed on May 27, up 9.4% to about 1,600 trillion KRW, roughly $1.07T, and up 258% year to date. The fundamentals under the valuations are the real story. Micron's entire 2026 HBM supply, HBM4 included, is sold out, and CEO Sanjay Mehrotra says the company can fill only 50–66% of core-customer demand. SK Hynix holds about 57% of the HBM market and has secured roughly 70% of the HBM for Nvidia's Vera Rubin platform, and guides DRAM supply-constrained through 2030. TrendForce projects conventional DRAM contract prices up 58–63% quarter over quarter and NAND up 70–75% in Q2; HBM alone is eating about 23% of global DRAM wafer capacity, and data centers consume roughly 70% of all memory produced. One correction to the newsletter framing: it named only SK Hynix and Micron, but Samsung crossed the line first, and these are genuine USD caps, not KRW figures misread as dollars. The exposure: UBS's $1,625 is a single outlier call, and SK Hynix's run embeds heavy multiple expansion that an early HBM oversupply would punish.

At its AI Now Summit in Paris on May 28, Mistral launched "Mistral for Industrial Engineering," built on simulation surrogate modeling, neural nets trained on the outputs of physics simulators to return airflow, thermodynamics, and deformation results in seconds. The capability comes from its May 22 acquisition of Vienna's Emmi AI. The genuinely new customers are Airbus (full product suite, on-prem or trusted cloud, spanning commercial, helicopter, defense, and space, including edge object recognition on aircraft) and BMW, which is training a "Large Industry Model" on its 1-petabyte-plus archive of crash simulations to predict structural-test outcomes in seconds instead of hours. EDF and ASML were also named. The moat logic is sound: a model trained on your proprietary crash data is far stickier than a generic API contract. One correction worth carrying: CMA CGM, which got lumped in as a new deal, is actually the €100M, five-year contract signed back in April 2025, repackaged under the new banner. On policy, CEO Arthur Mensch told the French National Assembly on May 12 that Europe risks becoming a "vassal state," cited a potential €1 trillion annual AI trade deficit, and pushed European-preference procurement clauses, with the lever being that about half of EU GDP flows through public spending.

SAG-AFTRA and the studios' AMPTP reached a tentative successor to the 2023 TV/Theatrical contracts on May 2, the national board approved it 89% on May 11, and about 160,000 members are voting now with ratification closing June 4. So it isn't finalized, despite recaps that say so. It's a four-year term worth over $700M, with 3% annual raises that trail inflation. The substance past 2023 is AI. The deal creates a new "synthetic performer" category for fully AI-generated characters, distinct from digital replicas of real actors, gated by a "significant additional value" standard backed by arbitration and damages that can exceed standard scale. Digital-replica consent now extends to replicas built from ordinary photography, from third-party sources, and for foreign-language dubbing, and a studio needs an "articulable business reason" before scanning anyone. Studios also can't use a performer's replica to cross a picket line, and there's an AI-training transparency clause requiring the union be notified of third-party licensing. The proposed "Tilly Tax," a fund payment each time a synthetic displaced a member, was rejected in favor of the value standard. The caveats: "significant additional value" is vague by design and will live or die on how arbitrators read it, the training clause is notice-and-meet rather than guaranteed pay, and the full text only went out shortly before the vote.

On May 21, at its Investor Day, Spotify and Universal Music Group announced paired licensing deals, covering both recorded-music and publishing rights, to power a forthcoming generative-AI tool that lets Premium subscribers create covers and remixes of songs by participating UMG artists. The structural point is that it covers masters and publishing in one framework, the split that usually stalls AI-music licensing, and it routes money back to rightsholders instead of pretending AI covers aren't already everywhere. It's built on opt-in, with a stated "consent, credit, compensation" frame, and it's a paid add-on for Premium only. Investors liked the direction: the stock rose 13% on the day, against a base of 761M monthly users and 293M subscribers. Temper it with what doesn't exist yet. There's no launch date, no price, no named participating artists, and "built with UMG" overstates the role, since UMG is licensing rights, not co-engineering the tool. Sony, Warner, Merlin, and Believe are reportedly in talks but none have signed, and an opt-in model could debut with a thin catalog if artists hang back.

On May 28 China's market regulator (SAMR) and economic planner (NDRC) jointly issued a "Guideline for AI Metrology System Development and Capacity Building," a roadmap to build national measurement infrastructure for AI, the equivalent of a standardized ruler for model performance, across six areas and framed under the 15th Five-Year Plan and the "AI Plus" initiative. The stated pivot is away from raw compute and scale toward quality benchmarking, domestically owned testing and calibration tech, national AI metrology research centers, and tools to address the "black-box" opacity problem. Two things keep this from being a regulation story. It's a guideline, not a binding national standard (GB) or law, with no enforcement mechanism in any of the source reporting, and the transparency angle is aspirational: it names the black-box problem but specifies no techniques, thresholds, or compliance requirements. The detail all comes through state media and SCMP, so the framing reflects official messaging. The real weight is geopolitical: sovereign benchmarks let China grade its models on its own terms rather than leaning on NIST or Western academic evaluations.

On the evening of May 28, during a static fire at Launch Complex 36 at Cape Canaveral, New Glenn's seven BE-4 engines ignited and the fully fueled two-stage vehicle exploded. It obliterated the transporter-erector, toppled a lightning tower, and badly damaged LC-36, the rocket's only operational pad. No one was hurt. The timing is harsh: New Glenn had flown three times (NG-3 in April put a satellite in the wrong orbit on an engine failure), the FAA had cleared a return to flight just six days earlier on May 22, and this test was prepping NG-4, a June 4 deployment of Amazon Kuiper satellites, now part of a frozen 24-mission manifest. Because the FAA treats static fires as outside its licensed activities, there's no federal grounding; Blue Origin runs its own investigation. The Artemis link is the real stakes. Blue Moon, the lander, launches on New Glenn, and the crewed Mark 2 is one of NASA's two Human Landing System providers, required for the Artemis 4 surface landing targeted around 2028. Corrections to the loose framing: the full stack was lost, not just the booster; the pad is heavily damaged, not destroyed; and the mission genuinely at risk is Artemis 4 (~2028), not Artemis 3 (~2027, an orbital demo). SpaceX's Starship, the other HLS, is untouched, so NASA keeps a path to the Moon, just a single-provider one now. If the root cause points at the BE-4, ULA's Vulcan, which shares the engine, could face scrutiny, though that's speculative. Bezos called it "a very rough day."

At the European Atherosclerosis Society Congress on May 25, with simultaneous publication in the New England Journal of Medicine, Eli Lilly reported Heart-2 Phase 1b data for VERVE-102. It's an in-vivo adenine base editor: a single IV infusion of a GalNAc-lipid nanoparticle carrying mRNA for the editor plus a guide RNA, which makes a permanent A-to-G change that inactivates the PCSK9 gene in liver cells. That's a precise DNA edit, not CRISPR nuclease cutting and not reversible RNA silencing, so "silences the gene" undersells it. Across 35 patients with familial hypercholesterolemia or premature coronary disease, the top 1.0 mg/kg dose cut PCSK9 by 88% and LDL cholesterol by 62%, with reductions described as durable out to 18 months. The headline win is safety. The predecessor, VERVE-101, caused Grade 3 liver-enzyme and platelet events on its old nanoparticle and was paused in 2024; the redesigned carrier showed no such signal here. VERVE-102 came to Lilly through its roughly $1B acquisition of Verve Therapeutics, whose founder Sekar Kathiresan is now a Lilly SVP, and Phase 2 is planned by year-end. The caveats are real: 35 patients is tiny and uncontrolled, the dose response is non-monotonic (the 0.7 mg/kg cohort managed only 33%), just 15 of 35 have a year-plus of follow-up, LDL is a surrogate with zero cardiovascular-outcome data, and off-target editing has only been checked preclinically. The 62% is comparable to existing PCSK9 injectables, but there's no head-to-head, and a permanent edit is a one-way door.