The Financial Times reported Sunday that AstraZeneca and Bristol Myers Squibb held preliminary discussions over the last few months about a combination worth nearly $400 billion; Reuters confirmed it through a person familiar with the situation and Bloomberg described the talks as early stage. Neither side would say whether anything is still live. AstraZeneca declined to comment. On the numbers alone it would be the largest pharmaceutical merger ever attempted, joining a roughly $264 billion company to a $133 billion one. London's verdict arrived within hours: AstraZeneca shares fell 6.7 percent to 11,804 pence on Monday while the FTSE 100 barely moved, meaning a single stock accounted for the index's entire drag. Jefferies analysts wrote that they were "perplexed," adding that "if there is one company that doesn't need financial engineering, it's AZ." The portfolio manager Markus Manns said the combination "does not make strategic or financial sense."

The logic on the Bristol Myers side is easier to state and harder to solve. Revlimid has already lost exclusivity and Opdivo and Eliquis could follow by 2028, a patent cliffConceptThe revenue drop when a blockbuster loses exclusivity and generics enter. It is the standard explanation for why a $133 billion company entertains being absorbed rather than defending its independence. steep enough to make absorption an option even after a Q2 beat and raised guidance. The obstacles are structural. Cancer treatments were about $25 billion of AstraZeneca's 2025 sales and more than forty percent of Bristol Myers' first-half revenue, and the two companies' immunotherapies compete directly. Andre Barlow of DBM Law Group expects "a Trump FTC to scrutinize the merger," requiring "meaningful divestitures" where late-stage pipelines overlap, and points to the Otezla divestitureConceptThe $13.4 billion forced sale of Celgene's psoriasis drug to Amgen as the price of FTC clearance for Bristol's 2019 Celgene deal. It is the template every pharma antitrust lawyer reaches for when overlapping franchises collide. that cleared Bristol's own Celgene deal in 2019. Then there is domicile. Pascal SoriotPersonAstraZeneca's chief executive since 2012, who rebuilt the company around oncology and rare disease and personally led the 2014 defence against Pfizer's hostile approach. The stock has more than quadrupled under him. has run AstraZeneca for fourteen years, saw off Pfizer roughly a dozen years ago, and last year announced a direct US listingConceptMoving a foreign company's shares to trade natively on a US exchange rather than through depositary receipts, to capture American multiples and index inclusion. AstraZeneca announced plans for one last year.. A British company buying an American pharmaceutical champion is a proposition this administration has spent a year arguing against.

Why it mattersEvery element a capital-markets associate ends up drafting around is present at once: an oncology overlap that guarantees a second request, a cross-border domicile question in an onshoring administration, and a target whose case for selling is a patent calendar. The shareholder reaction is the tell, because a bidder whose own register hates the idea rarely gets to make an offer.

Sony and Marvel Studios' Spider-Man: Brand New Day took an estimated $355.0 million domestically across 4,487 theatres, the second-largest opening ever and about $2.1 million short of Avengers: Endgame's $357.1 million from April 2019. It is the biggest opening in Sony's distribution history. Overseas it added $572 million from 66 territories and 73,500 screens for a $927 million global debut, second only to Endgame's $1.22 billion. China led all foreign markets at $121.0 million, reported as the biggest Chinese opening for a Marvel title since the pandemic and a real signal in a market where the import quotaConceptBeijing admits a capped number of foreign films for revenue-sharing release each year and controls their dates. Studios typically keep about twenty-five percent of Chinese gross against roughly fifty percent elsewhere, so the headline overstates what Sony banks. has functioned as an instrument of trade policy since 2018. The United Kingdom took $49.0 million, its third-biggest opening ever, and Mexico $38.3 million, its second. CinemaScoreConceptExit polling of opening-night audiences, reduced to a single letter grade. An A predicts a shallow second-weekend drop, which is what turns a record opening into a record run. graded it an A.

The frameConceptTrade shorthand for a single Friday-to-Sunday weekend counted across every film in release, rather than one title's gross. It is how exhibition measures the health of the business, and how theatre chains' lenders measure it too. around it was the story exhibitors will remember. Deadline puts the all-films North American weekend at $428.99 million, an all-time record; NBC News and wire coverage put it at $434 million, and both numbers were circulating Monday morning. Either way it clears the previous best, the $402 million Endgame weekend of April 2019, and it lifts 2026 domestic box office past $6 billion. Universal crossed $4 billion worldwide on Sunday, the first studio to do so in 2026 and only the fifth time any studio has, carried by Christopher Nolan's The Odyssey at $911.4 million in its third weekend and closing on a billion. Every figure here is a Sunday estimate tabulated by ComscoreInstitutionThe measurement firm whose box-office division supplies the ticketing data every trade cites on Sunday. Studio "estimates" are its projections; Monday actuals routinely move them by more than the $2.1 million separating this opening from the record., and the gap to the record is well inside the margin by which those estimates move.

Why it mattersA $927 million global debut against a reported $225 million negative and a $309 million promotional campaign is a reminder that the marketing line on a tentpole now rivals the production line, which is where participation and backend disputes are actually born. The China number matters more than its dollars: it is the first evidence in years that the quota window can still be opened for an American franchise.

Alibaba released Qwen3.8-Max on Sunday evening, a mixture-of-expertsConceptAn architecture that routes each token to a small subset of specialist sub-networks. Qwen3.8-Max carries 2.4 trillion total parameters but activates only about 95 billion per token, which is why a model this large can be served cheaply. model with 2.4 trillion total parameters, roughly 95 billion active, a one-million-token context window and the first multimodal input of any Qwen model above a trillion parameters. The announcement that matters is the second one: Alibaba says it will publish the weightsConceptOpen weights means publishing a trained model's parameters for download so anyone can run, fine-tune or inspect it. It is distinct from open source, since training data and code usually stay private and the licence can still restrict use. of both Qwen3.8-Max and a small Qwen3.8-27B next week, the first time it has open-weighted a Max-class flagship after a run of proprietary top-tier releases. Its own benchmarks put the model above Moonshot AI's Kimi K3InstitutionBeijing lab whose K3 was the first open-weight model above a trillion parameters. Alibaba's benchmark deck is aimed squarely at it, and conspicuously does not run a head-to-head on price. on several evaluations and comparable to or better than Anthropic's Fable 5. The coding-agent shop Cline measured a Terminal-Bench score two percent above Fable 5, and the model entered the Frontend Code Arena at fourth with a score of 1,668.

Two cautions travel with it. The licence does not exist yet, and a Max-class release is exactly where a company that has previously shipped Apache 2.0 might start bolting on usage restrictions. And at least one preview user found the model fast, "owing to the linear attention and sparsity," but "definitely not competitive with K3 in my experience. Much less thorough." What is not in dispute is the serving economics, roughly three times cheaper than Kimi K3 by independent observation, or the timing. It lands days after DeepSeek's V4-Flash and shortly after Kimi K3, and it lands the same weekend the Wall Street Journal reported that American open-weight startups including Arcee, Reflection AI and Poolside are running on thin capital because venture investors question the revenue model. The 235 companies that signed Microsoft's July "Open Weights and American AI Leadership" letter, including Nvidia, Amazon and the Linux Foundation, defended distillationConceptTraining one model on another model's outputs. Microsoft's letter defended it as "a long tradition of learning from, building upon, and improving existing technologies"; Anthropic, which did not sign, called for a crackdown on industrial-scale distillation. in exactly those terms. Anthropic did not sign.

Why it mattersThe American argument for open weights is a national-security argument about avoiding single points of failure. The Chinese argument is a pricing argument, and pricing is winning: a frontier-class model given away and served at a third of its nearest rival's cost sets the floor that every closed lab's enterprise contract has to clear.

Two preprintsConceptA paper posted publicly before peer review, timestamped on submission. Here the timestamps are the story: 10:35 a.m. Pacific, and three hours and eighteen minutes later. resolving the same open problem in unclonable encryptionConceptA quantum scheme in which a ciphertext cannot be split into two copies that both decrypt once the key is revealed, because an unknown quantum state cannot be perfectly duplicated. Prior constructions were inefficient or assumption-dependent. reached arXiv on the same day. Prabhanjan Ananth of UC Santa Barbara and Amit Sahai of UCLA filed at 10:35 a.m. Pacific; Seyoon Ragavan, an MIT doctoral student, filed three hours and eighteen minutes later. Both papers credit OpenAI's GPT-5.6 Sol Ultra with finding the core ideas. Both groups had heard the same question posed earlier that month at the Simons InstituteInstitutionThe UC Berkeley research centre for theoretical computer science whose programme talks function as the field's agenda-setting mechanism. The open question that seeded both papers was posed there. at Berkeley, and neither knew about the other until Yao-Ting Lin, a UCSB doctoral student, noticed the collision before either paper was public and emailed Ragavan: "We are definitely living in strange times."

The methods differed and the result did not. Ragavan drove Sol Ultra, which OpenAI says coordinates four agents in parallel by default, in two-hour stretches, checking and redirecting at each interval and then cleaning up the draft. AnanthPersonCryptography professor at UC Santa Barbara and co-author of one of the two papers, and the author of the line that the field's reflex is now to hand any new open problem to the model first. and Sahai used a bespoke UCLA harness built to have models pursue and critique candidate solutions; their paper says the model produced the construction and the main proof ideas, and that the authors take responsibility for the claims. Ananth described the new default: "if someone mentions an open problem, the first thing is to see if GPT solves it. A lot of problems that we didn't know how to solve are going to get solved in the very near future." Ragavan, who had failed at the same problem himself a couple of years ago, said the compressed timeline was the disorienting part: "It's like two weeks and a day since this idea even formed." Neither paper is peer reviewed, and the groups have discussed merging them for conference submission.

Why it mattersPriority disputes are the oldest governance problem in science and they have always been resolved by the calendar. When two independent groups can reach the same construction in an afternoon by querying the same system, the calendar stops discriminating, and every institution that allocates credit by first-filing, from tenure committees to patent offices, inherits the problem.

The Information reported Sunday evening that four states have rolled back or paused data-centre tax incentives and nine more are weighing repeal, a shift that could add seven percent or more to equipment costs and, in its estimate, several billion dollars to the price of every gigawatt of AI computing. The arithmetic is simple and brutal. In a state with a seven percent sales tax, killing the sales and use tax exemptionConceptThe core data-centre subsidy: a state waives sales tax on servers, networking gear and sometimes electricity. Because AI buildouts are equipment-heavy, the exemption is worth billions per gigawatt. moves the IT-equipment cost of a gigawatt-classConceptA campus drawing about a billion watts, the unit frontier labs now plan in. At roughly $40 billion of IT equipment per gigawatt, a seven-point tax swing is a three-billion-dollar line item. campus from about $40 billion to about $43 billion. Per the National Conference of State LegislaturesInstitutionBipartisan organisation serving all fifty state legislatures. Its tracking is the standard reference for how many states have live data-centre incentive bills., thirteen or fourteen states including Arizona, Connecticut, Georgia, Maryland, Michigan, Pennsylvania, Virginia and Texas have abolished, suspended or plan to suspend the exemption. Thirty-eight states currently offer some form of incentive and roughly two dozen considered rollback legislation this year.

The forms differ and the direction does not. Arizona enacted a three-year moratoriumConceptHere, a legislated pause rather than a repeal: Arizona suspended its exemption for three fiscal years, preserving the statute while denying the benefit. It is the compromise form most states are converging on. running from July 1, 2026 to June 30, 2029. North Carolina repealed the exemption on electricity purchases while preserving the equipment and capital-investment breaks, which targets the utility bill rather than the build. The driver is electoral rather than fiscal: backlash over electricity prices, water use, noise and disappointing tax receipts, with governors of both parties moving, and one person close to the shift observing that repeal "would've been unthinkable three years ago." The pressure is not American only. The Financial Times reported Sunday that London, Europe's largest data-centre hub, has more than twenty sites in development while Amsterdam has paused new construction until 2030, and that Irish households paid an estimated 715 million euros more for electricity between 2015 and 2023 because of data-centre expansion, with a further 1.43 billion possible by 2034.

Why it mattersEvery hyperscaler capex guide of the last two quarters was built on a subsidy the states are now withdrawing, and unlike power contracts or chip supply the exemption can vanish in a single legislative session. Watch the siting incentives in the announcements: a company that names its state before it names its gigawatts is telling you it has priced the risk.

At 8:00 a.m. London time today, Flutter Entertainment's admission to the London Stock Exchange's main market is cancelled and the shares trade solely on the New York Stock Exchange under FLUT. Friday was the last day of London dealing. The mechanics ran through the Financial Conduct Authority's Official ListInstitutionThe UK register of securities admitted to listing. Cancellation from it, not from the exchange, is the legally operative step in a delisting; the exchange's trading admission follows. rather than the exchange itself, with Listing Rule 21.2.17ConceptThe FCA provision requiring at least twenty business days' notice before a company cancels its listing. It is the minimum runway UK retail shareholders get to reposition. requiring at least twenty business days' notice, given on June 12. Flutter announced the review on May 7 alongside first-quarter results and gave two reasons: the low level of London trading activity, and the cost and regulatory burden of maintaining a second line.

The company leaving is not small. Flutter reported $16.38 billion of global revenue for 2025, up seventeen percent, and $4.30 billion in the quarter ended March 31, across FanDuelInstitutionFlutter's US sportsbook and the reason the group's centre of gravity crossed the Atlantic. It is the primary driver of revenue growth and of investor interest, and the commercial logic behind abandoning London., Sky Betting & Gaming, Sportsbet, PokerStars, Paddy Power, Sisal, Snai, Betfair and a dozen more. It already files a Form 10-K with the SEC, having converted to full US domestic-issuer reporting before dropping the London line. The friction, as ever, falls on retail: holders of Computershare depositary interestsConceptA UK instrument that lets investors hold a foreign-incorporated company's shares through CREST settlement. When the London line closes, these holders must convert or transfer or risk being stranded. were directed to a helpline to convert or transfer before cancellation. A dual listingConceptHaving the same ordinary shares admitted on two exchanges. The cost is duplicative regulators, filings and governance codes; the benefit is liquidity, and when liquidity migrates to one venue the second listing becomes pure overhead. is only worth its overhead while both venues supply liquidity.

Why it mattersThis is the migration argument stripped of rhetoric: not a domicile flip, not a tax play, just a company observing that its shares had stopped trading in one of its two markets. Every UK listing-reform proposal of the last three years was written to prevent exactly this filing.

Section 5 of Chapter IX of the EU AI Act became applicable on Sunday, and with it the AI OfficeInstitutionThe European Commission unit created to supervise general-purpose AI models across the bloc. As of August 2 it is an operating enforcement body rather than a policy shop. stopped being a body that engages informally with model providers and became one that compels. Article 91 lets it require any general-purpose AI modelConceptUnder Article 3, a model trained on broad data at scale showing significant generality and adaptable to many downstream tasks. Foundation models sit here; the systems built on top of them are regulated separately. provider to hand over technical documentation, training-data summaries and compliance materials, with incomplete or misleading responses themselves finable. Article 92 lets it commission evaluations by the Scientific Panel and request access to a model through APIs or other technical means, including source code. Article 93 lets it compel mitigation or, at the extreme, restrict, withdraw or recall a model from the EU market, and its third paragraph imports the commitment decisionConceptA device borrowed from EU competition law. A company offers binding remedies, the regulator accepts them and closes the file without ever finding an infringement. Article 93(3) brings it into AI supervision. from competition practice. Article 101 sets the ceiling at fifteen million euros or three percent of total worldwide annual turnover, whichever is higher.

The same date switched on the Article 50ConceptThe AI Act's transparency provision: users must be told they are interacting with AI, deepfakes must be labelled, and synthetic content must carry machine-readable marks so it can be detected. transparency duties. Chatbots operating in Europe must now disclose that they are not human, deepfakes must be labelled, and AI-generated or altered content must carry machine-readable marks. The Commission published a list of more than 180 organisations that have signed the Code of Practice on transparency of AI-generated content, opened a complaints channel for downstream providers, and issued guidelines. Enforcement is shared between national market surveillance authorities, the AI Office, and the European Data Protection Supervisor where an EU institution is itself the provider. The point most commonly got wrong is what the Digital OmnibusConceptThe Commission's simplification package, which pushes the Act's heaviest high-risk obligations to December 2027 and August 2028. It was widely misread as a blanket delay; it leaves the prohibitions, the transparency duties and the GPAI enforcement powers untouched. did: it moved the Annex III high-risk deadlines to December 2, 2027 and August 2, 2028, and it did not move any of this.

Why it mattersThe design lesson is jurisdictional. Fines are calculated on global turnover and imposed by the Commission directly rather than by twenty-seven national regulators, which is the deliberate structural fix for the fragmentation that made GDPR enforcement slow. The first real tell will be whether the AI Office issues Article 91 requests to every Code signatory as a baseline, or waits for a failure.

Solicitor General D. John Sauer applied to the Supreme Court on July 27 for an administrative stayConceptA very short pause a court grants purely to preserve the status quo while it reads the papers. It carries no view on the merits, which is why refusing even that is read as a signal. and a full stay of the order blocking the March 31 executive order "Ensuring Citizenship Verification and Integrity in Federal Elections," which directs the Postal Service to build voter lists and deliver mail-in ballots only to people on them. Justice Ketanji Brown Jackson, the Circuit JusticeConceptEach Justice is assigned one or more federal circuits and handles emergency applications arising there in the first instance. Jackson covers the First Circuit and may act alone or refer a matter to the full Court. for the First Circuit, denied the immediate administrative stay within hours and ordered briefing instead, with responses due today. Below, Judge Indira TalwaniPersonObama appointee to the US District Court for the District of Massachusetts, confirmed in 2014, a former union-side labour lawyer. Her June 25 order is the one the administration is trying to stay. had barred the mail-ballot and state-citizenship-list provisions for the November 3 elections in twenty-three states and the District of Columbia, which argue the order violates the Elections ClauseConceptArticle I, Section 4 gives state legislatures power to set the times, places and manner of federal elections, subject to congressional override. It names no role for the President.. A divided First Circuit panel refused to disturb her order, warning that letting the policy operate this autumn "would sow confusion and threaten disenfranchisement of many eligible voters."

The reason the application has a chance is a conflicting posture rather than a conflicting holding. A D.C. Circuit panel declined to block the same executive order, affirming Judge Carl Nichols's May ruling that an injunction was premature. That is a ripenessConceptThe doctrine that a court should not decide a challenge until the injury is concrete rather than speculative. Judge Nichols used it to avoid the merits in Washington, producing the split posture the Solicitor General now exploits. holding, not a merits one, and the Solicitor General is asking the Court to treat two circuits reaching opposite practical results as a conflict requiring resolution. With briefing closing today the full Court can act at any time; the next scheduled Summer Order List is August 17, but emergency-docket rulings do not wait for order lists.

Why it mattersWatch the procedural move rather than the politics. A Circuit Justice refusing even a temporary hold and setting a briefing schedule is a signal about the merits delivered entirely through calendar management, and it is the way most of the consequential constitutional law of the last two years has actually been made.

Volodymyr Zelensky confirmed personally on Sunday that Ukrainian forces struck two strategic facilities in Saratov simultaneously overnight, more than 600 kilometres behind the line: Rosneft'sInstitutionRussia's largest oil producer, majority state-owned and under Western sanctions. Its refining assets have become Ukraine's preferred deep-strike target because damaged units cut fuel supply and export revenue at once. Saratov oil refinery and the Engels-2PlaceRussia's principal strategic aviation base, near Saratov on the Volga. It hosts the Tu-95MS and Tu-160 bombers that carry the Kh-101 cruise missiles used in mass strikes on Ukrainian cities. strategic aviation base. The General Staff formally confirmed both, plus an oil depot at Lyudinovo in Kaluga Oblast and a drone storage and launch facility in Bryansk. The Saratov plant refines roughly seven million tonnes of crude a year and is among the largest in the Volga region; Ukrainian reporting says the strike hit its ELOU-AVT-6ConceptThe Russian designation for a combined desalting and atmospheric-vacuum crude distillation unit, the first and most essential stage of a refinery. Taking one out halts the whole train, and replacement components are sanctioned. crude distillation unit, rated at six million tonnes. A major fire was reported.

Two things distinguish this from the running campaign against Russian refining. The first is the pairing: hitting a refinery and the base that flies the Tu-95MS and Tu-160 in one coordinated night is a targeting decision about the platform, not just the fuel. The second is reach. Separately on Sunday morning the Bashneft refinery at Ufa in Bashkortostan burned after a long-range drone attack more than 1,700 kilometres from the Ukrainian border, and Ukraine has not claimed it. The context is the barrage Ukraine absorbed the night before: 35 missiles, mostly ballistic, and nearly 190 drones, killing at least nine including four children, with only one of 27 ballistic missiles intercepted. Trump has since walked back his pledge to license Ukrainian domestic production of PatriotConceptThe PAC-3 interceptor is among the only Western systems that can engage Russian ballistic missiles in the terminal phase. An interception rate of one in twenty-seven is the metric driving Ukraine's demand for domestic production rights. interceptors, telling the Financial Times the US has "not agreed to" it and that a recipient country could "someday turn on you."

Why it mattersAn interception rate of one in twenty-seven against ballistic missiles explains the target selection better than any doctrine statement: if you cannot stop the missile, you go after the aircraft and the fuel that move it. That logic points at deeper strikes, not fewer, whatever happens at the negotiating table.

Hadestown: The Musical grossed $2 million in its second weekend on 1,567 screens for a cumulative $18.2 million, passing Hamilton to become the highest-grossing theatrical release of a live captureConceptA multi-camera film of a stage production performed live, distributed theatrically or streamed. Distinct from an adaptation: no reshoots, no screenplay, the stage blocking intact, and budgets in the low seven figures against eight-figure film versions. ever made. It opened to $10.2 million at 1,949 locations, itself a record, edging Hamilton's $10.1 million from September 2025. Disney's Hamilton release topped out at roughly $17 million domestic and about $4 million overseas. The capture was filmed from the West End production with Reeve Carney as Orpheus, André De ShieldsPersonVeteran stage actor whose Hermes in Hadestown won the 2019 Tony for Best Featured Actor in a Musical, a career running back to the original production of The Wiz in 1975. as Hermes, Amber Gray as Persephone, Eva Noblezada as Eurydice and Patrick Page as Hades. The run has been extended and dates are being added.

The distributor is the part worth noting. Bleecker StreetInstitutionThe independent New York distributor founded by Andrew Karpen in 2014, which acquired and released Hadestown: The Musical and now holds the live-capture box-office record over Disney., an independent founded in 2014, now holds a record it took from Disney. A proshot clearing $18 million theatrically on a capture budget rewrites the arithmetic of stage rights licensingConceptThe bundle of rights a show's authors and producers control, including the separate and separately negotiated right to film a performance. Capture rights were historically withheld to protect live ticket sales., because it undercuts the premise on which capture rights were withheld for decades, and it competes directly with the assumption that a Broadway property's screen value lies in a conventional film adaptation costing twenty times as much.

Why it mattersTwo data points make a trend an argument. Producers negotiating capture rights now have a second comparable showing that filming the show does not cannibalise the show, and the party with the most to lose from that argument is the studio that used to buy the adaptation rights instead.