Issue #2

AnalysisAnthropic (the AI lab behind Claude) put out Opus 4.7 last week, leading SWE-bench Verified (a coding test that measures whether a model can fix real software bugs) and Aider polyglot benchmarks. Quietly held back from public release: Mythos, a roughly ten-trillion-parameter model Anthropic itself says can exploit decades-old vulnerabilities and run multi-step cyberattacks with minimal hand-holding. Only the White House and a handful of federal agencies are getting access. The split is the story. Anthropic gets to ship a polished assistant for paying customers and keep the dangerous one behind a curtain marked responsible disclosure. One model funds the company, the other proves the company is serious. Both arguments work better when nobody outside D.C. can audit Mythos directly.

AnalysisOpenAI committed more than twenty billion dollars over three years to Cerebras (a maker of giant single-wafer AI chips that competes with Nvidia), and threw in another billion to help Cerebras build data centers ahead of a planned summer IPO. In return, OpenAI gets an equity stake. The shape of it matters more than the dollar figure. Microsoft's Azure used to be the supply line; Nvidia is the bottleneck everyone is trying to widen; now OpenAI is buying into a chipmaker that does not yet ship at hyperscale (the volumes used by AWS, Google Cloud, and Azure). The bet is that paying ahead, on equity terms, locks in compute the open market will not have. Whether Cerebras can deliver on its promised volumes is the only question that matters, and the answer is months away.

AnalysisMozilla's for-profit arm released Thunderbolt last Thursday, an open-source AI client (free to download, modify, and run on your own machines) aimed at companies and developers who do not want their prompts living inside an OpenAI or Anthropic account. The tool routes between local models, self-hosted servers, and commercial APIs from a single interface. The play is the same one that worked for Firefox in 2004: arrive late, target the people allergic to the dominant vendor, lean on open source as the wedge. Thunderbolt will not knock ChatGPT off the consumer stack. It might give a regulated bank or a German hospital a way to use AI without handing every internal document to a US lab. Distribution to people who already trust Mozilla is the only moat that matters here.

AnalysisOpenAI introduced GPT-Rosalind last Thursday, a reasoning model trained on biochemistry, genomics, and the messy back end of pharma research. The pitch: knock years off the ten-to-fifteen-year drug discovery timeline. The reality: the model is gated to vetted labs and pharma partners, and Decrypt's blunt headline ("you probably can't use it") captures the tone. OpenAI is following Google DeepMind into specialized scientific models, where the buyers are Pfizer-sized customers signing six- and seven-figure annual contracts. The interesting part is what it signals about OpenAI's revenue mix. ChatGPT subscriptions hit a ceiling; pharma, defense, and finance verticals do not. Expect more domain-specific models with lab coats this quarter, and fewer demos that fit in a tweet.

AnalysisAllbirds, the wool-shoe company that IPO'd at four billion dollars in 2021, raised fifty million on Tuesday to gut the remaining business and become NewBird AI, a GPU rental service. The stock jumped roughly six hundred percent in two days off a market cap of twenty-two million. Shareholders will vote next month to strip the company's public-benefit-corporation status, formally ending its sustainable-footwear charter. There is a clean reading of this and a queasier one. The clean reading: a dying retailer found a buyer for its listing and is recycling capital where demand exists. The queasier one: any tradable shell with a SaaS-shaped backstory can now reprice itself by adding AI to the ticker. We have seen this movie. It ended in dot-Bob and pets.com.

AnalysisManycore Tech, a Hangzhou software company that pivoted to spatial intelligence training data for robots, debuted on the Hong Kong exchange on Wednesday and closed up 187 percent after raising $156 million in its IPO. The pop matters less than what it signals. Manycore is now priced at roughly $4.5 billion on the bet that the bottleneck for humanoid robots is not the hardware (BYD and Unitree have that) but the volumetric, physically grounded data needed to train them. There is no comparable US listing. The closest analog, Niantic's Pokémon Go-trained spinout with thirty billion images of streets and storefronts, sits inside a private company. Beijing built the public-markets pipeline first. Whether spatial AI is a real category or a narrative wrapper, China is the place pricing it.

AnalysisGoogle is in active talks with the Pentagon to deploy Gemini inside classified networks, the Information reported last week. This is the same Google that pulled out of Project Maven in 2018 after employee protests over Pentagon imagery work. Eight years and one capability gap later, the math has shifted. OpenAI signed with the Department of Defense in 2024, Anthropic followed with a CIA-flavored Mythos preview this month, and Google was the only frontier lab without a flag-saluting public reference customer. The deal would put Gemini in spaces where the answer cannot leak and the audit trail has to satisfy a clearance review, which is a much harder problem than the consumer chat product. Internal protests, if they come, will be smaller news than they were last time. The labor leverage moved.

AnalysisNature published results last week from a Stanford-led study that ran top AI agents (autonomous systems that string together model calls, tools, and browsers to complete multi-step tasks) against PhD researchers on real scientific workflows: literature review, dataset building, statistical analysis, the boring middle of science. The agents scored roughly half as well as the humans on tasks designed by domain experts to mirror real work, and failed in characteristic ways: hallucinated citations, broken pipelines, surface-level summaries that look credible to anyone who does not already know the answer. Vendors are selling these agents as PhD-equivalent at $200 to $20,000 per seat per year. The study is a useful pin in the marketing balloon, and a reminder that benchmark performance and end-to-end usefulness diverged sometime in 2025 and have not reconverged.

AnalysisAn Illinois bill working through committee would establish a liability framework for critical harm caused by AI systems, and the two leading labs have taken opposite sides in public. OpenAI is backing the bill, which sets a high bar for claims and limits exposure to developers who follow stated safety practices. Anthropic is opposing it, arguing the threshold is so high that real-world harms (privacy breaches, financial fraud, biased denials of service) will not qualify. This is the first time the two companies have publicly disagreed on a US state bill, and the disagreement is the story. OpenAI wants legal certainty before scale. Anthropic wants regulatory pressure that keeps competitors honest, including itself. Whichever version Springfield passes will be copied. State capitols are where AI liability law is actually being written.

AnalysisApple is sending Siri engineers to a multi-week internal bootcamp on AI-assisted coding tools, 9to5Mac and Bloomberg reported last week. The company that has spent two years explaining why it is taking its time on AI is now teaching its own staff how to use the productivity tools its competitors shipped in 2023. Two things are happening at once. Apple's overhauled Siri is overdue and the engineering org needs to accelerate. And the company that built its brand on you-don't-need-to-understand-the-technology is admitting its own engineers need training to keep up with what GitHub Copilot and Cursor users have been doing for two years. The catch-up cost shows up in salary, schedule, and the awkward visual of the world's most valuable consumer-tech company running remedial classes.

AnalysisOrbital, an Andreessen Horowitz-funded startup, set a launch date for the first test mission of a space-based AI data center, aiming to put compute hardware into low Earth orbit where solar power is constant and cooling is the vacuum itself. The pitch makes more sense than it sounds. Power, water, and grid permits are now the binding constraint on terrestrial AI buildouts, and waiting for new transmission lines takes a decade. Orbit dodges the line item entirely. The catches are launch cost, radiation hardening, repair logistics, and the question of who fixes a fried GPU 400 kilometers up. None of those are solved. But AI compute is energy-bound is the consensus view, and the moment that is true, every reasonable place to build a data center on Earth gets contested. Space is unreasonable, which is the point.

AnalysisThe EU AI Act's high-risk provisions for hiring tools enter enforcement on August 2 (about 105 days from now), and any company using algorithmic systems to screen, score, or rank candidates inside the bloc has to meet documentation, bias-audit, and human-oversight requirements or face fines up to seven percent of global revenue. US companies hiring into European subsidiaries are in scope. So are the resume-screening vendors most HR teams forgot they were using. The interesting bit is who is unprepared. Large enterprises started compliance work last summer. Mid-market employers and the hundreds of recruiting SaaS tools embedded in their workflows mostly did not. Brussels is about to discover how many off-the-shelf hiring tools cannot pass an audit, and a year of paused contracts is the realistic outcome for any vendor that cannot show its work.

AnalysisOn April 16 Anthropic, the AI lab behind Claude, shipped Opus 4.7 with image input up to 3.75 megapixels and unchanged pricing. A day later it released Claude Design, a research preview that turns prompts into slides, prototypes, and one-pagers. The interesting footnote is in Anthropic's own launch materials: an unreleased model codenamed Mythos already outscores Opus 4.7 on its internal benchmarks. Shipping the second-best model while holding back the first is the habit of a company managing pricing, compute capacity, and safety review, not one bursting with abundance. With Anthropic now running at a $30 billion annualized revenue pace driven largely by coding, the design release is about fencing off the slide-and-mockup corner of enterprise work before Figma, Canva, or OpenAI get to it.

AnalysisOne of OpenAI's own investors just compared it to Netscape, the 1990s browser pioneer later crushed by Microsoft and swallowed by AOL. The occasion: Anthropic's annualized revenue jumped from $9 billion at the end of 2025 to $30 billion by March, driven by enterprise coding demand, while OpenAI's $122 billion round at an $852 billion valuation looks harder to justify. Secondary-market buyers want Anthropic shares and are discounting OpenAI's. CFO Sarah Friar told the Financial Times the round itself was proof of confidence, which is the kind of argument CFOs make when the other arguments have stopped working. Justifying today's price tag assumes an IPO valuation north of $1.2 trillion. The loudest AI brand in consumer does not automatically win the enterprise wallet.

AnalysisSnap CEO Evan Spiegel announced 1,000 layoffs and 300 closed roles on April 16, roughly 16% of its global workforce, citing advances in artificial intelligence. The stock rose 11% pre-market. The footnote the press release buried: activist investor Irenic Capital had sent a letter pushing for 21% cuts, arguing AI can and should replace many existing roles. The AI framing is doing a lot of work here. Irenic wanted profitability; AI-generated code, now reportedly 65% of Snap's new code, supplies the story. U.S. staff get four months of severance. The company forecasts $1.5 billion in Q1 revenue and more than $500 million in annualized savings by H2 2026. This is the Block and Uber playbook, and it is becoming a Q2 earnings-season template.

AnalysisThe Nebraska Supreme Court suspended Omaha attorney Greg Lake on April 15 after 57 of 63 citations in an appellate brief turned out to be defective, including 20 outright fabrications: cases that do not exist and quotes from statutes that do not exist. Lake repeatedly denied using AI; the court said his explanation lacks credibility. The suspension is indefinite. This is the most severe professional sanction yet for AI-hallucinated filings in the U.S., and it lands as courts imposed at least $145,000 in AI-citation sanctions in Q1 2026 alone. The norm forming under all this: AI drafts are fine, AI as the final word gets you disbarred. Courts have stopped debating whether AI belongs in legal practice. They are deciding how expensive it is to skip the checking step.

AnalysisThe EU AI Act's August 2 deadline for high-risk AI systems is 105 days out, and regulators have now published the specific audit scope and documentation requirements for two categories: resume-screening tools under Annex III, and AI agents logging actions under Article 12. For a 100-person company running AI-based hiring, compliance now means a written risk assessment, a paper trail on training data, a human-review step for rejected candidates, and a signed engagement with a certified auditor (of which there are not yet enough in Europe). The agent-logging rule is the quietly heavier lift, since no finalized technical standard exists yet; the draft frameworks are still in review. Most vendors cannot produce Article 12-compliant logs today. Companies that started AI hiring in the last year have a paperwork problem they did not budget for.

AnalysisOn April 16 Siemens and Humanoid, a startup building industrial robots, announced that the HMND 01 Alpha, a wheeled humanoid running on Nvidia's physical-AI software stack, had successfully completed autonomous logistics tasks at Siemens' electronics factory in Erlangen, Germany. The pitch is familiar; the detail is that the robot actually ran in a live production environment, not a stage demo. Industrial humanoids have spent three years on trade-show floors and investor decks. Siemens operating one inside its own plant is the first credible signal that Nvidia's physical-AI framing is moving from Jensen Huang keynote to purchase order. The tasks most exposed are the easiest parts of warehouse and factory logistics, the ones that were already hardest to staff. Which is, mechanically, why this is moving first.

AnalysisFlorida Attorney General James Uthmeier opened an investigation into OpenAI on April 10, citing national security and ChatGPT's alleged role in the April 2025 Florida State University shooting that killed two people. Court documents show the suspect exchanged more than 200 messages with ChatGPT before the attack, including asking how the country would react and what time the student union was busiest. OpenAI said it will cooperate. This is the first state AG probe of a frontier AI lab tied to a specific violent crime. It also arrives a month after OpenAI's $122 billion raise and as the company is reportedly preparing for an IPO. The legal theory is still taking shape. The political appetite is obvious. Expect more filings like this, not fewer.

AnalysisIn the same two weeks OpenAI closed its $122 billion round, it quietly bought personal-finance startup Hiro and media podcast TBPN, while Ronan Farrow and Andrew Marantz published a 16,000-word New Yorker piece documenting a two-decade pattern of alleged deception by Sam Altman. A man was arrested at OpenAI's San Francisco headquarters threatening to burn the building down. Altman said in a late-night post that he had underestimated the power of words. TBPN buys friendly media; Hiro tries to stretch ChatGPT into more hooks than a chatbot. Both are acquihires that also solve PR and product-depth problems the company cannot solve organically. A near-trillion-dollar narrative is fragile in a way a near-billion-dollar one was not. Every news cycle now costs OpenAI enterprise trust it used to have by default.

AnalysisAt the Morgan Stanley tech conference Wednesday, Nvidia CEO Jensen Huang said his company's recent stakes in OpenAI and Anthropic are likely its last. Nvidia's pledged $100 billion OpenAI investment last year shrank to $30 billion in the final round, and an MIT Sloan professor publicly called the structure kind of a wash, since OpenAI was turning around to buy Nvidia chips with the same money. Circular financing is the polite term. Regulators use less polite ones when a bubble pops. Huang's stated reason is that IPOs close the window. The real reason is probably that Nvidia's income statement no longer needs the equity upside from customers it already sells every chip it can make. Pulling back now also insulates the balance sheet from a writedown if either lab wobbles.

AnalysisMastercard partnered this week with Crossmint's Lobster.cash to let AI agents (software that takes actions on a user's behalf, including making purchases) run real card transactions using real consumer Mastercards. The initial rollout is early-access via OpenClaw. This closes a gap agent builders have complained about for a year: until now, an agent that wanted to buy something had to use a prepaid wallet, a crypto rail, or a developer's own card and a lot of hope. A mainstream payments network green-lighting agent transactions is a soft but real threshold. The implicit rules (who is liable when an agent buys the wrong thing, and how chargebacks work when the customer is a language model) are being written in contracts rather than in regulation. The lawsuits will follow.

AnalysisGoogle released Gemma 4, its latest family of open-weight models (free to download, run, modify, and ship in commercial products) under the Apache 2.0 license, with 2B, 4B, 26B, and 31B parameter sizes. The pitch is agentic workflows and video understanding on modest hardware. The Gemmaverse, Google's name for community forks of its open models, has more than 400 million downloads and over 100,000 variants. The calculation is clear. Giving away capable models commoditizes the rest of the stack, which is exactly where Google Cloud (a hyperscaler competing with AWS and Azure for enterprise workloads) wants the fight. Meta's Llama has spent a year looking adrift, and Gemma 4 under a genuinely permissive license squeezes both sides of the open-versus-closed pitch. The free tier is the weapon, again.

AnalysisShield AI, the San Diego startup behind the Hivemind autonomous-pilot software, raised $1.5 billion in Series G at a $12.7 billion valuation, part of a broader $2.25 billion package. The valuation roughly doubled in nine months. Hivemind was selected for the U.S. Air Force's Collaborative Combat Aircraft program, the first time mission-autonomy software has been decoupled from the aircraft itself. The raise also funds Shield AI's acquisition of Aechelon Technology and its next-generation X-BAT jet fighter drone. Founders Fund (Peter Thiel's firm), Blackbird, DCVC, BOND, and Bessemer Venture Partners are in. Projected 2026 revenue is $540 million, up 80% year on year. The broader pattern: defense-AI rounds are now Silicon Valley's safest capital, with the Pentagon as an anchor customer and Congress in no mood to cut AI-for-warfare budgets.