Friday Briefing — May 22, 2026

Newsletter Cover image:

🎉 Good morning. SpaceX just opened the books on rockets, satellites, AI, Mars, and a public-market appetite that apparently still has room for a little madness.

Today we have the SpaceX IPO filing, AMD shipping the next chip race, Waymo discovering water is still undefeated, a developer-tool breach with teeth, an AI policy reversal, and a science story that makes quantum physics feel slightly less like a prank.

Let's get into it.

Listen on Apple Podcasts or Spotify.

SpaceX filed its public S-1 on Wednesday, May 20, which means the rocket company that trained a generation of founders to say "vertical integration" with a straight face is finally showing Wall Street more of the receipts. The filing says SpaceX completed about 650 orbital launches by March 31, 2026, operates roughly 9,600 Starlink broadband and mobile satellites, and now treats xAI as an integral pillar after acquiring it earlier this year. The official SEC filing reads like a rocket company, a satellite carrier, and an AI infrastructure company all got locked in a conference room with a Mars poster.

The financials are the hook. SpaceX reported about $18.7 billion in 2025 revenue and an operating loss of roughly $2.6 billion, while the filing left the actual IPO pricing blank as early filings do. Various reports put the potential raise around $75 billion and the valuation discussion as high as $1.75 trillion, which would put this thing in the "largest IPO ever" conversation before the roadshow even starts, per AP and The Guardian.

The strangest part is that rockets may no longer be the entire pitch. SpaceX is bundling launch dominance, Starlink distribution, AI infrastructure, Grok risk disclosures, space-based compute dreams, and the civilizational Mars stuff into one public-market object. Its filing even claims a giant AI-heavy total addressable market. Translation: buy the rockets, stay for the satellites, maybe let Wall Street fund the weirdest data-center adjacency on the planet.

Here is why this matters beyond Musk theater. IPOs reset what investors will tolerate, what competitors can raise against, and what categories get a fresh narrative halo. If SpaceX sells public investors on "space plus AI plus national infrastructure" at mega-scale, every late-stage company with a defense, compute, energy, robotics, or orbital angle is going to redraw its deck by Monday.

AMD says its next-generation EPYC "Venice" server chips are already in production on TSMC's 2-nanometer process, and the company is pitching the move as an early claim on the next era of data-center performance and efficiency. The company said the platform will reach up to 256 cores and targeted a major jump in performance per watt as AI, cloud, and enterprise workloads keep demanding denser compute, per AMD and The Register.

That matters because the AI boom is not only about GPUs. CPUs still coordinate the mess. They feed the data center, host the enterprise workloads nobody brags about on stage, and get judged by power bills that already look like national infrastructure. When a chip company can squeeze more useful work into less power, that is not a nerd trophy. That is gross margin, rack density, procurement leverage, and a better story for every cloud customer trying to fit another workload into the same energy budget.

Why it matters: Nvidia owns the loudest part of the AI infrastructure conversation. AMD is making sure the quieter server layer does not become an afterthought, because "the model is amazing" loses its charm when the rest of the stack is slow, hot, or expensive as hell.

Waymo paused service in four U.S. cities after severe weather and flooding exposed an unglamorous robotaxi edge case: some vehicles kept entering waterlogged streets that humans already knew were trouble. The company said the pause covered Austin, Atlanta, Phoenix, and San Francisco while it worked through the response, and the episode landed just after U.S. regulators opened a review tied to vehicles behaving badly around flooded roads, according to CNBC and NHTSA.

This is funny right up until you remember a vehicle is a two-ton decision engine with tires. Autonomy demos usually happen in bright sunlight with crisp lane lines and a product manager nearby. Real cities deliver standing water, trash bags, construction tape, hand signals, parking-lot chaos, and one guy in a reflective vest improvising traffic law with his palms.

Why it matters: Physical AI has a much meaner eval suite than software. A chatbot can be wrong in a paragraph. A robotaxi can be wrong at curb speed in a thunderstorm while the insurance team learns a new word.

Security researchers say a malicious update to the popular Nx Console developer extension spread through a compromised maintainer environment and exposed a supply-chain path into thousands of repositories and developer machines. The coverage is still moving, but the core lesson is already familiar: when attackers get into the tools developers trust, they do not need to knock on every customer door one by one, per SecurityWeek and the Nx incident update.

This is the kind of breach normal companies underreact to because it feels "inside baseball." It is not. Developer tools sit near source code, tokens, package managers, cloud credentials, and the automation scripts that ship production changes while everyone else is in meetings. A compromised extension can turn a laptop into a distribution channel.

Why it matters: The new security perimeter is not just your app. It is the whole build path, from editor extension to dependency registry to CI runner. If the tools that make engineers faster also make attackers faster, speed without provenance is just a prettier liability.

The Trump administration was reportedly preparing an executive order that would have asked frontier AI companies to share certain advanced models with the federal government before public release. Then, late Thursday, the order was postponed amid internal disagreement over how much review belongs in the launch path for the most capable systems, according to Axios and Semafor.

That sounds like Washington process sludge. It is actually the center of the AI business argument right now. Labs want speed because speed compounds. Government wants a look before models become useful for cyber offense, critical infrastructure attacks, or national-security surprises. Everybody says they want responsible innovation. Nobody agrees who gets to hold the clipboard right before launch.

Why it matters: AI regulation is moving from philosophy to workflow. What gets logged, who signs off, and which evals count are about to matter as much as the keynote.

Researchers say they have developed a way to observe complicated quantum dynamics without leaning on the same assumptions that made some earlier "quantum advantage" claims hard to verify. The work focuses on extracting useful information from interacting quantum systems more efficiently, which could matter for simulation, materials research, and future quantum benchmarking, according to EurekAlert and the underlying Science paper.

The newsletter translation is simple: quantum computing is still in the awkward phase where hype, genuine progress, and painfully hard measurement problems all share a conference badge. Better measurement tools do not sound sexy. They are how a field stops arguing by press release and starts proving what is actually happening.

Why it matters: Science businesses get interesting when the instrumentation gets better. The shovel is often the breakthrough before the gold rush notices it.

The European Commission opened consultation on guidance for high-risk AI obligations under the EU AI Act. The regulation phase is now getting translated into forms, controls, and deadlines, which is where consultants and compliance software start licking their chops. European Commission

Nvidia reported $81.6 billion in quarterly revenue and said data-center demand kept ripping as Blackwell shipments scaled. That is not a quarter. That is a scoreboard hung over the entire AI capex trade. Nvidia

NASA researchers say lightning on Jupiter may be generated deeper in water-rich storms than expected. Earth has thunderstorms. Jupiter has weather that sounds like a metal album. Phys.org

CISA's latest exploited-vulnerability activity remains a reminder that security teams do not get to wait for every incident to become a headline before patching. The boring weekly hygiene list is often where tomorrow's emergency gets defused early. CISA

AMD also surfaced fresh Ryzen AI Max+ 400 chatter for high-memory local AI PCs. The market keeps asking the same question in different hardware shapes: what can run near you instead of renting a data center every time? NotebookCheck

Markets are still treating chip news like macro news. Reuters noted tech futures, Treasury yields, and AI earnings expectations all tugging on one another this week, which is a polite way of saying Wall Street has tied a lot of feelings to server racks. Reuters

Nvidia reporting $81.6 billion in quarterly revenue is not merely a chip-company flex. It is the clearest signal that "AI infrastructure" has become a market mood, a capex plan, a power-grid argument, a cloud sales pitch, and half of Wall Street's personality before breakfast.

That makes the company more important and weirder. A normal winner can post a monster quarter and enjoy the applause. Nvidia posts one and immediately resets expectations for data centers, semiconductor suppliers, utilities, startups, sovereign AI projects, and every investor asking whether the boom still has legs. The scoreboard is enormous now. So is the pressure.

The European Commission opening consultation on high-risk AI guidance is not thrilling content. That is exactly why it matters. The next fight is not just which model is smartest. It is whether companies can turn obligations, records, tests, and disclosures into a workflow humans will actually use.

Bad compliance makes teams hide from the process until launch week. Good compliance behaves like a product: clear states, useful defaults, evidence where it belongs, and fewer surprise meetings with legal. Somebody is going to make governance feel less like punishment and more like shipping. That somebody will get paid.

NASA researchers saying Jupiter's lightning may form deeper in water-rich storms has no SaaS pricing page, no waitlist, and no founder thread explaining the moat. It still makes you stop. Giant planet. Hidden storms. Lightning below the clouds. Come on.

That is a reminder for media people and tech people alike. Curiosity is not a sideshow to the business of news. It is the whole motor. If every update starts sounding like a changelog with investors attached, the strange beautiful stuff gets an opening.

AI teams are about to get squeezed from both directions. Leadership wants faster launches because competitors are moving. Security, legal, procurement, and regulators want proof that new model features are not being thrown into production with a shrug and a release note.

Here's the play: Build a productized service for AI product teams that turns messy launch-risk work into a lightweight release desk. Charge a setup fee, then a monthly retainer for policy mapping, launch checklists, evidence packs, vendor reviews, and incident-ready documentation.

What you'd actually do: Start with B2B AI companies selling into regulated or security-sensitive buyers. Interview product, engineering, security, and legal once. Turn their model and feature release path into a one-page operating map. Create reusable eval checklists, launch sign-off templates, disclosure language, risk registers, and a customer-facing evidence packet. Then offer a live pre-launch review for every material release so the team can ship faster with fewer last-minute "wait, did anyone ask legal?" surprises.

Why it works: The market gap is between policy theater and shipping reality. Big firms can hire entire governance teams. Small AI companies cannot, but their customers still ask enterprise-grade questions. You become the adult supervision layer that keeps momentum alive and leaves a paper trail that does not look like it was invented five minutes before procurement asked.

That's the briefing. Now go build something.

— Michael

P.S. The next AI moat might not be one more dazzling demo. It might be being the company that can ship the dazzling demo without setting off every alarm in the building.

Until next time,

Michael Benatar