Wednesday Briefing — May 20, 2026

🎉 Good morning. While you were sleeping, Google quietly rewrote the next ten years of computing, Meta started swinging axes at 8,000 jobs, and a bunch of humanoid robots ran a 24-hour package-sorting livestream that Silicon Valley is treating like prestige TV.

Today's the day Nvidia tells us whether the AI economy is still real or just a really expensive group hallucination. Buckle up.

Let's ride. 🤠

Sundar Pichai walked on stage Tuesday in Mountain View and didn't bother pretending. Google is done playing nice with the post-iPhone world. They announced a new flagship model (Gemini 3.5 Flash), a world model that simulates physics for video generation (Gemini Omni), a general-purpose agent that can act across your connected apps (Gemini Spark), an autonomous coding platform (Antigravity 2.0), and — the headliner — Android XR smart glasses launching this fall in partnership with Samsung, Gentle Monster, and Warby Parker.

That last one is the story. Google is shipping actual, normal-looking glasses with all-day Gemini access, real-time translation that matches the speaker's tone, and visual translation on menus and street signs. The keynote demo was a single voice command while walking past a cafe: the agent queued a DoorDash coffee order on the phone, the user kept walking, and the only step left was a confirmation tap. That's it. That's the future they just demoed.

Google claims Gemini 3.5 Flash is 4x faster than rival frontier models on agentic coding, and Omni is rolling out today to AI Plus, Pro, and Ultra subscribers globally inside the Gemini app and Google Flow. Spark is in beta, going to Ultra subscribers next week. The pace is what matters here — Google launched five product lines in 90 minutes that any other company would each call a flagship reveal.

Why this hits so hard: the smart glasses category has been a graveyard. Google Glass, Snap Spectacles, Humane AI Pin, the Rabbit R1 — all hilarious failures. But Google brought brands people actually wear, hardware Samsung helped build, and an AI system that finally has something useful to do. Meta is selling Ray-Ban Display glasses with a 600×600 screen in the right lens right now. The fight is officially on, and your face is the battleground.

Heads up: Google also said Gemini 3.5 Pro, the heavyweight model, is being used internally but won't ship for "wider distribution until next month." Translation: there's a second wave of this announcement coming in late June. Don't relax yet.

Jensen Huang's quarterly check-in lands at 4pm ET. The Street wants roughly $78 billion in revenue and $1.77 EPS, with data center revenue around $73 billion. The options market is pricing an implied move of 8 to 10 percent, which is roughly the size of Costco's entire market cap moving in one direction in one night.

What everyone's actually watching: Blackwell shipment cadence, GB300 Ultra production timeline, and any China commentary. Bears think AI capex is peaking and gross margins are about to crack. Bulls think the GB200 NVL72 racks are landing on schedule at every hyperscaler and Nvidia is going to print a fourth straight blowout.

Why it matters: Every other tech stock you own is downstream of this print. If Nvidia stumbles, the entire AI infrastructure trade — Anthropic, OpenAI, CoreWeave, AMD, Broadcom, the whole capex frenzy — gets a pricing event tonight. If they beat, expect the AI bull market to extend through summer.

Meta begins cutting 8,000 jobs today, with more rounds planned for the second half of 2026. The total expected hit is nearly 20% of headcount. CNBC's reporting calls it Zuckerberg's "harsh AI reality" — meaning Meta is reorganizing around the assumption that their next flagship model "Avocado" (which got pushed from March to May after underperforming Google, OpenAI, and Anthropic) needs a leaner, more aggressive team to compete.

Zoom out: tech has already shed 113,863 workers in 2026, averaging 819 jobs lost per day. The official narrative is "AI productivity gains." The unofficial narrative is everyone is panicking that they overhired during the ZIRP era and AI is the most socially acceptable excuse to fix it.

Why it matters: If you're an operator, the labor market just got more brutal and the AI talent market got more expensive. Engineering managers from Meta's reality labs are about to flood LinkedIn. And if you're an indie builder, this is the most fertile hiring window since 2009.

Figure AI put three Figure 03 humanoid robots on a conveyor belt last week and let them run autonomously for 8 hours. Then 24. They sorted over 30,000 barcoded packages with zero human intervention and pulled more than 3 million live viewers on YouTube.

CEO Brett Adcock said the robots matched human speed (3 seconds per package) using on-board cameras and their Helix-02 software. Their new BotQ factory now claims one humanoid robot rolling off the line per hour. Meanwhile BMW already has Figure robots loading 90,000 sheet metal components on the X3 line, and Japan Airlines just deployed two Unitree humanoids at Tokyo's Haneda Airport for baggage and cabin cleaning.

Why it matters: "Robots taking jobs" went from a future-tense conversation to a real-time spectator sport this month. The ROI math is starting to work — payback periods drop to ~6 months under high utilization. Your warehouse-adjacent business model assumptions have a shelf life. Plan accordingly.

Yesterday cybersecurity researchers flagged a compromised version of the Nx Console extension published to the Microsoft Visual Studio Code Marketplace. Version 18.95.0 of "rwl.angular-console" was packed with what researchers called "a multi-stage credential stealer and supply chain poisoning tool" that harvested developer secrets and exfiltrated them via HTTPS, the GitHub API, and DNS tunneling.

2.2 million installs. If you use VS Code with an Angular workflow, you should probably stop reading this and go check your extensions list right now.

Why it matters: Supply chain attacks on developer tools are no longer the exotic edge case — they're the new normal. Every AI-coded plugin and "vibes-checked" install is a soft attack vector. If you've ever told a junior dev "just npm install it, it's fine," now's a great time to revisit that policy.

OpenAI and Dell Technologies announced a partnership Monday to bring Codex to hybrid and on-premises enterprise environments. Translation: large companies who refuse to put their proprietary code in someone else's cloud can now run Codex locally on Dell PowerEdge infrastructure.

This is the part of the AI race nobody covers because it sounds boring. But on-prem AI is where the next $100B of contracts get signed. JPMorgan, Lockheed, Pfizer — every Fortune 100 with regulatory paranoia is allergic to public cloud AI. OpenAI just bought a permission slip to walk into all of them.

Why it matters: If you sell to enterprise, the AI vendor selection process now has a real "we'll come install it" option. The competitive moat just shifted from model quality to deployment optionality. Watch which other vendors strike Dell, HPE, or Lenovo deals in the next 60 days.

A new Apple plan internally called "Extensions" would let iOS 27 users select third-party AI providers like Google and Anthropic to power Apple Intelligence features for text generation, editing, and image tasks. Models from Google and Anthropic are reportedly already in testing.

For Apple, this is a quiet capitulation. They couldn't catch up on frontier models, so they're going to be the open marketplace instead — your iPhone becomes a neutral switchboard for whichever AI is winning that month.

Why it matters: This is brilliant defensive judo. If Apple can't be the smartest, they'll be the most flexible. And it means every AI company is now also a "default iOS provider" candidate, which is going to drive bidding wars Apple will profit from for years.

Bitcoin is sitting at $76,565 as of Tuesday morning, down about $29,000 from a year ago. Deribit's CCO is calling $76-77K the critical support line — below that, $70K is the next floor and $60K is the lava beneath it.

The Bitcoin-focused DeFi protocol Echo got drained for about $77 million yesterday when an attacker minted 1,000 unauthorized eBTC on the Monad blockchain. Another week, another DeFi exploit.

Anthropic just launched Claude for Small Business, a connector pack with ready-to-run workflows inside QuickBooks, PayPal, HubSpot, Canva, Docusign, Google Workspace, and Microsoft 365. They're also debuting 20+ new legal MCP connectors. The pitch: Claude as a deployed teammate, not a chatbot.

SpaceX launched 24 more Starlink satellites from Vandenberg last night (booster 1103 stuck the landing), and pushed the maiden Starship V3 launch to tomorrow, May 21, after a one-day slip from yesterday.

Tesla has 22 fully driverless cars operating in Austin, 6 in Houston, and 2 in Dallas. Musk is forecasting FSD Unsupervised rollout in Q4 2026. Meanwhile, Tesla just clocked 93,000 autonomous miles inside Giga Berlin — in a country that hasn't approved FSD for public roads.

Forza Horizon 6 dropped on Xbox Game Pass yesterday for Ultimate and PC subscribers, and a Stegosaur skull — the best-preserved one ever found in Europe — was just unearthed in Spain. That's not a tech story. I just wanted you to know.

A group of major academic publishers including Elsevier, Cengage, Hachette, Macmillan, and McGraw Hill filed a class-action against Meta for allegedly using millions of copyrighted books and journals to train Llama. The discovery in this lawsuit is going to be wild.

Sereact closed a $110M Series B and Profluent inked a $2.25B drug discovery deal with Eli Lilly — pharma is now openly bidding for foundation-model talent.

I've been bearish on smart glasses for a decade and I am calling it now: Google's Android XR demo flipped me. Not because the hardware is finally good (though it is — Gentle Monster and Warby Parker collabs are not a joke). Not because Gemini Live is finally fast enough. Because the use case loop finally closes. You see a thing → you ask about a thing → an agent does a thing on your phone → you keep walking.

The phone is being demoted to the iPod-of-the-pocket — a battery and antenna with a chip in it. The interface migrates to your eyeballs and ears. This always sounded like a TED Talk fantasy, but the DoorDash demo was so banal and so seamless that I actually believe it now. By 2028, asking somebody for their phone number is going to feel as quaint as asking for their fax number.

The official line is "AI is making us more efficient." The unofficial line: Meta's Avocado model got punted from March to May because internal benchmarks showed it underperforming Google, OpenAI, and Anthropic on coding and reasoning. When you're losing the model race, you need cheaper unit economics to compete on price. And when you can't beat your competitors on quality, you cut headcount to make sure you can outlast them on burn rate.

Zuckerberg isn't optimizing — he's restructuring around the assumption Meta is going to be a fast follower for the next 24 months. Every CEO with a flagship model that's 6 months behind frontier is about to do the same math.

There is no middle outcome. The implied move is 8-10%, and the AI capex narrative is so finely balanced right now that any soft signal from Jensen — a margin compression hint, a China caveat, a delay on the GB300 Ultra — will get punished. Conversely, a clean beat plus aggressive Q2 guidance will trigger a 6-month melt-up across every AI-adjacent ticker.

My bet: revenue beat, margins fine, but the China commentary gets coded ambiguously enough that the stock trades sideways for a week while everyone tries to figure out if "softness" means "we're losing customers" or "we just shipped the GB300 a quarter late." Buy the boredom, not the fireworks.

Setup: Google's Gemini Spark, Anthropic's Small Business connectors, and OpenAI's Codex-on-prem all just made deploying a vertical AI agent a 2-week build instead of a 6-month build. Meanwhile every founder is chasing horizontal "AI assistant for SMBs" — competing for the same 14 plumbers in San Antonio.

Here's the play: Pick a non-sexy professional services vertical that has weird workflows, expensive labor, and zero AI competition. Auto body estimators. Court reporting prep. Insurance subrogation. Wedding vendor coordination. Then build a Claude-based agent that ingests the customer's existing tools and runs the entire repetitive intake-to-quote workflow autonomously.

What you'd actually do: Spend a week interviewing 10 people in the vertical you picked. Build an MVP using Claude for Small Business connectors plus a custom intake form. Charge $300-500/month per seat. Pitch it to one regional trade association — they have email lists and conferences and zero AI vendors talking to them right now. Sign 20 customers in 90 days. The whole thing is repeatable across 50 verticals and almost nobody is going to compete with you because the markets look "too small" to VC-backed competitors. Which is exactly why you win.

Why it works: The AI tools just hit "good enough" for vertical agents. The big players are chasing enterprise logos. The infrastructure is cheap. The buyers are starved for AI-anything but allergic to setup. Your competitive moat is domain knowledge plus the patience to talk to humans who hate technology.

That's the briefing. Now go build something.

— Michael

P.S. If you're watching Nvidia tonight at 4pm ET, do yourself a favor and put your phone in another room. Anything bigger than a 4% move is noise; anything smaller is a non-event. Either way you don't need to be refreshing CNBC live.

Until next time,

Michael Benatar