Brutalist

This week Google wrote the biggest cybersecurity check in history, closed a billion-dollar deal with Merck, and pulled Sergey Brin out of retirement to fix the one place the brutalist strategy keeps tripping.

THE NUMBER: ~~6~~ 7

Six Google products have more than a billion monthly users: Search, Maps, Gmail, YouTube, Android, Chrome. The seventh is the iPhone, which Google pays Apple roughly twenty billion dollars a year — and by some 2024 estimates closer to twenty-six — to stay on as the default search engine. Which means the most beautifully designed consumer device in human history is, functionally, a Google distribution channel with an aluminum chassis. Nobody else has two products with a billion users. Microsoft has one (Windows, stretched). Meta has three (Facebook, Instagram, WhatsApp), but none of them carry calendar, commerce, location, or document context. Apple has several (iPhone, App Store, iMessage), but their privacy posture deliberately keeps the signal thin. Google has seven, and pays cash for the one they don’t own. That’s the moat. Everything else in AI — the models, the harnesses, the benchmarks, the funding rounds, the coverage cycle — is happening on top of a distribution graph that was won a decade before anyone heard the word agent.

Apple sells you the penthouse. Google pours the concrete. Guess which one the whole city is built on.

None of Google’s products has ever been praised for its design. Gmail’s UI is the same set of gray rectangles it had in 2014. Google Maps gets more cluttered every time you zoom in. Search results look like a printed phonebook stapled to an ad server. Android is where aesthetic ambition goes to die. And yet: 1.8 billion Gmail accounts, 3 billion active Android devices, 3.5 billion Chrome installs. Google ships brutalist software — load-bearing, indifferent to how it looks, built to last — and lets everyone else compete to make it pretty. Superhuman charges thirty dollars a month to put a better wrapper on Gmail. HEY, Shortwave, Spark, Apple Mail, every AI email assistant shipping this quarter — all harnesses on top of Gmail. Google doesn’t care. Google already has the accounts, the message graph, the spam filter, and the storage economics nobody else can match.

This week Google pressed the advantage on every front at once.

💲 Alphabet closed its acquisition of Wiz for $32 billion — the largest cybersecurity deal in history, larger than Canada’s annual defense budget, larger than any acquisition Google has ever made. In the same week three other AI products got publicly breached by classical security failures that had nothing to do with AI.

💲 Merck signed a multi-year, $1 billion agreement with Google Cloud spanning R&D, manufacturing, and commercial operations. Three hundred and thirty Google Cloud customers are now running more than a trillion tokens a year through Gemini. Forty are over ten trillion.

🧠 Thomas Kurian confirmed at Google Cloud Next that roughly three-quarters of new code at Google is now AI-generated and reviewed by engineers, up from fifty percent in February. The company that sat in second place on the coding benchmark leaderboard for a year is writing most of itself with its own models.

📉 And Sergey Brin un-retired from Google, specifically — his words, leaked from an internal town hall — to “out-code Anthropic.” The founder came back to fix the one place the brutalist playbook kept failing: a fragmented coding product line that even Google’s own DeepMind engineers quietly abandoned for Claude Code.

The frontier labs are fighting over who makes the best model. Google is playing a completely different game. Make sure every AI harness worth shipping has to reach into Gmail, Calendar, Drive, Maps, Chrome, and Android to function. Let somebody else win the benchmark of the week. Charge the toll on the way in, harvest the signal on the way out, run it again. We’ve written before that models already exceed what the average user needs. Google figured that out first, and built the only business model that doesn’t care who wins the race.

Apple Is Google’s Largest Paid Distribution Channel

💲 When the DOJ’s antitrust trial against Google cracked open in 2023, the most expensive single fact to come out of it was the payment stack Apple was sitting on. Google had been paying Apple $1 billion a year in 2014 to be the default search engine on iPhone. By 2018 it was north of $9 billion. By 2021 the number read into court testimony was $18 billion. By 2022 it was $20 billion. The most recent estimates put 2024 at somewhere between $20 billion and $26 billion — around a quarter of Apple’s Services revenue, which is the line item the market credits with saving Apple from being a cyclical hardware story.

Read that out loud. Google pays Apple twenty billion dollars a year to make the iPhone a Google distribution channel. Apple takes the money. Every time a user picks up the most beautifully engineered consumer product on earth and types anything into its address bar, Google gets the query, the session, the signal, and the opportunity to show a search result next to an ad. Apple’s take is the cash, and maybe the story that their privacy policy keeps the user safe. Google’s take is 1.4 billion devices worth of live context.

This is the part of the Apple-versus-Google framing that nobody wants to say out loud: Apple is not Google’s rival in AI. Apple is Google’s largest paid distribution channel. The stack that wins consumer AI is Apple’s chassis running Google’s context. It’s already the stack most people are using. The only question is who captures the economics when the search bar becomes an agent bar.

Apple Intelligence has been a quiet bust for eighteen months now. The software is beautiful. The context is thin. Apple deliberately doesn’t collect the signal that makes an AI assistant actually useful — they’ve built a privacy brand that’s incompatible with the thing they’d need to ship to win. Google built the opposite brand, for the opposite reason, and the breakpoint between the two strategies is exactly where this AI cycle is going to get decided. Siri will keep being a polished dud. ChatGPT will keep borrowing your calendar through a brittle integration. Google’s agent will already know you, because Google already has you — and, for twenty billion a year, so does Apple’s hardware.

The other four AI companies are running much thinner plays. OpenAI has one product people love and zero context (and is reportedly paying Apple to put ChatGPT in iOS — a much smaller version of the exact deal Google already has). Anthropic has one product enterprises love and a Pentagon feud. xAI has a compute advantage and a Cursor option that expires in April 2027. Microsoft has three of the four stack layers but rents the model from a partner that’s visibly breaking up with them. None of them are playing the same game Google is playing, because none of them can. You can’t retroactively win twenty-seven years of default-setting.

The signal: If your AI strategy assumes your competitors are the labs, you’re fighting the wrong war. Google’s competitors are the other companies with consumer distribution at scale, and there’s only one of those — sitting in Cupertino, taking Google’s twenty billion a year, and showing up to AI with the thinnest context in the group. That race already ran.

$32 Billion Says the Substrate Gets Hardened

🔒 The Wiz acquisition closed this week at $32 billion. That is not a normal number. It is the largest cybersecurity deal ever written — above Cisco’s $28 billion Splunk buy, which held the record since 2024 — and by a wide margin the biggest Alphabet has ever paid for anything. Wiz was founded six years ago. The four founders are the same Israeli team that built Adallom, which Microsoft bought in 2015 for $320 million. A hundredfold return on the same founders in about a decade is what an incumbent pays when it has decided the acquisition is load-bearing to everything else.

Why now is the cleanest question. The answer was written in red ink across the week’s news feed. Anthropic launched Mythos, its frontier cyber-defense model, and got breached on day one by a private Discord that exploited a contractor with easy URL access. Vercel went down when one compromised OAuth token from an AI coding tool cascaded into GitHub and npm. Lovable exposed user data; the CEO apologized publicly. Three AI products. Three classic 2005-grade security failures. One week. Aligned News ran the headline: Two Launches. Three Breaches. One Week. The framing is the right one.

The message to every enterprise buyer of every AI tool this week was: the model can be brilliant and the stack around it can still be on fire. That is a problem Google sees with particular clarity because Google already has the stack the AI tools plug into. Wiz runs cloud security posture management at something like 50 percent of the Fortune 100. Colgate-Palmolive reports external exposure reduction of 44 percent. Deloitte reports 60 percent analyst efficiency gains. Shell says it now catches critical vulnerabilities in “near real-time” versus three-to-fourteen days before. Bolt it onto Google Cloud, and every Gemini deployment in a regulated industry ships with a defensive perimeter Google wrote the contract for.

💲 The other half of the same play closed the same week. Merck signed a $1 billion multi-year agreement with Google Cloud across R&D, manufacturing, and commercial operations. Pharma is the cleanest enterprise bellwether in the market — regulated, cautious, and very good at picking the vendor that will still exist in fifteen years. Merck picking Google over Microsoft is a tell. Novo Nordisk went OpenAI. Eli Lilly signed a $2.75 billion deal with Insilico. The pharma industry is dividing its AI bets between vendors because nobody wants the single-vendor lock-in of the cloud era, and Google is running first in that bucket at the biggest customers.

Underneath the deal is the math Thomas Kurian wanted everyone to see. 330 Google Cloud customers run more than a trillion tokens a year on Gemini. 40 are over ten trillion. Sixteen billion tokens a minute flow through Google first-party APIs, up 60 percent quarter-over-quarter. The Gemini Enterprise Agent Platform shipped this week. The Agentic Data Cloud shipped this week. Chrome Enterprise added Auto Browse and Skills and a Gemini Summary surface this week. All of it is designed to be the rail that other companies’ agents ride on. Peter FitzGibbon of Insight Enterprises said at the conference: “Agentic development has absolutely gone mainstream. There is no more tire-kicking going on like we had in 2024 and ’25.” Read that line with a finance brain. Tire-kicking is the pilot phase. The phase after tire-kicking is where the invoices get big.

Why it matters: The $32 billion Wiz number is the tell, not the Merck deal. Merck is what winning looks like. Wiz is what spending to keep winning looks like. If you are a CIO staring at an AI vendor map right now, the question isn’t which model you use. It’s which platform’s security team is going to be holding your liability when an AI product in your stack gets breached the way Mythos, Vercel, and Lovable all got breached this week. Google just wrote that team a $32 billion check. Pick your platform in Q2. The decision goes to the vendor underwriting the floor, not the one with the best demo.

The One Seam: Sergey Came Out of Retirement to Sew It Shut

📉 The brutalist playbook has always had one weakness. When a product category is about craftsmanship — not function, but opinionated taste — ugly loses to beautiful. Search worked brutalist. Gmail worked brutalist. YouTube, Maps, Android, Chrome — all brutalist, all winning. Code editors are different. Code is where engineers spend eight hours a day, and engineers are the one audience in tech that has always paid for aesthetic opinion in tooling. Vim won by being ugly with a strong opinion. TextMate won by being beautiful. Sublime won by being both. Cursor won the same way.

The LA Times wrote the autopsy on Tuesday. Google has six overlapping coding products with six different brandings — Antigravity (the platform Google built on top of the $2.4 billion Windsurf acquisition), Gemini Code Assist, Gemini CLI, AI Studio, Firebase Studio, Jules. Even inside Google, engineers quietly use Claude Code. Teams at DeepMind working on the Gemini model itself use Claude Code. Most employees are formally banned from competing tools for “security reasons” and granted exceptions when they can demonstrate a business case. Kathy Korevec, who ran Jules, left for OpenAI this month. Brian Saluzzo, who oversaw internal adoption, departed the same week. Chief AI Architect Koray Kavukcuoglu is now consolidating the six products under the Antigravity banner. It is a classic late-2020s big-company org chart problem with an early-2000s small-company feel: too many parallel efforts, not enough taste at the top, no single owner.

Which is why Sergey Brin un-retired. The quote that leaked was “out-code Anthropic.” That is not the sentence a VP of Engineering says. That is the sentence the founder says when he has decided the business risk is existential. When Brin last dove back into a product it was to rebuild Google’s AI infrastructure after the Bard launch faceplanted in early 2023. That one worked. Gemini 3 shipped in late 2025 and was briefly, objectively, the best model on benchmarks. The coding problem is harder because the fix is not primarily technical — it’s editorial. Somebody has to decide which of the six products gets killed, and that somebody has to have enough authority that nobody sues the email chain when it happens. Brin has that authority. Sundar doesn’t, not like that.

This is the place the thesis is vulnerable, so it’s worth saying out loud. If the coding experience keeps rotting inside Google while Anthropic and OpenAI keep compounding on top of Claude Code and Codex, the brutalist bet has a real hole — because code, more than any other layer, is the workflow that drags context into an agent. If a developer at an enterprise uses Claude Code all day, that developer’s commit history and repo context feeds Anthropic, not Google. Multiply by a million enterprise developers and you have a counter-flywheel building inside the one category Google has historically lost. Brin coming back is the move that says management already figured this out and is pulling the founder lever. The question is whether he has eighteen months before the flywheel matures.

Anthropic’s read on this is different from ours, obviously, and worth acknowledging. Their bet is that Claude Code becomes the Cursor for enterprise, and the enterprise developer carries Claude into the building the same way Figma developers carried design into it a decade ago. That bet is not crazy. It’s also not what Google is betting. Google is betting that developers will use whatever is in front of them inside the IDE, and that Google — with Antigravity, Gemini Code Assist, Firebase Studio, and whatever survives the consolidation — can ship “good enough” code tooling that sits one click away from every other Google product the developer’s company already runs on. Which is the whole brutalist thesis compressed into one product category.

The action item: If you run an engineering org, your coding vendor decision for 2026 is the one place the Google thesis is least obvious. Claude Code has a real taste advantage today. The Google stack is messier and playing catch-up. But the consolidation is happening, Brin is focused, and the integration with Workspace, Chrome, and Vertex is the leverage nobody else has. Don’t lock in. Re-evaluate at the end of Q3 and watch who’s still around.

What This Means For You

Google spent twenty-seven years building the context graph the entire consumer internet runs on. The rest of the AI industry spent the last twenty-seven months arguing about who makes the best model. This week was the moment those two timelines visibly intersected — with a $32 billion security acquisition, a $1 billion pharma deal, a founder coming out of retirement, and three non-Google AI products getting breached by classical security failures in the exact same news cycle. The narrative about who is winning AI just got reordered on the only axis that ever mattered.

Assume Google is your AI vendor whether you picked them or not. Your enterprise already runs Workspace or touches GCP. Your employees already use Gmail, Calendar, Docs, and Chrome. Every agent you’re evaluating from a startup will eventually plug into those rails. Budget accordingly.

Pick your platform in Q2 based on liability, not demo quality. The models will converge. The security posture won’t. The vendor underwriting the floor — the one writing $32 billion checks to harden the stack — is the one your CISO will thank you for in two years.

The frontier labs are your vendors, not your strategy. Anthropic, OpenAI, and xAI are selling capability. Google is selling distribution. Your business needs both. Don’t confuse which one is which — and don’t let the benchmark leaderboard decide your vendor map for you.

The substrate always wins. Cisco won the dot-com build. ARM won mobile. Snowflake won the data cloud. The company that ships the picks and shovels to the gold rush ends up worth more than the miners, because the miners change and the picks don’t. In AI, the picks are the context graph. Google owns it. The miners rent it.

The penthouse gets the Architectural Digest spread. The concrete holds up the building.

Three Questions We Think You Should Be Asking Yourself

Which Google product is your AI strategy already silently depending on? Every AI agent you’re evaluating reaches into at least one of Gmail, Calendar, Drive, Maps, or Workspace. Pick the one that matters most to your business, and ask what your roadmap looks like if Google changes the API terms, deprecates a surface, or launches a competing agent inside the product. If the answer is “we’re done,” you have a concentration risk masquerading as a partnership.

If Apple Intelligence never gets good, does your AI distribution strategy still work? A lot of 2026 consumer-AI plans implicitly assume Apple will eventually ship something that closes the context gap. It hasn’t in eighteen months, and the structural reason — the privacy posture — isn’t going anywhere. The companies that are winning assume iPhone is a Google device running a Chrome-shaped window. Does yours?

Is Sergey Brin coming back to Google the best news of the week for Anthropic, or the worst? The obvious read is that Google panicking is bullish for the current coding leader. The less-obvious read is that founder mode at the one company with the full stack is the only thing that can actually catch Claude Code. Which way you answer this question decides whether you’re long Anthropic at an $800 billion IPO valuation, or whether you think the window closes by Q4.

A house is a machine for living in.”

— Le Corbusier, Toward an Architecture, 1923

Le Corbusier lost the argument with the critics for fifty years. The critics said brutalism was ugly, inhuman, unfit for the way people actually live. The architects said function was the point, and the machine would still be standing when the decoration finished peeling off the Victorian facade next door. Most of the brutalist buildings are still standing. Most of the Victorian facades aren’t. Ask the people who live in the buildings which one they’d rather own.

— Harry and Anthony

Sources

• Aligned News — Wednesday Afternoon Analysis: “Two Launches. Three Breaches. One Week.”
• LA Times — Google’s Internal Struggle Is Handing the AI Coding Race to Anthropic, OpenAI
• The Rundown — Sergey Brin’s Plan to Out-Code Anthropic
• Bloomberg — Mythos Breach Coverage
• Google Cloud — Official Merck Agentic AI Partnership Announcement
• Google — Wiz Acquisition Close
• Google Cloud Next 2026 — TPU 8t/8i Technical Deep Dive
• DOJ US v. Google — Search Default Payment Testimony (Apple Deal)
• Implicator.ai — Musk Options Cursor. Anthropic Drops the Indies. OpenAI Wins the Arena.
• The Information — OpenAI and Anthropic Rethinking Reasoning Architecture
• Anthropic — “What 81,000 People Told Us About the Economics of AI”
• Semafor Technology — April 22, 2026
• TheNeuron — April 22, 2026 Daily
• TLDR AI — April 22, 2026
• The Deep View — Cursor-SpaceX Deal Signals Coding Agents’ Victory