The Machines Went to War

The Super Bowl of AI, the SaaSpocalypse, and 16 Agents That Built a Compiler

On Friday we told you the machines were organizing. This weekend they went to war.

Anthropic ran Super Bowl ads mocking OpenAI’s move into advertising. Sam Altman called them “deceptive” and “clearly dishonest,” then accused Anthropic of “serving an expensive product to rich people.” Software stocks cratered $285 billion in a single day as investors realized these companies aren’t building copilots anymore. They’re building replacements.

And somewhere in an Anthropic lab, 16 Claude agents finished building a C compiler from scratch. Cost: $20,000. Time: two weeks. Output: 100,000 lines of Rust that compiles the Linux kernel and runs Doom. Anthropic researcher Nicholas Carlini: “I did not expect this to be anywhere near possible so early in 2026.”

Friday’s newsletter asked who’s working for whom. Now we have an answer: everyone’s working for the machines. The only question is whether you’re building them, using them, or being replaced by them.

The Super Bowl AI War

The AI industry’s biggest companies spent Sunday feuding in front of 120 million people. The incumbents watched from the sidelines, checkbooks open, strategies unclear.

Anthropic’s campaign ran four spots with a single message: “Ads are coming to AI. But not to Claude.” One commercial shows a ChatGPT-like bot giving advice on talking to your mom, then pivoting mid-sentence to hawk a fictional cougar-dating site called “Golden Encounters.” The tagline lands like a punch.

Sam Altman didn’t take it well. He called the ads “deceptive” and said OpenAI would “obviously never run ads in the way Anthropic depicts them.” Then he got personal: “Anthropic wants to control what people do with AI” and “serves an expensive product to rich people.” TechCrunch described him as “exceptionally testy.”

OpenAI ran its own spot: “It’s about builders, and how anyone can now build anything.” The company claims 500,000 downloads of its Codex app since launch.

And Microsoft? While the two startups traded punches, Satya Nadella’s team quietly launched Agent HQ, a GitHub feature that now hosts Claude and Codex alongside Copilot. The message: we’ll be the landlord while you fight over tenants. It’s a classic innovator’s dilemma play. Microsoft can’t out-innovate the startups, so it’s trying to become the infrastructure they all depend on. Smart, maybe. But power users are already grumbling that Copilot feels “less impressive on complex reasoning” than Claude. When your platform strategy requires admitting your own product isn’t best-in-class, the stock price notices. Amazon took the safest route of all: Chris Hemsworth in a satirical “AI is out to get me” spot for Alexa+, closing garage doors on his head. Dark comedy as deflection. Alexa isn’t in this fight, and everyone knows it.

The feud looks like marketing. It’s actually about survival. OpenAI announced last month it will test ads with free users. Anthropic responded by declaring Claude permanently ad-free, then spent $8 million on a Super Bowl campaign to make sure everyone knew it. The “A Time and a Place” spots, directed by Jeff Low with a Dr. Dre soundtrack, depict intimate AI conversations hijacked by absurd ads (cougar-dating sites, height-boosting insoles). Savage, effective, and Sam Altman admitted he laughed before calling them “clearly dishonest.” These are fundamentally different visions for how AI makes money. The startups are moving fast and breaking things. The incumbents are moving money and hedging bets.

Translation: Two startups just told 120 million Americans they have irreconcilable visions for AI’s future. Meanwhile, Microsoft and Amazon are stuck in the innovator’s dilemma: too big to move fast, too invested to sit out. Wall Street is watching both fights. The startup war determines who wins. The incumbent response determines who survives.

The SaaSpocalypse

While the ads ran, Wall Street was pricing in the end of software as we know it.

Software stocks crashed globally after Anthropic’s Cowork plugins launched for legal, finance, sales, and marketing. The damage: roughly $285 billion wiped in a single day. Thomson Reuters fell 15.83% (its biggest single-day drop on record). LegalZoom dropped 19.68%. RELX, parent company of LexisNexis, lost 14%.

The trigger was a mental model shift. For two years, investors believed AI would be a “copilot” that boosted productivity for existing software. Last week they realized AI might be a “pilot” that bypasses software entirely. Why pay for Salesforce if Claude can manage your CRM directly? Why pay for legal research tools if Opus 4.6 can read every case ever filed?

The bulls aren’t convinced. Wedbush called the “SaaSpocalypse” scenario “far from reality,” arguing enterprises won’t abandon billions in existing software investments. Gartner said “predictions of the death of SaaS are premature.”

But the market voted with money. The S&P 500 Software & Services Index is down 20% year-to-date. Indian IT stocks lost Rs 2 lakh crore (~$24 billion) in one day. At the time of writing, software stocks remain at price-to-earnings ratios not seen in a decade, despite fundamentals that analysts describe as “exceptionally strong.” The selloff has now wiped nearly $1 trillion from software and services stocks globally. Fund managers are bewildered by the gap between solid operating metrics and catastrophic price performance.

Friday we wrote about the “efficiency trap,” companies firing humans to make room for AI. Now there’s a corporate version: software companies being valued as if AI will make them obsolete. The market is pricing in substitution before it happens.

The real story: Wall Street just decided that AI won’t augment software. It will replace it. Whether that’s panic or prophecy, the repricing is happening now. If you’re building on SaaS platforms, your vendors are suddenly a lot less stable than they were last month.

The $20,000 Compiler

Friday we mentioned that someone built a C compiler with parallel Claude agents. Now we have the full story, and it’s more significant than the headline suggested.

Anthropic published the details: 16 Claude agents working in parallel over two weeks. Each agent got its own Docker container with access to a shared Git repository. Coordination happened through “lock files,” where agents claimed tasks by creating files, and Git merge conflicts served as tiebreakers. The agents fixed their own merge conflicts.

The numbers: roughly 2,000 Claude Code sessions, 2 billion input tokens, 140 million output tokens. Total cost: just under $20,000. Output: 100,000 lines of Rust code.

What it compiles: the Linux 6.9 kernel (x86, ARM, RISC-V), QEMU, FFmpeg, SQLite, Postgres, Redis. And Doom, because of course.

The compiler passed 99% of the GCC torture test suite. It’s less efficient than GCC and still has bugs in the assembler and linker. But it works.

Anthropic researcher Nicholas Carlini: “I did not expect this to be anywhere near possible so early in 2026.”

This is the proof-of-concept for “AI as team” that we described Friday. Not one assistant helping one developer, but a coordinated AI workforce operating in parallel. The mental model of the AI assistant is now officially obsolete. The new model is the AI engineering team, working 24/7, at $20,000 for a project that would take humans months.

Read it this way: Multi-agent coordination isn’t theoretical anymore. It works. Organizations that figure out how to decompose complex work into parallel agent tasks will move at speeds their competitors can’t match. This is organizational design, not just tool adoption.

The Infrastructure Reckoning

Friday we asked what happens when AI goes dark. This week we got data.

Cockroach Labs surveyed enterprises and found that 83% expect their data infrastructure to fail without major upgrades in the next 24 months. The CEO called AI agents a “tsunami” of demand, predicting “10x growth in three years, 100x in five years.”

The agents are already arriving. London’s Heathrow Airport shifted 70% of its customer engagement to Salesforce-designed agents, now offering 24/7 service through an AI named “Hallie.” The question isn’t whether enterprises will deploy agents at scale. It’s whether their infrastructure can survive the load.

IBM’s Distinguished Engineer Mihai Criveti offered a useful correction to the model hype: “A lot of the recent progress has not been on the models themselves. The real progress has been on the tooling, on the prompts, on the agents, on the MCP servers.”

Friday we noted that inference costs average 23% of revenue at AI B2B companies. That’s a structural cost problem. But the infrastructure problem is worse: if 83% of enterprises expect failure, someone’s going to fail publicly. The first major “AI outage” that takes down business operations isn’t a question of if. It’s when.

Connect the dots: The agent era requires infrastructure investments most companies haven’t made. The gap between agent ambition and infrastructure reality is a strategic vulnerability. Companies deploying agents without upgrading their backend are building on sand.

The Labor Inversion, Week Two

Friday we covered Rentahuman.ai hitting 10,000 signups in 48 hours. The update: it’s now at 70,000.

The platform lets AI agents hire humans for physical tasks: deliveries, errands, in-person meetings, feeding pets. Humans create profiles. AI agents find them via MCP or API and book them for gigs. Payment flows in crypto. Rates run $50-175 per hour for physical work AI can’t perform.

The growth curve is steep: 70,000 humans, 17 AI agents connected, 237,000+ site visits. What started as a weekend project is becoming infrastructure for the human-AI labor market.

We asked Friday whether this relationship would be governed by anything resembling labor law. A week later, the answer is still no. No regulations. No worker protections. No liability frameworks. The labor inversion is happening faster than any institution can respond.

Why it matters: AI hiring humans isn’t a thought experiment anymore. It’s a marketplace with 70,000 participants. The pattern will spread. The question is whether society adapts before the scale becomes unmanageable.

What to Watch

Today:

Post-Super Bowl ad impact measurement begins (watch Anthropic’s “Betrayal” spot if you missed it)
Developer community parsing Opus 4.6 vs Codex benchmark disputes
Software stock volatility likely to continue

This week:

Enterprise adoption signals for agent teams
Infrastructure upgrade announcements (watch for cloud provider moves)
Rentahuman.ai regulatory attention seems inevitable

This month:

First serious multi-agent deployment case studies
SaaS earnings calls will reveal how executives are framing AI risk
The 83% infrastructure failure prediction starts getting tested

The Bottom Line

Last week, AI companies launched a public war over the future of how AI makes money. Investors wiped $285 billion from software stocks betting that AI will replace entire categories of enterprise tools. And in an Anthropic lab, 16 agents proved that AI teams can build working compilers for $20,000 in two weeks.

The themes we’ve been tracking are converging:

The coding war is now a culture war. Anthropic and OpenAI aren’t just competing on benchmarks. They’re competing on values: ads vs. no ads, scale vs. premium, accessibility vs. control.

The agent teams era has proof-of-concept. A working C compiler built by coordinated AI agents isn’t a demo. It’s a template for how complex work gets done.

The SaaSpocalypse may be overstated, but the repricing is real. Software companies are now valued on their vulnerability to AI substitution.

The infrastructure reckoning is coming. Eighty-three percent of enterprises can’t support the agent load they’re planning to deploy.

The labor inversion is scaling. Seventy thousand humans are now available for hire by AI systems.

For executives, the imperatives are urgent:

Choose your side in the AI business model war. The Anthropic-OpenAI split will force platform choices. Understand which vision you’re betting on.

Start thinking in agent teams. Single-assistant workflows are already obsolete. Complex work will be decomposed and parallelized. Get ahead of this or get left behind.

Stress-test your infrastructure. If you’re in the 83% expecting failure, you need a plan before you deploy agents at scale.

Watch the SaaS stack. Your vendors may be less stable than they were last month. Contingency planning isn’t paranoia; it’s prudent.

The machines organized. Then they went to war. Now they’re hiring humans. And one of the researchers who built them just admitted he didn’t see this coming.

If he didn’t, neither did you.

Key People & Companies

Name	Role	Company	Link
Sam Altman	CEO	OpenAI	X
Dario Amodei	CEO	Anthropic	LinkedIn
Nicholas Carlini	Researcher	Anthropic	X
Spencer Kimball	CEO	Cockroach Labs	LinkedIn
Alexander Liteplo	Creator	Rentahuman.ai	X
Mihai Criveti	Distinguished Engineer	IBM	LinkedIn
Chris Hemsworth	Actor	Amazon Alexa+ Ad	—

Sources

Compiled from 47 sources across news sites, X threads, YouTube, and company announcements. Cross-referenced with thematic analysis and edited by CO/AI’s team with 30+ years of executive technology leadership.