Hey — It’s Nico.

Welcome to another Failory edition. This issue takes 5 minutes to read.

If you only have one, here are the 5 most important things:

• Marin Software, an OG in the digital marketing industry, has shut down — learn why below.

• You should be playing with AI agents for marketing

• Cloudflare launches a marketplace that lets websites charge AI bots for scraping.

• xAI raises $10B in debt and equity.

• Microsoft has made a super intelligent doctor AI — learn more below.

Lets jump in!

This Week In Startups

🔗 Resources

How we built Ramp by taking asymmetric risks.

The State Of MCP

You should be playing with AI agents for marketing

The Vertical Review: Voice AI in Healthcare

📰 News

Google rolls out its new Veo 3 video-generation model globally

ChatGPT referrals to news sites are growing, but not enough to offset search declines.

Cloudflare launches a marketplace that lets websites charge AI bots for scraping.

Grammarly acquires AI email client Superhuman

💸 Fundraising

Castelion is raising a $350M Series B to scale hypersonic missile business.

Remark raises $16M to build out human-powered expert models for e-commerce.

xAI raises $10B in debt and equity.

UK-based AI-driven parcel delivery startup Hived raises $42M

Fail(St)ory

A Digital Marketing OG

A few days ago, Marin Software—once the darling of ad-tech IPOs—quietly filed for Chapter 11 protection in Delaware.

The move caps a decade-long slide for a company that, in 2013, had investors buzzing about a new era of cross-channel campaign management. 

What Was Marin Software:

In 2006 paid search looked like a war room: a laptop, twelve open tabs, and an Excel sheet that only one brave soul would edit.

Marin showed up with a cleaner idea—one screen where you could set bids, swap copy, and watch results across Google, Yahoo, and a brand-new thing called Facebook Ads. 

Agencies loved the time saved, brands loved the clear ROI, and venture dollars poured in. By 2012 Marin’s platform was steering more than $10 billion in yearly ad spend for names like AT&T and Apple, and the company had 650 people spread across a dozen offices.

Wall Street bought the story too. When Marin rang the NYSE bell in March 2013, the stock popped and the market cap touched $425 million—not bad for seven years of work.

Then the ground shifted. Google kept making Ads Editor better, Facebook rolled out its own Business Manager, and the “single dashboard” Marin charged for started to look like a paid version of tools everyone now got free. Revenue peaked in 2015 and slid every year after. Marin never recorded an annual profit and ended up $350 million in the red.

The team tried one last swing in 2024, wiring OpenAI copy suggestions into the product, but most customers were already living inside Google’s and Meta’s native consoles. Cash ran down to about $5.7 million, headcount shrank to 40, and this week Marin filed for Chapter 11, planning to sell what was left to ESW Capital.

A 19-year run that began with “kill the spreadsheet” ended with a blunt lesson: if the platform you sit on copies your best feature and gives it away, the clock starts ticking.

The Numbers:

• 📅 Founded: April 2006

• 💰 Peak valuation: $425 M at IPO in March 2013

• 🧑‍💼 Headcount: 650 at peak, 40 by early 2025

• 📈 Revenue peaked in 2015; every year thereafter declined

• 🔥 Cumulative losses: $350 M

Reasons for Failure: 

• The landlord ate the tenant. Google and Meta kept upgrading their free campaign managers. “Good enough” plus a $0 price tag beat Marin’s paid seat every time. Once big customers started using in-house tools, sales cycles stretched to infinity.
  

• No Moat Beyond UI: Marin’s secret sauce was convenience rather than proprietary data or exclusive inventory. Once convenience got commoditized, the company had nothing defensive to lean on—no unique targeting graphs, no must-have algorithms.
  

• Late-Cycle Pivot to AI: In 2024 the team wired OpenAI into the product to write ad copy and suggest bids. Useful, sure, but rivals shipped similar AI features within months. When differentiation arrives that late, it looks more like parity than innovation.
  

• Death by a Thousand Dilutions: Marin tried every page of the survival playbook—layoffs, debt refinancing, and eventually an asset sale to ESW Capital. Each cut bought months, not momentum, and employee morale drained alongside the balance sheet.

Why It Matters: 

• Building on top of a gatekeeper is a rental agreement, not a moat—landlords can (and will) raise the rent to free.

• “Good enough” from a giant beats “slightly better” from a startup when the giant bundles it for $0.

• Shrinking headcount can extend runway, but without a credible path to profitability it merely prolongs the inevitable.

Trend

Medical Superintelligence

Last week  Microsoft quietly published research on a new system called MAI-DxO (short for Medical AI Diagnostic Orchestrator).

In 304 real-world medical case records, the AI hit the correct diagnosis ≈85 % of the time, while a group of seasoned physicians, forced to work without textbooks or teammates, landed at ≈20 %. Mustafa Suleyman, the new head of Microsoft AI, calls it a step toward “medical super­intelligence.”

Why It Matters:

• Faster, cheaper answers: MAI-DxO orders only the tests it needs, trimming the diagnostic bill instead of padding it.

• Dr House, on demand: The tool acts like a panel of specialists who volley ideas until the puzzle clicks—great news for edge-case patients who usually bounce between clinics for months.

• A playground for builders: Startups that wrap data pipes, workflow glue, or regulatory guardrails around this tech will create the next wave of “picks and shovels” for healthcare.

From Trivia Champ to Clinician

Large language models already ace the USMLE (think the SAT for doctors). But multiple-choice trivia isn’t real medicine. A solid diagnosis follows a sequence: listen → probe → test → narrow → decide.

To see whether an AI could handle that dance, Microsoft fed MAI-DxO a stack of NEJM Case Records—essentially real-life House M.D scripts published by The New England Journal of Medicine. Each case starts with a cryptic symptom list, then marches through labs, imaging, and frantic hallway consults until the final “aha!”

The AI mirrors that process.

1. Analyze the clues – fever, cough, mystery rash.

2. Pick the next move – ask a follow-up or order a targeted test, always watching the budget.

3. Hold an internal debate – several frontier models (OpenAI o3, Gemini, Claude, Grok, Llama, etc.) argue like sleep-deprived residents.

4. Call the ball – the orchestrator tallies votes, prints the diagnosis, and shows its receipts.

On this Dr House gauntlet, MAI-DxO landed the right answer about four times as often as a solo physician working blind.

So, No More Human Doctors?

Microsoft’s team is careful to say this isn’t a robo-clinician. Bedside manner, consent conversations, and the sheer messiness of real-world data remain very human chores. 

But if MAI-DxO can handle the cognitive heavy lifting—much like autopilot handles cruise flight—clinicians get to focus on empathy, context, and edge-case judgment. 

Early deployments will sit inside decision-support dashboards, quietly suggesting “Order a D-dimer” or “Have you ruled out sarcoidosis?” before the doctor clicks Submit.

Looking ahead, the team expects near-perfect accuracy within a decade. Once hospital networks plug in petabytes of anonymized scans and visit notes, MAI-DxO can fine-tune on edge-case reality—and every fresh data point sharpens the blade.

Roadblocks & Red Tape

• Hallucinations still happen. Even a 1 % error rate looks benign until you multiply it by millions of annual visits.

• Regulators will want a paper trail. In both the U.S. and EU, any “black-box” diagnosis tool faces FDA/EMA scrutiny plus new AI-safety rules. Build in audit logs from day one.

• Liability is murky. If a model suggests the wrong drug and the doc rubber-stamps it, who pays? Expect new insurance products—and courtroom test cases—to emerge fast.

• Data deals decide the winners. Hospitals won’t hand over records for free. Startups that craft revenue-sharing or privacy-preserving training loops will lock in unfair advantages.

Help Me Improve Failory

That's all of this edition.

Cheers,

Nico