The $2B Mirage

How a $2B “hotel Airbnb” collapsed overnight.

Author

Nicolás Cerdeira
November 13, 2025

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:

Today’s sponsor, Redrob, is helping teams replace generic AI outreach with real human insight through its Human Intelligence Lab. Also, shoutout to Cloudways, grab their Black Friday deal for effortless site migrations and big savings.

Your Prospects Can Spot AI From a Mile Away AD

That’s why your outreach feels invisible.

Everyone’s using the same AI tools, trained on the same data, writing the same “personalized” messages. Authenticity got lost in automation.

Redrob’s Human Intelligence Lab brings real researchers back into your process—uncovering paywalled data, social signals, and human context that make your outreach actually relevant and stand out.

And the best part? You can try it free, and keep using it for less than the cost of a single bad cold email.

Because real connection starts where automation ends.

This Week In Startups

🔗 Resources

This Black Friday, scale smarter. Cloudways gives 50% off for 3 months + 50 free migrations — move all your sites effortlessly. Get the Deal today! *

📰 News

Google augments AI shopping with conversational search and an AI that calls stores for you.

Wikipedia urges AI companies to use its paid API, and stop scraping.

💸 Fundraising

Former Twitter CEO Parag Agrawal’s Al startup Parallel raises $100M.

* sponsored

Fail(St)ory

Hotel-Airbnb Hybrid

For years, Sonder looked like the smarter, sleeker cousin of Airbnb, a company that could finally bridge hotels and home-sharing. 

Designer apartments, digital check-ins, no hosts to text about missing towels. It was the dream version of hospitality: scalable, tech-driven, asset-backed.

And then, a couple of days ago, it all ended in a bankruptcy filing.

What Was Sonder:

Founded in 2014 in San Francisco, Sonder started with a seductive pitch: what if you got the consistency of a hotel and the personality of an Airbnb?

They leased entire apartment buildings, turned them into chic, uniform suites, and managed them like a hotel chain, but through an app. Sonder raised hundreds of millions and expanded to 37 cities worldwide, operating 9,000 apartment-style units from New York to London.

By 2021, Sonder went public via SPAC at a $2.2 billion valuation, riding the pandemic rebound and the travel tech hype. Airbnb’s stock was soaring, and Sonder seemed like the next logical bet.

And then, reality hit. By mid-2022, its stock dropped from $200 to under $50. Losses piled up. The company faced delisting threats, postponed shareholder meetings, and late financial filings. By early 2025, both the CEO and CFO had resigned, leaving a vacuum at the top.

Desperate for a turnaround, Sonder struck a deal with Marriott International in 2024, a licensing partnership that rebranded its apartments as “Sonder by Marriott.” The deal promised $126 million in potential liquidity.

But the integration became a disaster. Sonder’s tech couldn’t align with Marriott’s legacy systems. “Unexpected challenges” ballooned costs and delayed listings, slashing revenue instead of boosting it.

By 2025, Marriott declared Sonder in default and pulled its 140 properties and 7,700 rooms off its platform entirely.

The Numbers:

  • 💸 Valuation peak: $2.2B (2021)

  • 🏙️ Cities operated: 37

  • 🏘️ Units managed: 9,000

  • 📉 Stock price: $200 → <$1 (2022–2025)

  • 💀 Q2 2025 net loss: $44.5M (3x prior year)

Reasons for Failure: 

  • The asset-heavy model killed flexibility. Sonder’s “hotel-meets-Airbnb” pitch hid a brutal truth: owning or leasing space doesn’t scale like listing it. Every market expansion meant new leases, furniture, staff, and maintenance. When travel demand dipped or occupancy slipped, losses multiplied.

  • The Marriott partnership backfired. The deal was supposed to legitimize Sonder and drive bookings. Instead, it became a tech nightmare. Integrating a nimble, cloud-native system with Marriott’s decades-old infrastructure caused endless sync failures and delayed listings. The result: months of downtime, lost bookings, and rising integration costs that drained its last reserves.

  • Leadership drifted as the company spiraled. When your CEO and CFO both walk out before a liquidity crunch, investors smell smoke. Sonder’s executive churn reflected a deeper problem: a company with no clear path to profitability. It kept chasing growth metrics even as revenue per unit fell. 

  • Venture-style growth economics clashed with hospitality reality. The VC world loves “blitzscaling.” Hotels don’t. Sonder tried to apply tech startup speed to a business where every unit requires physical upkeep, permits, and staff. Rapid expansion only amplified inefficiency. By the time investors realized the margins looked more like Marriott’s than Airbnb’s, the market had lost patience for asset-heavy dreams.

Why It Matters: 

  • Tech doesn’t erase physical risk. Asset-heavy startups crash when they pretend they’re software.

  • Partnerships can be poison. One bad integration can wipe out an entire turnaround plan.

  • Cash flow still rules hospitality. Sonder learned that great branding can’t cover ugly unit economics.

Trend

Spatial Intelligence

Large language models can write symphonies in text, but can’t tell you how far the chair is from the table. They speak fluent abstraction, yet fail at physics.

The next big AI leap isn’t about bigger models or cleaner prompts, it’s about giving machines a sense of the physical world.

That’s the argument from Fei-Fei Li, one of the world’s most famous AI researchers. Her recent post, From Words to Worlds, frames “Spatial Intelligence” as the next frontier. 

The idea: today’s AI can describe the world, but not inhabit it. Even the best multimodal models can’t estimate distance or predict what happens if you push a cup off a desk. 

Fei-Fei’s fix is “world models”: generative systems that understand space, time, and physics. Where LLMs predict the next word, world models predict the next frame.

Why it Matters

  • It turns AI from a text tool into an operational tool. Once models understand environments, you can automate workflows that were off limits because they involved physical context.

  • It makes real world testing cheaper. When you can rehearse failures in simulation instead of on hardware or in the field, you cut cost and downtime immediately.

  • New products become viable the moment consistency improves. AR, robotics, planning tools, and 3D content all get unlocked by models that keep a scene stable over time.

World Models

According to Fei Fei, a World Model is simply an AI system that can generate and maintain a consistent 3D environment over time. Instead of predicting the next word, it predicts the next moment.

The model should remember where objects are. It should understand solidity. It should know that if you open a drawer, the drawer stays open. And if you place a cup on a table, it doesn’t teleport away three frames later. Consistency is the key.

This sounds obvious until you look at what today’s models actually produce. A recent demo of an AI generated game built with Sora 2 made the rounds on X, and you could see the problem in seconds. The world kept slipping out from under itself.

The map warped unpredictably. Objects blinked in and out of existence. Nothing held together across frames. A proper world model wouldn’t behave like that because it would track physics, remember previous states, and update the environment in a stable way. It would run the simulation instead of hallucinating it.

Fei-Fei’s startup, World Labs, is the cleanest illustration. You feed their AI a photo and it generates a 3D scene you can explore. It makes up the missing geometry while keeping the whole thing stable. You walk around and nothing collapses. 

The Race to Build World Models


World Labs is not the only company chasing World Models. Here are some of the biggest players in the field:

  • Google DeepMind’s Genie 3 hit in August. It generates minutes of interactive 3D at 24 fps, remembers what happened earlier, and lets agents poke the environment. Their main goal is to use generative video like a training ground for embodied agents.

  • Earlier this year, Wayve released GAIA-2, a model that generates realistic driving footage to train autonomous driving systems. The goal is simple: create endless, controllable scenarios that would be impossible to collect reliably in the real world.

  • And finally, Meta has also entered the World Model race with V-JEPA 2, a 1.2 billion-parameter model trained on video that enables AI agents to understand, predict and plan in the physical world.

The Trend

Spatial intelligence keeps showing up in places you wouldn’t expect. After digging through all these releases, it feels like the market is warming up for products that assume models can understand space even if they’re not perfect yet.

A few opportunities stood out:

  • Workflow tools for teams adopting spatial AI. Companies will need scene-level debugging, data cleaning for 3D sensors, and simple ways to stitch world models into existing pipelines. 

  • Simulation-heavy vertical apps. Warehousing, construction, and energy all need safer planning environments. A world-model driven simulator that plugs into their actual operations could have immediate ROI.

  • AR and field service upgrades. Companies want on-device spatial reasoning to guide workers and verify tasks. Stable scene understanding unlocks real money here.

  • Content workflows for persistent 3D. Agencies, retailers, and studios all need coherent virtual spaces, not just assets.

Most of these ideas don’t require training world models. They require meeting companies where spatial AI actually becomes useful.

Help Me Improve Failory

How useful did you find today’s newsletter?

Your feedback helps me make future issues more relevant and valuable.

Login or Subscribe to participate in polls.

That's all for today’s edition.

Cheers,

Nico