The AI Boom's First Casualty: When Unicorns Go Bankrupt

The generative AI gold rush just claimed its first billion-dollar victim, and the details should terrify anyone who raised on a "we'll figure out monetization later" pitch. Meanwhile, the survivors are making moves that reveal what it actually takes to build sustainable AI businesses—and it's not what most founders assumed in 2023.

Today's briefing cuts through the noise: a unicorn bankruptcy that rewrites the playbook, Microsoft's stunning admission about OpenAI's limitations, Epic's reckless AI-first bet, a ghoulish new market in dead startup data, and why VCs are becoming infrastructure companies. If you're building anything AI-related, these five stories explain the gap between hype and survival.

1. Character.AI Files Bankruptcy, Sells Assets to Disney

Character.AI, the chatbot darling that hit unicorn status faster than almost any consumer AI company, filed Chapter 11 bankruptcy and is selling its assets to Disney. The company—which let users roleplay with AI characters—faced mounting lawsuits over chatbot safety but kept shipping features until the money ran out.

Here's what nobody's saying loudly enough: Character.AI's collapse isn't about safety lawsuits or product-market fit. They had 20 million monthly users at their peak. The problem was simpler and more existential: every conversation lost them money.

Inference costs for real-time chatbot interactions are brutal. Users would have hour-long conversations with AI boyfriends and fictional characters, burning through compute with zero revenue. The company tried subscriptions, but conversion rates couldn't cover the GPU bills. This is the dark side of product-market fit in AI—you can have rabid users and still go bankrupt because the unit economics are underwater from day one.

Disney's acquisition through bankruptcy means they're paying cents on the dollar for whatever IP and user behavior data Character.AI accumulated. They're not buying a going concern; they're buying training data and possibly some character AI tech for Disney+ or theme park applications.

🔥 Spark's Hot Take: Character.AI's bankruptcy is generative AI's first major reckoning with reality. In 2023, VCs threw money at anything with engagement metrics and a foundation model API key. Nobody asked the obvious question: "At what price per conversation does this break even?" The answer for Character.AI was apparently "never." If you're running an AI consumer app on venture capital, your board is now asking: What's our inference cost per user? What's our realistic revenue per user? And when do those numbers meet? If you don't have good answers, update your resume and your mental model of how much runway you actually have.

2. Microsoft's Lamanna Pitched Using Anthropic to Fix Copilot

Microsoft's Charles Lamanna, newly promoted to run the entire Office division, got there by doing something described internally as "heresy": he told CEO Satya Nadella that Copilot was buggy and needed Anthropic's Claude to work properly in Excel and PowerPoint. The fact that Microsoft even considered this—and that suggesting it got Lamanna promoted rather than fired—tells you everything about where the company's priorities are.

Microsoft has invested $13 billion in OpenAI. It has exclusive rights to GPT models. It has infinite Azure credits to throw at the problem. And yet its own AI chief went to Nadella and said: We might need to pay our competitor because our exclusive partner's models aren't good enough for our enterprise customers.

That's not a small admission. That's Microsoft publicly acknowledging that model lock-in is a fantasy and that even with the deepest partnership in AI, you need a multi-model strategy to ship products that work.

Lamanna apparently fixed many of Copilot's issues (whether using Anthropic models or not remains unclear from reporting), and his promotion to Office chief suggests Nadella values results over OpenAI loyalty. The message to the rest of the industry: making your AI product actually work matters more than which foundation model partnership your PR team announced.

For enterprise buyers, this is validation. Stop letting vendors lock you into single-model architectures. If Microsoft needs the flexibility to swap models for different tasks, you definitely do. Negotiate for multi-model support in every AI contract.

3. Epic Goes AI-First: Claude Writes All Code for Disney Shooter

Epic Games made a bet that will either revolutionize software development or become a legendary cautionary tale. The company shifted its $1.5 billion Disney collaboration to an "AI-first" model where Anthropic's Claude writes all code and generates all art assets. Human developers become managers of AI output. The company's internal messaging: "No more hand-written code."

The result so far? The Marvel and Star Wars shooter that was supposed to launch in March 2027 is now delayed indefinitely.

This is the most aggressive test of AI-native development at scale, and Epic deserves credit for trying it. But the early returns suggest we're hitting the limits of what current models can do. Claude is exceptional at generating code snippets, refactoring functions, and creating variations on existing assets. What it can't do—yet—is architect complex systems, make coherent design decisions across thousands of interdependent files, or maintain consistency in a AAA game's visual style across hundreds of characters and environments.

Epic is essentially betting that AI can compress game development timelines from 3-5 years to under 2 years. If it works, every studio follows and the game development workforce shrinks by half. If it fails, Epic wasted $1.5 billion and at least two years proving that AI augmentation works but AI replacement doesn't.

🔥 Spark's Hot Take: Epic's experiment will fail in its current form, but the lessons will be valuable. The mistake isn't using AI—it's going "AI-first" instead of "AI-augmented." The optimal workflow isn't humans reviewing Claude's output; it's humans architecting systems and Claude filling in the implementation details. Epic's approach inverts this, essentially asking Claude to be the architect and humans to be the QA team. That's backwards. The real breakthrough will come when someone figures out the right division of labor: humans own vision, architecture, and creative direction; AI owns repetitive implementation, variations, and grunt work. That's a 3x productivity boost. What Epic's trying—AI owns everything, humans just supervise—is how you get indefinite delays.

4. Dead Startups Selling Workspace Data to AI Labs for Training

The tech undertaker industry has found a new revenue stream: selling failed startups' internal data to AI labs. SimpleClosure's Asset Hub platform brokers sales of workspace data—Slack conversations, Notion docs, Google Drive files, Figma designs—to foundation model companies hungry for training data. One founder netted $40,000 total from his crypto startup's code and workspace archives.

This is darkly brilliant and deeply troubling. Startup failures are up 26-56% annually, creating thousands of potential data sellers. AI labs get authentic workplace communication and problem-solving data that's far more valuable than scraped internet text. And desperate founders get five figures for data they assumed would be deleted.

Think about what's in your company's Slack: strategic discussions, competitive analysis, pricing negotiations, technical approaches, failed experiments, customer complaints, internal debates about product direction. That's catnip for training models, especially for startups building workplace AI or vertical SaaS tools. Your competitors' failed attempts are now training data for the models they'll use against you.

The ethical questions are obvious: Did employees consent to having their conversations sold? Do customers know their support tickets might end up in training data? The legal questions are murkier: Who owns this data after a company dissolves? Can former employees sue?

For operating companies, the implications are clear: Assume everything you write in company tools will eventually be public or used to train models. Update your data retention policies. Consider what happens to your data if you shut down. Include data destruction clauses in your dissolution plans. And maybe think twice before discussing your actual competitive moat in Slack.

5. Andreessen Horowitz Expands GPU Cluster, VCs Become Infrastructure Providers

Andreessen Horowitz is expanding beyond its initial 20,000-GPU cluster for portfolio companies, potentially through two data center companies it backs. General Catalyst is building its own data center. Multiple VCs are using brokers to secure chips for startups. Venture capital is becoming infrastructure-as-a-service.

This started as a marketing gimmick—a16z's GPU cluster was announced with great fanfare in 2024 as a way to attract AI founders. Now it's table stakes. If you're raising a Series A for anything AI-related, your choice of lead investor increasingly depends on whether they can get you H200 access before your competitors ship.

The implications cascade: VCs with compute clusters can offer better terms because they're bundling capital with scarce infrastructure. Their portfolio companies train bigger models faster. They attract better founders who know compute access is the bottleneck. This creates a flywheel where compute-rich VCs win the best deals, which generates more returns, which funds more infrastructure.

For founders, this changes your fundraising calculus entirely. That Series A term sheet with the highest valuation might be worth less than the one that comes with guaranteed GPU allocation. That corporate venture arm that seemed boring suddenly looks attractive if they can offer Azure or GCP credits at scale.

We're also seeing the emergence of a two-tier startup ecosystem: Those backed by compute-rich VCs can iterate faster, experiment more, and train larger models. Everyone else is stuck in queues, rationing experiments, and making do with smaller models. This isn't just about money anymore—it's about access to the physical infrastructure that determines what you can build.

Bottom Line

The AI boom's first bankruptcy arrived not with a dramatic implosion but with a whimper: great product, huge usage, impossible unit economics. Character.AI proved that engagement means nothing if every interaction loses money. Meanwhile, the survivors are making hard choices—Microsoft admitting OpenAI isn't enough, Epic betting the company on AI-first development, VCs building data centers, and an entire industry emerging to sell dead startups' data exhaust. The question that ties all five stories together: When does the AI business model actually work? Not in 2028. Not when models get cheaper. Not when usage scales. Right now, with current costs and current prices, who's making money? Because the companies that can't answer that question just became the training data for the ones that can.