AI in 15 — April 02, 2026

Investors are dumping OpenAI shares like they're radioactive, and nobody's buying. Six hundred million dollars worth of stock sitting on secondary markets with zero takers. Meanwhile, Anthropic shares? Banks are literally fighting over them.

Welcome to AI in 15 for Thursday, April 2, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Happy Thursday, Marcus. Today we're following the money as Wall Street pivots hard from OpenAI to Anthropic. Forbes published a brutal inventory of OpenAI's abandoned products and dead deals. Oracle just laid off thirty thousand people with a six a.m. email. Salesforce is turning Slack into an AI surveillance platform. A hospital CEO says he's ready to replace radiologists with AI right now. Meta open-sourced a model that makes concrete cure faster and pollute less. And non-technical clients are using AI to rewrite production codebases over the weekend. Let's get into it.

Wall Street dumps OpenAI shares as Anthropic becomes the hottest ticket in AI.

Forbes catalogs the OpenAI graveyard, and the Sora numbers are staggering.

And Oracle fires thirty thousand people by email at six in the morning.

Marcus, as we reported yesterday, the Claude Code source leak was embarrassing for Anthropic. But apparently Wall Street doesn't care about embarrassment when the revenue numbers look like this. What's happening on secondary markets?

It's a dramatic reversal, Kate. About half a dozen institutional investors, hedge funds and venture capital firms, have approached secondary marketplace Next Round Capital trying to offload roughly six hundred million dollars in OpenAI shares. Last year, those would have been grabbed within days. Right now, there are literally no buyers. Zero.

And on the Anthropic side?

The opposite problem. Too many buyers, not enough sellers. Hiive alone has registered over one point six billion dollars of demand for Anthropic shares. Secondary market bids are valuing Anthropic at roughly six hundred billion, more than fifty percent above its last formal round at three hundred and eighty billion. And here's the detail that tells you everything. Goldman Sachs and Morgan Stanley have waived their carry fees on OpenAI shares to try to move them, while Goldman is still charging the standard fifteen to twenty percent carry on Anthropic equity.

When the banks give up their cut just to move the paper, that's a signal.

A loud one. Anthropic is reportedly considering an IPO as soon as Q4 this year, with bankers expecting it could raise more than sixty billion. Their revenue has grown fourteen X year over year to a fourteen billion dollar annualized run rate. And ironically, the source code leak may have actually helped the narrative. Investors saw serious engineering depth in that codebase, even if they weren't supposed to see it.

So the leak was accidental marketing?

Unintentional due diligence. The anti-distillation systems, the autonomous agent framework, the forty-four unreleased features. That's a product roadmap investors would normally pay consultants to estimate. They got it for free.

Now, the OpenAI narrative gets worse. Forbes published what they're calling the OpenAI Graveyard. This went viral on Hacker News, two hundred and twenty-five points, nearly two hundred comments. Marcus, what's in the graveyard?

The headline casualty is Sora, OpenAI's video generation tool, shut down March 24. The economics were catastrophic. Fifteen million dollars a day in inference costs at peak usage. Total lifetime revenue from in-app purchases? Two point one million dollars. Not per day. Total. Downloads peaked at three point three million in November and crashed sixty-six percent by February.

Fifteen million a day in costs against two million total in revenue. That's not a business model, that's a bonfire.

And the planned billion-dollar Disney deal for Sora? Dead. No money ever changed hands. OpenAI cited "compute reallocation" and a pivot to world simulation for robotics. The Hacker News discussion was pointed. Multiple commenters identified a pattern where OpenAI optimizes for the demo, not the sustained interaction. Making products impressive for five minutes but failing to build lasting user value.

And this comes alongside projections that OpenAI will lose fourteen billion dollars in 2026.

Despite generating roughly twenty-five billion in annualized revenue. The compute costs, infrastructure spending, and employee compensation are just enormous. That hundred and twenty-two billion dollar fundraise we covered yesterday? It's financing a burn rate that would make most CFOs physically ill. The counterargument is real, ChatGPT is approaching a billion weekly active users. But the gap between revenue and profitability raises serious questions about whether first-mover advantage translates into a durable business.

The graveyard plus the burn rate plus the secondary market collapse. That's a rough week for OpenAI's narrative.

Especially when your main competitor's stock is trading at a fifty percent premium to its last round.

Let's talk about Oracle, because this story is about the human cost of the AI infrastructure race. Thirty thousand layoffs, Marcus. How did they handle it?

Badly. Workers received a six a.m. email from "Oracle Leadership" with no prior warning from HR or direct managers. No meetings, no phone calls. Just an email at dawn telling you your job is gone. The cuts are roughly seventeen to twenty percent of Oracle's global workforce, spanning the U.S. and India, with nearly five hundred jobs lost in the Seattle area alone. Restructuring costs are expected to hit two point one billion in fiscal year 2026.

And the reason is AI infrastructure spending.

Directly. Analysts at TD Cowen had forecast earlier this year that Oracle would need to cut twenty to thirty thousand positions to finance its expanding AI datacenter buildout. The company is making a massive bet on Oracle Cloud Infrastructure becoming a major player in AI compute, a market currently dominated by AWS, Azure, and Google Cloud. They're firing the present workforce to fund the future one.

Atlassian did the same thing this week, right? Sixteen hundred workers.

Ten percent of their staff, redirected to AI. It's a pattern across the industry. The AI boom is creating jobs, but it's destroying different ones at a faster pace. And the way Oracle handled the notification, that cold six a.m. email, has drawn fierce criticism. When you're making people's lives worse to fund your AI ambitions, the least you can do is tell them in person.

Salesforce is going the other direction, pouring AI into everything. They just unveiled thirty new AI features for Slack. But Marcus, one of these features has me concerned.

You're talking about the desktop monitoring. Slackbot can now operate outside of Slack and observe desktop activities, drawing on data from deals, conversations, calendars, and user habits to make suggestions and draft follow-ups. Your workplace chat app is now watching what you do on your computer.

Even if it's opt-in, that's a significant escalation in enterprise surveillance.

It is. But the technically interesting part is the MCP integration. Slack is now a Model Context Protocol client, meaning it connects to external services using the same open standard Anthropic developed. Slackbot can coordinate with Salesforce's Agentforce platform and potentially any MCP-compatible tool. They've also introduced reusable AI Skills, where you define a task like "plan an event" and Slackbot executes it across multiple systems automatically.

So Salesforce is positioning Slack as the AI command center for enterprise work.

Directly competing with Microsoft Copilot. The battle for enterprise AI is now fundamentally about which platform becomes the hub where AI agents live and operate. Salesforce is betting that's Slack. Microsoft is betting that's Teams and Office. And both are willing to push privacy boundaries to win.

Here's one that's going to spark debate. The CEO of the largest public hospital system in America says he's ready to replace radiologists with AI. Marcus, what exactly did he say?

Mitchell Katz, who runs NYC Health and Hospitals, said at a forum, quote, "We could replace a great deal of radiologists with AI at this moment, if we are ready to do the regulatory challenge." He proposed AI handling first reads on mammograms and X-rays, with radiologists only double-checking abnormal results.

And radiologists were not pleased.

A San Diego radiologist called it "undeniable proof that confidently uninformed hospital administrators are a danger to patients." The Hacker News discussion was fascinating because multiple commenters pointed out that predictions about AI replacing radiologists have been made since the late nineteen eighties. One commenter's brother-in-law was told not to enter radiology in the early nineties because computers would take over.

Thirty-five years later, still waiting.

The regulatory question is the real barrier. Who bears liability when AI misses a cancer? That's not a technology problem, it's a legal and institutional problem. And it's worth noting the motivation here. Katz runs a public hospital system. He's talking about cost savings. The clinical accuracy debate is almost secondary to the economic pressure driving this conversation. But I'd want to see a lot more validation before trusting first reads to any AI system, no matter how good the benchmarks look.

Let's pivot to something genuinely positive. Meta open-sourced an AI model for designing better concrete. Marcus, concrete and AI, not an obvious combination.

But a brilliant one. The model is called BOxCrete, Bayesian Optimization for Concrete. It designs optimized concrete mixes that reach full structural strength forty-three percent faster while reducing embodied carbon by thirty-five percent. Meta developed it with the University of Illinois and Amrize, the largest cement manufacturer in North America. They've already deployed it at a Meta data center in Minnesota.

And concrete is a massive carbon emitter.

About eight percent of global CO2 emissions. If you can cut that by a third at scale, that's genuinely meaningful for climate. Meta releasing the model and training data as open source means any concrete producer can use it. The Hacker News discussion was enthusiastic, with commenters noting how underappreciated concrete science is. This is AI applied to atoms, not just bits, and it's one of the most tangible real-world applications we've covered.

Last story, and this one's a wild ride. A Hacker News thread about non-technical clients using AI coding agents to rewrite production codebases. Marcus, how bad is it?

A developer posted that their client remodeled half the codebase over a weekend using Claude Code. Ten thousand plus lines of AI-generated code pushed without review. Another commenter reported a client who did the same thing and then deleted their production database.

Deleted the production database.

The practical advice in the thread was solid. Enforce pull request workflows, limit direct push access, require code review regardless of who or what wrote the code. But the philosophical point is bigger. AI coding tools are now good enough that non-developers can produce functioning code, but not good enough to guarantee it's maintainable, secure, or won't destroy your data.

One commenter said "prepare to make absolute bank on maintenance charges."

The developer's role is shifting from "person who writes code" to "person who ensures code doesn't burn the house down." Whether a human or an AI wrote it.

Thursday big picture. Investors are fleeing OpenAI for Anthropic. Oracle fires thirty thousand people to fund AI data centers. Salesforce wants Slack to watch your desktop. And clients are rewriting production code with AI over the weekend. Marcus, what's the pattern?

The AI industry is sorting itself into winners and losers, and the criteria aren't what people expected. It's not who has the most users or the biggest fundraise. It's who has sustainable economics and genuine technical depth. OpenAI has a billion users and can't find buyers for its stock. Anthropic has a leaked codebase and investors lining up around the block. Oracle is sacrificing thirty thousand jobs on a bet that may not pay off. And non-technical clients are proving that access to AI tools without engineering judgment is a loaded gun. The easy money phase of AI is over. What comes next rewards execution, not hype.

Execution over hype. A refreshing change if it holds.

The market is finally doing what markets do. Separating signal from noise. It just took six hundred billion dollars to get here.

That's your AI in 15 for Thursday, April 2, 2026. See you tomorrow.