Mythos Leaked, Sora Shut Down, SoftBank's 2x Leverage: A Week of AI Shockwaves

Source: 20VC with Harry Stebbings | Published: 2026-04-02T14:01:24Z

Anthropic is building Mythos, a 10-trillion-parameter model so powerful they didn't dare release it publicly — until a security lapse leaked it anyway, proving that in the agent era, safety holes will only multiply.

A model that supposedly rewrites the rules of cybersecurity, leaked because of a security lapse. The irony is almost too perfect.

In the Vibe Coding Era, Leaks Will Only Multiply

Jason Lemkin's take: get used to it — this will happen more and more. His logic is straightforward. Vibe coding has sent shipping velocity through the roof. When speed goes up, corners get cut. Security is the easiest corner to cut. Mountains of code are already being pushed to unsecured GitHub repos; Supabase databases default to public and nobody changes them. Once AI agents start making decisions for humans — where code lives, what security level to use — leaks become the norm.

He did the math: Agents are a thousand times more productive than humans. Even if their error rate is one-tenth of a human's, the absolute number of errors is still a hundred times greater.

He cited his own time at Adobe. Back then, Adobe treated source code like the crown jewels — strictly forbidden from the cloud. His team won the first exemption to use GitHub. With local code management tools, a single release took a month; on the cloud, they could do sixty a day. The tradeoff between speed and security has always existed — the agent era just amplifies that tension by orders of magnitude.

OpenAI Kills Sora: A Long-Overdue Strategic Correction

OpenAI officially shut down Sora. Months ago they were trumpeting how amazing it was; now they've pulled the plug. The numbers tell the story: Sora's revenue was in the low single-digit millions, while it burned compute measured in the billions. Video generation is one of the most compute-hungry AI applications, with monetization potential that's essentially zero.

Rory O'Driscoll frames this in a bigger picture: available compute is still a scarce resource, and the economists and accountants have entered the room. Their logic is simple — scarce resources should go to the highest-value use case. Code generation also eats compute, but at orders of magnitude less than video, and there's real enterprise willingness to pay behind it.

Killing Sora says one thing: OpenAI's all-in bet on the consumer market had a problem. Meanwhile, Anthropic never touched image or video generation from the start. This is a massive strategic retreat — though the retreat itself may be the right call.

OpenAI Has Only Two Paths Left

Jason's read on OpenAI's situation boils down to a brutal minimum: the company has exactly two things it can do, and should only do those two things.

First, make advertising carry the consumer side. OpenAI has roughly 500 million unique users with about a 5% paid conversion rate — enough to support a $10–15 billion consumer subscription business. But to justify its current valuation, subscriptions alone won't cut it. Facebook and Google each pull in over $200 billion a year in digital advertising. If OpenAI can't reach $20 billion in ad revenue within a few years, it's not even in the game. To truly match its market cap, it probably needs $50–70 billion. Advertising isn't a "nice to have" — it's existential.

Second, catch Anthropic in coding and the enterprise market. They're already late, but there's no alternative.

Jason says at least they've finally moved from "wandering around the forest thinking they're cool and building a bunch of useless stuff" to "we have two things to do, let's do them." Better late than never.

The Wall Street Journal Lifts the Curtain on OpenAI's Internal Drama

The Wall Street Journal this week published a detailed account of Dario Amodei's departure from OpenAI. The internal tension was staggering: Dario and his sister refused to work with Greg Brockman — wouldn't even speak to him, wouldn't let him touch the LLM or GPT teams. Sam Altman had to repeatedly tell everyone "you're the boss" — telling Dario he was in charge, then telling Ilya and Greg they could fire Sam anytime. Dario's condition for staying was reporting directly to the board, not to anyone else. Then Sam got fired and brought back. Then Sora. Then "we don't do coding."

"This man spent enormous amounts of time mediating dramatic conflicts between extremely intelligent people — that alone can consume the vast majority of a CEO's bandwidth."

Rory pointed out that this level of executive and board-level turmoil, at any non-founder-led company, would have prompted the board to sit the CEO down for a serious talk long ago. By contrast, Anthropic's founding team has been stable and aligned. The gap in execution is almost inevitable.

Masa Son's Big Bet: 2x Leverage on AI

SoftBank secured a $40 billion bridge loan to buy OpenAI shares. How deep will Masa Son go? The answer: as deep as it gets.

SoftBank Group currently runs at roughly 1.5 to 2x leverage. That means a 30–40% decline in holdings could wipe them out. Think of it this way: take a $900 million venture fund, borrow $1.8 billion, deploy it all — win and you double your returns, lose 30% and you're done.

That said, Masa Son has survived before. In 2002, the Nasdaq fell 85% — "if you haven't lived through an 85% index drawdown, you haven't really lived." His two core holdings are OpenAI and ARM, both world-class assets. But even the best assets can drop 30% — that's well within the realm of possibility.

The Cybersecurity Selloff Is Panic, Not Logic

After the Mythos leak, CrowdStrike, Palo Alto, and Zscaler dropped an average of 6%. Okta and Netskope fell 7%. SentinelOne dropped 9%.

Rory broke down each company's business: Anthropic's security capabilities are primarily about finding vulnerabilities during the code development phase, which does pose a real threat to application security and code review companies. But Okta does single sign-on and identity management. CrowdStrike does real-time perimeter defense. Those are entirely different businesses from what Anthropic does. Both categories dropping together means the market is having a knee-jerk reaction.

Jason's view is even more bullish: this is cybersecurity's golden age. In the agent era, security threats are growing by orders of magnitude. Agents build applications in unpredictable ways, developers don't read the code, features ship at unprecedented speed, and corners are being cut everywhere. Any cybersecurity company — application layer, perimeter defense, or otherwise — with a capable engineering team should benefit from this wave.

"If you have any established brand and you call saying 'we have a new agent security product to help defend against this threat,' you'll get a meeting that same afternoon."

Anthropic itself is capitalizing on this fear for marketing — offering the Mythos model first to enterprise CISOs, essentially saying: "Look, here's the terrifying new weapon we built. Give us a million dollars and we'll help you defend against it."

The ARR Definition Game and the Infinite Token Resale

Anthropic's and OpenAI's ARR definitions are actually quite clear: take the average actual revenue from the past four weeks and multiply by 13 (there are 13 four-week periods in a year). This is real GAAP revenue, not commitments or contracts. By this metric, Anthropic is around $19 billion and OpenAI around $25 billion.

But there's a key difference: when OpenAI sells through Microsoft, it reports net revenue (after Microsoft's cut), while Anthropic reports gross revenue when selling through AWS (and then pays Amazon back 20% as cost of sales). Same type of revenue, two different accounting treatments.

Even more interesting is the token resale problem. Anthropic sells tokens to AWS. AWS sells them to Cursor. Cursor sells them to users. Every layer books the same tokens as its own ARR. As long as everyone accepts near-zero gross margins, an infinite number of companies can keep reselling the same tokens to each other. How long does this game last? Until someone actually has to turn a profit.

Does Emergent Labs' "8 Months to $100M" Hold Up?

Emergent Labs claims it hit $100 million ARR in eight months — the number is splashed across its homepage. Jason actually tried the product and found it genuinely good. In his six-criteria benchmark test, it even beat all major competitors on the homepage replication test, including Replit, Lovable, and v0.

What he didn't like was the billing. The product nudges you into clicking a Stripe link for "$0 first month, $20/month after" while you think you're on a free tier. That means they're almost certainly booking the full $240 annual ARR in the first month when the user has paid zero. This isn't unusual among AI startups, but if you put that number on the biggest banner on your website, you'd better be ready for scrutiny.

Tranched Rounds: The Game Theory of VC Valuations

Harry Stebbings voiced clear frustration with tranched pricing rounds. Here's how it works: a hot company raises money where the lead investor puts some capital in at a $250 million valuation and the rest at a $1 billion valuation, for a blended cost of roughly $600 million. But the number announced publicly is always "a $1 billion valuation."

Rory unpacked it: if you're a follow-on investor who can only get in at $1 billion while the lead's blended cost is $600 million, you're paying 50% more for the same asset. That's the price of admission — the premium for getting a seat at the table.

The deeper problem: by accepting the blended price, the company implicitly admits it's only worth $600 million but wants the optics of $1 billion. If the next round doesn't clear $1.5 billion, it's technically a down round on paper — and for a company that runs on optics, that could be worse than having honestly reported $600 million in the first place.

Oura and Whoop: Can Consumer Hardware Deliver Venture-Scale Returns?

Oura and Whoop both made news this week — Whoop raised $500 million at a $10 billion valuation. These wearable health products have real subscription revenue and genuine consumer value, and they're unlikely to be replicated by someone with Claude Code on a Friday afternoon.

But Peter Thiel's framework creates anxiety here: competition is for losers; monopoly drives innovation. Can the wearables market be monopolized? Jason says he's been running for ten years, five miles a day — if a better device comes along, he'll switch. Users are loyal but replaceable.

Rory's counter is practical: not every business needs a five-year lock-in. Coca-Cola faces the risk of consumers switching to Pepsi every single day. Amazon Retail could lose to Walmart on any given purchase. These companies have endured for decades. Capitalism is hard — if you want to create $10 billion in value, you have to keep delivering value to consumers. That said, he concedes that manufacturing a modular electronic device worn on a finger that measures blood oxygen has a significantly higher technical barrier than making a pair of sneakers. The number of companies capable of building wearables is far smaller than those that can make athletic shoes.

Manus Founders Trapped in China: The End of Singapore Washing

Manus — the company acquired by Meta, originally founded in China, later redomiciled to Singapore, backed by Benchmark, and effectively operating as a US-Singapore entity — now has its two core founders barred from leaving China by the Chinese government.

Jason's take is blunt: this kind of "Singapore washing" is over. There won't be another deal like this. Rory added another dimension: maybe not many people in America noticed this, but every Chinese founder currently considering a similar move did. Their calculus is now: "If I do this deal, I might never go home again. If my family is still in the country, how much exposure do I have?"

For Meta, $2 billion isn't a huge number, and the product still works. But when you lose the founders, you lose the soul of the company — even if you can't see the impact in the short term.

California's Billionaire Exodus: Killing the Golden Goose

Steve Jurvetson left California and bought the most expensive house in Incline Village. Howard Schultz left Washington State. Washington passed a 9.9% state income tax on millionaires this week, with the governor's reasoning being "they should pay a bit more."

Rory made the decision framework concrete: if you have regular taxable income year after year, relocating means uprooting your life — not worth it for most people. But if you have a one-time massive capital gains event — say, selling all your SpaceX stock, realizing $2 billion in gains, with California taking an extra 13%, or $260 million — you might sit down with your spouse and say: "Let's spend 165 days a year in Incline Village for the next two years. I'll cover the flights. I'll come back every weekend. You won't lose a single friend. We save $260 million."

Jason's worry is the longer-term consequence. California's billionaire tax proposal assumes it can collect huge sums from Larry Ellison — but Ellison moved away five years ago. When you try to collect $200 million and end up with zero, someone in Sacramento will eventually cut a line from the budget. That line won't be teacher salaries. It won't be firefighter pay. It'll be marginal services for marginal populations.