20VC x SaaStr This Week: Anthropic Sues the Government, The Death of the Junior, and Why “Gentle Deceleration” is Over

The latest from the 20VC x SaaStr collaboration with Harry Stebbings, Jason Lemkin, and Rory O’Driscoll

Anthropic sued the federal government this week. Oracle and OpenAI scrapped plans to expand their flagship Stargate data center. Meta immediately said “we’ll take it.” CrowdStrike beat expectations and traded down anyway. And we made our public market stock picks for the year.

But the thing that matters most isn’t any one headline. It’s the convergence of three forces that every B2B founder, investor, and operator needs to understand right now: the economics of AI compute are colliding with the reality of who pays for it, the junior employee is being eliminated from every function at the same time, and the public markets have completely abandoned companies that aren’t reaccelerating.

The era of gentle deceleration is dead. Here’s what replaces it.

Top Takeaways

1. Anthropic Will Probably Win the Law. They Won’t Win the Fight.

Anthropic filed two federal lawsuits against the Trump administration after the Pentagon designated the company a “supply chain risk,” a label typically reserved for foreign adversaries. At its core: Anthropic refused to let Claude be used for autonomous weapons or mass domestic surveillance. The Pentagon said that’s not the company’s call to make.

The legal consensus is that the government overreached. Designating one of America’s fastest-growing tech companies as a supply chain risk because it exercised its First Amendment rights is constitutionally shaky. And the most extreme version, where the government told Microsoft and Amazon that using Anthropic at all would disqualify them from government contracts, was so aggressive that the hyperscalers pushed back and it was walked down.

But here’s the B2B reality that matters for founders: the $200 million Pentagon contract is noise. Anthropic is reportedly doing $1.5 billion in run rate and 10x-ing. The DoD revenue is roughly 1% of the business. The real damage is in B2B sales cycles. As Anthropic stated in their own complaint, they’re struggling to close deals because prospects are worried about any federal government exposure. And competitors (read: OpenAI, which signed its own Pentagon deal the same day Anthropic was blacklisted) are using it to steal deals.

This is a textbook B2B sales problem. Risk doesn’t need to be real. It just needs to be ambiguous. If your prospect has any federal government exposure and your competitor can say “we don’t have these problems,” you lose. Period.

The most likely outcome: Anthropic wins on the law, bends the knee on the politics, and the two sides reach a settlement where the company keeps its broader B2B business but gives up on selling directly to the Pentagon. It won’t show up in the numbers at the current growth rate, and the IPO (which Polymarket now says won’t happen this year) probably proceeds whenever the legal overhang clears.

The broader lesson for every B2B founder: your principles can cost you a lot more than the contract they’re attached to. The blast radius of a government dispute extends far beyond government revenue.

2. The Stargate Expansion Died, Meta Immediately Absorbed It, and We’re Definitely Overinvesting. Nobody Cares.

Oracle and OpenAI scrapped plans to expand the Abilene, Texas Stargate data center from 1.2 gigawatts to 2 gigawatts. Financing talks dragged, OpenAI’s needs shifted to next-gen Nvidia Vera Rubin chips, and the power grid wasn’t going to be ready for over a year.

Within hours, Meta moved in to lease the surplus capacity, with Nvidia paying a $150 million deposit to broker the deal.

Is this the beginning of the end of the capex cycle? More no than yes. Because the demand side of this equation is barely getting started.

Here’s the framework: the hyperscalers are betting on a world where your AI runs 24/7, persistent and infinite. Right now, most of us use a couple hours of ChatGPT or Claude per day, maybe some coding compute. But the world where your AI debriefs you after every meeting, coaches you on every decision, and runs 10 to 50 agents in parallel for code review, QA, and outbound sales? That requires orders of magnitude more compute than exists today.

Anthropic’s new Claude Code review feature illustrates this perfectly. It spins up 10+ agents in parallel, runs for 20 minutes, finds every bug in your codebase, and costs $15 to $25. The internet lost its mind at the price. But the response from the team that built it was essentially: we’re giving you automated code review that humans literally cannot do this fast, and you’re complaining about $20?

The real question isn’t whether $20 per code review is expensive. It’s that you’d ideally run this after every single commit, not episodically. That’s orders of magnitude more compute than any developer uses today. And code review is just the beginning. Who’s doing the QA? Who’s doing the security audit? We haven’t even scratched the surface.

The math on overinvestment is simple. $600 billion in capex. 150 million workers in the United States. That’s roughly $4,000 per head. A lot of those workers make coffee and don’t need $4,000 of AI compute. At some point, the chickens come home to roost.

But game theory says it doesn’t matter yet. Every hyperscaler has explicitly said: “This is probably going to go wrong, but I’d rather ante up and have a chance of winning than sit out like Apple and know I’m going to lose.” When six or seven players all articulate that logic, overinvestment is the guaranteed outcome. The question is who folds first. Oracle, with the weakest balance sheet and the least compelling use case, is the obvious candidate. Meta can borrow. Google has a real business. Microsoft is quietly stepping back from the table.

This game goes on for at least another year or two.

3. The Death of the Junior Is Where Data Center Budgets Come From

The enterprise is willing the replacement of humans into existence. And it’s accelerating because they want it to accelerate. The category getting hit hardest and fastest: junior roles.

No one wants to hire junior developers. No one wants to train junior lawyers for two years before they can work on a real case. No one wants junior SDRs who don’t know the tools. And the evidence is everywhere.

At Penn State (not a top-10 school, but a large, well-respected state university), there are essentially zero tech recruiting visits for CS and math grads. Six students in the class have offers, because they’re publishing research on Jensen fine-string theory that most humans can’t comprehend. Everyone else? Nothing. Not even the C-tier companies are showing up.

This isn’t just anecdotal. The “hire no juniors” movement is being willed into existence across every function. Support? You want experienced humans who know the product cold to handle escalations. Everything else gets automated. Marketing? You don’t need junior content writers when agents can generate and optimize campaigns. Sales? SDR teams are being replaced by AI agents that text prospects, pitch the product, and set up meetings that are fully ready to close.

The economic connection to data centers is direct: when you hire no juniors, you suddenly have budget for $200 to $500 per month in AI compute per senior employee. That senior developer doesn’t need a junior to review their code. They need Claude Code at $200/month plus $20 per code review. The junior’s salary becomes the AI budget.

This should worry everyone. History (including the French Revolution) tells us that dispossessed urban poor can exist indefinitely without political consequence. But when you produce masses of college graduates who played by the rules for 20 years, spent $400,000 on education, and face unemployment? That cohort causes trouble. In 2026 this is going to be an economic issue. In 2027, it’s going to be a political issue.

The technology takes longer to diffuse than Silicon Valley allows, and people are more adaptable than we give th