Anthropic vs The Pentagon: Who Wins? | The Data Center Arms Race | The Ultimate Stock Picks

Summary

• Anthropic probably wins a good slug of the law, but may lose the fight. The panel’s consensus on Anthropic suing the US government over its “supply chain risk” designation: “Anthropic would probably win a good slug of this case, which is different than saying they’re going to win the fight.” Jason’s tariffs analogy — even if a court reverses the designation, the Department of War “will just come back with another issue… it’ll be endless until they bend to the will.” Expected endgame: probably win in court, then “bend the knee and be slightly supplicant” in exchange for being left alone in B2B.
• The real damage is B2B sales, not the $200M contract. Anthropic’s own complaint says deals are being cut in half because prospects fear federal exposure while OpenAI and xAI say “we don’t have these problems” — “the risk only has to be ambiguous in a B2B deal for you to lose.” But financially it’s noise: $200M ≈ $17M/month against ~$1.5B, roughly 1% for a company 10x-ing. On the IPO, Rory dismisses the panic: “Until it shows up in the numbers, I don’t think fear overcomes greed.”
• The Stargate Texas cap is not the capex turn — “more no than yes,” because Meta instantly offered to absorb the capacity and Jensen’s demand “is insatiable.” Rory still thinks we’re over-investing, but calls it tautological: six or seven players are explicitly running that logic — “I’d prefer to ante up and be in the game… than not ante up like Apple and know I’m going to lose” — so over-investment continues through 2026. “The only person that’s even struggling slightly to cover their nut is Oracle,” which will “throw 20 or 30,000 people over the side to make the P&L work,” while “Microsoft is quietly stepping back from the table.”
• Compute demand is still orders of magnitude under-built, and juniors are paying for it. The panel’s thesis: 24/7 persistent AI agents plus Claude’s new $15–25 code review (10+ agents in parallel for 20 minutes, ideally run after every commit) imply compute needs “orders of magnitude” beyond today — the constraint is affordability, not capability. And the budget source is explicit: “I think getting rid of juniors is where we get budget for these data centers in part.” Rory predicts the graduate-unemployment squeeze becomes “a political issue” in 2026.
• “The era of gentle deceleration has ended. It’s dead.” Public markets have stopped tolerating managed slowdown: CrowdStrike beat and traded down on 27%→23% guidance while Cloudflare re-accelerated 27%→34%. Rory’s framing — markets finally “corrected to fundamentally what they’re worth with no pixie dust,” with 8–9x EBITDA examples for value stocks. Jason extends it to venture: “you got to start wondering this year when you should just give up on your portfolio if they’re not accelerating.”
• Stock picks. Jason, momentum-only: Palantir, Cloudflare, Shopify, CrowdStrike (Snowflake dinged “just on not being founder led”; Figma killed by his own product test). Harry: CrowdStrike, Nubank (28% growth, entering the US), Nvidia (“I don’t [__] sell Jensen”), and Reddit down 40% as “a data layer for a lot of the LLMs.” Rory, in three buckets: value Salesforce at 8–9x EBITDA and team at 9x EBITDA, GARP Toast and Intuit, and “with fear” CrowdStrike and Palantir — the latter “one more year of growth away from being normalized” with an administration that “is going to spend for 2 years.”
• Incumbents face a lab-experiment moment. Wix’s Base44 hit $100M ARR but the core is shedding paid customers at 1.2%/year — “if Wix can’t cross-sell the new AI product to 6.11 million customers, what hope is there for you?” Jason’s Figma Make test was “the worst vibe coding experience I’ve had in 6 months,” and the general verdict is brutal: “best efforts quarterly releases — that’s death today,” with “1.something trillion dollars worth of public companies and another trillion of private” facing the same existential crisis.

Deep dive

1. Anthropic v. the Pentagon: “like all good disasters, both sides are doing things wrong”

• Rory’s context: Anthropic sued the federal government in California and DC on procedural grounds after being designated a supply chain risk over its refusal on a ~$200M Department of War contract. He grants the government’s starting position — “You want $200 million of revenue from us, and you want to get in our shorts and tell us what to do… you’re not getting your $200 million. That’s totally fair” — but the designation carries escalating consequences: no vendor selling to the DoW can use Anthropic, then exclusion from every government contract, and in one floated version, “if you use Anthropic at all, Microsoft or Amazon, then we won’t use you at all in the US government” — “almost ludicrous overreach.”
• Jason’s read is bleaker: “It’s like tariffs.” Even if Anthropic wins the case, “the Department of War will just come back with another issue… a different type of supply chain risk. It’ll be endless until they bend to the will.” He adds the startling framing of the government’s position: “We used Palantir without Anthropic’s knowledge to conduct a war in Venezuela and depose a government… you’re either going to agree to this, or we are going to freeze you out of the US economy.”
• The complaint leans on the First Amendment — Anthropic arguing it’s being punished for articulating “less than love” for military use cases. Rory, critical of Anthropic last week and standing by it, still calls the response un-American: “It’s probably a mistake to pound the crap out of one of the two leading companies in that space simply cuz at the margin you find them slightly sanctimonious… stop trying to cut them off at the knees across the entire B2B infrastructure.”

2. Probably win the law, may lose the fight — then bend the knee

• The consensus line of the episode: “In law, Anthropic would probably win a good slug of this case, which is different than saying they’re going to win the fight.” Rory’s predicted endgame, from watching other institutions tangle with this administration: probably prevail on the law (Anthropic has moved for immediate redress), then “some kind of acquiescence… in return for leave us alone to pursue the rest of our daily business.” “I’m sure Anthropic deeply regrets getting down in the mud pit wrestling with this particular opponent.”
• Jason identifies the real wound as classic B2B panic: the complaint itself says deals are being cut in half because prospects with any federal exposure are spooked, and competitors are saying “we don’t have these problems.” “The risk only has to be ambiguous in a B2B deal for you to lose.”
• On the IPO (Polymarket now predicts Anthropic doesn’t go public this year), Rory thinks existential-IPO-risk talk is media overstatement: “IPOs are very binary” — massively oversubscribed or barely done — and “until it shows up in the numbers, I don’t think fear overcomes greed.” Pushed by Harry on whether it shows up: no. The math: $200M is ~$17M/month against ~$1.5B — about 1% for a company 10x-ing. “It’s lost in the noise.” “If you’re the sales rep for the Pentagon and Anthropic, it’s going to be a slow week” — but PLG on the Claude app “is exploding.”
• Jason’s zoom-out worry — these dilemmas generalize: every next-generation CRM logs every keystroke because “the next generation CRM just does not work if you have any privacy in the workplace at all,” the Gong pause all over again. “As we chase these faster and faster growth companies… we’re going to discard more and more of our previous moral standards. Maybe that’s the AI life, but I have a little bit of anxiety here.”

3. The Stargate cap isn’t the capex turn — game theory guarantees over-investment

• Oracle and OpenAI reportedly capping the Stargate Texas expansion at 1.2 gigawatts. Early sign of the cycle ending? Rory: “More no than yes” — “instantly Meta said, ‘if you guys don’t want this data center, we’ll take it,’” and per Jensen’s comments, “the man is insatiable.” Reading a one-off Oracle dynamic as the turn “would be overextrapolation.”
• Yet Rory still believes in over-investment — he just refuses “to call the turn 3 years before it hits.” His mechanism is game theory: Rory’s logic is “this is probably going to go wrong, but I’d prefer to ante up and be in the game and have a chance of winning than not ante up like Apple and know I’m going to lose.” With only six or seven players who matter all playing that game, “it is almost tautological to say we’re going to over invest… And I don’t think it’s happening in 2026.”
• The balance-sheet sort: “The only person that’s even struggling slightly to cover their nut is Oracle” — weakest balance sheet, least compelling use case, which is why “at some point they’re going to throw 20 or 30,000 people over the side to make the P&L work. Meta can borrow, Google actually has a business, and Microsoft is quietly stepping back from the table.”
• Harry’s exhibit for over-investment already happening: the rumor that Alex Wang is being sidelined at Meta, with a new lab director reporting to Boz and the Scale deal looking hasty. Rory pockets it: “I paid 15 billion dollars for an asset that 12 months later I put on the bench. That is the very definition of over investment.” But Meta’s shrug-and-try-something-else response shows “we’re not yet at the over-investment being terrifying stage.”

4. The compute thesis: 24/7 persistent AI, affordable to no one yet

• Rory’s core claim: Meta is betting on AI that is “24/7 persistent and infinite” — an agent that sees everything you do and debriefs you the second a call ends (“Hey Jason, that was something pretty dumb you just said with Rory”). The challenge — what tense are you using? — draws the crux: “That could be true today in everything you do in life. It’s just we cannot afford for the moment the level of compute it takes.” Multiply by parallel agents (Cursor, Claude Code, Replit V4 running 5–50 agents) and “we’re talking about multiple orders of magnitude of compute” — Rory concedes, half in humor, it “might well require data centers in space.”
• Rory’s structural diagnosis of the whole bull-bear stalemate is worth keeping verbatim: “The people who say it’s not talk technology, and the people who say it is talk economics — and we kind of just don’t quite overlap.” His back-of-envelope: $600B of capex against 150M US workers is ~$4,000 per head — “there’s a bunch of people serving coffee; they probably don’t need four grand’s worth of AI… I stand by my comment. I think we probably are over investing.”
• Rory believes Meta (with TikTok) is well positioned to be “the AI that lives with you 24/7” — “unless Claude, which is playing its own end game, or ChatGPT can figure out its social side” — and Oracle simply “doesn’t have the cash to keep going.”

5. The $20 code review: where technology meets economics

• Anthropic shipped an automated Claude code review at $15–25 a run and “the internet blew up” over the price. Jason’s defense: it spools up 10+ agents in parallel for 20 minutes to find every bug — “humans can’t even do this” — and the right cadence is after every commit, 10 times a day, “not getting up your nerve to spend 20 dollars worth of tokens.” That’s orders of magnitude more compute per developer than before: “we’ve just scratched the surface.”
• Rory supplies the answer Claude “was too nice to give”: “You’re automating an entire developer… and you won’t even pay 20 bucks to check it before you put in production. For God’s sake, man, pony up.” The general law: “The idea that you can just have all this [__] for free is at some point going to stop, cuz someone’s going to have to cover their nut.” Demand is infinite; the open question is what people will “pay full boat for.”
• The pricing architecture Rory sees emerging: metered API, an under-priced $200 Claude Code subscription for hackers, and managerial-budget line items on top — “you, Mr. Hacker, can have all the free code you like for 200 bucks. But if you’re rolling this out into Bank of America systems, you’re going to want the $250 security-proofed code review.” (Rory discloses they are happy investors in CodeRabbit — Claude’s entry “validates the space,” even if competition from Claude is bad.)

6. The death of the junior is the data center budget — and 2026’s political issue

• Jason’s blunt accounting: “I think getting rid of juniors is where we get budget for these data centers in part.” Enterprises are willing AI replacement into existence — “hire no juniors” in code, sales, support, marketing, law (“who the hell wants to wait for an associate to scale up for 2 years?”). His own anecdote: his son at Penn State reports “zero jobs for anyone in CS or math” — he has a job offer because he’s publishing research, and six people in his class have tech offers. “It’s happening in front of us, the death of the junior. Bad for society, but it’s a reality.”
• Rory’s evolution, explicitly flagged: he now concedes real unemployment impact in targeted demographics — entry-level comp-sci, customer support, legal associates — while rejecting the macro catastrophe. His sizing: ~800,000 US software developers; halve them and it’s 400,000 people, ~0.2% unemployment. “I have 200 years of facts on my side, and you have [__] — but maybe you’re right.” Technology diffuses slower than Silicon Valley allows, and people adapt.
• The historical warning cuts the other way: “You can have dispossessed urban poor forever, and nothing happens. But if you piss off the 20-something-year-old middle class, the overeducated elites, they tend to cause trouble.” Graduates who “played by the rules for 20 years” and hit mass unemployment are the French Revolution cohort — “I don’t know if it’s a moral issue or an economic issue, and I predict in ‘26 it’s going to be a political issue.” As someone tweeted about AI’s negative NPS: “You’ve been telling us you’re going to blow up the world and make us all unemployed, and you’re surprised to discover we don’t like you.”
• On colleges: unemployment still ranks college grads (3.something%) below high-school grads (4.4%), but “you take into account the net present value of that 400 grand, you’re not getting your money back on anything other than the best degrees.” Jason: universities are becoming “expensive babysitting for 4 years” — though he grants Harry’s pushback that even Andrej Karpathy took 6 months to go from 20% usage, so expecting curricula to track Claude’s skill curve is unreasonable.

7. People would rather buy an agent than hire a human — but everyone’s still a services business

• Jason’s accelerating theme: in mixed human-agent support environments, agent CSAT “is always top 10%… never number one,” and the deeper truth is preference, not just cost: “We’d rather hire agents, we’d rather buy an agent… It doesn’t even matter if it’s cost-effective.” Harry’s 3,000 EA applicants prove the point — “if he could have an agent, he’d pick an agent.” His best specimen: Monaco, Sam Blond’s GTM agent, “closed seven figures in their first 5 days, and they have 60 days of people lined up” — because it texts the prospect, pitches, and sets up a close-ready meeting. When you deliver that, “the budget magically comes out of nowhere… the TAMs for some of these products don’t even make sense to us.”
• Jason went further at home: “We built our own VP of marketing literally yesterday. This week was the first time it led our human staff meeting.” And on Clay: “as great as Clay is for enrichment — if it could do that, I’d give you an extra $50,000 tomorrow.” His indictment of public B2B: “I don’t want to make my humans 8% more efficient with your AI. I don’t want these humans — I want an agent to do this work.”
• The refinement of the narrative: until recently, Harvey/Legora-style adoption was task-level augmentation, “not a wholesale replacement story except very much at the margin.” Only in the last 6–9 months, starting with software, has true agentification arrived — “the next 12 months will be B2B companies saying, ‘Now I finally can automate this task entirely.’ And then we’ll find out what the demand is.”
• Jason’s under-appreciated constraint: forward-deployed engineers are the limiting factor. “No one has enough fully trained FDEs that can get a customer up and running fast enough,” so every vendor — Sierra, Decagon, Intercom — is picking lanes; “in that sense, they’re all services businesses” for now. The diagnostic for losers: “How many folks do you have that can spool up a customer in 30 days? ‘Well, I’ve got three.’ Well, you’re going to lose.” (The week’s funding tape: Legora at $500M on $5.5B led by Accel [likely Accel, spoken as “Excel”], Harvey at ~$11B, Intercom’s $250M in debt, Databricks tender at “4 and 1/2 billion” as spoken, Founders Fund closing on $6B having deployed $3.3B in 11 months.)

8. “The era of gentle deceleration has ended. It’s dead.”

• Jason’s public-markets thesis, triggered by CrowdStrike crushing the quarter and trading down (27% growth, 23% guidance): from ‘22 to late ‘25 the job at scale was managing “a genteel deceleration with higher net margins” — no longer. Cloudflare accelerated from 27% to 34% growth with net new customers up 40% YoY, and that’s now the bar: “you’ve got to be Cloudflare or better.” Echoing Monday’s Eran on 20VC: “We’ve got to accelerate to get back credibility with the public markets.”
• The precision, “cuz you know me, I’m that guy”: “Gentle deceleration is like entropy — the end state of the universe, cuz everything decelerates to GDP growth.” What actually changed is the multiple regime: markets spent 3 years thinking the slowdown was temporary, then “corrected to fundamentally what they’re worth with no pixie dust” — technical-obsolescence and free-cash-flow haircuts, with 8–9x EBITDA examples for value stocks. The pushback stands: at scale, 25% growth should be enough — Microsoft and Google grow at 13 — but Jason’s retort is empirical: “the markets have given up on folks not accelerating. They’ve just entirely abandoned it.”
• Jason’s venture translation is the darkest line of the section: “You got to start wondering this year when you should just give up on your portfolio if they’re not accelerating. If the public markets won’t tolerate gentle deceleration, how will the private markets tolerate it?”

9. Wix and Figma Make: the incumbent lab experiment is failing in public

• The Wix math, worked live: ~$2B revenue growing 13%, core paid customers declining 1.2% a year, and Base44 at $100M ARR inside it. Harry’s compounding case: 100→300→600 adds 10 then 15 points to Wix’s growth — “you can’t just say it’s not big enough; the nature of a small thing is it’s small and then it exponentially compounds” — but only if it keeps pace with Lovable (at 300, projecting $1B by year-end) rather than being “hobbled by the wider company or the bureaucracy.” Jason’s rhetorical dagger: “If Wix can’t cross-sell out to 6.11 million customers — people who would rather vibe-code a website than deal with the crappy templates — what hope is there for you?”
• Jason’s Figma Make test — his standard “go to saster.ai and make me a better version” prompt, which Replit, Lovable, and even V0 handle — produced “the worst vibe coding experience I’ve had in 6 months. It didn’t even know what was on my website. It didn’t even try.” Asked directly whether it changed his view as a shareholder or potential shareholder of Figma: “Yeah… it’s rough.” His generalization: “Best-efforts quarterly releases — that’s death today. That world does not exist any longer.”
• Harry zooms out — Figma’s people were “the best people out there” a year or two ago, which is exactly the point: adopting new architecture inside an existing product at needle-moving scale is the existential crisis for “literally 1.something trillion dollars worth of public companies and another trillion dollars worth of private companies.” His prescription is ironic given section 6: hire the young — not the 2-year veteran “indoctrinated” into old tools, but “the 18-year-old who’s on every subreddit and knows every intro to a TikTok” — or acqui-hire failed YC teams. Otherwise “you’re going to trade at eight times EBITDA… pretty freaking unsatisfactory for the people who paid 30 times.”

10. The picks: momentum, GARP, and buying with fear

• Jason, unchanged and single-thesis: Palantir, Cloudflare, Shopify, CrowdStrike — “I’m betting on momentum is the only thing that’s going to save us in the age of AI”: overpriced Palantir, reaccelerating Cloudflare, share-gaining Shopify, and CrowdStrike over IGV as the reluctant fourth. Snowflake was dinged “just on not being founder led — that’s just my bright line,” and Atlassian intrigues him (“I don’t know why it hasn’t gotten any credit for reaccelerating”) but the market “doesn’t believe it will sustain.” And: “After using Make, I cannot do Figma. This is not a world-class product.”
• Harry: CrowdStrike, Nubank, Nvidia, Reddit. Nubank is “dramatically underappreciated” — 28% YoY growth, David still at the helm, entering the US. Nvidia is the boring inference winner — “acquisition of Grok [likely Groq], movement into the inference layer. I don’t [__] sell Jensen.” Reddit, down 40%, is “a data layer for a lot of the LLMs” at a good price. Jason grades the homework leniently: Nubank and Nvidia don’t really fit the beaten-down-software brief, “but it’s fun and it’s all good by me.”
• Rory, two of each across three buckets: cheap value — Salesforce at 8–9x EBITDA and team at 9x EBITDA; GARP where “the AI fear is overdone” — Toast and Intuit, software “with a huge slug of transactions that aren’t going to be replaced by AI”; and expensive story stocks picked “with fear” — CrowdStrike (mid-high-teens NTM revenue, ~50x EBITDA, 23% growth, “scary but it’s an enduring company” — and it’s already recovered almost the whole SaaS-apocalypse hit) plus, “reluctantly, reluctantly,” Palantir: “one more year of growth away from being normalized… you’re forward-paying 2 years, and I think this administration’s going to spend for 2 years.” He also confirms the earlier group bet: the full $250k went into World Cloud [likely WCLD], up ~4% so far. His Atlassian caveat: the software development lifecycle is changing so fast that a tool coordinating 2022-era engineering workflows may not be “the thing” in 2027.
• The case for Salesforce, from Jason of all people: “unlike a lot of folks on this list, they have more demand than they can serve” on Agentforce — “literally turning away customers because they can’t service” them, the same squint-test he applies to startups. The closing self-reminder tempers the whole exercise: the average SaaS company used to trade at 6x revenue growing 30%; today’s 30–40x EBITDA on 20% growth “looks great relative to Salesforce, but it looks [__] relative to Anthropic.” “Good growth at a pricey price is always tricky… the great thing with 100% growth is it covers everything.”