Citrini Research Breakdown: Agents, "Ghost GDP", Consumer Spend | Figma Earnings Beat

Summary

• Anthropic’s security-review release wiped ~$20B off Cloudflare, CrowdStrike and other cyber names — and Jason Lemkin’s point is that the capability is “months old news”: he ran a full security audit inside Replit from an airplane last week. The real lesson is pricing, per Jason: CrowdStrike still traded at 16x revenues on 22% projected growth after Friday’s hit — “when you are priced for perfection, anything less than perfection will be a kick in the nuts.”
• Jason’s organizing metaphor for all of software: Claude is the Fortnite circle — “your surface area is shrinking because of Claude and AI, and the question is how much.” Claude Code added in-product app preview last week (“you don’t need Replit and Lovable anymore” for many product people), and two of Jason’s own last-year investments “have no reason to exist today.” Agents owning the value routing can mean “terminal decline” even while customers keep renewing.
• The flat assertion of the episode: among all public B2B companies, “there’s only one that has a competitive agent. It’s Palunteer” (likely Palantir). Incumbents fail because every agent today is custom — trained, onboarded, data-cleansed by forward-deployed engineers they don’t have and can’t afford — so “they’re going to get killed by the startup that does it.” But the revenue doesn’t all accrue to foundation models: both agree the software mediation layer built on top of them is the opportunity, “just as there was in SaaS.”
• On the Catrini (likely Citrini) piece that shook markets, Jason’s framework: settle the micro before you get to macro — if you don’t believe the ~2-year adoption cycle for everything from ServiceNow to DoorDash to Amex, “you can literally ignore the rest of the piece.” Rory’s evidence is the DoorDash CTO himself: “we need to earn the right to service customers’ agents end to end.” Verdict: incumbents don’t get destroyed, they get maimed — SaaS apps “become dumb databases” that miss the incremental value.
• Ghost GDP is directionally real but slower than advertised: Jason’s team went from 12 people to two (well, three) while generating eight figures, and his 10 SaaStr agents buy “nothing except tokens”; Shopify has added zero net headcount in three years while growing 50% to $12B. His prediction: a SaaS leader “does an Elon Musk” and cuts half its staff in one day within 12 months, and levered PE-backed SaaS ends in Frankenstein mergers at 1.5–2x revenue that “will try to IPO in 2027.”
• OpenAI is doubling spend to $665B by 2030 while raising its revenue forecast 27% to $280B on products that mostly don’t exist — “slides for the believers,” a startup board deck “with three zeros attached to every number.” Rory’s scoreboard: ‘23–‘24 were OpenAI’s years, ‘25 was Anthropic’s — “if it’s a 10-year race, two years up, one year down, bring on the next year.”
• Figma’s epic quarter ($1.22B ARR accelerating to 40%, 136% NDR, stock +15%) gets no credit — “we’ve just given up on the present.” Jason’s answer is momentum: only five public names are up over 12 months (Palunteer, likely Palantir; Figma, Cloudflare, Shopify among them), and he’s buying winners. Rory’s value counter: Atlassian, down 74.85% yet accelerating to 23% growth — “as armchair value investors, you couldn’t do better” — while the Klaviyo bargain is a trap because Shopify “has to kill you in order to survive.”
• Jack Altman’s move after ~$400M raised in two years to join Benchmark reflects Benchmark’s classic pick-proven-talent playbook; per Jason, he “gave up the dream of 95% of folks stuck in venture” — made-whole offers come “with asterisks and daggers” — and the fact he took it “says a lot about 2026.”

Deep dive

1. Anthropic wipes $20B off cyber — panic over months-old features

• Jason Lemkin’s deflation of the sell-off: he ran “a detailed security audit” inside Replit on a plane last week — Claude Code already does full static review and, on live Replit/Lovable/Vercel apps, penetration testing — “better security audits and testing than a mediocre board engineer will ever do.” Markets clipped $20B off Cloudflare and CrowdStrike for “features that have already been in Claude for months” — same reflex as Cobol code review dropping IBM 10%. The panic is “worth panicking about” but “a total nothing burger” if you’ve been paying attention.
• Jason’s aside, worth keeping verbatim: Anthropic “markets itself as the nice AI company. And for a nice AI company, it sure creates a lot of damage and kills a lot of stock portfolios. Maybe the Pentagon is wrong and they need to buy more Anthropic and just point it at the enemy. It’ll bring China to its knees.”
• Harry’s substantive call: Anthropic won’t eat CrowdStrike’s business — capabilities like code scanning diffuse into the enterprise through incumbents. He cites an HSBC report line approvingly: “software is the means by which AI will diffuse into the enterprise,” and expects the same for security.
• The real vulnerability was pricing: even after Friday’s correction CrowdStrike showed 22% projected revenue growth, 31% cash margins — at 16x revenues. “When you are priced for perfection, anything less than perfection will be a kick in the nuts.” Jason would “err to a bit of value” instead (reserving the right to change his mind next week): Toast and DocuSign at 7–8x are “absurdly cheap,” and his construction is a basket — “buy 20 stocks at an average of 3x revenues, 8x EBITDA, I think you’ll do just fine” — because single names are too idiosyncratic.

2. The Fortnite circle: Claude is shrinking everyone’s island

• Jason’s metaphor of the episode: Claude is Fortnite’s shrinking circle — “your surface area is shrinking because of Claude and AI, and the question is how much.” Claude Code added in-product app preview just since last week’s show (“for a lot of product people, you don’t need Replit and Lovable anymore”), the same day Anthropic launched an enterprise agent solution while its AI head insisted “we’re best friends” with software.
• DocuSign is his nuance case: it won’t be destroyed inside a Claude chat — it’s “a very complex enterprise workflow system… a partial system of record” — but an agent that autonomously runs a company’s commercial contracts “could take enough of the value away that these companies are maimed.” Even growth slipping from 31% to 20% “is a big deal”; owning less of your space can be “terminal decline.”
• His own book proves the point: two investments made last year “have no reason to exist today” because of Claude — “it’s not going to bankrupt the fund,” but the circle closed on them. Figma isn’t exempt either: Claude Code design is laughable today (same icons, same purple), but “I would be shocked if that’s not possible by the end of the year.”
• The flip side is the prize: Replit and Lovable “built a billion dollars of revenue building the agent that didn’t exist” on top of Claude Code. “If you can do something that is extremely high-value that could not be done before, you can close millions of revenue your first week. It’s never happened before in the history of software.”

3. Only Palantir has a competitive agent

• Jason’s flat assertion: of all publicly traded B2B companies, “there’s only one that has a competitive agent. It’s Palunteer (likely Palantir).” No one else has seen a single ounce of revenue acceleration due to their AI agents — because winning “is not sprinkling AI dust on top of their analytics software.”
• Why incumbents fail, two practical reasons: every agent today is custom — trained, onboarded, data-cleansed — “a vast amount of work for organizations that already think they’re overworked”; and it requires forward-deployed engineers who don’t exist (“the average customer success person with a green-yellow-red dashboard cannot train and tune an agent”) and can’t be afforded at Shopify/Monday/HubSpot/Toast price points. Hyper-niche agents work; Monday’s 100 verticals — “churches and basketball courts and refrigerator businesses” — can’t share one. “They’re going to get killed by the startup that does it.”
• Where the two converge: intelligence reaches the enterprise one of four ways — buy from Claude directly, build in-house, incumbents integrate, or new companies built on foundation models. Harry favors routes three and four, while Jason agrees that “there will be software opportunities to build compelling software companies in most of these verticals just as there was in SaaS.” Jason’s rule: “intelligence-led applications are the only applications that are going to sell and grow quickly over the next 10 years” — non-intelligent ones are flat at best.
• The LinkedIn tell: leaders posting “we’re adding AI to our email feature” are performing lip service — “you’re going to go out of business… your customers are going to renew, but your growth is going to fall so far that you become irrelevant in two years.” That, Jason says, is why the founders Harry interviews “are stressed as f*.”

4. Catrini (likely Citrini): conquer the micro before you’re allowed macro

• Jason’s framework for the piece that wiped billions (Noah Smith called it “scary bedtime reading,” and Jason opens with “I’m going to call bullshit”): first decide at the micro level whether each disruption actually happens — is AI replacing coding, DoorDash, Amex — then, only if you believe the hypothesized ~2-year adoption cycle, graduate to global macro. “If you don’t think the adoption is going to be that quick, you can literally ignore the rest of the piece.”
• The DoorDash fight is the episode’s best exchange. Jason’s pizza bit: no one delegates dinner — “Good news, I saved you two bucks. Bad news, you like the high-end pizza, but I got you the crappy little pizza.” Rory’s evidence: DoorDash CTO Andy Fang — “we strongly believe agentic commerce will be transformative to our industry” and “we need to earn the right to service customers’ agents end to end.” Jason adds he has data from 10,000+ restaurants: “I already know the answer is yes.”
• Jason’s counter to Harry’s Netflix-recommendation skepticism: YouTube is the #1 way we consume video and is entirely recommendation-driven — “the best recommendation company on planet Earth… followers don’t even matter anymore.” And AI-generated Star Wars shorts “better than the crappy last three movies” are why Netflix is “so panicked they have to buy a studio.”
• The convergence: nobody vibe-codes a rival DoorDash (“the core idea he started with is stupid… he’s trying to get millions of views”) — but if the agent, not the first party, owns the customer, incumbents don’t die, they get maimed. SaaS apps “become dumb databases,” Toast “just becomes a POS system” — “they don’t capture enough of the incremental value. It’s the incremental value we’re investing in.”

5. Ghost GDP: 12 people to two, and agents buy nothing but tokens

• Jason as the laboratory: his team went from 12 people to two (well, three) while generating eight figures of revenue — “that is ghost GDP. Those folks that are gone… the profits that are left go to two people.” His 10 SaaStr agents (Repley, Art, Qualie, Monty — “they’re good kids”) produce millions but “they buy nothing… except tokens. They buy millions and millions of tokens.”
• The wealth-concentration math, off Anthropic lining up a $5–6B employee tender at $350B (following OpenAI’s): Nvidia already has 20,000 decimillionaires, so Jason’s wag is “we’re going to produce 100,000 decimillionaires out of these AI leaders” — while everyone else gets leaner. His hedge on Jevons paradox creating tech employment, as spoken: “I believe we will need more engineers than ever. I don’t know if it will create more employment.”
• The 200-year defense: ghost-GDP doom implies productivity gains are bad — but farming went from 80% of workers to 4% “and we sell so much food that we’re all fat… productivity is freaking awesome. It’s the only thing that’s made us rich.” The onus is on doomers to name the short-term mechanism.
• The counterargument names it, with precedent: softening consumer spend plus concentration — Japan’s 1990s productivity gains never dispersed to the population, hence Abenomics. The dystopian corroboration: a VIP dinner of Japanese B2B founders last November, all describing seat bases that shrink every year. The concession: if displacement is extraordinarily fast — 6 million programmers on the street in a month — “you will have… definitely some form of recession,” “while I’m still correct over the medium and long term.”

6. Diffusion speed is the whole debate

• Jason’s estimate: customer support, legal, bookkeeping, plus Waymo and self-driving over ~4 years — “I think you could see 30 to 40 million” displaced of ~150M US workers. Rory’s counter: Waymo is doing $350M with single-digit thousands of vehicles after a decade of Catrini (likely Citrini) it’s-over predictions — “almost all diffusion takes longer than you think… I think we massively overestimate the pace of adoption here.”
• Jason’s lived rebuttal: everything is faster than he’d have believed — the airplane security audit again (“I would have said, guys, it deleted my whole database”) — and “we could delete some of our podcasts from four weeks ago. They’re so dated.” Meanwhile “almost all the B2B software we use today is terrible now… because AI software is so good” — leaders can’t keep up with how dated their products look.
• Rory’s proportionality check: US software/tech is 1–1.5M workers — nuking the entire industry is ~1–1.5% of jobs versus the 5–7% that churn every year. “We didn’t bleed in Silicon Valley when the car industry went down the toilet. Don’t hold your breath thinking they’re going to come for us… the rest of the world would go, ‘Yeah, I’m willing to lose those guys.’”
• Jason’s predictions anyway: Shopify has added zero net headcount in three years while growing 50% to $12B — already “an economic loss to the tech lifestyle we lived.” And “one of the leaders in the next 12 months is going to do an Elon Musk and cut half their team in one day.” He asked Claude to model 50% tech headcount cuts: $600–900B GDP impact, 4–5M jobs with multipliers, devastation in five to six cities — “one of the largest economic shocks in US history outside of a world war or pandemic. I’m not saying it’s true.”

7. The levered-SaaS endgame: Frankenstein mergers at 2x

• Rory: the draconian cuts land first at levered PE-backed SaaS — bought at 8–9x EBITDA with 6x of debt, now marked at 4x and growing single digits: “the math doesn’t solve any other way.” At 20%+ growth you can run the Toby playbook and hold headcount flat; at 6% with debt, the equity is gone, debtors extend rather than crystallize losses, and it’s “a long 5-year grind,” not a Friday cataclysm — starved R&D, unattractive employer, dying slowly on contract inertia.
• Jason’s endgame: five-to-eight startups at $50–200M revenue “mashed together at nominal prices” of 1–2x revenue — Frankenstein B2B constellations with professional management and 20 products. “We’ll see 20 unicorns merge into one thing that will IPO in 2027.” Rory: “Agreed. Except the last sentence. They’ll try to IPO in 2027.”

8. OpenAI’s $665B: slides for the believers

• The numbers: spend doubling to $665B by 2030 while the revenue forecast rises 27% to $280B — resting on products that mostly don’t exist (hardware, ads, agentic products, $30–77B of “consumer monetization beyond subscription”) and requiring another $110B of capital. Jason: it felt like a startup board meeting — “a stacked chart that looks beautiful, but three of the colors have never been done yet… with three zeros attached to every number.” The leaked slides were “for the believers” — SoftBank and friends; skeptics “take your marker out and just delete those bars.”
• Rory’s sentiment read: Claude now gets the benefit of the doubt and OpenAI gets none — “the truth is probably no one is ever as good or as bad as they seem.” OpenAI remains “still the clear winner in the consumer space.” And the meta point: “if you believe you’re the thing that can kill everyone else, then the only rational response as an investor is, ‘Oh my god… I better get me some models’” — the fear-mongering funds both companies.
• The Sam-vs-Dario scoreboard: if you were 10x ahead and are now 3x ahead, “you’d have to say it wasn’t the right play” to sprawl across hardware, ads and health while Anthropic took the enterprise. But: “‘23 and ‘24 were good years for OpenAI and ‘25 was a good year for Anthropic. If it’s a 10-year race, two years up, one year down, bring on the next year.” Jason’s caveat in Sam’s favor: “he can dump the hardware business in a heartbeat… they will ruthlessly dump it if it doesn’t work” — and these are “the two fastest growing companies in history.”

9. Figma fights back; Jason goes momentum, Rory finds Atlassian

• Figma’s Q4 2025: $1.22B ARR accelerating to 40% YoY (from 38% in Q3), 97% GRR, 136% NDR on $10k+ customers, stock +15%. Harry’s framing: “this is what fighting back looks like” — a generational founder pushing from design into code in a sector where AI-native disruption is “here right now” (accounting, by contrast, is “5 years away”). His Vegas card: Dylan Field, heavyweight champion, versus Lovable and Replit as the marquee bout.
• Jason: “epic company, epic quarter” — and no credit: “we’ve just given up on the present. We’re all panicked about the future.” He’d still be shocked if within 8–18 months Claude Code can’t produce designs “as elegant, as beautiful as a designer can” — it “has ingested every single website and mobile app on planet Earth” — and Figma citing Claude Code integration as a top growth driver is the Fortnite overlap risk in miniature.
• His buy-or-sell answer is momentum: only five public names are up over 12 months — Palunteer (likely Palantir), Figma, Cloudflare, Shopify and a fifth bleeped in the audio — “I’m going to bet on whoever has the gravitas, and momentum is gravitas.” He’s lost money bargain-hunting before (GitLab: −59.62%). Rory calls it “utterly coherent”: momentum wins over 6–18 months, value over 5 years — “the trick is to figure out when you’re transitioning… trees don’t go to the sky” (Palantir is already −27% in 3 months, de-rated from 70x to 46x revenues).
• The value showdowns: Jason would buy Klaviyo (−58%) over Shopify (+63%) as “the greatest dislocation” — Rory disagrees: Shopify “probably has to kill you in order to survive,” so the bargain is a trap. Atlassian is different: down 74.85% while accelerating from 20% to 23% growth at $6.3B — “as armchair value investors, you couldn’t do better than Atlassian.” Jason’s dark caveat: “if there’s any humans left to buy the product.”

10. Jack Altman gave up the dream of 95% of venture

• On Benchmark’s move: “a clever move by a very shrewd firm” running its 15–20-year MO — “we’ll make you broadly equal in a very successful partnership with a lot of autonomy… you’re not in the growing-talent business, you’re in the picking-talent business.” Jack is just “an extreme version of that.”
• Jason on what may have been surrendered: ~$400M raised in two years, essentially as a solo GP, with the last $250M fund possibly handed back to LPs — “not a minor give,” because made-whole offers come “with asterisks and daggers. You’ve got to stay. You’ve got to deliver.” He recalls turning down his own mega-firm make-whole offer after a $70M first fund: “I didn’t sell my last company to go work for somebody.” The takeaway: “Jack gave up the dream of 95% of folks stuck in venture… it says a lot about 2026.”
• Why Jason wouldn’t take Harry’s hypothetical $500M three-GP fund: “I don’t think I would be successful” — venture rewards “a certain genericism” of Monday partner meetings and “weird consensus-driven outcomes,” and he’s done performing: “I’m done with this performative all-day circus of an AGM.” His tie-back to the episode’s theme: if you want off-the-charts talent, “you’ve got to let them do their thing and nothing else” — easier at Anthropic, where “they’re going to find you your niche,” than at a venture firm.