In August I wrote a thesis I never published. The funds I was warning were key Crossover Research clients, so I stayed quiet. Since then, 𝗦𝗼𝗳𝘁𝘄𝗮𝗿𝗲 𝗺𝘂𝗹𝘁𝗶𝗽𝗹𝗲𝘀 𝗮𝗿𝗲 𝗱𝗼𝘄𝗻 𝟱𝟬%+. Salesforce $CRM, ServiceNow $NOW, Adobe $ADBE, Workday $WDAY all off 40% from highs. Thomson Reuters $TRI dropped 16% in a single session on the Anthropic legal agent launch. The SaaSpocalypse arrived. So here's the follow-up. Not commentary on what happened, but where I think this goes next. Most vertical SaaS companies aren't underperforming because their software is bad. 𝗧𝗵𝗲𝘆'𝗿𝗲 𝘂𝗻𝗱𝗲𝗿𝗽𝗲𝗿𝗳𝗼𝗿𝗺𝗶𝗻𝗴 𝗯𝗲𝗰𝗮𝘂𝘀𝗲 𝘁𝗵𝗲𝘆 𝗻𝗲𝘃𝗲𝗿 𝗯𝘂𝗶𝗹𝘁 𝘁𝗵𝗲 𝘀𝗲𝗰𝗼𝗻𝗱 𝗯𝘂𝘀𝗶𝗻𝗲𝘀𝘀. And the first business is under attack. For twenty years, one of the biggest SaaS moats was engineering complexity: deep technical talent, long roadmaps, compounding codebases that were genuinely hard to replicate. 𝗔𝗜 𝘂𝗽𝗲𝗻𝗱𝗲𝗱 𝘁𝗵𝗮𝘁 𝗮𝗹𝗺𝗼𝘀𝘁 𝗼𝘃𝗲𝗿𝗻𝗶𝗴𝗵𝘁. Product development is democratizing to operators with no code background but strong product vision. Look at Anthropic: they've built the engine and are shipping lookalike products at a cadence that would have taken a legacy SaaS vendor three years of roadmap, with a fraction of the headcount. That pace can kill legacy businesses overnight. 𝗜𝗳 𝘁𝗵𝗲 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 𝗺𝗼𝗮𝘁 𝗶𝘀 𝗴𝗼𝗻𝗲, 𝗳𝗼𝘂𝗿 𝗺𝗼𝗮𝘁𝘀 𝗿𝗲𝗺𝗮𝗶𝗻: 𝗱𝗶𝘀𝘁𝗿𝗶𝗯𝘂𝘁𝗶𝗼𝗻, 𝗽𝗿𝗼𝗽𝗿𝗶𝗲𝘁𝗮𝗿𝘆 𝗱𝗮𝘁𝗮, 𝘄𝗼𝗿𝗸𝗳𝗹𝗼𝘄 𝗯𝗿𝗲𝗮𝗱𝘁𝗵, 𝗮𝗻𝗱 𝗿𝗲𝗴𝘂𝗹𝗮𝘁𝗼𝗿𝘆 𝗶𝗻𝘀𝘂𝗹𝗮𝘁𝗶𝗼𝗻. The first three are moats the company builds. The fourth is a moat the company captures, and it's the one most resistant to AI disruption. 𝗥𝗲𝗴𝘂𝗹𝗮𝘁𝗼𝗿𝘆 𝗰𝗼𝗺𝗽𝗹𝗲𝘅𝗶𝘁𝘆 𝗰𝗿𝗲𝗮𝘁𝗲𝘀 𝘀𝘄𝗶𝘁𝗰𝗵𝗶𝗻𝗴 𝗰𝗼𝘀𝘁𝘀 𝘁𝗵𝗮𝘁 𝗵𝗮𝘃𝗲 𝗻𝗼𝘁𝗵𝗶𝗻𝗴 𝘁𝗼 𝗱𝗼 𝘄𝗶𝘁𝗵 𝗽𝗿𝗼𝗱𝘂𝗰𝘁 𝗾𝘂𝗮𝗹𝗶𝘁𝘆. Once a vendor is embedded in a compliance workflow, ripping them out means re-attesting, re-auditing, and re-certifying every downstream process. The buyer isn't paying for software, they're paying for the accumulated paper trail. Tyler Technologies ($TYL) is the clearest version of the pattern. State and local government software across courts, public safety, assessment, and ERP. Every module is married to statutory process, FIPS, CJIS, audit trails, and procurement cycles that take years. TYL is down 42% TTM and 2026 guidance came in soft, but the moat didn't break. Revenue still compounded, and government procurement runs on five-year cycles, not five-week news cycles. Veeva is the sharper version. Revenue up 16% in FY26, Q4 beat, the stock still down 25%. The market is selling execution, not weakness. Guidewire in P&C insurance, where regulatory filings and rate approvals anchor the stack, sits in the same setup: still compounding ARR, still winning cloud conversions, multiple reset anyway. Same pattern across all three: multiples compressed, fundamentals intact. The moat is the regulatory surface area itself, and it compounds because the rules get more complex, not less. 𝗜 𝘄𝗮𝘀 𝗹𝗼𝗻𝗴 𝗣𝗮𝗹𝗮𝗻𝘁𝗶𝗿 𝗮𝘁 $𝟭𝟯 (read that here: xitter.cf/blyons151/status/17920…). 𝗡𝗼𝘁 𝗯𝗲𝗰𝗮𝘂𝘀𝗲 𝗼𝗳 𝘁𝗵𝗲 𝗺𝗼𝗱𝗲𝗹 𝗼𝗿 𝘁𝗵𝗲 𝘁𝗼𝗼𝗹𝗶𝗻𝗴. 𝗕𝗲𝗰𝗮𝘂𝘀𝗲 𝗼𝗳 𝘁𝗵𝗲 𝗼𝗻𝘁𝗼𝗹𝗼𝗴𝘆. Palantir is the proprietary-data version of the regulatory thesis. Once Palantir sits between the customer and their own data, ripping it out means rebuilding the data model from scratch. Snowflake and Databricks never had that entrenchment layer. AIP bootcamps then turned the data moat into a distribution moat: 660 bootcamps in a single quarter, 94% y/y US customer deal growth, bookings at 1.9x sales. Own the data, ship functional AI on top of it, let the GTM compound. Every vertical incumbent has a version of this available. The question is whether they'll build it before a challenger does. But regulatory insulation is necessary, not sufficient. Plenty of vendors inside regulated verticals are still getting squeezed because they never became AI-native. BlackLine ($BL) and Trintech are feeling it in close and reconciliation as Numeric, Maximor, and Stacks build AI-native from day one. nCino ($NCNO) in banking faces the same challenge. The regulatory moat buys you time. It doesn't buy you the decade. 𝗧𝗵𝗲 𝘄𝗶𝗻𝗻𝗶𝗻𝗴 𝗳𝗼𝗿𝗺𝘂𝗹𝗮 𝗶𝘀 𝗱𝗮𝘁𝗮 𝗼𝗿 𝗿𝗲𝗴𝘂𝗹𝗮𝘁𝗼𝗿𝘆 𝘀𝘂𝗿𝗳𝗮𝗰𝗲 𝗮𝗿𝗲𝗮 𝗽𝗹𝘂𝘀 𝗳𝘂𝗻𝗰𝘁𝗶𝗼𝗻𝗮𝗹 𝗔𝗜, 𝗻𝗼𝘁 𝗼𝗻𝗲 𝗼𝗿 𝘁𝗵𝗲 𝗼𝘁𝗵𝗲𝗿. Look at why Claude is winning. Anthropic isn't competing on model benchmarks, they're competing on functional workflow. Building for the user, not the leaderboard. That's the playbook vertical incumbents need to run. Take the moat you already have, whether it's regulatory or data-entrenchment, layer genuine workflow AI on top, and the challenger can't catch you. The vendors that do both win the decade. The ones that rely on inertia alone get caught. The ones that ship AI without an anchor get commoditized. You need both. 𝗧𝗵𝗲 𝗯𝘂𝘆𝗲𝗿 𝗶𝘀 𝘁𝗲𝗹𝗹𝗶𝗻𝗴 𝘆𝗼𝘂 𝘁𝗵𝗶𝘀 𝗽𝗹𝗮𝗶𝗻𝗹𝘆. A study we ran with Battery Ventures on AI adoption in the Office of the CFO (battery.com/blog/first-codin…) surveyed 129 finance leaders at companies from $50M to $5B+ in revenue. 77% said they want to uplevel existing systems with AI from new vendors that layer onto existing systems. Only 15% want to replace their current system of record with an AI-native platform. The incumbent wins if they ship AI. The AI-native challenger wins only if the incumbent doesn't. The signal shows up in our VoC data too. In regulated verticals, mission criticality scores cluster above 9, and NPS doesn't track satisfaction, it tracks switching friction. Customers will tell you the product is mediocre and still score it 9 on "would not switch" because the compliance team vetoes any alternative. 𝗧𝗵𝗮𝘁'𝘀 𝘁𝗵𝗲 𝘀𝗶𝗴𝗻𝗮𝘁𝘂𝗿𝗲 𝗼𝗳 𝗮 𝗰𝗼𝗺𝗽𝗹𝗶𝗮𝗻𝗰𝗲-𝗶𝗻𝘀𝘂𝗹𝗮𝘁𝗲𝗱 𝘃𝗲𝗻𝗱𝗼𝗿, 𝗮𝘀 𝗹𝗼𝗻𝗴 𝗮𝘀 𝘁𝗵𝗮𝘁 𝘃𝗲𝗻𝗱𝗼𝗿 𝗶𝘀 𝗮𝗰𝘁𝗶𝘃𝗲𝗹𝘆 𝘀𝗵𝗶𝗽𝗽𝗶𝗻𝗴 𝗮𝗴𝗮𝗶𝗻𝘀𝘁 𝘁𝗵𝗲 𝗔𝗜 𝗰𝘂𝗿𝘃𝗲. Which brings us back to the second business for everyone outside the regulated or data-entrenched moat. Seat ARR got them to $100M. But with the shift to agentic workforce structures, partial human capital replacement, and pricing pressure compressing margins, the traditional SaaS model has to transform fast. The next $500M comes from monetizing the installed base: marketplace rake on demand they generate for their own customers, capital products underwritten by their own transaction data, supplier monetization, brand partnerships, group buying. The assets are already sitting there. Captive SMB audience. Proprietary transaction and behavioral data. A distribution pipe (the UI itself) that delivers new products at near-zero CAC. 𝗪𝗵𝗮𝘁'𝘀 𝗺𝗶𝘀𝘀𝗶𝗻𝗴 𝗶𝘀 𝗼𝗿𝗴𝗮𝗻𝗶𝘇𝗮𝘁𝗶𝗼𝗻𝗮𝗹 𝘄𝗶𝗹𝗹. Monetizing the installed base requires a different org than the one that got you to scale. Different GTM, P&L optics, and talent. Founders and boards under-invest because year one looks worse before it looks better, and public markets punish any SaaS multiple that starts to look like fintech or marketplace. So the second business never ships. The round prices in the optionality. The multiple compresses. The exit underwhelms. 𝗧𝗵𝗿𝗲𝗲 𝗱𝗶𝗹𝗶𝗴𝗲𝗻𝗰𝗲 𝗾𝘂𝗲𝘀𝘁𝗶𝗼𝗻𝘀 𝗻𝗼𝘁 𝗲𝗻𝗼𝘂𝗴𝗵 𝗶𝗻𝘃𝗲𝘀𝘁𝗼𝗿𝘀 𝗮𝗿𝗲 𝗮𝘀𝗸𝗶𝗻𝗴: 𝟭. 𝗪𝗵𝗮𝘁 𝗽𝗲𝗿𝗰𝗲𝗻𝘁 𝗼𝗳 𝗿𝗲𝘃𝗲𝗻𝘂𝗲 𝗰𝗼𝗺𝗲𝘀 𝗳𝗿𝗼𝗺 𝘀𝗼𝘂𝗿𝗰𝗲𝘀 𝗼𝘁𝗵𝗲𝗿 𝘁𝗵𝗮𝗻 𝘀𝘂𝗯𝘀𝗰𝗿𝗶𝗽𝘁𝗶𝗼𝗻 𝗮𝗻𝗱 𝗽𝗮𝘆𝗺𝗲𝗻𝘁 𝗽𝗿𝗼𝗰𝗲𝘀𝘀𝗶𝗻𝗴? Under 5%, they haven't started. 10 to 20%, thesis is live. Over 20%, it's working. 𝟮. 𝗛𝗼𝘄 𝗵𝗮𝗿𝗱 𝘄𝗼𝘂𝗹𝗱 𝗶𝘁 𝗯𝗲 𝘁𝗼 𝗿𝗲𝗰𝗿𝗲𝗮𝘁𝗲 𝘁𝗵𝗶𝘀 𝗰𝗼𝗺𝗽𝗮𝗻𝘆 𝗳𝗿𝗼𝗺 𝘀𝗰𝗿𝗮𝘁𝗰𝗵 𝘄𝗶𝘁𝗵 𝗔𝗜 𝘁𝗼𝗱𝗮𝘆? If a well-funded team with Claude and six engineers could rebuild the functional product in nine months, the software isn't the moat. The moat has to live somewhere else: proprietary data, a network, integrations, or regulatory surface area the challenger can't clear. If you can't point to at least one, you're underwriting a melting ice cube. 𝟯. 𝗪𝗵𝗮𝘁 𝗽𝗲𝗿𝗰𝗲𝗻𝘁 𝗼𝗳 𝘁𝗵𝗲 𝗯𝘂𝘆𝗲𝗿'𝘀 𝘀𝘁𝗶𝗰𝗸𝗶𝗻𝗲𝘀𝘀 𝗶𝘀 𝗿𝗲𝗴𝘂𝗹𝗮𝘁𝗼𝗿𝘆, 𝗮𝗻𝗱 𝘄𝗵𝗶𝗰𝗵 𝘄𝗮𝘆 𝗶𝘀 𝘁𝗵𝗲 𝗿𝘂𝗹𝗲 𝘀𝗲𝘁 𝗺𝗼𝘃𝗶𝗻𝗴? A regulatory moat evaporates if the regulation simplifies. Underwrite the direction of travel, not just the current state. 𝗔𝗻𝗱 𝘁𝗵𝗲 𝗰𝗹𝗼𝗰𝗸 𝗶𝘀 𝘁𝗶𝗴𝗵𝘁𝗲𝗿 𝘁𝗵𝗮𝗻 𝗺𝗼𝘀𝘁 𝗿𝗲𝗮𝗹𝗶𝘇𝗲. Retention in enterprise SaaS has largely been defined by the pain of systems replacement, not genuine moat. If the stickiness isn't backed by proprietary data, a harvesting flywheel, or regulatory surface area, those vendors are about to get disrupted. Pure seat-based pricing is dying unless vendors embrace agent-seat models, and LLM providers have been subsidizing the market on token cost, with recent pricing shifts signaling cash reserves aren't infinite. 𝗛𝗲𝗿𝗲'𝘀 𝘁𝗵𝗲 𝘂𝗻𝗱𝗲𝗿𝗮𝗽𝗽𝗿𝗲𝗰𝗶𝗮𝘁𝗲𝗱 𝗽𝗼𝗶𝗻𝘁: 𝗔𝗜-𝗻𝗮𝘁𝗶𝘃𝗲 𝗰𝗼𝗺𝗽𝗲𝘁𝗶𝘁𝗼𝗿𝘀 𝗵𝗮𝘃𝗲 𝘄𝗼𝗿𝘀𝗲 𝗴𝗿𝗼𝘀𝘀 𝗺𝗮𝗿𝗴𝗶𝗻𝘀 𝘁𝗵𝗮𝗻 𝗦𝗮𝗮𝗦 𝗶𝗻𝗰𝘂𝗺𝗯𝗲𝗻𝘁𝘀, 𝗻𝗼𝘁 𝗯𝗲𝘁𝘁𝗲𝗿. Inference costs haven't collapsed, and burning VC cash to subsidize unit economics is a bridge, not a business model. The incumbents should be winning on P&L. They're losing on product velocity and AI-readiness. That's a solvable problem if the board has the will to ship. Vendors without a second business, without a data moat, and without regulatory insulation will still lose, despite having better margins than their AI-native challengers. Customers switch on features and speed, not on unit economics. 𝗘𝗻𝘁𝗲𝗿𝗽𝗿𝗶𝘀𝗲 𝗮𝗻𝗱 𝗿𝗲𝗴𝘂𝗹𝗮𝘁𝗲𝗱 𝘃𝗲𝗿𝘁𝗶𝗰𝗮𝗹𝘀 𝗮𝗿𝗲 𝘁𝗵𝗲 𝗹𝗮𝘀𝘁 𝘀𝗮𝗳𝗲 𝗵𝗮𝗿𝗯𝗼𝗿, 𝗮𝗻𝗱 𝗼𝗻𝗹𝘆 𝗯𝗲𝗰𝗮𝘂𝘀𝗲 𝗼𝗳 𝗱𝗮𝘁𝗮 𝗯𝗿𝗲𝗮𝗱𝘁𝗵 𝗮𝗻𝗱 𝗰𝗼𝗺𝗽𝗹𝗶𝗮𝗻𝗰𝗲. Everywhere else, the premium is about to get competed away. Any fund underwriting vertical SaaS exposure right now should be asking the second-business question before the next check clears. DM me, email me [email protected], or let's chat about your portfolio/underwriting process (book.crossoverresearch.com). Crossoverresearch.com

Cloud is a useful template for the cohort point in this thread. AWS growth went from 33% in Q2 2022 to 12% by mid-2023, and revenue was roughly flat sequentially in Q1 2023, even as new workloads kept landing. Optimization was attenuating by Q4 2023 and growth was back to 17% by Q1 2024. It was still 17.5% in Q2 2025, then printed 36.7% last quarter as AI scaled. The Notion chart looks like phase one of that cycle for an early cohort.

Introducing Crossover Intelligence. It started as the internal tooling we used to run our own research. Over about 18 months it grew into a full suite (40+ tools consolidated) covering discovery, diligence, analysis, decks & outreach in one interface, built and run by a team of 5. The presentation builder is in the video. Marketing module tomorrow, full platform Friday. $IGV $MSFT $ANT $CRM

We agree! After running thousands of diligence engagements over the years, I came to realize that the slowest part was always the same: turning analysis into a deck. So we decided to rebuild our stack from scratch. Here is our presentation builder. It’s one piece of Crossover Intelligence, the suite we built to run our own research process, now in one interface from analysis through outreach. Multi-channel marketing module tomorrow. Full platform launches Friday.

CoreWeave dropped 15% yesterday. Healthy EBITDA margins, a $104 billion backlog, and the stock still fell off a cliff because the cost of money moved. 𝗧𝗵𝗮𝘁 𝗶𝘀 𝘁𝗵𝗲 𝗿𝗶𝘀𝗸 𝗶𝗻 𝗔𝗻𝘁𝗵𝗿𝗼𝗽𝗶𝗰'𝘀 $𝟮 𝘁𝗿𝗶𝗹𝗹𝗶𝗼𝗻 𝗜𝗣𝗢, 𝗼𝗻𝗲 𝗹𝗮𝘆𝗲𝗿 𝘂𝗽 𝘁𝗵𝗲 𝘀𝘁𝗮𝗰𝗸. We run Claude deep in our own workflows, and we're the kind of customer this IPO is counting on. We're also the kind that watched the bill climb and quietly diversified to other providers. 𝗧𝗵𝗮𝘁 𝗽𝗮𝗿𝘁 𝗼𝗳 𝘁𝗵𝗲 𝘀𝘁𝗼𝗿𝘆 𝗶𝘀𝗻'𝘁 𝗶𝗻 𝘁𝗵𝗲 𝗱𝗲𝗰𝗸. 𝗧𝗵𝗲 𝗴𝗿𝗼𝘄𝘁𝗵 𝗶𝘀 𝗿𝗲𝗮𝗹. Yesterday's investor day put Q3 revenue at $16.8 billion, up 45% from Q2, with an 85% adjusted gross margin and positive cash flow. The $965 billion private mark from May looks fair against that. The $2 trillion ask is 30 times the current run rate. And the number the market flinched at: Q2 added $6.9 billion of revenue, Q3 added $5.2 billion. Still enormous. 𝗡𝗼 𝗹𝗼𝗻𝗴𝗲𝗿 𝗮𝗰𝗰𝗲𝗹𝗲𝗿𝗮𝘁𝗶𝗻𝗴. At 30 times revenue, that distinction is the whole trade. 𝗧𝗵𝗲 𝗯𝗲𝗮𝗿 𝗰𝗮𝘀𝗲 𝗲𝘃𝗲𝗿𝘆𝗼𝗻𝗲 𝗿𝗲𝗽𝗲𝗮𝘁𝘀 𝗶𝘀 𝘄𝗿𝗼𝗻𝗴. Tokens aren't a loss leader. An 85% gross margin is a software margin. 𝗧𝗵𝗲 𝗿𝗲𝗮𝗹 𝗽𝗿𝗼𝗯𝗹𝗲𝗺 𝗶𝘀 𝘁𝗵𝗲 𝘁𝗿𝗲𝗮𝗱𝗺𝗶𝗹𝗹. A business that profits on every sale and has to spend all of it, plus more, to stay in the race. Anthropic has committed $518 billion to compute over the next decade, 80% of it owed whether they use it or not. Meanwhile nearly a quarter of last year's revenue came from two customers, and many of the largest aren't locked into long-term contracts. 𝗧𝗵𝗲 𝗯𝗶𝗹𝗹𝘀 𝗮𝗿𝗲 𝗳𝗶𝘅𝗲𝗱. 𝗧𝗵𝗲 𝘀𝗮𝗹𝗲𝘀 𝗮𝗿𝗲 𝗻𝗼𝘁. 𝗧𝗵𝗲 𝗳𝗿𝗼𝗻𝘁𝗶𝗲𝗿 𝗶𝘀 𝘁𝗵𝗲 𝗳𝗮𝗰𝘁𝗼𝗿𝘆, 𝗻𝗼𝘁 𝘁𝗵𝗲 𝗽𝗿𝗼𝗱𝘂𝗰𝘁. Fable 5 launched in June at $10 per million tokens. Six weeks later Opus 5 matched it within half a point at half the price. Fable 5.1 launched September 1 at $10. Three weeks later Opus 5.5 beat it at $4. On Vercel's AI Gateway, Fable fell from 13% of spend to 5% in a month while Opus 5 rose to 22%. Anthropic kept the money. 𝗪𝗵𝗮𝘁 𝗶𝘁 𝗰𝗮𝗻'𝘁 𝗸𝗲𝗲𝗽 𝗶𝘀 𝘁𝗵𝗲 𝗽𝗿𝗶𝗰𝗲. A flagship holds its premium for weeks, and the cheaper tier only stays ahead of open-weight models, now 56% of tokens, while the next flagship keeps shipping. Stop training and the margin goes with it. 𝗖𝗼𝗿𝗲𝗪𝗲𝗮𝘃𝗲 𝗶𝘀 𝘁𝗵𝗲 𝘀𝗮𝗺𝗲 𝘁𝗿𝗲𝗮𝗱𝗺𝗶𝗹𝗹 𝘄𝗶𝘁𝗵 𝘁𝗵𝗲 𝗱𝗲𝗯𝘁 𝗮𝗹𝗿𝗲𝗮𝗱𝘆 𝗮𝘁𝘁𝗮𝗰𝗵𝗲𝗱. 59% EBITDA margins, then $640 million of quarterly interest on $52 billion of debt and leases, borrowed at around 9% to earn what one analyst puts at 1%, against a 4 to 5% target. CPI printed hot, rate-hike odds jumped, and the market did the math in one session. The data centers Anthropic has committed to are being built with the same borrowed money. 𝗢𝗻𝗲 𝗹𝗼𝗼𝗽. 𝗗𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝘁 𝘀𝗲𝗮𝘁𝘀. 𝗙𝗼𝘂𝗿 𝘁𝗵𝗶𝗻𝗴𝘀 𝗵𝗮𝘃𝗲 𝘁𝗼 𝗴𝗼 𝗿𝗶𝗴𝗵𝘁 𝗳𝗼𝗿 $𝟮𝗧. Revenue triples to Anthropic's own forecast of $200 billion by 2028. Cash profit reaches 30 cents on the dollar, which no lab has shown. All $518 billion of capacity gets used. The big customers stay. Give each one a generous two-in-three chance and all four together come out around 𝗼𝗻𝗲 𝗶𝗻 𝗳𝗶𝘃𝗲. 𝗧𝘄𝗼 𝗾𝘂𝗲𝘀𝘁𝗶𝗼𝗻𝘀 𝗳𝗼𝗿 𝗮𝗻𝘆𝘁𝗵𝗶𝗻𝗴 𝗼𝗻 𝘁𝗵𝗲 𝘁𝗿𝗲𝗮𝗱𝗺𝗶𝗹𝗹: How much of each revenue dollar goes straight back into compute? Over a dollar, you're funding the treadmill. Under thirty cents, it's a software company. And how long does a flagship hold its premium before the tier below catches up? Fable 5 got six weeks. Fable 5.1 got three. If that window keeps shrinking, every training run has to pay for itself faster than the last. 𝗢𝗻𝗲 𝗹𝗮𝘀𝘁 𝘁𝗵𝗶𝗻𝗴. 𝗜 𝗮𝘀𝗸𝗲𝗱 𝗙𝗮𝗯𝗹𝗲 𝟱.𝟭 𝘁𝗼 𝘃𝗮𝗹𝘂𝗲 𝗶𝘁𝘀 𝗼𝘄𝗻 𝗽𝗮𝗿𝗲𝗻𝘁 𝗰𝗼𝗺𝗽𝗮𝗻𝘆. Fair at $965 billion, overvalued at $2 trillion, two-in-three odds it trades below the IPO price within 18 months. Then it reminded me it's an Anthropic product giving a bearish read on Anthropic. "𝗚𝗿𝗲𝗮𝘁 𝗽𝗿𝗼𝗱𝘂𝗰𝘁" 𝗮𝗻𝗱 "𝘄𝗼𝗿𝘁𝗵 $𝟮 𝘁𝗿𝗶𝗹𝗹𝗶𝗼𝗻" 𝗮𝗿𝗲 𝘁𝘄𝗼 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝘁 𝗰𝗹𝗮𝗶𝗺𝘀. If your fund has AI exposure anywhere in the stack, let's pressure-test it. DM me or email [email protected]. $ANT $QQQ $META $MSFT $SPCX

🎯 Harvey’s gross margin fell from about 50% to -50% by June as agent token use spiked twentyfold on rented OpenAI and Anthropic models, Bloomberg reports.

Harvey at $11B. On the surface, the story does not justify the valuation, and the part everyone argues about does not even matter. The only figure it has printed is about $190M in ARR, from the CEO at the end of 2025, and the $11B mark was set in March against that number, about 58x the last ARR Harvey had disclosed. Independent trackers put revenue near $300M by mid-2026, which would still be 37x. Either way the multiple is north of 35x, several times what durable software earns. Everyone wants to argue about revenue quality, how much of that ARR is durable license and how much is a markup on tokens. Set it aside. Grant that every dollar is pristine, durable SaaS. The valuation still fails on two things you can check from the outside: the market is too small to grow into the mark, and the price Harvey can charge is being competed down. Start with what the multiple requires. To hold $11B at even a rich multiple, you need on the order of $1B in durable revenue to grow into. That is the bar no matter what the revenue is made of. 𝗧𝗵𝗲 𝗺𝗮𝗿𝗸𝗲𝘁 𝗶𝘀 𝘁𝗼𝗼 𝘀𝗺𝗮𝗹𝗹 𝗳𝗼𝗿 𝘁𝗵𝗮𝘁. Size it from primary data, not analyst forecasts. Total US legal services revenue is about $430B a year, per the Census Bureau. The two dominant legal incumbents book roughly $2.9B and $3.4B in total legal revenue, Thomson Reuters and LexisNexis, per their latest filings, and the overwhelming majority of that is legacy research and workflow, not AI. The legal-AI software actually being sold today, Harvey plus CoCounsel plus everyone else, is on the order of one to two billion dollars, and on any reasonable build the whole category reaches roughly $5B to $6B by 2030, with a bull case near $10B. So Harvey's private mark is about twice the entire market it leads five years out, and nearly double the combined legal revenue of both incumbents right now. Then look at what Harvey actually charges, not on the rate card but in the revenue. Its disclosed $190M ARR across the 100,000-plus lawyers on the platform is about $1,900 per lawyer a year, and even the $300M estimate across 142,000 lawyers is only about $2,100. To reach $1B in durable revenue at that price takes close to 500,000 premium seats, roughly a fifth of the entire base-case 2030 market, against an AmLaw 100 that totals 129,000 lawyers and incumbents embedded in the content Harvey licenses from them. There is no version of the seat math that gets to $1B. 𝗔𝗻𝗱 𝘁𝗵𝗲 𝗽𝗿𝗶𝗰𝗲 𝗶𝘀 𝗺𝗼𝘃𝗶𝗻𝗴 𝘁𝗵𝗲 𝘄𝗿𝗼𝗻𝗴 𝘄𝗮𝘆. On the rate card Harvey lists about $1,000 to $1,200 a seat a month, several times CoCounsel bundled with Westlaw at roughly $225 to $430. But the realized figure above, about $2,000 a lawyer a year, is roughly $160 a month, a fraction of the list and at or below what the incumbent bundle itself costs, rather than multiples above it. The premium lives on the rate card, not in the revenue, which tells you how much of the book is discounted, piloted, or only partly deployed. Industry trackers report initial quotes are commonly negotiated down 40 to 60%. And Harvey's own CEO says it is moving off per-seat pricing toward usage and revenue-share deals, which is what a company does when it cannot hold the seat price. 𝗛𝗲𝗿𝗲 𝗶𝘀 𝘁𝗵𝗲 𝗯𝘂𝗹𝗹'𝘀 𝗯𝗲𝘀𝘁 𝗮𝗻𝘀𝘄𝗲𝗿, 𝗮𝗻𝗱 𝗶𝘁 𝗱𝗼𝗲𝘀 𝗻𝗼𝘁 𝘀𝗮𝘃𝗲 𝗶𝘁. Harvey does not sell software, the argument goes, it replaces lawyers, so the market is not the few-billion-dollar legal-AI software line, it is the roughly $430B US legal services market, more than $1T globally. Capture 1% and you are at several billion in revenue and $11B looks cheap. Grant the frame entirely. It changes the shape of the risk, not the size of it. Billing against labor means replacing billable hours, which means work product a firm will send out the door unsupervised, and that is exactly what today does not support, the tools still hallucinate and still sit under a human reviewer. Whatever Harvey does capture, it captures at the realized price above, a price the incumbents and the base models are compressing, through revenue-share deals where it keeps a sliver of the work value, not the whole spend. A trillion-dollar labor market sets the size of the pond. It does not tell you Harvey lands the fish, or at what price it holds onto it. A bigger denominator with no pricing power and no lock-in is not a bigger business. It is a bigger number to divide by. So on any split, the ceiling holds. Size Harvey against the software market or the labor market, and the durable, defensible, priced revenue it can actually keep is capped well below what a multiple north of 35x needs. The mix is a detail. The ceiling is the argument. The growth everyone cites does not close the gap either. Tokens are up 14.6x in six months, revenue about 1.6x. The usage is real, but it is throughput the customer is pulling, not durable revenue racing to fill the market. Even if you credit all of it as durable, the ceiling is the same. And there is no moat to defend the price with. Displacing Thomson Reuters and LexisNexis is a product lead, not a moat, and leads in AI compress by the quarter. The clean test is replicability, not polish, and most of what Harvey does is search and synthesis that general-purpose agents are closing on fast, with open models trailing the frontier by months, not years. The defensible sliver is enterprise trust and workflow embedding, not the model, and that sliver does not hold a rate-card premium the revenue already is not collecting. 𝗔𝗻𝗱 𝗮𝘀𝗸 𝘄𝗵𝗼 𝗯𝘂𝘆𝘀 𝗶𝘁 𝗮𝘁 𝘁𝗵𝗲 𝗻𝗲𝘅𝘁 𝗺𝗮𝗿𝗸. A private mark this size quietly assumes an acquirer shows up, but the natural ones are building competitors, not bidding. Thomson Reuters is rolling out agentic CoCounsel, LexisNexis its Protégé assistant, both on the content Harvey licenses from them. Microsoft ships Copilot. The model labs Harvey buys tokens from are not paying a huge markup for a reseller of their own inference. LexisNexis even licensed its content into Harvey rather than buy the company. And a round this size carries liquidation preferences, so the $11B headline flatters the common equity beneath it. Strip out the acquisition premium the mark leans on, and the number has to be carried by an IPO into public multiples or by durable ARR on its own. Two objections, and neither clears the ceiling. Seat count is reportedly doubling inside a year, which sounds like the durable bucket finally arriving. But doubling off a small base is the easy part. Holding the price while you do it is the hard part, and the realized price says that is the part already failing. And yes, AI-native companies trade at 30 to 50x, so this looks like the going rate, until you notice that a whole category clearing 40x is not a floor under 40x. It is one untested assumption, that the growth is durable and the ceiling is far off, repeated across every name at once. So my read: on any split, a multiple north of 35x has no room. The market is too small, the price is compressing, and there is no natural buyer at the next turn. The mix nobody will disclose is not even the issue. The ceiling is. And Harvey is just the clean example. Across the late-stage AI market, marks are being set on revenue growth without anyone pricing the ceiling underneath it. Multiples are in nosebleed territory, the TAMs below a lot of these rounds are too small to carry them, and the exit only works if someone pays the next turn. Right now the next buyer is mostly the same funds bidding each other's assets higher. That is not a path to liquidity. It is a closed loop. So the discipline is the same on every one of these. Size the real market, test whether the durable revenue can fill it at a price that holds, and name who buys it at the next mark. The growth rate is not the thesis. The ceiling is.

In 2 weeks we're launching Crossover Intelligence OS. This is the first unified application of the kind Alex Karp has pointed to as the missing gap in the market. We've been building it for the past 18 months, and we're finally ready to unveil it. We collapsed the entire stack into one interface. Coding, documents, survey research, the database, the CRM, accounting, live public-market data, web research, outreach, and contact enrichment, the work that used to live across Cursor, Microsoft Office, Qualtrics, Supabase, Alphasense and a dozen point tools, now happens in one system that runs the full institutional research workflow end to end. Point it at a market and it sizes a defensible TAM from primary sources only, Census and SEC, zero aggregators, reconciled top-down against bottom-up and stress-tested through adversarial bear, bull, and regulatory lenses before a single number ships. It flags what it can't defend as Undefended instead of papering over the gap. Then it turns that market into a buy list: founder-owned targets screened clean of PE and public ownership, with named leadership and verified emails. It reverse-engineers any company from the outside in, revenue model, product map, moat, and confidence-scored gaps. It runs Voice of Customer studies from survey design through synthesis in a session, not weeks. And it builds itself, with an in-house coding agent that reads our repos, writes changes, and ships pull requests. 60+ specialized skills, one orchestration layer, primary-source discipline throughout. The market is in for a rude awakening. Technical engineering is no longer the moat. Closing the gap between raw model capability and deployed institutional work takes ingenuity and vision, and that is exactly what we built. Crossover Intelligence OS isn't a set of tools we use. It's a system that extends itself. We can produce detailed firm profile builds. We can reverse engineer businesses. We can do full TAM deep dives. We can help you run proprietary competitive intelligence studies. Want Access? Sign Up Here: crossoverintelligence.com/au…

Quite the move since this inflection $MU

Crazy how well this trade has worked out. Sept 25, 2023: “grid constraints forcing their hands to use nuclear.” The basket since: • Talen +670% • Vistra +370% • Oklo +360% (post-IPO) • Constellation +170% • NuScale +140% • Cameco +100% S&P 500 same period: +73%.

$NTDOY - was cycling through old charts I've done and the Nintendo price action I mapped out in 2023 panned out exactly as technicals were suggesting. Pairing fundamentals + with chart framework = a good way to improve win rate.

Revenue segmentation is the whole game. ARR isn’t ARR if 40% sits in the exposed bucket, and the multiple shouldn’t treat it like it is. The cut that matters for any software name you own: Non mission critical (engagement layer): massive trouble. Easiest to displace with an agent or thin wrapper. No data gravity, no compliance moat. Net retention is the tell, watch it quarterly. Enterprise systems of record: insulated, in some cases growing with AI add-ons. Rip-and-replace cycles run 18 to 36 months, and CIOs prefer AI from the incumbent over a new vendor relationship. Pricing power holds. Mid market: price chasing. AI-native challengers undercut on seat price, and customers use those quotes to negotiate incumbents down even when they don’t switch. Margin compresses either way, and consensus models aren’t pricing it. SMB: huge risk. Lowest switching costs, no procurement friction, free or cheap AI tools clear the bar. Logo churn shows up first, then ARPU. System of record vs system of engagement is the frame. Engagement layer is exposed. Record layer is sticky.

The AI Investment Supercycle Hypothesis Here is my original post from August 11, 2025. Releasing the full version after multiple requests. Since I wrote this the math has only gotten worse. The thesis in one line: the current AI boom cannot clear its capital stack. We are heading into a severe crunch and a mass consolidation event. Growth equity will absorb most of the damage. In slightly longer form: Too many companies for the available spend. Too much capital chasing too small a market. Too much dependency on unprofitable infrastructure. Something has to give. What gives is the middle of the cap table. This is not a thesis against AI. The technology is revolutionary. The mistake is assuming every AI company is. Growth equity is going to take the worst of it. The investors who wrote the largest checks at the highest valuations will be the ones marking down hardest. Most LPs do not yet know how exposed they are. The era of indiscriminate AI hype investment is winding down. What follows will separate the enduring players from the rest. The math has the final say. It always does. And here is the case, broken down... 1. Too many companies, not enough market > ~70,000 AI startups globally. Most are "AI-enabled" software vendors riding the same handful of foundation models. > The total AI hardware and software market is forecast at $780B to $990B by 2027. > Venture and corporate investors have already deployed $500B to $600B into AI companies. > To clear typical 10x return hurdles, those companies would need to generate $5T to $6T in revenue, roughly 6x the entire expected market. > Sequoia's David Cahn pegged the gap between AI capex and AI revenue at a "staggering $500 billion" annual hole that has to be filled to justify the spend. > AI-powered is the new .com. The structural setup looks like 1999. The math doesn't add up. There is not a pie large enough to justify the cap table. 2. Funding has surged into a market that cannot absorb it > Generative AI startups raised a record $56B from VCs in 2024 across 885 deals, nearly double the prior year. ~33% of all global VC funding in 2024 went to AI. > Most of those rounds priced in future dominance with little proven revenue. > One mid-2025 read: 90% of "AI companies" are just expensive wrappers around the same 5 foundational models. > Too many startups, same customers, undifferentiated tech, all assuming each captures outsized share. 3. The economics are inverted by subsidy The core platforms are running at massive operating losses to grab share, and that is what is keeping the rest of the stack alive. > OpenAI: ~$4B revenue in 2024 against ~$9B in spend. Around $5B in losses for the year. > One read on the unit economics: OpenAI spends $2.25 for every $1 it earns. > Anthropic: $5.6B cash burn in 2024 against well under $1B in revenue. Projected ~$3B loss on ~$3.7B revenue in 2025. > API prices are artificially low. Investors are subsidizing AI compute to drive adoption. When the subsidy ends, the entire downstream collapses with it. 4. The wrapper economy is built on sand > An estimated 30,000+ companies are "AI wrappers" that call foundation model APIs, repackage the output, and resell it. > Same engines means interchangeable products. No IP, no moat, just a well-structured API call, some markup, and marketing. > Many charge $50 to $100 per month for what a power user could replicate with direct API calls for a few dollars. > Every token sent through a wrapper, paid or not, earns OpenAI money. > Wrapper startups are effectively unpaid distribution arms, subsidizing OpenAI's growth while bleeding out. The house always wins. Until the wrappers run out of cash. 5. The commoditization spiral The vicious circle for everyone except the foundation model leaders: > Same models everywhere means undifferentiated products. > No real differentiation means pricing wars. GPT-4 price cuts already undercut Anthropic by up to 7x on cost per token in some configurations, forcing the industry to match. > Lower price per customer means collapsed gross margins for any startup reselling AI. > Lower revenue makes it impossible to support previous valuations or absorb still-rising compute and energy costs. > Unit economics flip upside down. More users actually means more losses. > These companies cannot operate without continual investor subsidy. A race to the bottom is great for end users in the short term. It is lethal for thousands of me-too vendors. 6. Growth equity is the most exposed pool of capital This is the part the LP letters are not yet ready to write. > Growth funds wrote $25M to $100M checks at $400M+ valuations on the assumption of 1 or 2 failures out of 10. > Most companies funded during the 2021 to 2023 boom had 18 to 36 months of capital. Many will run dry by late 2025 or early 2026. > Industry observers now predict 90%+ of AI startups fail within 5 years. > Conservative scenario, 60% to 70% failure: ~$400B of the ~$600B deployed gets wiped. > Realistic scenario, 80% to 90% failure: $500B+ in losses. > PitchBook already shows AI deal count down 42% in 2024 as reality starts to filter in. > A growth fund with 40% of its book in AI could see an 80% write-down on that slice. That would be historic. Early-stage VC tolerates 6 or 7 losses out of 10. Growth equity does not. The model breaks. 7. The exit window is closing too > The 2025 IPO window for tech is selective at best, and the listings that have priced are not delivering exit multiples that clear late-stage entry. > M&A is the realistic outcome, but acquirers will wait until valuations crumble. > A $50M check at a $500M post will recoup pennies in a fire sale. 8. Timeline: from bubble to shakeout 2021 to 2023, build-up: GPT-3, DALL-E, generative AI breakthroughs trigger a flood of startup creation. AI is the new electricity. FOMO drives indiscriminate funding. 2024, peak froth: $56B into genAI. Headlines everywhere. Cracks underneath: GPU bottlenecks, startups with negligible traction, and Big Tech racing ahead. Even the leaders are unprofitable. OpenAI doubles ARR from $6B to $12B in H1 2025, still expects ~$14B loss for the year. 2025, saturation: Every vertical is crowded. Sales cycles lengthen as CIOs get fatigued by the 50th similar pitch. CAC climbs. Big Tech bundles AI into existing platforms, often free, undercutting standalones. Investors get selective. Late 2025 to 2026, the crunch: Boom-era runways exhaust. Funding environment is harder. LPs are nervous about AI overexposure. Sharp pullback for everyone except the top 5%. Down rounds and outright failures pile up. Sentiment flips from FOMO to caution. The mere mention of AI no longer secures a premium. 2027, mass extinction: The global liquidity squeeze hits. Growth equity and crossover capital largely retreat. Thousands of AI startups fold in a 12 to 18 month window. Direct analog to the 2000 to 2001 dot-com collapse. 2028 to 2029, reset: Survivors fall into two camps. Foundation model and infrastructure leaders. Specialists with truly defensible domain or data moats. With less crazy competition, pricing power returns. API rates rise to profitable levels. A few big winners emerge with public-market validation. Many VC funds report poor returns on the bubble cohort. 9. The endgame Three things hit at once: > Valuation collapse. Multiples compress dramatically. AI startups that raised at 100x forward revenue trade at 5 to 10x if they survive at all. Private valuations could fall 70% to 90% before finding a floor. > Mass closures. Not dozens. Thousands of companies disappear in a short window. Talent and IP get absorbed by larger players. > Industry reset. Survivors capture larger shares. Pricing rises as subsidies fade. Focus narrows from gimmick AI features to core uses that deliver real ROI. A handful of mega-winners dominate and finally make money on AI. The truly useful AI applications and companies remain and thrive under more rational economics. The middle of the cap table does not.

𝗧𝗛𝗘 𝗔𝗜 𝗣𝗟𝗔𝗬𝗕𝗢𝗢𝗞 𝗙𝗢𝗥 𝗢𝗣𝗘𝗥𝗔𝗧𝗢𝗥𝗦 𝗔𝗡𝗗 𝗜𝗡𝗩𝗘𝗦𝗧𝗢𝗥𝗦 The technical stack matters more than ever. If the vendors in your stack aren't shipping aggressively, your competitors are already outpacing you. Every C-suite and CTO should be running a full evaluation of vendor product development timelines right now. 𝗦𝗽𝗲𝗲𝗱 𝗶𝘀 𝘁𝗵𝗲 𝗳𝗮𝗰𝘁𝗼𝗿 𝘁𝗵𝗮𝘁 𝗺𝗮𝘁𝘁𝗲𝗿𝘀 𝗺𝗼𝘀𝘁, and the delta between internal tooling and your existing vendors is what determines whether you're a 𝗹𝗲𝗮𝗱𝗲𝗿 𝗼𝗿 𝗮 𝗹𝗮𝗴𝗴𝗮𝗿𝗱. Before I get into it, three questions every founder and C-suite team needs to answer: 1) When did you last audit your vendor stack for shipping velocity? 2) If Anthropic or OpenAI shipped your core capability tomorrow as a native feature, what would you still own? 3) Can you, the CEO, explain your product's AI architecture in under 5 minutes without deferring to your CTO? 𝗔𝗡𝗧𝗛𝗥𝗢𝗣𝗜𝗖 𝗕𝗔𝗦𝗘𝗟𝗜𝗡𝗘 In 30 days: Claude Opus 4.7. Claude Design. Project Glasswing with Mythos Preview. Cowork GA on Mac and Windows. Computer use in Cowork and Claude Code. Interactive apps on mobile. Auto mode and /ultrareview in Claude Code. 𝗠𝗼𝗿𝗲 𝗽𝗿𝗼𝗱𝘂𝗰𝘁 𝘁𝗵𝗮𝗻 𝗺𝗼𝘀𝘁 𝗦𝗮𝗮𝗦 𝘃𝗲𝗻𝗱𝗼𝗿𝘀 𝗵𝗮𝘃𝗲 𝘀𝗵𝗶𝗽𝗽𝗲𝗱 𝗶𝗻 𝘁𝘄𝗼 𝘆𝗲𝗮𝗿𝘀. Imagine an Olympian showing up to your after-work sports league. Cool to watch, not a fair fight. And they sit in the 𝗿𝗶𝘀𝗸-𝗳𝗿𝗲𝗲 𝗶𝗻𝗰𝘂𝗯𝗮𝘁𝗼𝗿 𝘀𝗲𝗮𝘁. Every business building on Anthropic's platform to stay competitive is also showing them which vectors are worth attacking. Claude Design came for Figma the week it shipped. Claude Code came for Cursor. 𝗬𝗼𝘂 𝗯𝘂𝗶𝗹𝗱 𝘁𝗵𝗲 𝗺𝗮𝗿𝗸𝗲𝘁. 𝗔𝗻𝘁𝗵𝗿𝗼𝗽𝗶𝗰 𝗿𝗲𝗮𝗱𝘀 𝘁𝗵𝗲 𝘀𝗶𝗴𝗻𝗮𝗹. 𝗧𝗵𝗲𝘆 𝘀𝗵𝗶𝗽 𝗻𝗮𝘁𝗶𝘃𝗲. Reading the signal only works if you can act on it faster than anyone else, and that's why 𝗔𝗻𝘁𝗵𝗿𝗼𝗽𝗶𝗰'𝘀 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝘀 𝗿𝘂𝗻 𝘄𝗶𝘁𝗵𝗼𝘂𝘁 𝗮 𝘁𝗼𝗸𝗲𝗻 𝗰𝗲𝗶𝗹𝗶𝗻𝗴. The NYT reported one engineer spent over $150,000 on Claude Code in a single month. 𝗧𝗵𝗮𝘁'𝘀 𝗻𝗼𝘁 𝗮 𝗯𝘂𝗴. 𝗧𝗵𝗮𝘁'𝘀 𝘁𝗵𝗲 𝗶𝗻𝗰𝗲𝗻𝘁𝗶𝘃𝗲 𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲. Uber rolled Claude Code to 5,000 engineers and 𝗯𝗹𝗲𝘄 𝗶𝘁𝘀 𝗲𝗻𝘁𝗶𝗿𝗲 𝟮𝟬𝟮𝟲 𝗔𝗜 𝗯𝘂𝗱𝗴𝗲𝘁 𝗶𝗻 𝗳𝗼𝘂𝗿 𝗺𝗼𝗻𝘁𝗵𝘀. CTO Praveen Neppalli Naga: "I'm back to the drawing board, because the budget I thought I would need is blown away already." He's replanning upward, not pulling back. 𝗧𝗵𝗮𝘁 𝗮𝘀𝘆𝗺𝗺𝗲𝘁𝗿𝘆 𝗶𝘀 𝘁𝗵𝗲 𝗴𝗮𝗽: Anthropic ships without a cost ceiling. Every buyer ships with one. 𝗦𝗣𝗘𝗘𝗗 𝗜𝗦 𝗧𝗛𝗘 𝗠𝗢𝗔𝗧 Every SaaS vendor in your stack should be benchmarked against internal dev velocity. If your team with Claude can ship in weeks what your vendors take quarters to deliver, those vendors aren't buying you time anymore. 𝗧𝗵𝗲𝘆'𝗿𝗲 𝗰𝗼𝘀𝘁𝗶𝗻𝗴 𝘆𝗼𝘂 𝘁𝗶𝗺𝗲. The question on every renewal is no longer "is this vendor reliable." It's 𝘄𝗼𝗿𝘁𝗵 𝘁𝗵𝗲 𝘃𝗲𝗹𝗼𝗰𝗶𝘁𝘆 𝗴𝗮𝗽 𝗶𝘁 𝗰𝗿𝗲𝗮𝘁𝗲𝘀. If the answer is no, they're the thing killing you. 𝗩𝗜𝗕𝗘 𝗦𝗟𝗢𝗣 𝗜𝗦 𝗔 𝗦𝗞𝗜𝗟𝗟 𝗜𝗦𝗦𝗨𝗘 Most people dismissing AI coding as slop are reacting to bad implementations, not AI itself. The slop they're seeing is real. 𝗜𝘁'𝘀 𝗮 𝘀𝗸𝗶𝗹𝗹 𝗮𝗻𝗱 𝗮𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲 𝗽𝗿𝗼𝗯𝗹𝗲𝗺, 𝗻𝗼𝘁 𝗮 𝗺𝗼𝗱𝗲𝗹 𝗽𝗿𝗼𝗯𝗹𝗲𝗺. Good engineers put proper bounds in place: modular architecture, shared layers, fallback patterns, drift detection. 𝗗𝗼𝗻𝗲 𝗿𝗶𝗴𝗵𝘁, 𝘃𝗶𝗯𝗲 𝗰𝗼𝗱𝗶𝗻𝗴 𝗶𝘀 𝟭𝟬𝘅 𝗹𝗲𝘃𝗲𝗿𝗮𝗴𝗲. The architecture you build on determines whether AI compounds or corrupts. Proof point: Boris Cherny, head of Claude Code, said Anthropic's AI coded "pretty much all" of Cowork, 𝗯𝘂𝗶𝗹𝘁 𝗶𝗻 𝗿𝗼𝘂𝗴𝗵𝗹𝘆 𝘁𝘄𝗼 𝘄𝗲𝗲𝗸𝘀. 𝗧𝗛𝗘 𝗦𝗧𝗔𝗥𝗧𝗨𝗣 𝗪𝗜𝗡𝗗𝗢𝗪 Startups with no legacy architecture have a 𝗻𝗮𝗿𝗿𝗼𝘄, 𝗿𝗲𝗮𝗹 𝘄𝗶𝗻𝗱𝗼𝘄 𝘁𝗼 𝗵𝗶𝘁 𝗔𝗥𝗥 𝘁𝗵𝗿𝗲𝘀𝗵𝗼𝗹𝗱𝘀 𝗻𝗼 𝗼𝗻𝗲 𝘁𝗵𝗼𝘂𝗴𝗵𝘁 𝗽𝗼𝘀𝘀𝗶𝗯𝗹𝗲. Disadvantage: data breadth. Advantage: data capture design. 𝗔𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁 𝗔𝗜-𝗻𝗮𝘁𝗶𝘃𝗲 𝗳𝗿𝗼𝗺 𝗱𝗮𝘆 𝗼𝗻𝗲 𝗮𝗻𝗱 𝘆𝗼𝘂 𝗰𝗼𝗺𝗽𝗼𝘂𝗻𝗱 𝗽𝗿𝗼𝗽𝗿𝗶𝗲𝘁𝗮𝗿𝘆 𝗱𝗮𝘁𝗮 𝗳𝗮𝘀𝘁𝗲𝗿 𝘁𝗵𝗮𝗻 𝗲𝗻𝘁𝗲𝗿𝗽𝗿𝗶𝘀𝗲𝘀 𝗰𝗮𝗻 𝗽𝗼𝗿𝘁 𝘁𝗵𝗲𝗶𝗿𝘀. 𝗔𝗜-𝗙𝗜𝗥𝗦𝗧 𝗖-𝗦𝗨𝗜𝗧𝗘𝗦 𝗦𝗛𝗢𝗨𝗟𝗗 𝗕𝗘 𝗔𝗡 𝗨𝗡𝗗𝗘𝗥𝗪𝗥𝗜𝗧𝗜𝗡𝗚 𝗥𝗘𝗤𝗨𝗜𝗥𝗘𝗠𝗘𝗡𝗧 𝗙𝗼𝘂𝗻𝗱𝗲𝗿𝘀 𝗮𝗻𝗱 𝗖-𝘀𝘂𝗶𝘁𝗲𝘀 𝘄𝗵𝗼 𝗱𝗲𝗹𝗲𝗴𝗮𝘁𝗲 𝗔𝗜 𝗳𝗹𝘂𝗲𝗻𝗰𝘆 𝗵𝗮𝘃𝗲 𝗮𝗹𝗿𝗲𝗮𝗱𝘆 𝗹𝗼𝘀𝘁. Secondhand information moves too slowly in a market where architecture decisions compound daily. C-suites set vision, and successful AI deployments come from leaders who can 𝗰𝗼𝗻𝗳𝗶𝗴𝘂𝗿𝗲, 𝗻𝗼𝘁 𝗷𝘂𝘀𝘁 𝗮𝗽𝗽𝗿𝗼𝘃𝗲. If you're deferring to someone more technical than you on strategic AI calls, you're reacting to a market someone else is reading in real time. 𝗪𝗛𝗔𝗧 𝗔𝗖𝗧𝗨𝗔𝗟𝗟𝗬 𝗦𝗟𝗢𝗪𝗦 𝗘𝗡𝗧𝗘𝗥𝗣𝗥𝗜𝗦𝗘𝗦 𝗧𝗵𝗲 𝗿𝗲𝗮𝗹 𝗸𝗶𝗹𝗹𝗲𝗿𝘀: procurement cycles, security review, SOC 2 / HIPAA / FedRAMP regimes, change management, and RAG that doesn't scale to real corpus volume. Most enterprises are managing AI like a side project inside a pre-AI codebase. 𝗧𝗵𝗲 𝗽𝗿𝗼𝗱𝘂𝗰𝘁𝗶𝗼𝗻 𝗯𝗿𝗲𝗮𝗸𝘀 𝗮𝗿𝗲 𝘁𝗵𝗲 𝘀𝘆𝗺𝗽𝘁𝗼𝗺, 𝗻𝗼𝘁 𝘁𝗵𝗲 𝗱𝗶𝘀𝗲𝗮𝘀𝗲. 𝗧𝗵𝗲 𝗳𝗶𝘅 𝗶𝘀𝗻'𝘁 𝗮𝗻𝗼𝘁𝗵𝗲𝗿 𝗽𝗶𝗹𝗼𝘁. Stand up a dedicated AI review track so security and procurement don't restart from zero on every vendor eval. Put a single C-suite owner on AI budget with authority to spend, not a committee. Build AI-native services parallel to the legacy codebase, not bolted onto it. 𝗧𝗵𝗶𝘀 𝗶𝘀 𝘄𝗵𝗲𝗿𝗲 𝗖𝗿𝗼𝘀𝘀𝗼𝘃𝗲𝗿 𝘀𝗽𝗲𝗻𝗱𝘀 𝗺𝗼𝘀𝘁 𝗼𝗳 𝗶𝘁𝘀 𝘁𝗶𝗺𝗲. We score AI Capability and AI Resilience across PE portfolios and operators, and most organizations are further behind than they think. 𝗧𝗛𝗘 𝗪𝗥𝗔𝗣𝗣𝗘𝗥 𝗕𝗟𝗢𝗢𝗗𝗕𝗔𝗧𝗛 The majority of companies marketed as "AI-native" are repackaged base models with a UI on top. 𝗙𝗼𝘂𝗻𝗱𝗲𝗿𝘀 𝗮𝗻𝗱 𝗶𝗻𝘃𝗲𝘀𝘁𝗼𝗿𝘀 𝗻𝗲𝗲𝗱 𝘁𝗼 𝗴𝗲𝘁 𝗺𝘂𝗰𝗵 𝘀𝗵𝗮𝗿𝗽𝗲𝗿 𝗮𝘁 𝘁𝗲𝗹𝗹𝗶𝗻𝗴 𝘁𝗿𝘂𝗲 𝗺𝗼𝗮𝘁𝘀 𝗳𝗿𝗼𝗺 𝗮 𝗵𝗲𝗮𝗱 𝘀𝘁𝗮𝗿𝘁 𝗱𝗿𝗲𝘀𝘀𝗲𝗱 𝘂𝗽 𝗮𝘀 𝗱𝗲𝗳𝗲𝗻𝘀𝗶𝗯𝗶𝗹𝗶𝘁𝘆. Some wrappers will monetize fast and sell to strategics in the next 2-4 years. Most won't. Two forces compress the field: AI participation costs (infrastructure plus token burn) keep rising, and customer wallet share for AI tooling is finite. Only so many winners emerge per category. 𝗧𝗵𝗲 𝗿𝗲𝘀𝘁 𝗿𝗲𝘁𝘂𝗿𝗻 𝗰𝗮𝗽𝗶𝘁𝗮𝗹 𝗮𝘁 𝗯𝗲𝘀𝘁, 𝘇𝗲𝗿𝗼 𝗮𝘁 𝘄𝗼𝗿𝘀𝘁. 𝗪𝗛𝗢 𝗦𝗨𝗥𝗩𝗜𝗩𝗘𝗦 Three archetypes make it through the cycle. Vendors with proprietary data, engineered workflows, regulated-vertical lock-in, or hardware-attached integrations deep enough that a native platform release can't replace them. SaaS incumbents that rebuild before their slot gets benchmarked out of the stack. Founders fluent enough to dictate AI strategy, not delegate it. Configured, not approved. 𝗜𝗳 𝘆𝗼𝘂 𝗰𝗮𝗻'𝘁 𝗮𝗿𝘁𝗶𝗰𝘂𝗹𝗮𝘁𝗲 𝗵𝗼𝘄 𝗟𝗟𝗠𝘀 𝗳𝗶𝘁 𝘆𝗼𝘂𝗿 𝗽𝗿𝗼𝗱𝘂𝗰𝘁 𝗶𝗻 𝘂𝗻𝗱𝗲𝗿 𝟱 𝗺𝗶𝗻𝘂𝘁𝗲𝘀, 𝘆𝗼𝘂'𝗿𝗲 𝗻𝗼𝘁 𝗶𝗻𝘃𝗲𝘀𝘁𝗮𝗯𝗹𝗲. Email [email protected] with "investor checklist" or "founder checklist" and I'll send you the 20 questions every investor or every founder should be asking right now.