9 min read · 15 sources

Investment committees are hitting a wall on AI startups, and the numbers finally explain why. AI made it 10x easier to build a $10M ARR company, as one founder-investor put it, but did not make it 10x easier to build a $1B one. Carta’s data shows why: only 15.4% of Q1 2022 seed companies reached a Series A within two years, down from 30.6% for the 2018 vintage.

The result is what one seed and Series A investor at FPV Ventures calls “the new valley” - the $5M to $25M ARR stall zone where AI companies with great revenue curves but no differentiation grind to a halt. Every viral product has copycats within weeks, and investors can’t tell who’s building something durable. It’s not that revenue isn’t real; it’s that “why now” is just AI itself, and that’s not a moat.

A seed and Series A investor calls the stall zone the new valley, $5M to $25M in ARR, where 15.4% of seed companies now die.

The New Valley: Why $5M-$25M ARR Is Where AI Companies Stall

The thesis is uncomfortable because the revenue is real. Carta’s numbers show Series A conversion for seed companies nearly halved between 2018 and Q1 2022, and the AI boom made it worse by flooding the market with similar products. Investors pause weekly on companies with great revenue curves to ask what’s truly unique, and most can’t answer.

The four escape routes, per the FPV investor: talent M&A and “seed-strapping,” raising small, getting profitable fast, and keeping 80% of the company. Notice what’s missing - no path involves out-executing copycats with the same playbook. The difficulty is inherent to the venture job: identifying whether a startup can expand beyond its initial niche before the market does.

The Harness, the Horse, or the Hay: Pick Your Category or Die

Source: queener.substack.com ↗

There are three ways to make money in AI, and the strategies are close to inverted. The horse is the foundation model; the harness is the application layer; the hay is everything the model and harness need to stay healthy at scale.

The full essay argues that most startups should not build the horse - the capital and compute requirements are fatal for all but a handful. The playbook that works for a harness company (move fast, own the UX, churn through features) is wrong for a hay company (reliability, SLAs, boring infrastructure). Misapplying one category’s strategy to another is how founders burn capital and credibility. The horse is powerful but unpredictable; it requires skilled riders, and most teams aren’t experienced enough to control it.

The New Hiring Filter: Pedigree Dies, Judgement Wins

Source: iconiq.com ↗

Iconiq Growth’s report on AI hiring makes a claim that should unsettle every hiring manager: traditional signals like years of experience and prestigious employers are becoming less predictive of success. AI lowered the cost of execution, so the differentiator now is judgment - learning velocity and the ability to operate in unfamiliar situations.

The report advises founders to hire for the company’s biggest uncertainty and to value proof of capability over pedigree. This isn’t a feel-good shift; it’s a response to AI compressing the execution gap between a 5-year veteran and a sharp newcomer. If ChatGPT can write the boilerplate, the premium is on people who know what to build and when to build it. The caveat is that pedigree still matters, but it’s becoming a floor rather than a ceiling - and the report admits the evidence for what replaces it is still emerging.

The Company Does Not Live in the LLM

Source: hoeijmakers.net ↗

A sharp technical essay describes a failure mode that will become more common: an AI-heavy company’s knowledge became embedded within LLMs, reducing oversight. The fix was extracting durable organizational knowledge into explicit domains, often using Git, so knowledge belongs to the organization rather than any specific model.

The principle: deterministic systems establish context; LLMs reason over that context. Models should never establish whether something has happened or what the current state of a system is. Systems should assemble that information into a context document, and models should interpret, connect, reason, and synthesize. This is the difference between a company whose knowledge is trapped in a prompt cache and one that can swap models without losing its operational memory. If your knowledge lives only in the model, you don’t own your company.

Citation Share Is Not a Metric Yet

Source: agenticlandmark.com ↗

The citation share debate got the numbers it deserved, and they’re damning. A McKinsey study shows brand-owned sites account for 1-2% of LLM citations overall - but that figure varies threefold (3-10%) depending on the platform. ChatGPT, Gemini, and Perplexity produce wildly different numbers, so two monitoring tools will give you wildly different answers.

The argument: the discovery layer gets all the attention and spend, but its numbers are contested, produced by interested parties, and unstable across instruments. The infrastructure layer - whether agents can transact with you at all - returns a clean pass or fail on every check. If you’re building a strategy around citation share, you’re building on a number that moves depending on which tool you use to measure it. The article’s blunt conclusion: infrastructure is what gets you transacted with.

"I'm Churning": A CFO's Playbook for Killing Your SaaS Renewal

Source: onlycfo.io ↗

This is the person on the other side of your renewal itemizing how he decides to cancel. A working CFO lists what he’s cutting and why: a workflow tool that quoted him double what he pays for Claude, and a $50K-a-year point solution whose function a larger vendor now bundles for free. His playbook is opening renewal conversations four or more months early and spending 30-60 minutes demoing alternatives to build negotiating power.

The backdrop makes this worse: the Bessemer cloud index is up 55% in six months, its highest since early 2022. SaaS is rebounding, but CFOs are pushing hard on pricing and value. Increased AI spend is squeezing budgets, and the benchmark set by tools like Claude is resetting everyone’s price expectations. If your product can be replaced by a prompt and an API key at a tenth of the price, you’re not a SaaS company; you’re a feature that hasn’t been bundled yet.

The Player-Coach Role Is Breaking

Source: newsletter.eng-leadership.com ↗

Linear’s report on AI usage in software teams has a number that explains the burnout epidemic: teams using coding agents open 6x more pull requests than traditional teams, and 3x more than they did in 2024. Review load is climbing faster than anyone’s model of the job, and it lands hardest on engineering managers and tech leads - the people expected to write code and manage at once.

Managers already see only about 40% of what a report does in a day. AI was supposed to make the “player” part easier, but it’s actually making it harder by adding complexity and review burden. The article’s argument that this dual expectation is unsustainable reads like a cry for help, and the data backs it up. If you have a player-coach, you need a plan for them that doesn’t assume infinite capacity.

WorkOS Relay: Credentials Never Touch the Agent

Source: workos.com ↗

WorkOS’s Relay solves a problem everyone running AI agents should be paranoid about: bearer credentials in agent context are the target of prompt injection attacks. Relay proxies an agent’s third-party API calls and injects credentials at the boundary, so the agent never holds a token it could leak.

The design is elegant - the credential isn’t in the context, so no amount of prompt injection can exfiltrate it. Requests are proxied through WorkOS, where credentials are swapped in based on user and organization context. It’s in early access and requires contacting support, which is a sign the credential-swapping logic is still being hardened. For anyone building agents that call third-party APIs on behalf of users, this is the pattern to watch.

Cloudflare's Adaptive Intelligence: Making Bot Attacks Not Worth Running

Source: blog.cloudflare.com ↗

Cloudflare’s new bot detection engine reframes the problem honestly: determined attackers will always find a way through, so the goal is to make it not worth running. The engine analyzes over a trillion requests daily and adapts to changing tactics, making sure nothing attackers learn stays true. Each new attempt costs more than the last.

The key shift is treating detection as a statistical judgment across many signals rather than deterministic logic that an attacker can isolate and beat. It recognizes bots without visibly reacting to them, so bots can’t learn from the response. This undermines the economics of bot attacks - if every attempt gets more expensive and less effective, the campaign stops being profitable. That’s a better metric than block rate.

Why Adding More AI Agents Makes Your Team Slower

Source: blog.mempko.com ↗

Max Mempko has actual data that the “more agents = more output” crowd doesn’t want to hear. Going from 4 to 8 agents reduced his output from 70k to 67k lines of code per month. The Universal Scalability Law applies to AI agents the same way it applies to database transactions: coordination overhead and context switching costs eventually negate the benefits of adding more workers.

The implication is that there’s an optimal agent-per-codebase ratio, and it’s low. Coding is the easy part of writing software - it’s everything else that takes most of the time, and agents stepping on each other’s work makes it worse. If you’re adding a fourth agent because three isn’t enough output, the problem is probably not the number of agents.

OpenClaw's Accidental 2.0

Source: openclaw.ai ↗

OpenClaw 2.0 shipped after a seven-week pause that followed 106 releases in 230 days. The release carries 16,000+ merged pull requests from 933 contributors - about half of every PR ever merged into the project. The slower cadence was deliberate: rebuilding the foundation, simplifying installation, and reworking the browser app, messaging, memory, and skills.

The “accidental” framing is honest - the pause was supposed to be shorter, but foundation work expands to fill the time available. For users, the goal is getting to a useful “Claw” faster and letting it grow with simple workflows. For maintainers, the lesson is that 106 releases in 230 days is a treadmill, not a strategy.

Shopify + Stripe Projects, Figma's Generative Plugins, and the Rest

Source: x.com ↗

Shopify now works with Stripe Projects, letting users spin up a store from the command line with a single Stripe account for setup and billing across all providers. Figma’s generative plugins and shaders now support community publishing, animated shaders, code access, and MCP updates - a reminder that design tools are becoming code generators too. And a 16-minute read on enshittification argues the theory is wrong, pointing to rising user numbers and time spent on Instagram. Whether you buy it depends on whether you trust engagement metrics from the platforms under critique.