7 min read · 15 sources

SpaceX spent $60 billion on Cursor in the first half of the year. That single deal is roughly half of every dollar changing hands in software M&A so far in 2026. Pull it out and the rest of the market looks like a decade low.

That’s the headline from Kroll’s Summer 2026 Global Software Sector Update, which projects 2,672 software transactions this year, second only to one prior peak. The count looks healthy; the dollars do not, once you adjust for the SpaceX/Cursor outlier. For anyone running a $10M-$50M ARR B2B business, the practical read is that buyer appetite for small-cap deals remains weak even when headlines scream about a hot market.

Rebuilding isn’t the moat; the sales motion is, and AI hasn’t made phone calls cheaper.

Rule of 40 is half dead, and category beats it

Kroll’s numbers also punch a hole in the Rule of 40 framework that has governed SaaS valuations for a decade. Engineering software and HCM both score exactly 46% on the Rule of 40 metric (growth plus margin). They trade at 5.2x and 3.0x revenue respectively, a 73% premium for identical fundamentals, all attributable to category framing. Engineers deciding where to build should internalise that “what you sell” now beats “how well you sell it” at the same score.

Growth itself is doing more of the explanatory work than it used to. Companies growing above 20% trade at 7.2x revenue; the 10-20% cohort sits at 4.1x. Margin gains past 25% barely move the dial. Cursor alone drove 64% of Q2 software deal value, the most concentrated single-deal impact Kroll has recorded in ten years of tracking.

Pre-revenue is a strategy, and it's working

Source: galratner.substack.com ↗

If you can’t be judged on revenue, you can’t be judged at all. That is the throughline of Gal Ratner’s piece on early-stage AI valuations. Safe Superintelligence has roughly 50 staff, no released product, and a $32B valuation, about $640M per employee. humans& announced a $480M seed at $4.48B three months after founding with about 20 people, or $224M per person. Kleiner Perkins wrote a $75M Series A into Instinct at $500M; 21 days later Instinct raised a $250M Series B at $2.5B. That is $66,000 per minute of mark-up with nothing shipped.

TPG president Todd Sisitsky publicly described early-stage AI valuations of $400M to $1.2B per employee as “breathtaking.” Engineers should not mistake these marks for fundamentals. They reflect compressed expectations of future value; the moment a metric arrives, the scaffolding falls away.

Short drama has cracked mobile acquisition

Source: revenuecat.com ↗

Quibi died in 2020 trying to do mobile-native serialized video. Short drama, the format that grew out of China’s micro-drama market (576 million viewers by June 2024, over half the country’s internet users, formalised as a distinct format in 2020), has now done it properly. RevenueCat’s analysis puts short-drama apps at 6.5% of the highest-grossing entertainment apps and 38% of downloads. Within 35 days, 5.5% of installs become paying customers against 1.6% for the rest of the entertainment category.

The reason is structural. The ad, the cliffhanger, the free episodes, and the paywall are designed as one continuous experience rather than four separate steps owned by four separate teams. Stories are engineered around base desires (revenge, romance, wealth) with each episode ending on a payment trigger. The weakness is retention: 31.1% renew after the first purchase against 44.3% across the broader entertainment set. For engineers, the lesson is that aggressive paid acquisition plus immediate monetisation can work if retention is treated as a separate, second-act problem.

Slackbot grows up: Skill Sets and Deep Research

Source: slack.com ↗

Slack’s Slackbot update moves the agent from a one-on-one assistant to team-scale infrastructure. Skill Sets bundle related skills, like the “Helpdesk AI Kit” with triage, ticket creation, and escalation, so an admin can ship them across an org in a single click rather than turning on each capability by hand. Deep Research synthesises organisational data and the public web into cited reports. Big Mode hands the agent a full-screen workspace for long-form research without tab-switching.

The operational consequence for IT teams is that Slack’s deployment model is shifting from “individual AI features” to “agentic packages with default behaviour.” Worth piloting on a small org before rolling out company-wide.

ChatGPT Work is two products, not one

Source: simonwillison.net ↗

OpenAI’s ChatGPT Work, available to $20+/month subscribers only, splits into Work Cloud (web/mobile) and Work Local (the desktop app formerly known as Codex). Work Local has direct access to the filesystem and can execute code with internet, run a headless Chrome, and persist state across sessions. Sub-agents can publish ChatGPT Sites and run scheduled prompt automations.

The persistent filesystem plus code execution plus headless browser is the real differentiator from standard ChatGPT. Engineers building agentic workflows should map their state-management assumptions against Work Local’s model, not against the cloud variant.

PMF is a treadmill and the speed just changed

Source: thegtmnewsletter.substack.com ↗

GTM Newsletter argues that product-market fit was always a treadmill requiring continuous re-earning, but the AI innovation cycle has shortened the loop from years to quarters. Teams pulling ahead are building for the intelligence they expect to have in six to nine months, and compounding usage into model performance. Andy Rachleff’s original definition still holds (a value proposition that resonates with customers who buy repeatedly), but Marc Andreessen’s “good market with a good product” version is now a moving target. Founders should ask not only “do we have PMF” but also “are we compounding fast enough to keep it.”

Apps are losing their grip on customers

Source: blog.mastykarz.nl ↗

Mastykarz’s piece on agent UX makes a sharp operational point: your app now competes with “not needing an app at all.” For deterministic tasks like reviewing a familiar pipeline view, the UI still wins. For repetitive multi-record work like moving 12 opportunities to the next stage, agents win because the cost of delegating is now lower than the cost of clicking through. The implication for product teams is that interfaces need to match the job; some tasks deserve a UI, others deserve an agent handler, and the choice between them should be deliberate.

Mintlify ships mint signup and exposes the deeper problem

Source: apievangelist.com ↗

Mintlify’s CLI signup creates accounts from the terminal for coding agents. 66% of Mintlify’s 250M+ July 2026 queries came from Claude Code, Cursor, and Devin, so a browser-based signup flow was effectively blocking onboarding for their largest user segment. The command writes credentials to ~/.config/mintlify/config.json, but it blocks until the user clicks a verification email link, documented as taking “several minutes,” so background execution is required.

The underlying problem is that agents cannot install 27,464 CLIs or read 27,464 blog posts to discover each one’s syntax. A proprietary CLI command fixes Mintlify’s onboarding but does not fix cross-platform agent onboarding. The standardisation question stays open.

The scanner that said "don't"

Source: petervijeh.com ↗

A scanner run against Constellation Software’s 1,074-company portfolio (vertical SaaS in dental labs, municipal billing, funeral homes, transit fare) scored rebuildability by one engineer plus pricing transparency. Only ~1% of the 442 English-language sites publish prices; 9 of 442 clear the pricing screen at all; 2 clear it well. The thesis failed because Constellation deliberately buys phone-sold, contract-renewing software. Rebuilding is trivial. Selling by phone is the actual hard part, and AI has not made phone calls cheaper.

Founder-market fit beats age and experience

Source: zlatkov.substack.com ↗

Zlatkov’s data shows founders with 3+ years in their target industry are roughly twice as likely to build a top 0.1% company, and the mean founder age among fastest-growing ventures is 45. Passion outweighs experience though: decade-long veterans who are tired of an industry underperform passionate newcomers who learn fast. Engineers starting companies should assess whether they genuinely care about the problem space, not just whether they can ship.

Miscellany worth noting

Source: growthunhinged.com ↗

A one-year solopreneur retrospective on Growth Unhinged: ~$1M annual revenue, 88,000 subscribers, 700 premium, zero employees, 55% brand partnerships, 25% consulting, AI referrals up from 6% to 15.6% of new subscribers in six months. A $100 ticket experiment at Basecamp doubled show-up rate from 45% to 90% with the same two-person team. A 1997 realtor story captured 20% market share in a 2,000-home neighbourhood by mailing one useful postcard a week for eight weeks before the fall selling season. The lesson is the same in 1997 and 2026: show up during the consideration phase, not at the decision.