Quanta Bits - February 21, 2026

Hi everyone,

Another busy week of setting up and configuring OpenClaw as my COO and advisor. I've shifted all things that matter to me operationally to OpenClaw now. Obsidian is my source of truth, a markdown note-taking app where files sit on your machine and sync across devices; you own all your notes and I highly recommend it. OpenClaw is the daily driver. Even book and movie list management and curation, based on my past ratings and tastes, are fully in OpenClaw. Still figuring out the right setup. If you are thinking about it or using it already, would love to hear from you and share learnings!

The major theme this week continues to be the impact of AI on enterprise buyers and their automation practices. We are shifting from hype to practice. With SaaS software companies taking a beating in the market recently, that has been the source of many conversations about whether AI and build-it-yourself are replacing the old paradigm of SaaS.

I've been advocating for "Earn Your Complexity" as a technology philosophy: experiment and get your hands dirty to understand the AI paradigm, start small, prove value and learn about your data and capabilities. Let that inform your decision to build or buy later.

Four patterns from this week's reading that matter for operations leaders.

Private equity firms spent a decade and trillions of dollars on a simple thesis: enterprise software subscriptions are sticky, predictable, and nearly impossible to displace. Now two forces are hitting that thesis at the same time, and the fallout is reshaping how every enterprise thinks about software.

Firms like Thoma Bravo and Vista Equity Partners paid 20x+ revenue multiples for niche enterprise software companies, layering massive debt on the belief that software contracts were "better than first-lien debt." Software takeovers accounted for roughly 40% of all PE deal activity over the past decade. Then rising interest rates crimped cash flows on those leveraged deals. And AI made investors ask a question nobody wanted to hear: could AI recreate the specialized software these firms acquired for billions?

The public markets moved first. Salesforce and ServiceNow dropped by a fifth this month as investors priced in AI replacement risk across the software sector. $2 trillion wiped from software market cap. That repricing is now flowing downstream to PE portfolios, where the pain is worse: firms that bought private software companies at peak multiples with heavy leverage can't exit into a market that no longer values what they paid for. Funds raised between 2019 and 2022 by Thoma Bravo, Vista, and Insight Partners have returned a third or less of investors' initial capital, and most of the largest deals haven't been sold yet.

The responses split into two camps. Apollo is cutting software exposure and betting against the debt of AI-vulnerable companies, essentially wagering that some of these firms won't be able to service the loans PE stacked on them as their products lose value. John Zito, co-president of Apollo's asset management, calls it "a logical repricing of terminal value."

Vista is making the opposite bet. Robert Smith has built an "agentic factory" to embed AI agents across his $150bn portfolio. It's arguably the largest enterprise AI adoption experiment right now, and it's top-down mandated. If it works, incumbent vendors prove they can absorb AI rather than be disrupted by it. If it fails, bolting agents onto legacy products won't save a deteriorating value proposition.

No question in my mind that AI is disrupting the software world and the build vs. buy calculus with it.

The big infrastructure players, Salesforce, Oracle, ServiceNow, will survive. They store enterprise data and connect systems together. That plumbing stays valuable.

The squeeze is in the middle. Companies like Gainsight and similar mid-market vendors selling $200K-$1M specialized apps are in a harder spot. Large enterprises at the top of the pyramid have the talent and budget to build those tools internally, and an internal team can often do it for a fraction of the cost. That leaves mid-tier vendors needing to move down-market, targeting SMB and mid-enterprise where the build-it-yourself calculus doesn't work yet.

This thesis changes each week with new developments. But the direction is clear: the middle layer of enterprise software is getting compressed from both sides.

Sources:

96% of companies miss AI ROI because they buy tools instead of building capability
PwC found only 4% of companies achieve significant financial benefits from AI. The reason is consistent: most organizations spend the majority of their AI budget on technology and a fraction on training, enablement, and process redesign. The companies getting results flip that ratio. Buying the platform is the easy part. Getting people to use it effectively is where the ROI lives.

Mercor's AI agents failed at consulting, and the CEO said so out loud
Mercor, the $2B AI talent marketplace backed by Thiel and Horowitz, tested its agents on real consulting work and published the results: they failed. The agents could pull data and draft decks but fell apart on ambiguous, multi-step analysis. CEO Brendan Foody predicts agents will handle standard consulting engagements by end of 2026. The gap between "impressive demo" and "billable work product" is where the $300B consulting industry lives.

Marketing teams are now optimizing for chatbots, not just Google
With 800M weekly ChatGPT users, AI chatbots are where buyers increasingly research solutions - before their buying decision. Companies are responding by flooding the zone with content designed to be absorbed by AI models. Athenahealth published 250,000 words in six months to reshape how chatbots describe their products. One luxury brand went from 5 to 100 content pieces a month. The key insight: Reddit is cited in over half of ChatGPT's solution-seeking responses, ahead of YouTube, Wikipedia, and every news site. That's making Reddit strategy and old content cleanup urgent priorities for marketing teams that never had to think about what an AI model reads.

Japan's biggest toilet company is an AI stock

Toto, maker of Washlet bidets, has been making electrostatic chucks for NAND memory chips since the 1980s using ceramics expertise from toilet manufacturing. That segment now generates 40% of operating profit. Stock is up 60% in a year. And Toto isn't even the strangest crossover: Ajinomoto, the umami soup stock company, makes chip insulation material. The AI supply chain has interesting dependencies!

Hands-on lessons from Quanta Lab and the field. What practitioners are building, what it costs, and what actually works.

A few weeks into running my OpenClaw setup at full scale, I hit the wall I'd been expecting: cost. The Anthropic Max plan ($200/week) ran out a day early this week. I spent another $100 on overflow. And that's with aggressive model switching already in place.

Here's what's consuming tokens at that rate: email triage, task management, relationship management across thousands of LinkedIn connections, nightly industry research scans, and a security audit agent running every night. Each is useful individually. Together, they eat through a quota designed for heavy individual use.

The fix I'm testing: offloading non-real-time tasks to local models. I have QwQ 32B running on a Mac mini. It's slow, five to ten minutes per task, but the output quality is comparable to Sonnet for some batch work. Real-time interactions stay on Anthropic's Sonnet or Opus. Background research, scoring, and analysis shift to local.

This is the same pattern enterprises will hit at scale. The answer isn't less AI. It's tiered workloads: frontier models for real-time, judgment-heavy tasks; smaller or local models for everything else. The organizations that figure out this tiering early will spend a fraction of what their competitors spend for equivalent output.

"TokenOps" is about 18 months away from being a job title.

Chatbots have no temporal bias. Google's algorithm has always favored fresh content, creating a natural incentive to update or retire stale pages. ChatGPT and Gemini don't share that preference. They'll surface an outdated case study from 2015 or a pricing page from a discontinued product with the same confidence as your most recent content.

Athenahealth's CMO discovered this when she started querying chatbots about her own company and found them pulling from "insidery software websites that hadn't been updated in years." Fixing the AI presence wasn't just about publishing new content. They also had to go back and audit everything old.

This is a governance problem most teams haven't framed yet. Old content persists invisibly in chatbot answers, and you only find out when a prospect tells you what ChatGPT said about you. Content governance now means tracking what an AI says about your brand, not just what ranks on page one. In most organizations, nobody owns it yet.

Danny Boyle | 2025 | 109 min
Aaron Taylor-Johnson, Jodie Comer, Ralph Fiennes, Alfie Williams
★★✪☆☆

The promos for 28 Years Later - The Bone Temple look great and they have one of my favorite actors, Ralph Fiennes, in them. That got me to try to catchup with the series, watching the 28 Days/Weeks/Years Later over the past few weeks. So far, it has been a "meh", with the last one "28 Years Later" being the best. May be a trend upward?

28 Days Later (2002) started the "fast zombie" genre and Danny Boyle's empty London shots are iconic, but the plot was too predictable. 28 Weeks Later (2007) had a great opening farmhouse sequence, but too many silly plot holes, even for a zombie movie, to forgive. 28 Years Later (2025) was the best of the three. It had actual thematic direction beyond "escape the zombies," centering on a new generation of survivors that doesn't remember the old world and has accepted this life as normal.

Ralph Fiennes's role was more marginal than I expected. I'm hoping he gets more to work with in The Bone Temple, or that Cillian Murphy's return gives the series the anchor it needs for a strong finish. And I'm skeptical this is really the "final" trilogy. Final Destination taught us that "final" is a relative term in horror franchises. But I'll be there for The Bone Temple either way.