PROMPTED: The Claw Economy

Last night I wanted a more fun survey page for Fintech Nerdcon. Custom sounds, branded buttons, the works.

I built it and deployed it to Vercel in 30 minutes.

I'm not a developer.

That's a one-way door. Once you walk through it, everything changes. In an alarming, magical, scary and hopeful all at once kind of way.

If you haven't had that moment yet, this piece is for you. If you have, you already know what I'm about to say.

AI is moving 3x faster than Moore's Law. The task length AI can handle is doubling every four months. By year end, 20% of all GitHub code commits will be by AI.

And most people have no idea.

You're in a bubble. Just not the one you think.

This chart is the best reality check I've seen in a while.

Each dot represents 3.2 million people. The entire grid is 8.1 billion humans.

• 84% have never used AI. 6.8 billion people.

• 16% have used a free chatbot. The green strip at the bottom.

• 0.3% pay $20/month for AI. That tiny yellow sliver.

• 0.04% use coding scaffolds like Cursor or Claude Code. The barely visible red dots in the corner.

There's a good chance you're in the yellow or red. Which means your entire perception of how "mainstream" AI is comes from a bubble so small it's almost invisible on this chart.

And if you’re not in the red, what those 0.04% are experiencing is mind-blowing.

At the beginning of 2025, AI coding was a fancy auto-complete. By December 2025, Spotify engineers didn't write a single line of code.

Sure, you could argue that will be constrained to software engineering for the foreseeable.

But what if it isn't?

The METR task length chart lives rent-free in my mind. This is an independent research organisation that benchmarks frontier AI models. Claude Opus 4.6 has a 50% time-horizon of around 14.5 hours on software tasks.

Half the time, it can work independently for over 14 hours before needing human input. And the curve keeps going exponential.

When you hear "AI is a bubble" takes, remember. Your experience of AI can be wildly different from what's actually happening at the frontier.

The SaaSpocalypse was just the beginning

We're already seeing it play out.

Every SaaS vendor is finding contract renegotiations harder now it's possible to build more in-house with AI. Systems of record like ServiceNow are losing seats as companies cut staff.

That's today.

A scenario by Citrini Research lays out what happens next — and it's worth your time (link). They model the dominoes with brutal clarity.

• Personal AI agents begin to re-shop insurance renewals and opt out of low-usage subscriptions automatically.

• This drives job cuts in those industries, which hits consumer spending, which hits the credit card sector.

• Professional services — law, tax prep, financial planning — see custom agent solutions displace those jobs.

• That further hits consumption. 75% of American GDP. Which hits consumer brand companies held by private equity.

• PE and private credit are massively exposed to SaaS and recurring revenue. Revenue that is no longer recurring in an AI world.

Citrini calls it "Ghost GDP": output that shows up in the national accounts but never circulates through the real economy. A single GPU cluster in North Dakota is generating the output previously attributed to 10,000 white-collar workers in Manhattan.

Their core insight is devastating.

The US economy is one long daisy chain of correlated bets on white-collar productivity growth. The top 20% of earners account for roughly 65% of all consumer spending. When those workers lose their jobs or take 50% pay cuts, the consumption hit is enormous relative to the number of jobs lost.

Each company's individual response would be rational. The collective result becomes catastrophic. Every dollar saved on headcount flows into AI capability that made the next round of cuts possible.

And it's not just SaaS anymore. The contagion is spreading.

Last week, Anthropic released Claude Code Security.

Claude now does security audits (or claudits, if you will) — scanning codebases for vulnerabilities and suggesting targeted patches, catching issues that traditional tools miss.

The market reaction was immediate.

$15bn wiped from cybersecurity stocks. One product launch. One tweet.

Not SaaS. Cybersecurity. Palo Alto Networks, CrowdStrike, Cloudflare, Okta, Rubrik, SailPoint — almost every major name down in a single day.

Every team building a product has now delayed releases for a couple of weeks while they red-team their products with these new tools. That's a pattern. AI enters a category, reprices the incumbents overnight, and the affected companies respond by... adopting more AI. Which funds the next round of disruption.

If this seems far-fetched, consider the number of people posting about rebuilding SaaS dashboards and replacing marketing agencies with their own AI agents.

The issue isn't that this is impossible. It's that it's only the ultra-early adopters doing it.

For now.

The revenues confirm this is just the beginning

Anthropic has been growing at 10x per year since hitting $1B in annualised revenue. OpenAI at 3.4x. Anthropic could overtake OpenAI by mid-2026 if trends continue.

Anthropic's growth is driven by being the leader in coding models. Coding drives the flywheel into all other types of knowledge work. When people say "AI is a bubble" I say look at Anthropic's revenue growth, not OpenAI's.

OpenAI's financials look questionable — they've reset spend expectations to ~$600 billion by 2030, down from $1.4 trillion. Reality is biting.

And this raises a question. If the displacement scenario plays out, does the value only accrue to foundation model creators? Or does cost pressure push towards open weight models — consistently 3 months behind the frontier at 10x less cost?

That gap matters when AI becomes a significant line item.

Claws: Personal AI as the new operating system

So that's the disruption. Now here's what's enabling it.

Over the holidays the OpenClaw movement took over the internet. Nerds began buying Mac Minis and wiring them up to email accounts, Google Drive, and countless other services to do little daily jobs.

Andrej Karpathy coined the term "claws" for this. AI that runs on a device you own, connects to all your personal accounts and data, and works for you. Not provided by a 3rd party.

Yours.

I've been running one for a few weeks. Mine:

• Sweeps all 5 of my Google calendars for a daily summary

• Pings me to photograph food for macro tracking

• Takes voice notes and turns them into social posts with generated images for LinkedIn (still flaky, but you get the idea)

But here's the moment that got me.

My claw does a daily research sweep of the internet. I'd had both Perplexity and OpenAI doing this before. But this one learns from my feedback. Each time I tell it what was useful, what was noise, what I would have wanted to see — it gets a little better.

A little sharper.

A little more me.

It's gotten good enough that I'm now thinking about launching it as a separate daily email.

That's the thing about claws. You don't buy them as a product.

You raise them like children.

The pain is the point (for now)

I won't pretend it's easy. Honestly, it's frustrating a lot of the time. You have to put in the work to give feedback. Change your workflow. Adapt to a new way of doing things.

That pain is the gap for mass adoption right now. And it's the gap most sceptics point to. "This is too hard for normal people."

They're right. Today.

But that gap is closing fast. Because claws use skills.

Skills are the new apps

Skills are modular capabilities you can give your AI. Give it the "marketing agency skill" or the "financial modelling skill," and it goes from generic to really capable in one command.

Think about what this means.

Apps were to smartphones what skills are to claws. The App Store didn't just distribute software — it created an economy. A marketplace. An entire class of company.

Skills will do the same thing. Except the product isn't an app you download. It's a capability your AI absorbs.

Today, adding skills is too hard. You're fiddling with configs, writing prompts, debugging. That's temporary. Skills for adding skills will get easier. The tooling will improve. And when it does, the skill becomes the unit of value — not the software product.

This has profound implications.

If the skill is the product, then who builds skills? Who distributes them? Who quality-controls them? Is there a skill store? Do skills have network effects? (Probably — the more users refine a skill through feedback, the better it gets for everyone.)

The entire SaaS economy was built on the idea that you buy software products.

The claw economy is built on the idea that you install skills.

And this is already moving beyond hobbyists with Mac Minis.

The Airtable founder just announced Hyperagent — an agent platform where every session gets its own isolated computing environment in the cloud. No Mac Mini required. Real browser, code execution, data warehouse access, hundreds of integrations, and the ability to learn any new API as a skill.

But this line from that announcement is the one I can't stop thinking about:

Teach the agent how your firm evaluates startups or how your team runs due diligence — now anyone on the team gets output that reflects your actual methodology, not a generic template.

Read that again.

Your methodology. Your process. Your taste. Your judgment. That's what you install. That's what compounds. That's what your claw learns and carries forward.

Skills learn. Skills adapt. Skills compound over time. Products don't do that.

The unit of economic value is moving from the product to the skill.

That's what the Citrini disruption cascade, the SaaSpocalypse, the coding revolution, the claw movement — that's what they're all converging on. Products can be replicated. Methodologies — refined by your feedback, shaped by your judgment, compounding over time — are yours.

So where does this leave you?

The question isn't whether AI will reshape your market. That's happening.

The question is whether your methodology is worth teaching to an agent — or whether an agent can figure it out without you.

Drew Bent, after a year at Anthropic, made an observation that captures this: roles are becoming both more manager-like (directing agents) and more IC-like (everyone's a builder) at the same time.

Manager of agents. Builder with agents. Simultaneously. That's the new job description.

Start raising a claw. Not because it's easy — it's painful, and the tooling is rough. But the people who put in that work today are building something that compounds.

The feedback loop between you and your AI is personal.

Nobody else's claw will be quite like yours.

ST.