Software Is Dead, Long Live Software

Last week, roughly half a trillion dollars was wiped from software valuations.

The narrative was clean and dramatic: AI will do the work, software goes away, SaaS dies.

That same week, one of the most advanced AI companies in the world finished implementing enterprise HR software instead of replacing it.

Those two facts look contradictory only if you think software is mostly code.

It is not. Software is workflow, policy, memory, and accountability encoded into systems that organizations can trust.

I wrote recently that distribution wins in AI. I also argued in The Speed of Forgetting that labor markets reprice faster than people expect. This is the bridge between those ideas.

The Tier-1 Lens

If you want a clean way to think about this cycle, the best “reading stack” is surprisingly consistent.

Charlie Munger: incentives and social proof drive overreaction.

Ben Thompson and Benedict Evans: software is often a thin technical layer over very thick organizational process.

Peter Thiel: distribution, not product quality alone, determines who captures value.

Ronald Coase: when coordination costs fall, firm boundaries change and often expand before they contract.

Rockefeller: in every infrastructure transition, value accrues to the player who controls distribution and workflow, not necessarily to the inventor of the new tool.

These are not separate takes. They are one argument.

What “Software” Actually Is

From the outside, software looks like UI and code.

From the inside, enterprise software is usually a machine for institutional decisions:

• Who can approve what
• Which exceptions are allowed
• What gets logged for audit
• How compliance is enforced across regions
• How errors are handled when real life does not match the happy path

This is why the “AI writes code, therefore software disappears” thesis misses the center of gravity.

Code generation lowers build costs. It does not eliminate the need for institutionalized process.

AI can draft a workflow.

Software has to survive legal review, finance controls, security policy, and the Tuesday morning edge case no one remembers until it breaks.

That is why critical systems persist.

Where The Real Disruption Is

The disruption is not replacement. It is compression.

AI compresses the cost and time of feature development. Once that happens, category boundaries blur fast.

Startups can build more with fewer people.

Incumbents can copy obvious features quickly.

Products that looked differentiated at seed stage become checkboxes in larger platforms.

This is exactly where Thiel’s distribution point matters. If everyone can build, then the scarce asset shifts to reach, trust, installed base, and system-of-record position.

The question for software companies is no longer “Can we build this?”

The question is “Can we own the workflow before everyone else can ship the feature?”

Coase Was Early To This

Coase explained that firms exist because coordination has costs.

AI meaningfully lowers coordination costs inside organizations: drafting, analysis, synthesis, and execution all get faster.

The naive conclusion is that firms should splinter into tiny units.

Often the opposite happens first.

When coordination gets cheaper, the best-managed firms can absorb more scope. They can run more workflows through the same operating system, enforce standards faster, and spread improvements across a large base.

That is one reason large incumbents can get stronger during platform shifts, even while specific products inside them get disrupted.

Why Munger Still Matters Here

Munger’s psychology models explain the market behavior around this topic better than most technology forecasts.

Incentive-caused bias: AI vendors have every reason to frame timelines aggressively.

Social proof: once a narrative starts (“all white-collar work changes in 18 months”), everyone repeats it to avoid looking behind.

Contrast misreaction: markets overreact to visible model progress and underreact to the friction of real institutional adoption.

These forces create violent repricing.

Violent repricing is not the same as accurate long-term classification.

The Rockefeller Reminder

Rockefeller did not win because he predicted every technology change perfectly.

He won because he controlled distribution and operating infrastructure as the underlying technology shifted.

The AI-era parallel is straightforward.

The value pool will not be captured only by model builders. A large share will be captured by companies that own customer workflow and can continuously absorb new AI capability into existing trust relationships.

Tooling changes. Control points matter.

What To Do With This

For operators:

1. Build AI features, but prioritize workflow ownership over demo novelty.
2. Move toward system-of-record position in your niche.
3. Treat trust, auditability, and reliability as product features, not legal overhead.
4. Assume feature parity arrives quickly; design moats around distribution and embedded process.

For investors:

1. Separate “cool output” from “durable adoption under operating constraints.”
2. Underwrite companies that can absorb AI repeatedly, not just launch once.
3. Favor businesses with distribution and switching costs where AI increases throughput.
4. Be suspicious of businesses whose moat is only access to a model everyone else can also access.

Long Live Software

“Software is dead” is a clean headline.

Reality is messier and more investable.

Software is getting easier to build, harder to defend, and more critical to integrate.

The winners will pair AI speed with institutional depth.

Munger would call that the uncomfortable middle: two truths held at once.

AI is transformative.

And software, in its most important form, is still very much alive.