Why Software Is Eating Itself

Shel Silverstein wrote a poem called “Hungry Mungry” about a boy who eats everything. He starts with his dinner, works through his family and his house, then the town, the state, the country, the oceans, the planet, the sun, the galaxy. When the universe is gone he starts on his own feet and works upward until he finishes with his mouth. The last stanza reports the result. Nothing.

Silverstein’s poems often carry a little more than just entertainment, and in this case the poem happens to be particularly prescient about the future of SaaS. The appetite is not a response to the food. It’s a property of the eater. Removing the food doesn’t remove the appetite. It just changes what’s in reach.

Marc Andreessen published “Why Software Is Eating the World” in 2011. It was a triumphalist piece of writing and it was mostly correct. Fifteen years later the thesis has no serious opposition, because it has no serious work left to do. Retail is software. Taxis are software. Hotels are software with linens. Banking is software with a compliance department. Media is software with grievances, and your fridge probably has Twitter on it. The essay predicted that the software industry would consume the industries adjacent to it, and it did. But the meal is largely over.

And yet…Hungry Mungry is still hungry. Andreessen described an appetite and every startup for the better part of two decades has been treating it like a strategy.

Every dollar of B2B SaaS markup rests on one sentence: writing software is hard, so you rent ours instead.

Slack charges somewhere between eight and fifteen dollars per user per month for a chat interface. The technical content of a chat interface is not eight dollars a month of anything. Notion charges ten to fifteen for a document editor sitting on a database. Salesforce moves hundreds of dollars per seat for a CRM whose data model fits on a napkin: accounts, contacts, opportunities, activities.

None of those prices pay for code. They pay for the distance between “this exists, it works, and someone answers the phone” and “I could build this myself.” Software was hard, so the distance was wide, so the rent was collectible. The price has never been a function of engineering cost. It’s a function of your inability to do it yourself. This isn’t an accusation. The business model was a good one, but that’s literally all there is to it.

The gap is closing on the long tail. Not evenly, and not everywhere. But observably, inexorably, Hungry Mungry is running out of things to chomp. Today, a four-person team that needs an internal tool (approvals routing, inventory reconciliation, a customer portal that does one thing) can have a model write it over a weekend, run it on a ten-dollar VPS, and pay nothing in perpetuity for a function that cost five hundred a month per seat. The build is not good. It doesn’t even need to be good. It needs to be theirs, do the one thing, and do it well.

What does that tell you about where we’re headed? If the core of SaaS is “writing software is hard, so you rent ours instead.” What happens when writing software isn’t hard? When one developer and a few thousand in anthropic tokens can replace your service?

All of a sudden, SaaS is eating its own business model.

However, it’s worth noting this isn’t uniform. Many incumbents are insulated via network effects. Slack survives because the value is the other people in it. You can build a Slack over a weekend. You can’t build the coworkers or the institutional lock-in.

Regulatory surface. Salesforce inside a pharmaceutical company isn’t a just a CRM, it’s an audit trail with a CRM attached. Nobody validates a weekend build against 21 CFR Part 11.

Data gravity. Fifteen years of ticket history is not so much a schema problem as it is a hostage situation and you may want to continue meeting the demands.

All of these are real moats. They protect the top of the market. They don’t protect the long tail, and the long tail is where the seat count lives. The characteristic enterprise complaint of the last decade is “we pay six figures for this platform and use one feature.” That sentence used to be a joke about procurement. It’s a build spec now.

But the real irony here lies in the recursion. OpenAI and Anthropic are SaaS companies. Hosted, subscription, per-seat and per-token, gross margins defended by access control. Structurally identical to the companies they’re dissolving.

Their product is the dissolution. They are selling the ability to throw out half your SaaS stack and just build it yourself, or tell an agent to. Their pricing is rent extraction on access to a machine that destroys rent extraction. They sell the tool whose central promise is that tool vendors are less necessary than they were, and the valuations require both of those facts to hold at once, indefinitely.

That’s the Hungry Mungry move performed literally. Eating outward, for now. But the direction of an appetite is not a fixed property.

Short term, the model providers. They sit upstream of the disintermediation and collect the value as it falls. Every cancelled seat replaced by four hundred lines of generated Python converts into API spend. That’s a toll booth on a collapse, and toll booths are not moats.

Medium term, open weights eat the margin. The frontier is expensive and good-enough is not, and the long tail’s needs were never frontier needs. A model that can write an approvals workflow is a solved problem running on hardware the company already owns. API revenue on commodity work has a half-life.

Long term, models write models. Once tool-making is itself commoditized, the model layer is a layer like any other, with nothing structurally distinct about it. The apex is not exempt. That’s what apex means in this particular poem.

Timelines are unknown, and anyone handing you one is definitely selling something. The arrows point one direction at every stage, and nobody has yet clearly described the mechanism by which they stop.

The precedent for all of this is 2000s-era media. Distribution used to be capital. Presses, trucks, spectrum licenses, a loading dock. Then it went to roughly zero, and the industry didn’t die. The Times is fine. The Atlantic is fine. What died was the rent layer: the classifieds monopoly, the regional paper that owned a market because it owned the only press within eighty miles, the value that came from owning the pipe rather than the thing moving through it. Most of the captured value evaporated. It didn’t transfer to someone else. It stopped existing.

SaaS multiples price the assumption that software stays hard. Not explicitly. It sits upstream of the model as an axiom, which is why nobody defends it, and which is why a company with modest revenue and negative income carries a number that only resolves after fifteen uninterrupted years of per-seat rent. That resolution requires that for fifteen years, nobody builds it themselves.

The assumption is the inflation pressure under the whole bubble. Every multiple is downstream of it. If it fails only for the long tail (not Slack, not pharma Salesforce, just the middle of the distribution) the multiples compress anyway, because the multiples were priced on the middle. Not all the incumbents die but the rent layer flattens out. The mechanism is a business model consuming its own supply.

Andreessen wrote at the start of the meal, Silverstein wrote the ending.

‘Cause nothin’ was nothin’ was

Nothin’ was nothin’ was

Nothin’ was left to eat.

— Shel Silverstein, Hungry Mungry