AI Ends Software Rationing: Costs Shift to Tokens

We live in the greatest of times to build software. A $200 subscription gives you access to state-of-the-art coding models that can build almost anything. So why isn't our software 10x better? And why does most of our software still look like it was designed in 2015 but now with a chat box as its homepage?

The short answer is rationing. It's over, and we haven't caught up.

Before Claude Code (BC)1, the physics of building software looked quite different. Software was expensive to build and cheap to run. Writing code was bottlenecked on hiring and retaining expensive engineers, who needed expensive managers to keep things on schedule and expensive product managers to help figure out what to build. The cost of compute, memory, and storage fell every year. The median public SaaS company kept about 75 cents of every revenue dollar as gross profit.2

Engineering bandwidth was the scarcest resource in all tech companies, and the products we got were the result of rationing it. You can see it in the doctrines the industry's elders handed down: the Minimum Viable Product (MVP), the lean startup, "do one thing". The tradeoffs the winners made on their way up shaped each category, because the rest of the industry copied them.

Take GitHub's pull request. It started in 2008 as little more than a notification when you made changes on a fork of someone's repository and asked them to "pull" those changes. Later, the feature evolved to what we see today, with a diff, a discussion, and a merge button. The feature was built originally for open source collaboration, yet it was copied by every other git forge even if they were not building for open source, to the extent that even the ones who have vowed to build a better GitHub for agentic coding today can't seem to shake it off. Salesforce's lead/opportunity/pipeline object model from 1999 became the definition of what a CRM is. Every CRM company started since inherits it.

After Claude Code (AC), the economics that gave us every single piece of software we use today have inverted. Building software has become cheap, with the marginal cost of building a new feature approaching zero, while running it, now that AI sits in the critical path, has become expensive. Any product using LLMs now carries real COGS and real marginal cost per user. Software has flipped from high fixed cost (R&D teams) and ~zero marginal cost (cheap compute) to low fixed cost (leaner R&D teams, more agents) and substantial marginal cost (frontier models).

Software now looks more like manufacturing than SaaS.3

Our products were shaped by rationing, and the rationing is over. If you're building today, you're no longer bound by yesterday's constraints. A new product or a new feature is cheap and easy to build. Experiments are easy too, because code, however complex, is now disposable. So the software we get from here ought to look entirely different, and be orders of magnitude better. It doesn't, at least not yet.

Take a stroll down Market Street in San Francisco and you'll see billboards and Muni buses plastered with ads for companies building AI agents. If you squint, almost all of them are OpenClaw clones masquerading as verticalized agents. Everyone seems to be building the same thing, as if they're unaware that they can build almost anything now. We're getting more software, not different software.4

Our industry hasn't caught up with the end of rationing, for a few reasons:

• The BC wisdom of how to build successful software products is deep-seated and difficult to shake off, and anyone who talks of "tokenmaxxing" (spending tokens freely instead of rationing them)5 is greeted with ridicule.
• Selling new products that don't fit a category is hard, because most procurement teams buy categories. A CRM that doesn't look like one, even if it's 10x better, is a hard sell unless there's a champion who's willing to take the risk.
• Capital is still flowing to proven "X for Y" startups. OpenClaw was wildly successful, so "OpenClaw for X" is easy to fund.
• Lastly, large incumbents ship their org charts, which are built around headcount. Fundamental change in how software is built is next to impossible for them.6

Some constraints from the BC era still hold. Distribution is still the incumbents' biggest moat, and the hardest for newcomers to crack.7 Trust takes years to earn, especially with consumers. Code is cheap to write but not to maintain: a product that's scaling needs a different organization than one being shipped from zero to one, and most people who talk about tokenmaxxing only talk about the shipping. And every user now costs tokens. But none of these explain why the software we do build still looks like it's from 2015.

To better understand this lag, consider how it took about four decades for US manufacturing to see productivity gains after electricity was commercialized in the early 1880s. Factories first replaced steam engines with one big electric motor driving the same shafts and belts they had, so the productivity gains were marginal. The gains came only when factories redesigned the floor around a small motor on each machine. The software industry has yet to redesign its factory floor, and the companies that do this before others have a good chance of hitting escape velocity.8

Catching up will take two forms.

GitHub won't get replaced by a "new and improved" GitHub that's built with AI slapped on, but by a company that rethinks code collaboration for humans and agents, in a world where humans read less of the code every day.9 Nor will Salesforce be replaced by an "AI-first" CRM. It will be replaced by a company that's solving CRM from first principles, without the baggage of the existing object model.10 Many companies are being funded to build the AI-native version of an existing category, and very few are building what the category looks like without the rationing.

The second is less obvious: packaging. The "do one thing" doctrine is no longer relevant in a post-rationing world, and companies can now "boil the ocean", bundling more products than before, aggregating value and owning the customer relationship.11

Companies like Rippling and PostHog were the trailblazers, bundling several related products long before agents. Parker Conrad was early.12 We ought to see more companies shaped like this. We'll also see pairings that look strange until you see why, like Stripe acquiring OpenRouter and Ramp launching its LLM router. Both are finance companies following software's new cost line: tokens.

The rationing ended before our old habits did, and we're in this weird gap before everyone catches up and shakes off the old doctrines of building products. The advantage belongs to the people who never learned to ration.