Build Got Cheap. Judgment Didn't.

Build less, choose better

The right AI product strategy for this moment is to build less and choose better. When the cost of shipping software falls toward zero, making things stops being the constraint, and whatever was hiding behind that constraint takes over: knowing what deserves to exist, and where you can defend it once it does. The founders, ex-founders, and product leaders who win the next cycle will be the ones reading where money is moving, and building only to test that read.

The dominant instinct points the other way. Cheap build feels like a green light to build everything, immediately. Treat it instead as a signal to slow down at the one step that still pays: deciding what is worth building at all.

What actually got cheap

The collapse is real and measurable. In Y Combinator's Winter 2025 batch, about a quarter of startups had codebases that were 95% or more AI-generated, companies YC's own leadership said would have written that code by hand a year earlier. A custom MVP that historically ran $50,000 to $300,000 over months can now be produced by a small team in three to six weeks at a fraction of the cost. And the unit economics keep falling: Stanford HAI's AI Index found the inference cost for GPT-3.5-level performance dropped roughly 280x in two years.

Andrej Karpathy named the phase "vibe coding," then within a year called it already dated and pointed at what comes next: humans directing agents that write the code for them. Whatever you call it, the direction is one way. The act of construction is being commoditised in front of us.

When everyone can build, value moves upstream

Here is the logic the reflex misses. If you can build the thing cheaply, so can everyone else. When construction is nearly free and available to all, the thing you construct carries almost no advantage on its own. Value relocates. It moves to the decisions upstream of the build: which problem, which buyer, which position.

You can already see the bottleneck shifting in hiring. One industry survey found postings for pure implementation roles, the ones focused on translating a spec into code, fell 17% between early 2025 and early 2026, while postings requiring experience with AI tooling multiplied. Treat the exact figures as one survey's read, but the direction matches what the market is doing everywhere else: assembly is getting cheaper. Deciding what to assemble is holding its price.

Treat building as a probe

This is where the advice to "pause and observe" gets misread. In a world where a prototype costs a weekend, building is the cheapest way to observe there is: a probe you put in front of a real buyer to see if your read was right. Keep building constantly.

The pause applies to one thing only: the reflexive commodity build, the one you start because you can rather than because you have a reason. Demote building from achievement to instrument. Use it to test a judgment about what deserves to exist. Ship to learn.

Watch the reallocation

"Observe the macros" is useless as a platitude and sharp as an instrument. The move is to see where money is being reallocated, because reallocation is demand arriving before the market has a name for it. Forecasting the whole economy is someone else's job.

The signal is loud right now. Enterprise AI spending is on track to rise from around $340 billion in 2025 to roughly $3 trillion by 2035, moving from under 4% to nearly a quarter of all enterprise tech spend. In a16z's survey of enterprise buyers, one CIO put the pace bluntly: "what I spent in 2023 I now spend in a week." In that same survey, the share of AI budget coming from experimental "innovation" pots collapsed from 25% to 7% as the spend graduated into core, recurring line items. That is money changing its mind about what is essential.

Three concrete signals turn that into something you can act on this week:

• Budget migration. Whose line items are quietly moving from headcount into software, and into which category? That is demand showing up before anyone writes the market-map slide.
• Incumbent exposure. Which established players just had a core workflow commoditised by a model, and have not reacted? The gap between what is now possible and what they still charge for is the opening.
• Un-bundling pressure. Which all-in-one suites are suddenly vulnerable to a sharper single-workflow tool, precisely because building that tool got cheap? Bundles hold together on switching cost; cheap build lowers switching cost.

Own something after you enter

Now the uncomfortable part, and the reason "identify the opportunity" is only half a strategy. If you can spot the opening cheaply, so can everyone with the same dashboard. Identification gets you to the starting line at the same time as every other sharp operator reading the same reallocation. The race is decided by what you own after you enter.

Three things hold when code is free. Distribution you built before the crowd arrived. Proprietary data that compounds every time the product is used. And the workflow itself: becoming the place the work actually happens, so that even a better-funded clone has to pull users out of a system they already live in. The build-it-first moat dissolved the day building first stopped being hard.

For a product leader inside an org, the same test applies to every AI feature on the roadmap. The question worth asking is "what do we own once we have built it that a competitor cannot copy by Friday," because "can we build it" now answers itself.

Budget for the run, not just the build

One honest caveat, because the falling-cost story has a second half that is easy to miss. Cheap to build and cheap to run are different claims. Token bills scale with usage in ways seat-based software never did. A major Microsoft division reportedly told its engineers to drop Claude Code by mid-2026 after power-user costs ran into the thousands per engineer per month, and the same reporting says Uber exhausted its 2026 AI budget in four months. The cost of construction fell off a cliff. The cost of operating at scale kept its altitude.

There is a live debate about how much of the current spend is real versus froth. An MIT analysis cited by Forbes suggests AI is economically viable in only about 23% of roles today, and investors including Ray Dalio have called the moment bubble-like. Take that as a reason for discipline. It sharpens the same point: when everyone can build and the bills are real, the winners are the ones who picked a problem worth the operating cost.

Aim before you fire

The pause this moment rewards is not inaction. It is aiming. Build cost fell so far that execution can no longer carry your differentiation, which means your judgment about what to build and where to stand has to carry the weight execution used to. Spend your cheapest resource, building, freely, as probes. Spend your scarcest one, attention on the right problem and a defensible position, deliberately.

Concretely: before the next build, name the reallocation you are betting on, and name the thing you will own six months after launch that a fast follower cannot copy. If you cannot answer the second question, you have found a feature, not a company. Cheap build made the first question easy. It made the second one the whole job.