If your company builds anything with AI, you keep making one expensive call: which parts do you build and run yourself, and which do you just rent from someone else? This is a two-question test for that call, plus a warning. The pieces teams rush to build keep turning into cheap, rentable products a few months later, so most of that work gets undone for them.
You don’t need to know anything about AI. We’ll learn the whole thing through a samosa stall, then map it back. Meet Priya.
Priya’s year
Priya runs a samosa stall outside a busy train station. One plate of four samosas sells for thirty rupees, and on a good evening she clears two hundred plates. The heart of the stall is the cooking: the spiced potato filling and the hot oil. She has two ways to get that filling. She can cook it from scratch, peeling and boiling and spicing for hours. Or she can buy a ready-made filling base from a wholesaler in the market for about eight rupees a plate, scoop it, and fry.
Here is the first idea, and it’s the one everything hangs on. A commodity is something that stopped being special and became a standard, buy-it-anywhere product, like bottled water or a phone charger, so nobody pays extra for it. The wholesaler’s filling base is a commodity. So is the shiny new gadget every stall is talking about: an automatic fryer that spices, times, and fries on its own. Last year that fryer cost sixty thousand rupees. This year the same machine costs twenty thousand. Next year, dealers say, about seven thousand.
That fall has a name worth knowing, because it’s the engine of this whole story. When the price of a thing keeps dropping fast year after year while the quality holds, call it deflation. The fryer is deflating at roughly three times cheaper a year.
So Priya makes the smart call. She rents the cooking, in a sense: she buys the cheap ready-made base instead of cooking from scratch, and she pours her energy into the parts she figures will be hers. Over one long year, she builds three of them.
First, a supplier-switcher. Three wholesalers in the market sell the filling base, and prices wobble daily. So she rigs up a little system and a phone of running prices to always buy from the cheapest, and to jump to another the moment one runs out. Weeks of fiddling.
Second, her own taste-tester: a small bench where a helper checks every tenth batch, scores the crispness and the spice, and rings a bell when quality slips. More weeks.
Third, her own sauce factory. She tastes the wholesaler’s base, reverse-engineers it, and starts making a near-identical filling herself in a back room, most of the flavour at a fraction of the cost. Months.
Then one Tuesday at the market, she opens the big wholesaler’s new catalogue. There they are, all three. A ready-made price-comparison service. A “quality-checked batch” stamp you can just pay for. And tubs of exactly the filling she spent months learning to make, sold cheaper than she can make it. A year of her work, reduced to three line items in someone else’s price list.
Now the reveal. Swap “samosa stall” for “company building with AI,” and this is the exact choice every team faces today, down to the bad Tuesday.
Three quick translations, then we go back to Priya:
- The model is the AI itself, the thing behind ChatGPT that turns your words into an answer. In stall terms, it’s the cooking: the part that does the real work.
- Rent vs. build. You can rent the model the way Priya buys ready-made base: pay a provider like OpenAI, Amazon, Microsoft, or Google to run a ready-made AI, billed per use. Or you can build it: run your own on your own machines, like cooking from scratch.
- The shiny fryer that keeps getting cheaper is the AI itself becoming a commodity. Renting the cooking, owning the cooking: either way, the cooking is no longer the special part.
Why it stings: Priya did the smart thing
Here’s the cruel part. Priya played it well. For two years the standard advice was exactly hers: rent the cooking, own your edge. And the cooking really was getting cheaper. By one venture firm’s count (a16z), the price of running AI at a fixed quality level fell about ten times a year from 2021 to 2024. When a thing drops tenfold a year, only a fool buys the factory. You rent, and you build above it.
The advice had a missing line, though. It said own your edge, but it never asked whether the things Priya built were an edge. They weren’t. They became commodities too, just a few months behind the cooking.
There’s a word for a real edge, and it matters for the test coming up. A moat is the thing a rival can’t simply go buy: the water around your castle. For Priya, the moat is her secret family recipe, her regulars who walk past two other stalls to reach her, and her spot right by the station exit. A competitor can buy the same fryer and the same base tomorrow. He cannot buy her regulars. The supplier-switcher, the taste-tester, the sauce factory: none of those was a moat. Each was just effort the market was about to sell to everyone.
And building a falling thing is worse than merely wasteful. Priya spent real money and a year she can’t get back, at the exact moment the price was about to drop without her. Economists have studied this (Dixit and Pindyck, 1994; brought into business tech by Benaroch and Kauffman in 1999): when your spending is sunk, the future is uncertain, and you could simply wait, committing early throws away something real, the value of waiting. AI adds a twist, because here the cost isn’t just uncertain, it’s predictably falling. So Priya paid three times. She spent the money. She gave up the chance to wait. And she handed back the discount the falling price was about to give her for free.
“So nobody should build anything? Just wait forever?” asks Skeptika, arms crossed, the voice of your inner doubt.
No, and that’s the trap to dodge. Using the cheap stuff is not the same as owning it. Priya should absolutely sell samosas today, using the cheap base greedily. She just shouldn’t pour her year into a back-room factory whose product is about to show up in the catalogue. Ship today. Rent the cheap part. Don’t pour concrete on a floor that’s sinking.
How fast is AI really getting cheaper?
Skeptika wants the number, so let’s be honest about it. You’ll see scary headlines, ten times cheaper a year, even fifty. Most of that is a one-time clearance: when a new top model laps an old quality level, that level’s price drops once, hard, then settles. The rate Priya can actually plan around is gentler and steady, roughly three to five times cheaper a year for the same quality. It would only stop mattering if it slowed below about 1.3 times a year, and no forecast sees that before 2028. Either way, the floor under her keeps dropping, so building on it stays a bad bet.
Two cautions Skeptika insists on. The chips that run AI rent at swingy prices, not a smooth slide: the research firm SemiAnalysis clocked rates bouncing back about forty percent off their lows when demand spiked. And the very best barely gets cheaper at all. OpenAI’s o1, one of its top models, launched at about sixty dollars per million words of output, roughly what the best model cost back in 2021. You only collect the discount by accepting last year’s quality, the same way Priya only gets the cheap base by serving yesterday’s recipe.
Cheaper per word, pricier per answer
Here’s the twist that breaks every neat “build pays off after X plates” calculation. Yes, the base gets cheaper per scoop every year. But customers now want the loaded, elaborate samosa platter, and the fancier the dish, the more scoops it burns. So even as each scoop gets cheaper, the cost of one finished top-tier plate keeps rising.
AI does exactly this. You pay the AI by how much it reads and writes (roughly, per word). The price per word keeps falling, that steady three-to-five times a year. But the newest “reasoning” models answer by thinking out loud first, spilling huge streams of words before they reply, and the smarter they are, the more they spill. So the finished cost of one top-tier task is rising, about three times a year for light thinking and up toward eighteen times for the heavy kind. Two lines crossing, like the blades of a scissors.
This is why “pick a monthly volume, find the crossover, declare victory” misleads you. Cheaper food doesn’t shrink the town’s total food bill, it just means everyone eats more samosas, so the total spend climbs. Microsoft’s CEO Satya Nadella made the same point about AI in January 2025: cheaper AI doesn’t cut the bill, it explodes demand. Google now handles about 3.2 quadrillion words a month, up roughly seven times in a year. For a steady seller like Priya, renting wins and keeps winning. For someone running heavy made-to-order feasts at huge volume, the sums can tip back the other way. There’s no single answer, which is exactly why you need a test, not a calculator.
The two-question test
When a price falls, the profit doesn’t vanish. It moves next door, to the part that’s still hard. Cheap cars sell more petrol, as the programmer Joel Spolsky put it in 2002. An MIT research project studying AI at work found the same: how well the AI is wired into the real work, not how clever the model is, decides whether a project pays off, and the thin add-ons get swallowed by the big providers on their next release.
That gives Priya a sharper test than “own what makes you special,” which waves almost anything through. For each thing you could build, ask two questions, in order, and it has to pass both:
- Does it get MORE valuable as the gadgets get cheaper? Priya’s recipe and her regulars get more precious as fryers flood the market and every corner has one, because the cooking is no longer what sets her apart. Things like the fryer just get cheaper whether she owns one or not. If it compounds, it’s a candidate to own. If it deflates, rent it.
- Can you just buy it ready-made off the shelf? If the wholesaler can sell it in a tub, what you built turns into wasted effort the day it hits the catalogue.
Own it only if it compounds AND no one can sell it ready-made. Everything else, rent.
Run Priya’s three builds through the gates, and all three fail, each in a worse way than the last. Start with the easy one. The supplier-switcher fails twice over: the wholesaler now sells one, and honestly almost nobody switches anyway. A December 2025 survey from Menlo Ventures found teams overwhelmingly stay with the provider they wired in first, because the real cost of moving isn’t the price, it’s re-learning everything around it. The taste-tester is the next surprise: it’s now a stamp you can buy. And the sauce factory is the one that looks like a real asset and isn’t. It’s a treadmill: she copied the wholesaler’s base, so the day he improves it, hers is suddenly worse than the cheap tub, and she has to copy it all over again. That’s not an asset that grows. It’s a wheel you run on to stand still.
“If nobody switches, isn’t being stuck my problem? Shouldn’t I own the way out?” Skeptika asks.
Yes, but the way out is your own stuff, not a switcher. Your recipe, your regulars, your records, and a backup base you know how to cook yourself: that’s the escape hatch. The price-comparison gadget isn’t.
The few things worth owning
After both questions wash through, only a handful survive, and notice they’re all about you and your stall, never about the machinery.
- Your regulars. The people who choose your stall over the two before it. In AI terms, that’s your own data: the record of your customers and your business that a provider can’t sell to your rival, because it’s yours.
- Your recipe and your taste standard. Not just “having lots of ingredients,” that’s what a16z calls the “data moat” myth, because generic data stops helping fast. What compounds is your own recipe plus your sense of what a perfect plate tastes like, and the habit of fixing the recipe every time a batch disappoints. In AI terms, that’s your quality bar: the expert-checked set of right answers for your problem. It grows more valuable as models get cheaper and look alike, because it’s the only thing that tells you which cheap model is actually good enough.
- Freedom to switch, plus a base you can cook yourself. Your recipe, records, and supplier list in a form you control, plus a tested backup you can fall back on. In AI terms: keep your data and setup portable, and keep a free, openly available model you can run yourself ready in the wings. The real prize is never needing to switch in a panic.
- Owning your kitchen and your recipe book. The whole thing in your hands, not locked in one supplier’s back room where a price hike or a shutdown is his decision, not yours. In AI terms: control of your own data and code.
Everything else, the cooking, the fryer, the supplier-switcher, the bought taste-stamp, the sauce factory, Priya rents or buys ready-made. Renting isn’t losing. It’s refusing to pour concrete on a sinking floor.
Where Gracient fits
One last thing Priya noticed. Her real headaches were never any single machine. They were the joins between them: the moment the bought base met her own spice mix, the moment “it passed the taste-test” had to become “a regular would actually pay for this,” the moment one supplier’s tub had to fit her own recipe. Those joins are where things went wrong at the evening rush, and they’re the part no wholesaler sells in a tub, because they’re specific to her.
That’s the work we do at Gracient. We set up your whole AI kitchen inside your own place, and we handle the tricky handoffs where the bought stuff meets your own recipe and your own data.
Want the two-question test run on your real setup? Book a 45-minute working session. We’ll run your list through both questions live, mark what’s gone rentable, and hand you the short list of what’s genuinely yours, whether or not we end up working together. No pitch. Start the conversation here.
Skeptika gets the last word: “So the point isn’t ‘own less because you can’t win.’ It’s ‘own less because you finally know which few things are worth keeping, and you stopped paying to make the ones the wholesaler sells you cheaper.’” That’s the one.