Micron Just Locked In ~$100B of Take-or-Pay Memory Deals and Won't Say When the RAM Crisis Ends. Here's What That Does to Your Build Costs.

On June 24, Micron reported its fiscal Q3 2026 results and disclosed 16 strategic customer agreements representing roughly $100 billion in minimum contracted revenue. The deals lock in supply across three to five years with data center operators, hyperscalers, and automakers. About $22 billion of that is firm financial commitment from customers today, and CEO Sanjay Mehrotra said he expects roughly half or more of company revenue to fall under these binding structures. Asked when the memory shortage eases, he said he didn't know.

Strip the earnings-call language off that and here's what it means for anyone who builds: the company that makes a big share of the world's memory just signed multi-year deals designed to keep prices high, told its largest customers it'll fill their orders first, and declined to predict relief. If your stack assumes RAM gets cheaper over time, that assumption just got contradicted by the people who set the price.

Take-or-pay with a floor is the part that matters

The structure is the story. These aren't loose "we'll probably buy from you" arrangements. They're take-or-pay deals with pricing floors, which means two things at once. The customer is on the hook to buy a minimum volume whether they need it or not, and the price can't fall below an agreed level no matter what the spot market does. Micron's own framing is that the floors are set to guarantee gross margins above any level the company has ever achieved.

That second half is what you should care about. A pricing floor is, by design, a refusal to compete on price. When a memory maker locks in record margins on half its revenue for three to five years, it has removed its own incentive to flood the market and chase volume, the move that historically crashes memory prices and ends shortages. The classic DRAM cycle is boom, oversupply, glut, price collapse, repeat. Take-or-pay floors are an attempt to break that cycle in the supplier's favor. If it holds, the "the glut is coming, just wait it out" reflex that's saved hardware buyers for thirty years stops working.

And the hyperscaler deals come with a priority claim on supply. When the people who buy memory by the trainload have contracts that put them at the front of the line, the leftovers for everyone else (meaning the channel where you buy a stick of RAM or a cloud instance) get tighter and pricier.

Why a solo builder should read a memory earnings call

You don't buy memory by the wafer, so why does this land on you? Because nearly every cost decision a solo operator makes about AI infrastructure has a hidden bet on falling hardware prices, and this is the news that the bet is off.

Take the most common one: self-host versus API. The pitch for self-hosting an open-weight model is that you pay once for hardware and then run inference cheaply forever, eventually beating the per-token API bill. That math depends heavily on the cost of the box, and the cost of the box is mostly memory. Modern inference is memory-bound; you need a lot of fast RAM or VRAM to hold a useful model and serve it at a tolerable speed. If memory prices have a floor under them for the next few years, the upfront cost of the self-host box stays high, the break-even point against the API moves further out, and for a lot of solo workloads it moves past the point where self-hosting ever pays off.

The same floor shows up in places you don't immediately connect to a Micron press release. Cloud instances priced on memory don't get the steady annual discount you've quietly come to expect. The dev-machine upgrade (more RAM so your local builds and local models stop thrashing) costs more than it would have in a normal cycle, and "wait for prices to drop" may not be a winning strategy this time. Even consumer SSDs and laptop RAM ride the same supply, so the personal-hardware refresh gets more expensive too.

What I'd actually do

Reprice the self-host spreadsheet with a memory floor in it. If you've been telling yourself "I'll buy the GPU box next quarter when prices come down," stop waiting for a drop that the supplier just told you it's engineering against. Either the workload justifies the box at today's prices or it doesn't, and for most solo operators serving bursty, unpredictable inference, it doesn't. Rent the compute. Let the hyperscaler eat the memory-supply risk you can't hedge, and pay per token until your volume is so steady and so large that owning the hardware clearly wins. That threshold is higher now than it was a month ago.

For your own machine, the move is the opposite of panic-buying. If you don't need the RAM upgrade this quarter, don't make it on the theory that prices will be even worse later. That's how shortages get talked into being worse than they are. Buy memory when a real workload requires it, size it once for two years instead of upgrading in steps, and otherwise leave it alone.

The honest counter-take: I'm reading a supplier's incentive into a future that hasn't happened. Take-or-pay floors are what every commodity producer wants in a tight market, and the history of commodities is that floors hold right up until they don't: a demand wobble, an AI-capex pullback, or a competitor breaking ranks on price can crack the whole structure, and then the glut everyone keeps predicting finally arrives. Micron also has every reason to talk its own book on an earnings call; "record margins forever" is exactly what a CEO says to a stock. But the contracts are real, the priority claims are real, and the CEO declining to call an end to the shortage is a tell. Plan for memory to stay expensive through 2027 and be pleasantly surprised if it doesn't, rather than the other way around.

Author

Lukas

@lukcombinator

Sources

• Micron inks long-term supply agreements worth $100 billion — says it has no idea when RAM crisis will end — Tom's Hardware
• Micron (MU) earnings report Q3 2026 — CNBC
• Micron Q3 2026 slides: record margins, $100B customer agreements — Investing.com
• Micron Technology, Inc. Fiscal Q3 2026 Earnings Call Prepared Remarks — Micron Investor Relations