NVIDIA’s customers just got the memory bill

Bloomberg says builders of NVIDIA systems are preparing price increases of more than 15% in many cases as memory costs rise. That is the Micron thesis in one sentence.

Disclosure: I currently hold a long position in MU and in NVDA.

This piece first went out on Sector Playbook. If you want the next one in your inbox, subscribe there.

This week they belong in the same paragraph, because the bill inside the rack moved.

Bloomberg reported on Saturday that some of NVIDIA’s biggest customers have been told the prices of servers containing its AI chips are going up more than 15% in many cases, with memory costs soaring.

The increases would land on systems shipping early next year, including Vera Rubin and Grace Blackwell. The size depends on chip generation and memory configuration.

That is not NVIDIA publishing a new price list. Contract manufacturers who build the servers for Microsoft, Google, and Oracle notified the customers.

So the sentence I am willing to write is not “NVIDIA cannot eat the memory bill.” It is this: the supply chain around NVIDIA can no longer hide the memory bill.

What the report actually is

Bloomberg is citing unnamed people familiar with communications that have not been made public. The notices came from the companies that assemble the boxes, not from a keynote.

If you want a clean corporate communication, you do not have one. You have a system invoice that is about to get more expensive, according to people who say they have seen the notices, and a GPU vendor that has not commented. I still think the report is worth writing about.

The direction of the leak matches the memory name I own. Samsung, SK Hynix, and Micron are the three names Bloomberg points at. The two Korean firms and Micron account for most of the world’s DRAM. That is the oligopoly collecting the shortage.

NVIDIA reports fiscal second-quarter results on Wednesday, so Colette Kress, CFO of NVIDIA, will need to comment on the increasing memory costs. I wrote earlier about Kress, CUDA, and why older NVIDIA chips still look economically alive.

Shortage economics, not a cycle print

Micron’s last reported quarter is the other half of the same sentence. Fiscal Q3, ended May 28, 2026, printed $41.46 billion of revenue and an 84.9% non-GAAP gross margin. The company guided fiscal Q4 to $50.0 billion, plus or minus $1.0 billion, at roughly 86% gross margin. The growth quarter over quarter is huge. Given the supply constrained market, Micron can dictate the price at which they want to meet demand. Cloud Memory, the slice closest to the AI rack, did $13.8 billion at an 83% gross margin in the prior quarter.

As we all know, those are not “memory is cycling up” numbers. Bits barely moved, since DRAM bit shipments rose in the low-single-digit percent range sequentially. What is driving the revenue growth is the selling prices dictated by the memory giants. DRAM prices rose in the low-60s percent range.

You get a margin like that when the constraint is wafers and the customer cannot wait, not when a commodity is finding a new average.

Memory has enough downcycles in the archive that I distrust any sentence that starts with “this time is different.” The numbers can still be extraordinary and still be rent.

The question is how long the rent lasts, and whether the company has done anything to keep some of it after the shortage eases.

We can look at the contracts. They are better evidence than another demand adjective. Micron said it has signed 16 Strategic Customer Agreements. They are take-or-pay.

Customers have binding commitments to buy specific volumes over a multi-year term, typically calendar 2026 through 2030. The largest ones generally carry a ceiling for existing products at the calendar Q2 market price, and a floor through the term.

Fourteen of the sixteen have about $100 billion of cumulative revenue at the minimum contract prices. Micron also expects $22 billion of cash deposits and related financial commitments under the agreements signed so far. That is a customer prepaying for allocation. It is also a floor under a business that used to reprice every quarter.

If the Bloomberg notices are real, they are what the other side of those contracts looks like when the cost shows up in a finished NVIDIA server. The Crucial decision is the same story in a consumer aisle.

In December 2025, Micron exited the Crucial consumer business so it could “improve supply and support for our larger, strategic customers in faster-growing segments.” You do not kill a 29-year retail brand because you are mildly optimistic about data-center mix. You kill it because the wafers have a better home.

At Friday’s close of $366.78, Micron trades at 21.8 times trailing earnings. The forward multiple is 6.7 times. Enterprise value is 15.7 times EBITDA, which is what the market pays when it treats those margins as rent.

What would break this

The honest bear is not “custom silicon eats NVIDIA, therefore memory dies.” Micron said the relevant part in the same prepared remarks. AI systems are powered by GPU, ASIC, and CPU designs from a broadening set of suppliers. They share one characteristic: performance is architecturally dependent on the memory subsystem’s performance and capacity. ASICs do not eliminate the memory bill.

They can change HBM intensity, rack design, and how much DRAM sits next to the accelerator, but it is not a zero.

The primary bear is supply catching demand. Micron itself says tightness lasts beyond calendar 2027, that industry supply should improve gradually in 2028, and that it still has no line of sight to catch-up. Idaho’s ID1 fab is on track for first wafers in mid-calendar 2027. Greenfield memory takes years, tradespeople, permits, and power.

All of that can be true and 86% gross margins can still be shortage rent. The rent ends when enough cleanroom exists, or when the demand story that filled the cleanroom slows first.

The second bear is AI capex actually slowing. Bloomberg already listed the frictions: project delays, labor shortages, tightening capital markets, community resistance. A more expensive rack is not a problem if the rack still gets ordered. It is a problem if the extra cost is the excuse a CFO uses to slip a cluster from 2027 into 2028.

Memory pricing power is downstream of someone still building. If that someone pauses, MU does not get a special exemption because the last invoice was fat.

The third bear is architectural memory efficiency, and this is where ASICs belong. Micron already told you customers are trimming average server DRAM content to ship more units under a tight allocation.

That is a shortage workaround, not a demand collapse. It is also a preview. If custom silicon, tighter memory hierarchies, or “good enough” HBM stacks reduce bits per accelerator, the same AI dollar buys less Micron.

I would rather put that risk in the file now than pretend the only way this breaks is a classic overbuild.

Margins are shortage rent until proven otherwise. The SCAs are the attempt to prove otherwise. They have not been through a downcycle yet.

Wednesday

I am publishing this before NVIDIA’s August 26 call on purpose. The test is not whether NVIDIA is still selling a lot of accelerators. We know that.

The test is whether Kress, or anyone on the NVIDIA Q&A, talks about memory costs, system pricing, or supply constraints in a way that makes the Bloomberg notices look real, overstated, or irrelevant.

Silence would be its own data point. It would mean the memory cost in the rack is still being handled in private, which is how an increase of more than 15% can show up at the contract manufacturer first.

If the notices are real and the systems still ship, the memory bill is in the customer’s budget. That is the version of this story I own.

If NVIDIA or the builders eat the cost after all, or if hyperscalers slip deployments rather than pay, the shortage is still real and the equity path gets worse.

Anyway. I own Micron because even the NVIDIA system supply chain cannot keep soaring memory costs out of the customer’s bill.

If you build or buy these racks, I want to know whether the 15% is showing up in your quotes yet, or whether it is still a story that has not hit procurement. Comments live on the Substack post.

This newsletter is for educational and entertainment purposes only. It is not financial advice. I am not a registered investment adviser. The content reflects my own personal opinions and analysis. I may hold positions in securities discussed. Always do your own due diligence or consult a licensed professional before making investment decisions.

Subscribe to Sector Playbook on Substack if you want the Wednesday follow-up after the NVIDIA call.