The Memory Tax: Who Can Pass It On

DEEP DIVE

ISSUE 26 COMPANION · FRIDAY 28 AUGUST 2026 · R. LAURITSEN

The Memory Tax: Who Can Pass It On

A five-question test for sorting AI holdings by repricing power — and the ten names it scores.

Nvidia’s second quarter contained two numbers that pointed in opposite directions, and the market only cared about one of them.

The first: gross margin heading from roughly 75% down to 71–72% by the January quarter, with memory named as the cause. The second, buried in the CFO commentary: supply commitments up from $119bn to $279bn in a single quarter, primarily for memory procurement. One number says costs are rising. The other says Nvidia has pre-bought years of the thing that’s rising and can charge for it. Days earlier, Bloomberg reported the company had told its largest customers that server prices would climb more than 15% on systems shipping in early 2027.

That’s the whole story. The cost went up. Nvidia sent a letter. The stock rose 8.74%.

The interesting question isn’t why Nvidia got paid. It’s who else can send that letter — and how you work it out before their next earnings release rather than after.

This is a test for that. Five questions, ten names, and an honest account of where it stops working.

HOW BIG IS THE TAX?

Large enough that it has changed the shape of a data centre’s cost structure inside eighteen months.

Memory is roughly 25% of the bill of materials for a high-end AI server rack, on Deloitte’s estimates. A single Vera Rubin GPU carries 288GB of HBM4. TrendForce projects DRAM and NAND together will account for 68% of cloud providers’ hardware spending by 2027 — up from 47% this year. That is not a cost line drifting upwards. That is the centre of gravity of the entire capex budget moving from the processor to the memory beside it.

The cause is capacity allocation, not scarcity of sand. Samsung, SK hynix and Micron control over 95% of DRAM output, and they have redirected that capacity toward HBM, which carries several times the margin of conventional DRAM and consumes roughly three times the wafer area per gigabyte. Conventional DRAM has been squeezed as a direct consequence — hard enough that the three now face a US antitrust class action alleging the transition functioned as a coordinated production cut.

Gartner expects the crunch to persist at least into the first half of 2027. Nvidia’s own guidance assumes supply remains the binding constraint through fiscal 2028.

So this is not a quarter-long input shock. It is a multi-year transfer of value along the supply chain, and every company touching AI infrastructure now sits somewhere on it.

THE FIVE QUESTIONS

Each is answered from public disclosure. None requires a model.

1. Who sets your price, and when can it next change?

The single most predictive question. List pricing that can be revised by notice is a different business from contract pricing agreed at signature. Nvidia raised prices with a letter. Marvell’s custom silicon prices were set when the designs were won — potentially years before the memory bill arrived.

Look for: the pricing language in the MD&A, the average contract duration, and whether cost pass-through clauses are mentioned at all.

2. Is the constraint on your side of the table, or your customer’s?

If you can sell everything you can build, price is yours to set. If you can build more than anyone wants, it isn’t. Nvidia’s CFO said demand implies growth well above the 70% it guided, with supply as the limit. That’s a company describing its own constraint. A company describing bookings is describing its customer’s.

Look for: management language about capacity versus language about pipeline. They are not the same disclosure and shouldn’t be read as one.

3. Is memory in your bill of materials, or in your revenue?

The cleanest split in the whole exercise. Memory makers have it in revenue. Hardware companies have it in cost. Software companies have neither — which is a large part of why Salesforce and CrowdStrike could beat and be rewarded this week without anyone raising the subject.

Look for: cost of revenue as a share of sales, and whether it moved this year.

4. Have you pre-bought?

Locking supply forward converts a variable cost into a known one. It also converts an operating decision into a balance-sheet commitment, which is a genuine risk, not a free option. Nvidia’s $279bn is the extreme case; Micron has disclosed multi-year supply agreements carrying roughly $22bn of customer deposits and related commitments, around $18bn of it cash.

Look for: purchase obligations in the commitments note. It’s usually the least-read page and often the most informative.

5. Can policy reach your price?

New this quarter, and underweighted almost everywhere. Commerce Secretary Lutnick has signalled tariffs that could extend beyond chips to derivative products — servers, laptops, consoles — with relief possibly conditioned on US manufacturing commitments. A company whose pricing power depends on a scope document it doesn’t control has a different risk profile from one whose pricing power depends on physics.

Look for: import exposure by product category, and where final assembly happens.

THE LEDGER

Positions, not scores. I’m deliberately not producing a composite number here — the five questions measure different things in different units, and adding them up would create a false precision that hides the reasoning. A reader should be able to disagree with one row without the arithmetic collapsing.

In plain English: repricers raise prices when costs rise. Collectors are the reason costs rise but can’t be certain of keeping the proceeds. Absorbers pay, and take it out of margin.

Two rows deserve elaboration.

Broadcom is genuinely unresolved, and I’d rather say so than force it. Structurally it looks like Marvell: custom silicon, design-win pricing, customer-programme timing. On Thursday it traded like Nvidia, up 4.49% on no news of its own. Either the market has decided Broadcom’s contract structure differs materially from Marvell’s, or it simply rode the sector. Its September print is the resolution, and I’d treat any conviction before then as guesswork wearing a suit.

The hyperscalers are absorbers by choice, and that’s the point. Microsoft, Google and Amazon are the ones who receive the 15% increase. AWS committed to two million additional GPUs this week, taking its total to three million. They can absorb it because their AI revenue is growing faster than their input costs — for now. The first hyperscaler to flinch at the new price is the most important single event on the 2027 calendar, and nobody knows which one it will be.

THREE SIGNALS TO WATCH

30 September — Micron Q4 FY26. Does Micron disclose the share of fiscal 2027 output already contracted at fixed prices? A disclosed, high figure means tariff and spot risk sits largely with someone else. A second consecutive quarter without the number is itself an answer.

Early September — Broadcom Q3. The single row this framework can’t fill. If Broadcom’s custom silicon is graded the way Marvell’s was — good numbers, sold anyway — then design-win pricing is the shared weakness and the Absorber column gains its largest member. If it’s graded like Nvidia, the distinction is company-specific and this test needs a sixth question.

Whenever it lands — the Commerce tariff scope. Not the announcement. The scope document. Whether data-centre servers are exempted determines, in one paragraph, whether the entire Collector column keeps its pricing power and whether the Absorber column’s problem is chronic or acute.

WHAT TO DO WITH THIS

Take your largest AI holding. Answer the five questions from its most recent filing — pricing language, constraint language, cost of revenue, purchase obligations, import exposure. Write down which of the three positions it occupies, and the date its next disclosure could change that.

If you can’t answer question one from public documents, that is the finding. A company that doesn’t tell you when it can change its prices is a company betting you won’t ask during a year when the answer decides everything.

Disclaimer: This article is for informational and educational purposes only and does not constitute financial advice. iPrompt Signals is not a registered investment advisor. Always conduct your own research and consult a qualified financial professional before making investment decisions.

iPrompt Signals Deep Dive · The Memory Tax · 28 August 2026