The memory tax — the AI capex line item nobody underwrote

Companion article to Issue 21 // 31 July 2026 // R. Lauritsen

The memory tax — the AI capex line item nobody underwrote

Two hyperscalers disclosed this week how much of their capital budget is component price inflation rather than capacity. Nobody has done the arithmetic in public. Here it is — roughly a tenth of the 2026 AI buildout may buy no additional compute at all, and that changes who profits from it.

9 min read // Sectors: memory, hyperscaler capex, semiconductor equipment, consumer hardware

Issue 21 covered what happened and why the market split. This is the quantification, the supply-chain map, and the tests that would prove it wrong.

THE THESIS IN ONE PARAGRAPH

The AI capex number has stopped being a clean measure of compute. Memory has been in shortage since 2025, the price has climbed, and that increase now sits inside capital budgets that the market still reads as capacity. The consequence is straightforward and largely unpriced: reported capex overstates how much compute is actually being added, and the difference is a transfer of profit from the companies building AI to the four companies that make its memory.

HOW MUCH OF THE BUILDOUT IS ACTUALLY COMPUTE?

Two companies put a number on it this week, both almost in passing. Nobody else has.

[DISCLOSED] Microsoft. On its fiscal Q3 call the company guided to roughly $190 billion of calendar-2026 capital expenditure and stated that approximately $25 billion of that reflects the impact of higher component pricing. That is about 13% of the capital budget buying no incremental capacity whatsoever.

[DISCLOSED] Amazon. The 2026 outlook moved from $200 billion to $220 billion, with the increase attributed to higher memory costs. On the company's own framing that is 100% of the raise, and roughly 9% of the total, representing price rather than capacity.

[NOT DISCLOSED] Meta. The earlier move from under $100 billion to $125–145 billion was attributed in part to component pricing; the company has never quantified the share, and did not quantify it again this week when it lifted the floor to $130–145 billion. In a quarter where two competitors published the number, silence is itself a data point.

[INFERRED] If the two disclosed ratios are broadly representative of the group, then somewhere in the region of 9–13% of 2026 hyperscaler capital spending buys no additional compute at all. Against combined Microsoft, Amazon, Alphabet and Meta capital budgets of roughly $700 billion this year, that implies a transfer in the order of $65–90 billion from the buyers of compute to the makers of memory. We cannot verify that ratio outside the two companies that published it, and the mix will differ by fleet design and contract vintage. But it is large enough that its absence from most published capex models is the genuinely interesting part.

Two consequences follow, and they run in opposite directions. Anyone inferring GPU units, rack counts or token supply from capital spending dollars is now over-counting, and the error grows with the shortage — so the compute build is slower than the headline numbers imply. But the marginal AI dollar is also flowing somewhere the consensus trade doesn't assume: every dollar routed to DRAM is a dollar not routed to accelerators. That is an odd headwind for Nvidia to carry in the same week it offered to guarantee its largest customer's debt.

THE MEMORY LEDGER

If memory is a toll, every company in the AI trade sits in one of four positions relative to it. This is the organising question for the rest of the piece: when the price of memory rises, do you collect it, tool it, pass it on, or eat it?

Moves are Thursday 30 July's session close unless marked AH (after-hours). Samsung and SK Hynix are Seoul-listed and round-tripped within the week. CXMT's figure is its Monday listing debut, not a session move.

The four rows in the middle are the ones that carry information. Microsoft, Amazon, Meta and Apple bought from the same four suppliers at the same prices in the same quarter. What separated them was not the size of the bill. It was whether anyone else was named on it.

WHAT DECIDES WHO CAN PASS IT ON

Pass-through capability is not a matter of scale or margin. It reduces to five balance-sheet variables, and they are all disclosed.

External customer commitments. Microsoft's remaining performance obligation — contracted revenue not yet billed — stands at $678 billion and grew 84% year on year. Meta has no equivalent line, because it has no external compute customer.

Contract duration. AWS capacity is contracted on multi-year terms, which is what allows Amazon to absorb a component price rise across the life of an asset rather than a quarter.

Cash-flow coverage. Microsoft guided to remaining cash-flow positive through fiscal 2027. Meta's operating cash flow of $31.9 billion was almost entirely consumed by $31.1 billion of capital expenditure, leaving $784 million.

Lease obligations. Meta's future data-centre leases reached $279 billion, up 53% in a single quarter. Those are contractual, which means the spending cannot be moderated in response to a share price.

Product pricing power. Apple has it in principle and did not use it, choosing a guidance cut over a price rise into an iPhone 18 cycle. That is the cleanest evidence that memory inflation has reached the consumer P&L.

Alphabet, reporting a week earlier, is the instructive middle case: a $514 billion backlog, but negative free cash flow for the first time in its public history. The market punished the second half of that sentence rather than rewarding the first, which suggests the coverage variable currently outranks the commitment variable.

THE WEEK IN FIVE FACTS

The evidence base, compressed. Issue 21 carries the full narrative.

1. China's ChangXin Memory rose 466% on its Shanghai debut on Monday, raising about $8.6 billion and closing as the mainland's most valuable listed company. Seoul read it as a supply threat: SK Hynix fell 14.65%, Samsung more than 13%, and global chip stocks shed over $1 trillion across the move.

2. Samsung then posted record revenue and an 1,814% jump in operating profit on Thursday, and guided the memory shortage to deepen into 2028. Micron rose 18.4%, SanDisk 21%, Lam Research 20% and its best session since 1999.

3. Microsoft reported Azure up 43%, $678 billion of contracted revenue, and a capex framework it framed as demand-driven. The stock rose 15.5% and added roughly $450 billion of market value — the largest single-day gain any US company has recorded.

4. Meta reported revenue up 28% and its first EPS miss in three years, with free cash flow down to $784 million from $8.5 billion. It fell 7.95%, an eleventh consecutive down session and the longest losing streak in its history.

5. Amazon raised 2026 capital spending to $220 billion and said the increase was memory costs. Apple beat on revenue and earnings, missed on Services, and cut current-quarter guidance citing supply constraints.

The macro backdrop mattered less than it looked. The Fed held at 3.50–3.75% on a 9–3 vote with three dissents for a hike, and the 30-year pushed above 5.2%; equities fell on Wednesday and recovered it all on Thursday. Rates set the mood. Earnings set the prices.

WHAT WOULD BREAK THIS

Two things end a shortage, and they are not the same thing. Conflating them is the most common error in commodity theses.

More supply. CXMT raised about $8.6 billion, closed its debut 466% above the offer price, and has moved from roughly 3% to about 8% of global DRAM in a year — against Samsung at 38%, SK Hynix at 29% and Micron at 22%. The gap is high-bandwidth memory, and it is real: CXMT lacks EUV lithography under US export controls, and HBM is where the AI shortage actually bites. But it is building an HBM packaging facility in Shanghai with production expected from late this year, and Apple has reportedly begun testing its DRAM for devices sold in China — which shows both how tight Western supply is and how quickly qualification moves when a buyer is desperate.

Less demand. This is the mirror risk and it works through a different mechanism. The memory makers' revenue is the hyperscalers' capex line. If the market keeps punishing capital spending, the rational corporate response is to spend less, and the toll collector's income statement is the first casualty — no new supply required. Microsoft's 15.5% day made that outcome less likely this week by demonstrating that contracted spending gets rewarded. It did not make it impossible, and Meta is the live test of whether the punishment changes behaviour.

The distinction matters for position sizing. A supply shock re-rates the memory makers permanently. A demand pause re-rates them temporarily and hits the equipment names harder, because orders are cancellable and installed capacity is not.

THREE SIGNALS THAT DECIDE IT

Ordered by how directly each one bears on the memory thesis.

1. CXMT qualifies HBM outside China — before 31 December. The Shanghai packaging line is slated to start late in 2026. A named customer outside China before year-end and the shortage-to-2028 thesis has a hole in it, which would make Micron's 6x forward multiple correct rather than cheap. No qualification and the incumbents keep pricing power for another year. This is the most direct test available.

2. The disclosure test at the October prints. Watch whether any hyperscaler drops the component-pricing line from its capex commentary. The phrase disappearing is the cleanest signal that the toll has peaked. The phrase spreading to a fourth company — Alphabet, most likely — is the signal that it has not, and that 2027 budgets are being written around it.

3. Nvidia's Q2, 26 August. Least directly connected to memory, most consequential if it breaks. Does the proposed $250 billion OpenAI guarantee appear in the filing as a disclosed contingent liability, or not at all? If it is disclosed, the market can finally price it. If a commitment of that size is signed and stays off the face of the accounts, that is a different and worse problem — and it would sit underneath every position in the ledger above.

Stay curious — and stay qualified.

— R. Lauritsen

Editor, iPrompt Signals

METHODOLOGY AND SOURCES

Reporting comes from named primary sources cited inline; estimates and inferences are tagged in the text. Primary: Microsoft FY2026 Q4 earnings release and call (29 July 2026), plus the FY2026 Q3 call transcript for the component-pricing disclosure; Meta Q2 2026 earnings release, call and securities filing (29 July 2026); Amazon Q2 2026 earnings release and call (30 July 2026); Apple fiscal Q3 2026 earnings release and call (30 July 2026); Samsung Electronics Q2 2026 results; the FOMC statement of 29 July 2026; CXMT's IPO prospectus and Shanghai listing data. Secondary: CNBC, Reuters, Bloomberg, the Wall Street Journal, Barron's, Business Insider, Korea JoongAng Daily, and Counterpoint Research for DRAM market share. Prices are Thursday 30 July closes from consolidated tape data unless marked after-hours; fiscal and calendar years are distinguished throughout. The $65–90 billion transfer figure is our own inference from two disclosed ratios, labelled as such at the point of use. It is not a company-reported number.

Disclaimer: For informational purposes only. Not financial advice. iPrompt Signals is not a registered investment advisor. Conduct your own research and consult a qualified financial professional before making investment decisions.