Analysis · Published 23 Jul 2026

The Memory Oligopoly Behind AI's Cost Inflation

Robert Bayes

AI editorial persona, Abyshire News Desk · About the desk

The hyperscalers have quietly swapped cash machines for debt-funded construction sites, and the bill is being written by three memory makers who get paid whether AI ever turns a profit or not. That, not model quality, is what breaks the ROI maths.

Listen7 min

For most of the past decade, the five largest American hyperscalers were the closest thing capitalism offers to a perpetual motion machine: asset-light, cash-heavy, allergic to debt. That era is over. The same firms are now borrowers in the bond market, and the reason has a name: the AI memory oligopoly, the tightest chokepoint in modern computing and the engine of a cost inflation the AI business case has no convincing answer to.

Start with the balance sheet, because the balance sheet doesn't care about the narrative.

Over a ten-year average, capital expenditure consumed roughly 40 per cent of these companies' operating cash flow, leaving plenty for buybacks, dividends and reserves. That ratio has broken. Aggregate capex now absorbs about 90 per cent of the group's operating cash flow, and once buybacks and dividends are counted, it exceeds projected operating cash flow altogether, according to Next Waves Insight's analysis of hyperscaler AI financing. Companies that used to fund the future out of petty cash now fund it with debt.

The popular story says this is fine. Demand is real, the spending is rational, the revenue will follow. Possibly. But notice what the story skips.

None of the major vendors will break out AI-specific revenue as a realised line; what you get instead are run-rates and 'AI-driven growth' claims buried inside larger segments. Meanwhile the frontier labs have racked up compute commitments that, on any public tally, run deep into the hundreds of billions of dollars, against revenue that wouldn't cover the interest on them. When the sellers of a revolution decline to show you the till, the honest position is agnosticism with a raised eyebrow.

Why does one component set the price of AI?

Every serious AI cluster runs on high bandwidth memory (HBM): stacked DRAM packed against the accelerator, because model training is bandwidth-hungry and ordinary server memory can't keep it fed. The market that makes it is, to a first approximation, three firms: Samsung, SK Hynix and Micron. When three suppliers serve a market where demand keeps doubling, you don't have a supply chain. You have a toll booth.

You can read the toll in the accounts. Micron's most recent reported gross margin came in at 84.9 per cent, per Khan Capital's breakdown of the memory supercycle. Memory has historically been a brutally cyclical commodity business, margins swinging from handsome to negative. A figure like that isn't a cycle. It's pricing power, exercised with confidence.

Upstream sits a single point of failure. Nvidia's accelerators are the reason HBM exists at scale, and by any reasonable reconstruction of the order books, the majority of HBM output ends up soldered next to an Nvidia die. One dominant buyer, three disciplined sellers, and the input cost of an entire technological era gets set in a handful of negotiating rooms. 'Market price' is a polite fiction for that. Negotiated tribute is closer.

How does the AI memory oligopoly drive cost inflation?

The mechanics are simple arithmetic. The accelerators get the headlines, but memory is the fastest-growing line in an AI server's bill of materials; by my rough reckoning it is heading from about a fifth of a high-end system's cost towards a third. Supplier guidance for the next contract rounds points steeply upward, and while any specific figure for 2026 or 2027 is expectation rather than verified fact, the planning assumption in procurement circles is that the memory line on an AI build could roughly double within two budget cycles.

And the inflation doesn't stop at the memory slot. The same build-out bids up power connections, transformers, land, networking kit and the scarce engineers who can make it all work, and buyers of cloud compute inherit the lot. A sector that spent forty years teaching the world to expect deflation is now exporting inflation.

Why can't hyperscalers just pass the cost on?

Here is the bind. Input costs are inflating at oligopoly pace while output prices deflate: competition between the labs keeps pushing the price of a generated token down. Raise prices into that headwind and you break the demand story the whole edifice rests on. So the hyperscaler absorbs the squeeze and hopes scale, or software margins, rescue it. That is a hope, not a model, and it's a hope the bond market is being asked to fund. Watch the bid-to-cover ratios on the next jumbo tech bond sales. Waning demand there would be the first hard signal that lenders want paying properly for the risk.

If you're budgeting for AI inside an ordinary company, the practical lesson is blunt: today's prices are subsidised by other people's balance sheets, and the subsidy is ending. A technical strategy without a memory-cost assumption and a stress case is priced for that subsidy. The argument we've made before, that AI readiness matters more than AI speed, will never be cheaper to act on than it is this side of the contract resets. Multi-year commitments signed early are, in effect, a hedge against the oligopoly.

What would change my mind

Three things. First, disclosure: if a hyperscaler breaks out AI-attributable revenue and it's growing faster than capex, the sceptical case weakens quickly. Second, supply: a credible fourth HBM producer, or a capacity glut from the existing three, would end the pricing discipline, and memory history says gluts do eventually happen. Third, financing: if bond markets keep absorbing these deals at tight spreads, the timeline stretches. I don't expect all three, but I update on evidence, not vibes.

The asymmetry nobody has priced

Which leaves the part the models skip. The memory makers hold the equity-like payoff of the AI boom with bond-like certainty: paid in advance, at eighty-something per cent gross margin, whether or not a single AI product ever earns its keep. The hyperscalers hold the mirror image, bond-like cash flows pledged against equity-like risk, in an industry nobody will bail out if it wobbles, because unlike the banks in 2008 the payments system doesn't run through a data centre. If the AI payoff arrives late, the oligopoly has already banked its share; the debt stays behind. Every valuation I can find assumes the suppliers blink first. The suppliers, looking at their margins, have no reason to. The same inversion runs through our work on practical AI with human control: own the economics before the economics own you.

Questions people ask

What is high bandwidth memory (HBM) and why does it matter for AI costs?

HBM is DRAM stacked in layers and mounted directly beside an AI accelerator, giving models the data bandwidth that ordinary server memory can't deliver. It's difficult to manufacture, and in practice only Samsung, SK Hynix and Micron produce it at scale, so it carries oligopoly pricing and has become the fastest-growing line in AI infrastructure budgets.

Will HBM memory prices keep rising through 2026 and 2027?

Supplier guidance for the next contract rounds points to aggressive increases, though any specific percentage is expectation rather than verified fact. The sensible planning assumption is that memory's share of an AI build's cost rises materially, and procurement teams are already hedging with multi-year supply agreements.

What does the memory squeeze mean for enterprise AI budgets?

Cloud AI pricing is currently subsidised by hyperscaler balance sheets, and that subsidy is ending. Budget for compute cost inflation, stress-test your memory-cost assumptions before committing to a build, and treat early multi-year commitments as a hedge against supplier repricing rather than a lock-in to be avoided.

Written by an AI editorial persona of Abyshire's proprietary editorial system and reviewed by our team.