AI Infrastructure Narrative Shift: Pricing Power Moves from TSMC to Memory
0xBroomberg
A Digitimes analysis of seven companies' Q2 earnings shows pricing power in the AI supply chain has shifted from TSMC to Samsung and SK Hynix — memory leapfrogged on price hikes, not market share — and how long this configuration lasts depends on when prices mean-revert.
How did TSMC lose its spot at the top of the chain?
A year ago TSMC was the bottleneck, the price-setter, and the single largest revenue source in AI chips. Memory revenue ran at roughly two-thirds of TSMC's.
Four quarters later the ranking has flipped: at constant exchange rates, Samsung Device Solutions quarterly revenue rose from 0.62× to 2.11× TSMC's; SK Hynix went from 0.49× to 1.32×.
This means → neither memory company gained foundry share. The driver was pure pricing — a price event, not a share event.
How much did memory prices actually rise?
SK Hynix disclosed average selling prices up ~30% quarter-on-quarter, while shipment volume — measured in bits — grew only high single digits. Volume was the footnote; price was the headline.
Samsung memory revenue surged 471% year-on-year while describing capacity as "limited."
In plain terms = the critical difference between a price event and a share gain: share, once won, stays; price, once up, eventually comes back down. That difference determines how long the current configuration can hold.
Are cloud companies still spending?
Over five quarters the Big Four's combined capex rose from $87.5 billion to $163.9 billion — but not in a straight line. In Q1 2026 Meta's capex fell from $21.4 billion to $19.0 billion; Microsoft held flat for two consecutive quarters. A reasonable analyst could have concluded "the spending cycle has peaked."
One quarter later, Meta spent $30.1 billion; Microsoft, $35.8 billion.
This reflects a structural reality: quarterly capex is driven by delivery schedules, lease commencement dates, and component availability. It is inherently lumpy and cannot be read as a smooth signal of management conviction.
With margins this high, why aren't suppliers expanding?
SK Hynix posted a 76% operating margin but spent only 18% of operating profit on capex. Samsung's ratio: 16%.
Compare the buy side: Amazon's capex-to-operating-profit ratio hit 193%; Meta's, 160%; Alphabet's, 110%. Buyers are investing aggressively; sellers are collecting cash and sitting still.
This means → suppliers earning outsized profits while barely reinvesting are, by definition, sustaining the supply tightness that supports high prices. But it also means that when expansion finally starts — or demand softens — the price correction could be sharper for the delay.
Can long-term agreements break the old memory-cycle playbook?
SK Hynix disclosed roughly ten long-term agreements, with pricing structures explicitly designed to absorb volatility, including "financial mechanisms such as prepayments to support contract fulfillment." Customers are paying cash upfront to lock in supply.
Microsoft carries $678 billion in commercial contract obligations; Alphabet's cloud backlog stands at $514 billion — capex is backed by signed demand, not forecasts.
In plain terms = the historical memory crash script runs like this: suppliers expand into spot demand, then trample each other on the way down. Prepayment-backed long-term agreements change the game theory — either they give suppliers the visibility to expand without triggering the usual price collapse, or they simply lock in peak pricing for two years and defer the correction. Both readings follow from the same disclosures, and which one proves right is the single most important open question for the sector.
What is custom silicon worth in this cycle?
Google Cloud's operating margin rose 14.8 percentage points; Amazon's cloud unit gained 6.5 points. Microsoft's cloud gross margin, by contrast, fell 3 points — and Microsoft named the reason directly: capex up 70%, including "the impact of higher component pricing"; Intelligent Cloud cost of revenue up 42% against 32% revenue growth.
Amazon's custom-chip business hit an annualized revenue run rate above $25 billion with triple-digit growth. Anthropic and OpenAI both signed multi-year, multi-gigawatt Trainium — Amazon's in-house AI training chip — commitments. Alphabet's TPU — its tensor processing unit, a custom AI chip — program has an even longer track record.
This reflects a payoff years in the making: vertical integration pursued for strategic reasons is now functioning as a cost buffer at the exact moment component pricing pressure is highest. Whoever enters the next memory-pricing cycle with stronger custom-silicon capabilities will see a smaller margin divergence in their cloud business.
Content is for reference only, not financial advice.