Based on SK hynix’s Q2 2026 results and reporting by Investing.com.
When Nvidia raises AI server prices 15–17% and cites memory costs, this is the company that collects the toll — and its Q2 2026 margin tells you exactly how much pricing power the memory bottleneck currently holds.
SK Hynix Q2 2026 — The Numbers Behind Nvidia’s Hike
Q2 2026 Revenue — SK Hynix IR (late July 2026)
~KRW 79.3 trillion — up +257% year over year. HBM for AI servers is the primary driver across the portfolio.
Q2 2026 Operating Profit — SK Hynix IR (late July 2026)
~KRW 60.5 trillion — up +557% year over year. Operating margin: ~76%. This is the verified, clean figure — not the net margin, which one-off gains distort.
HBM4 / HBM4E — SK Hynix Product Milestones
HBM4 mass production began; HBM4E sampled. Roughly ten major customers locked into multi-year agreements with built-in pricing mechanisms.
This Week — Nvidia Price Signal
Nvidia notified customers of 15–17% price increases on next-generation AI servers, citing surging memory costs as the rationale.
Two data points landed within weeks of each other, and they are two ends of the same story. SK Hynix’s Q2 2026 results, reported in late July and covered by Investing.com among others, showed revenue of roughly KRW 79.3 trillion — up 257% year over year — and operating profit of roughly KRW 60.5 trillion, up 557%, for an operating margin of approximately 76%. That is the verified figure. The reported net margin sits above 100%, but it is distorted by one-off investment gains and should not be the headline number. The operating margin is the clean signal, and it is extraordinary by any standard in hardware manufacturing.
The driver is high-bandwidth memory — HBM — the stacked DRAM architecture that AI accelerators depend on for the memory bandwidth their matrix workloads require. This is a structural analysis pegged to this week’s news, not a breaking-earnings piece: the SK Hynix print is approximately four weeks old. What makes it newly relevant is Nvidia’s announcement this week that it is raising prices on next-generation AI servers by 15 to 17%, with surging memory costs cited as the reason. Read those two facts together and the picture resolves quickly.
On the product side, SK Hynix began mass production of HBM4 during the quarter and sampled HBM4E, the generation after that. Approximately ten major customers — the AI hyperscalers and GPU makers that define the market — sit on multi-year supply agreements with pricing mechanisms baked in. That contractual structure is not incidental. It is the instrument through which a cyclical commodity business behaves, for now, more like a software business.
The key insight: When Nvidia raises AI server prices and cites memory costs, it is not absorbing a supplier’s margin — it is passing a tax straight through to its customers. SK Hynix’s ~76% operating margin is where that tax pools. The bottleneck in the AI supply chain has moved from the GPU to the memory around it, and profit concentrates wherever the bottleneck sits. That chain is our analytical framing; it is not a claim any single company has made.
In its second quarter, SK Hynix grew revenue about 257% year over year and operating profit about 557%, on an operating margin of roughly 76% — one of the highest in technology — driven by high-bandwidth memory for AI servers. This is the supply side of the same squeeze that has Nvidia raising prices: memory is the bottleneck, and the memory makers are capturing the margin. But read it as a peak, not a run-rate. Memory is a deeply cyclical business, and 76% margins do not survive supply catching demand — the risk is an average-selling-price fade as HBM capacity ramps. Sources: SK Hynix IR.
The Structural Read
The AI stack, viewed through this week’s evidence, is a pricing-power ladder. Every rung has a different amount of it, and the distribution is not what most coverage implies.
At the bottom: memory. HBM is supplied by three companies — SK Hynix, Samsung, and Micron. That is an oligopoly, not a monopoly, and competition among the three will eventually matter. But right now, demand for HBM is outrunning the industry’s ability to produce it. When supply is genuinely constrained across all three players simultaneously, the oligopoly behaves like a toll booth: every AI cluster that trains or serves a model pays the toll. SK Hynix’s ~76% operating margin is what that toll looks like in a financial statement. The multi-year customer agreements with pricing mechanisms mean the toll is not purely spot-priced — it is partly locked in, which explains why the margin has proved more durable than typical memory cycles would predict.
Above memory: Nvidia. Nvidia’s response to higher memory costs is not to absorb them. It is to pass them through with a 15–17% price increase and retain its own margin. That is what a company with pricing power does: the cost flows up the chain, not down to the seller’s income statement. Nvidia is not a victim of HBM scarcity — it is a conduit for it. The labs and cloud providers at the top of the stack absorb the combined margin of both the memory layer and the GPU layer every time they procure a new cluster.
Above Nvidia: the labs, routing around the tax. The rational response to being taxed by two upstream oligopolists is to try to own more of the stack. OpenAI’s Jalapeño inference chip is explicitly designed to attack the memory bottleneck that makes GPU inference expensive. That is not a bet against Nvidia primarily — it is a bet that if memory and GPU together extract this much margin, vertically integrating into silicon changes the economics of running inference at scale. The custom-silicon wave at the labs is, in part, a direct response to the pricing-power structure visible in SK Hynix’s results.
At the top: the model layer commoditizes. Open weights, price cuts, and free routing define the model layer today. That is the mirror image of what is happening below. Scarce layers capture value; abundant layers give it away. The model layer is becoming abundant faster than any other part of the stack, which is why its economics look nothing like HBM’s.
The pricing-power stack runs from HBM (three-player oligopoly, ~76% operating margin, genuine supply constraint) through GPUs (Nvidia passes costs through, protects its own margin) to the labs (taxed customers, designing their own silicon to escape) to the model layer (open weights, price cuts, deliberate abundance). Each layer’s behavior is a function of its scarcity, not its technical importance. See the full framework in The AI Value Chain and the competitive dynamics in Beyond Nvidia’s Moat.
One hedge that must stay in frame: memory is one of the most cyclical businesses in technology. A ~76% operating margin is a supercycle peak, not a run-rate. The historical pattern is unforgiving — when HBM supply catches demand, average selling prices compress fast. SK Hynix’s own 2027 risk is an ASP fade as HBM4 and HBM4E capacity comes online across all three players. The multi-year agreements with pricing mechanisms soften that trajectory, but they do not repeal the cycle. This is the most profitable moment to be in HBM. It will not be the permanent one. Nvidia’s power and land acquisitions — the Cloverleaf and Lancium infrastructure bets — are partly a recognition that owning more of the supply chain above the memory layer hedges against exactly this kind of upstream pricing power compounding over time.
Three Implications
IMPLICATION 1 — The Memory Oligopoly Has Structural, Not Permanent, Pricing Power
SK Hynix, Samsung, and Micron jointly set the floor cost for every AI cluster built today. With demand outrunning supply across all three, the oligopoly behaves like a toll on AI infrastructure investment. The multi-year agreements with built-in pricing mechanisms extend that power further into the cycle than spot pricing alone would. But “structural” is not “permanent” — ASP compression is the historical endpoint of every memory supercycle, and 2027 HBM4 ramp-up is the credible timing risk for when it begins.
IMPLICATION 2 — Nvidia’s Price Hike Is a Pass-Through, Not a Margin Grab
Reading Nvidia’s 15–17% increase as pure pricing aggression misses the mechanism. Nvidia cited memory costs because memory costs are genuinely elevated — SK Hynix’s results confirm the supply-side. Nvidia is protecting its own margin by passing the tax to its customers rather than absorbing it. That is a mark of Nvidia’s pricing power too: it can raise prices without losing the order. But the origin of the cost pressure is upstream, in HBM, not in Nvidia’s own cost structure expanding.
IMPLICATION 3 — Custom Silicon at the Labs Is a Rational Tax-Avoidance Strategy
OpenAI’s Jalapeño chip, and the broader wave of lab-designed inference silicon, is best understood as a response to the combined margin extraction of the memory and GPU layers. When two upstream suppliers both have genuine pricing power, the rational long-run move for a large-scale buyer is to vertically integrate. The labs are not designing chips primarily because it is technically elegant — they are doing it because the cost of not owning more of the stack compounds at the rate of SK Hynix’s operating margin.
Business Engineer Framework
The Map of AI — Where the Value Actually Lives
SK Hynix’s margin and Nvidia’s price hike both make more sense when you can see the full nine-layer AI stack and which layers are genuinely scarce versus deliberately abundant. The Map of AI maps 200+ companies across the stack and shows you why the profit concentrates where it does — and where it is likely to migrate as supply constraints shift. The memory chokepoint is one chapter of that story.
Right now, the most profitable position in the entire AI economy is not the model, and it is not even the GPU — it is the memory the GPU depends on. SK Hyn
91,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
