Everything About AI
NextFin News - OpenAI Chief Executive Sam Altman met with roughly 50 executives from Samsung Group at a leadership seminar in Yongin, south of Seoul, on Saturday, as the world's largest memory-chip maker wrestles with a shortage that has already pushed its third-quarter operating profit to a record KRW 107.4 trillion. The visit lands two days after Altman told a podcast audience that memory manufacturers are sitting on a potential $100 billion a year in free cash flow each, yet remain too scarred by past busts to build new factories fast enough — a tension that now sits at the center of the AI boom.
Samsung said the session, held at its human resources development center, brought together presidents from its electronics, financial and biotechnology affiliates to discuss how artificial intelligence will reshape industries, corporate decision-making and business opportunities over the next decade. Altman's message to the room was that competitive advantage does not come from simply adopting AI, but from using it to move faster, make better products and out-execute rivals.
The setting is the story. Samsung is not merely an AI customer; it is the choke point. And the industry's most powerful buyers are now flying to Seoul to negotiate directly with the companies that control the memory.
The Visit, the Earnings and the Empty Warehouse
The seminar was framed by Samsung as part of a series of leadership sessions designed to advance the group's company-wide AI transformation, following a letter of intent the two companies signed last October on building global AI infrastructure. Samsung SDS agreed to cooperate on AI data-center construction and enterprise AI services, while Samsung C&T and Samsung Heavy Industries agreed to jointly develop offshore floating data centers. Samsung Electronics chairman Lee Jae-yong visited OpenAI's San Francisco headquarters in July, and the exchanges have continued since. Saturday's meeting was the latest link in that chain — and it arrived with the memory market at its tightest point in years.
On October 29, Samsung reported third-quarter revenue of KRW 195.3 trillion, up 127% from a year earlier, and operating profit of KRW 107.4 trillion, up 783% — the first time a global technology company has exceeded KRW 100 trillion in quarterly operating profit. Net profit rose 515% to KRW 75.1 trillion. The company has now posted four consecutive record quarters for revenue, and the engine was memory, specifically the high-bandwidth memory that AI accelerators cannot run without.
"Until right now, I don't think making memory has ever been a great business over a long period of time," Altman said in the October 29 interview. "And then all of a sudden, these guys are each going to make like $100 billion of free cash flow in a year."
Inside Samsung's memory division, the numbers confirm the shift. The company told investors that HBM4 sales more than tripled in the third quarter and now account for 60% of its total HBM revenue, while DRAM average selling prices rose 20% quarter over quarter and NAND prices climbed by high-teens percentages. Management went further: it said demand for AI-related memory will significantly exceed industry supply in 2027, and that its 2027 HBM supply is already booked and sold out at agreed prices. Samsung also said it achieved its goal of bringing its HBM market share in line with its overall DRAM market share ahead of schedule, with market data putting its HBM share at roughly 33%, about double where it stood a year earlier.
Across the industry, the physical cushion has nearly disappeared. KB Securities estimated in early September that Samsung and SK Hynix — which together account for about 64% of global DRAM revenue — held less than 10 days of finished DRAM inventory. That is a level that leaves almost no room for a demand surprise or a production hiccup. Samsung is targeting roughly a 47% increase in HBM production capacity to about 250,000 wafers per month by the end of 2026, concentrated at its Pyeongtaek P4 line. But KB Securities notes that no meaningful new capacity is expected until Samsung's Pyeongtaek P5 line and the Yongin cluster come online in 2028.
That gap between demand and deliverable capacity is why Altman was in Korea, why AMD's Lisa Su made two trips to Seoul in seven months, and why pricing forecasts keep moving up. TrendForce has raised its 2027 HBM price outlook and now projects blended average selling prices to rise 121% year over year, while expecting DRAM supply to remain tight through 2027. HBM bit shipments are projected to grow 50% to 60% in 2027 — and even that growth rate is the constrained outcome, not evidence that the gap is closing.
Why the Invisible Hand Is Failing
Altman's diagnosis cuts to the mechanism. After noting the profit opportunity, he offered an explanation rooted in industry memory rather than economics textbooks.
"I think the invisible hand is failing right now," he said, citing "how many times they almost went bankrupt by making too many fabs" during past boom-bust cycles.
The reason is history. Memory is the most brutally cyclical business in semiconductors. For two decades, Samsung, SK Hynix and Micron have been burned by the same pattern: build capacity during a boom, flood the market, watch prices collapse, and survive only because they were too large to exit. The trauma of the last bust is now stronger than the lure of the current boom, and the result is a supplier base that is generating record cash while refusing to deploy it.
This is a structural constraint wrapped in cyclical clothing, and separating the two is the key to reading what comes next.
The cyclical part is real and familiar. Memory prices and capital expenditure will eventually mean-revert, as they have in every cycle since the 1980s. When prices rise, capacity follows; when capacity follows, prices fall. A three-year demand forecast in a business with two-year demand visibility is a recipe for overbuilding, and every memory executive alive has lived through that lesson at least twice. On that reading, today's scarcity is simply the upswing of a cycle that will turn, and the correct posture is caution.
But the structural part is new, and it is what makes this cycle different. HBM is not a commodity like standard DRAM. It is a specialized, tightly qualified component whose demand is tied to a handful of hyperscale buyers and a small set of accelerator makers. Qualification cycles are long, yields on advanced stacks are hard-won, and the production process consumes engineering talent and cleanroom capacity that cannot be switched on overnight. Samsung's HBM4 implementation is rated at up to 3.3 terabytes per second of bandwidth per stack, roughly 2.7 times the HBM3E figure, and the technology doubles the per-stack interface width to 2,048 I/O connections. That is not a product line a company can spin up between cycles.
Every wafer redirected to HBM4 is a wafer not making the DDR5 and LPDDR5 that go into phones, PCs and servers — which is why a shortage that began in AI data centers is now lifting prices across the entire memory market, including consumer DRAM. TrendForce reported in September that even as broader DDR4 demand weakened, spot prices for DDR4 modules rose more than 2% in a week, a textbook sign of a supply-constrained market rather than a demand-driven one. The scarcity is propagating through the whole stack.
The second-order effect is what most investors are still missing. The shortage is not just raising memory prices; it is reordering the entire AI supply chain and turning procurement into a competitive moat. AMD's chief executive is personally flying to Seoul to secure supply, and she was direct about the pressure.
"Memory supply is very tight," Lisa Su said during her October 7 visit, adding that "AI demand is very high, so we are planning long-term supply chains with our partners."
Anthropic's October 26 IPO filing flagged a potential HBM shortfall from key suppliers as a risk factor — a rare public acknowledgment from inside the AI industry that its growth path is physically constrained. Hyperscalers are signing multi-year agreements that lock up capacity years in advance, because a signed contract today is worth far more than cash tomorrow when the alternative is no memory at all. A 36GB HBM3E module was trading at about $2,100 on the spot market in early September, roughly four to five times the $300 to $400 that buyers typically pay under long-term agreements for the same module. That premium is the price of certainty in a rationed market. In this environment, the companies winning the AI race are not necessarily those with the best models, but those whose procurement teams moved first.
There is also a cost-channel effect running back through the AI buildout. Altman estimated data-center costs at about $70 billion per gigawatt and said $20 billion would be a far healthier number; he also noted that memory utilization in today's systems sits around just 5%. If memory supply does not expand, the economics of AI infrastructure deteriorate not because models are unprofitable, but because the physical inputs required to run them are rationed by a traumatized supplier base. The constraint compounds: higher memory prices raise the cost per unit of compute, which raises the revenue each AI application must generate to justify its existence, which makes the demand forecast that memory makers are refusing to bet on even harder to hit.
The Counter-Thesis: Restraint May Be Rational
The strongest argument against Altman's view is simple, and it deserves to be taken seriously: the memory makers may be right to hold back. If they build aggressively and AI demand disappoints — if data-center capital expenditure growth slows, if model efficiency improves faster than expected, or if the current generation of AI applications fails to monetize at the pace investors assume — they will be left with tens of billions in stranded capacity and a price collapse that wipes out the very profits they are now enjoying. Memory has punished over-optimism more consistently than almost any industry on earth.
KB Securities, which raised its Samsung target price by 18% to 130,000 won and called the firm its top pick among Korean semiconductors, anchors its bullish case on data-center capital expenditure jumping fivefold by 2030. That is a large assumption, and it is the assumption the entire shortage narrative rests on. If it proves wrong, the shortage resolves itself quickly and painfully — and the companies that restrained capex will look like geniuses rather than cowards.
There is also a pricing-power argument: keeping supply tight may be a feature, not a bug. In an industry where three players control the vast majority of global DRAM revenue, limiting supply to raise prices is simply a better business model than racing to the bottom on volume. From that angle, the "invisible hand" is not failing; it is working exactly as a rational, concentrated industry would want, and Altman's complaint is really a description of successful oligopoly behavior.
But this defense underestimates the systemic risk. A memory shortage that constrains AI infrastructure deployment does not just transfer wealth from chip buyers to chip sellers. It slows the entire technology buildout that the equity market is already pricing in. Nvidia, down 4.2% on the day, and the broader AI complex are exposed to the same bottleneck. If memory supply does not expand, the AI capital-expenditure cycle hits a wall not because demand faded, but because the physical inputs ran out. That is a different kind of shock — one that earnings revisions have not yet fully captured, because analysts are modeling a demand slowdown when the actual risk is a supply ceiling.
The falsifying signal is specific and observable. If Samsung and SK Hynix announce meaningful HBM capacity additions that come online before 2028 — or if HBM contract prices fail to rise in 2027 despite management's statement that 2027 supply is already sold out — then the structural-constraint thesis is wrong, and the shortage is a shorter, cyclical spike that the market is over-extrapolating. Conversely, if spot premiums for HBM continue to widen against contract prices, or if lead times stretch beyond current agreements, the constraint is real and durable. The 2027 pricing print will be the verdict.
What Comes Next
Short term, the market is telling its own story, and it diverges from the fundamental narrative. Samsung Electronics closed at 225,000 won, down 3.0%, while SK Hynix fell 7.1% and Micron dropped 5.3%. Nvidia slipped 4.2%. That is a broad memory and AI selloff on the very day the shortage narrative reached the highest levels of the industry. The divergence is worth watching: either the market is looking through a cyclical peak and pricing a demand slowdown that the supply data has not yet confirmed, or it is discounting the possibility that today's record margins are the top of the cycle.
Medium term, the calendar is fixed. The industry is living off capacity decisions made years ago, and the next meaningful inflection points are Samsung's Pyeongtaek P5 and the Yongin cluster, both slated for 2028. Until then, supply is essentially inelastic. Any demand surprise — up or down — translates almost entirely into price, which is why the next two years will be dominated by pricing power rather than volume growth. For buyers, the strategic task is no longer forecasting the optimal purchase quarter; it is securing physical allocation before the window closes.
Long term, the question is whether AI demand proves durable enough to justify the capital expenditure that memory makers are currently refusing to make. If it does, today's restraint will be read as the mistake of the decade, and the companies that held back will have ceded the AI era to whoever was willing to build. If it does not, it will be read as discipline that preserved an industry.
The base case is continued tightness through 2027, with HBM contract prices rising by multiples and memory makers enjoying record profits while rationing supply. The upside case for buyers is a faster-than-expected demand slowdown that breaks the pricing power and refills the inventory cushion. The downside case is a true physical shortage that forces hyperscalers to delay data-center deployments and pushes AI infrastructure costs above current models.
The memory industry spent two decades learning that overbuilding kills you. The risk now is that underbuilding kills everyone else first.
Explore more exclusive insights at nextfin.ai.
