Samsung's HBM4 Crossed $1B in Revenue Just Four Months In — and This Year's Supply Is Already Sold Out

The real gold mine of the AI era was the memory bolted next to the GPU

Here's the deal: Samsung's 6th-generation high-bandwidth AI memory, HBM4, crossed $1B in cumulative revenue about four months after launch. The figure came in late June, alongside news that Samsung's chairman personally reviewed HBM4 supply. The more striking part is elsewhere — this year's HBM4 output is already fully sold out, and from Q3, HBM4 is expected to make up more than half of Samsung's entire HBM revenue. Sold out before it's even built.

What's HBM? It's the ultra-fast memory stacked right next to AI accelerators (GPUs). Running AI models means shuttling enormous amounts of data to the GPU fast, and HBM is that "data highway." That's why the latest AI chips from companies like NVIDIA require HBM, and as the AI boom grows, HBM demand explodes alongside it. HBM4 is the newest generation, and it's Samsung's vehicle for re-entering the fight for market leadership.

Backing the growth is enormous investment. Samsung is pouring roughly 110 trillion won (~$73B) into its chip business in 2026 — a record, and a sharp increase from 2025. A large chunk is allocated to HBM production lines (Cheonan, Pyeongtaek), with the goal of sharply raising AI-grade memory output by Q4 and lifting share in the premium server market. It's a loud signal of intent to reclaim the HBM throne it ceded to SK Hynix.

So here's today's story: why HBM4 matters this much, why Samsung is going all out at this moment, what these numbers mean for AI infrastructure overall, how past memory cycles played out, and what changes for investors and industry watchers.

The players — Samsung, the HBM market, and rival SK Hynix

Samsung Electronics first. It's the world's largest memory chipmaker. It led for years in commodity memory like DRAM and NAND — yet fell a step behind in HBM, the heart of the AI boom. In the early HBM generations, SK Hynix locked up NVIDIA supply and took leadership. For Samsung, "memory No. 1 but No. 2 in the hottest memory" was a sore spot, and HBM4 is the card to flip it.

Next, the HBM market itself. It's the highest-margin corner of semiconductors right now. Ordinary DRAM is a volatile market where prices swing as supply grows; HBM is bound to AI chips as one unit, with sturdy demand and high prices. As each AI accelerator takes more HBM, and faster generations, HBM is a market that "grows automatically as AI compute grows." Samsung, SK Hynix, and Micron fight over this yolk.

The third character is rival SK Hynix. By locking up NVIDIA supply in earlier HBM generations, it took the "HBM champion" title. Samsung's HBM4 sellout and massive investment are about shaking that structure. When one side gets ahead at a generation transition (HBM3 → HBM4), it shapes the next several years of share. That's why the changeover happening now is the most important inflection point for both companies.

Tie it together: the AI boom's real beneficiary isn't only the GPU but the HBM bolted beside it, and in that hottest market, Samsung is moving to reclaim leadership with an HBM4 sellout and record investment. That's the spine.

What's happening — by the numbers

Item	Detail
HBM4 cumulative revenue	Crossed $1B ~4 months after launch
Annual revenue outlook	Around $10B by year-end
This year's output	Fully sold out
Share within HBM	Expected to exceed 50% of total HBM revenue from Q3
2026 chip investment	~110 trillion won (~$73B) — a record
HBM line allocation	Large share concentrated at Cheonan and Pyeongtaek

The numbers are clear. $1B at four months and a ~$10B year-end outlook mean HBM4 is effectively "printing money the moment it ships." And "sold out" matters — selling everything you can make means the HBM market is in a state where demand exceeds supply. In such a market, pricing power sits with the seller and margins are fat. That's exactly why Samsung is spending 110 trillion won to grow capacity — if it sells as fast as you can make it, making more directly means more revenue.

Two cautions, though. First, some figures mix reporting with estimates. Numbers like "$1B in four months," "$10B by year-end," and ">50% in Q3" blend market forecasts and company guidance, so they may adjust as quarterly results firm up. Second, memory is inherently a cyclical industry. Demand outrunning supply looks like a "super-cycle" now, but history says that when everyone piles into capacity, oversupply eventually bends prices down. The 110-trillion-won spend is also a bet that demand stays strong.

Still, the direction of this phase is clear. As long as AI compute demand shows no sign of slowing, HBM is likely to stay a "can't-make-it-fast-enough" market for a while. Samsung's HBM4 sellout is the sharpest signal of that trend.

Who wins

Samsung gained both pride and substance. It set up a springboard to shed the "memory No. 1, HBM No. 2" label, and locked in a sellout in the highest-margin market at the same time. The chairman personally checking supply is a message that "this is the company's top priority." Getting ahead at a generation transition makes the next several years easier, so this bet means more than a single quarter's results.

AI chip companies like NVIDIA benefit indirectly too. With two strong HBM suppliers in Samsung and SK Hynix, they aren't at the mercy of one and keep negotiating leverage. One of the biggest bottlenecks for AI chips is HBM supply, so Samsung's expansion opens a "supply valve" for chipmakers — easing a chokepoint for AI infrastructure overall.

Every company that has to build AI infrastructure benefits in the long run. More HBM output means more AI accelerator supply, which ultimately pushes AI compute costs down. Cloud companies scaling data centers and startups needing to run AI both get the indirect upside of an easing HBM bottleneck. When the lowest layer of the AI boom (memory) thickens, every layer above benefits.

Precedents — successes and failures

The memory industry's history is a history of cycles. The success side is "getting ahead at a generation transition and seizing leadership." When a new memory standard arrives, the company that masters mass production first and lands key customers tends to enjoy high margins throughout that generation. SK Hynix did that in earlier HBM generations; Samsung is aiming for that seat in HBM4. Pre-empting a generation shift is that valuable.

The failure and bubble side is just as clear. Memory has repeatedly seen everyone pile into capacity during a boom, only for supply to overshoot demand a few years later and prices to crater. Several "super-cycles" were followed by brutal down-cycles. Today's HBM boom may eventually feel the same gravity. The huge 110-trillion-won bet carries the risk of becoming an overbuild boomerang if the timing is misread.

The lesson: to win in memory you need balance — get ahead at the generation transition, but don't overreach at the top of the cycle. Samsung's HBM4 sellout is a clear success signal on the former (pre-emption). How the latter (cycle management) plays out depends on whether AI demand holds this pace for several more years. For now, demand is outrunning supply, and that's working in Samsung's favor.

Competitor counter-plays

SK Hynix's counter is "defend leadership." Armed with the NVIDIA relationship, mass-production know-how, and trust it built in earlier generations, it will try to hold core HBM4 volume. The premium of being a "proven supplier" doesn't crumble easily across a generation shift. Expect a back-and-forth of Samsung's sellout and expansion against SK Hynix's own HBM4 ramp and capacity additions.

Micron is a variable too. As a U.S.-based memory maker, it can court customers who want a non-Korean source for geopolitical reasons. If the HBM market hardens from a Samsung–SK duopoly into a three-way race, supply rises and pricing power disperses. For chipmakers, a third source is welcome.

AI chip companies' counter is "diversify supply and pre-buy." Knowing HBM is the bottleneck, they secure volume across multiple suppliers in advance and co-develop next-gen standards. Samsung's HBM4 sellout was, in the end, underpinned by these chipmakers' "buy-ahead" demand. Supply and demand pull each other and grow the market together.

So what changes

If you watch the chip/AI industry — this signals that the AI boom's spoils are spreading broadly from GPUs to memory. As AI compute grows, HBM demand automatically follows, and that market is supply-short with fat margins right now. When you look at AI infrastructure, don't watch only the GPU — watch the HBM supply chain beside it (Samsung, SK Hynix, Micron) to complete the picture.

If you're an investor or market watcher — the HBM4 sellout and 110-trillion-won spend are strong boom signals, but don't forget memory is cyclical by nature. Demand is outrunning supply now, favoring Samsung. But everyone piling into capacity today could be the seed of oversupply a few years out. "Does AI demand hold this pace?" is the key variable.

If you're a company building AI infrastructure — the direct effect is slow, but the direction is friendly. More HBM output eases the AI accelerator bottleneck and, over time, works to lower AI compute costs. If you're planning data centers or AI infrastructure now, factor in how expanding HBM supply may affect future availability and pricing.

🥄 Three Things You're Probably Wondering

— Are you saying HBM matters more than GPUs? No. The GPU (compute) and HBM (memory) are one unit — you need both for AI to run. It's just that attention has been on GPUs, and these numbers show HBM is equally a bottleneck and a beneficiary. Think of it as "the hidden gold mine next to the GPU."

— If it's sold out, isn't that just unambiguously good? Good for now — demand outrunning supply means fat margins. But memory is cyclical, and if everyone expands, oversupply could bend prices a few years out. Read "today's boom" alongside "tomorrow's cycle" for a balanced view.

— So has Samsung beaten SK Hynix? Too early to say. A strong HBM4 start is clear, but SK Hynix is countering with its earlier-generation relationships and know-how. Pre-empting a generation shift is high ground, not a finished match. The next few quarters are the real watershed.

References

• Samsung chairman reviews HBM4 supply as revenue tops US$1 billion — DigiTimes
• Samsung's HBM4 Sold Out for Year, to Exceed Half of HBM Revenue from Q3 — Seoul Economic Daily
• Samsung Reportedly Plans 50% HBM Capacity Surge in 2026 — TrendForce
• Samsung Electronics Announces First Quarter 2026 Results — Samsung Global Newsroom
• Samsung HBM4 Revenue and the $9.7B Annual Era — BigGo Finance

Numbers are as of announcement/forecast and may change. Investment calls are yours to make!