A Korean Semiconductor Engineer Explains: Why HBM4 Is the Real Inflection Point

 

If you've been tracking AI hardware headlines, you've probably seen "HBM4" mentioned as just the next number after HBM3E. It isn't. HBM4 rebuilds the interface and the base die from the ground up, and it just became the default memory standard for Nvidia's next GPU platform. At the same time, SK hynix and its rivals have already said a few concrete things about HBM5 — the generation after this one — in public roadmaps and conference talks. Here's what's confirmed, what's still an analyst's guess, and why the gap between the two matters.

So Why Is HBM4 a Structural Shift, Not Just a Speed Bump?

HBM4 doubles the interface width from HBM3E's 1,024 bits to 2,048 bits. Bandwidth also depends on the per-pin data rate, so a generation name alone does not define one universal bandwidth figure. Samsung lists its HBM4 at up to 13 Gbps and 3,300 GB/s, or approximately 3.3 TB/s per stack. This is a vendor product specification for HBM4, not a figure that should automatically be attributed to HBM4E, and it is not a measured application throughput result.

The bigger change is underneath the interface: the base die that connects the memory stack to the processor. Foundry logic technology allows vendors to add functionality, but HBM4 does not imply one universal 3–5nm process. Samsung specifies 4nm foundry technology for its HBM4. SK hynix announced that it would adopt TSMC logic technology for its HBM4 base die, without naming a process node in that announcement. The process choice should therefore be attributed to the particular vendor and product, not treated as a shared HBM4 specification.

This redesign is also why HBM4 has become the practical starting point for Nvidia's next-generation "Vera Rubin" platform. Rubin adopted HBM4 as its baseline memory from the outset, and that decision reportedly pushed both Samsung and SK hynix to move up their HBM4 mass-production timelines to as early as February 2026. SK hynix delivered finalized HBM4 samples to Nvidia in early January 2026, and the two companies' relationship was visible enough that SK Group chairman Chey Tae-won made his first-ever appearance at Nvidia's GTC event in March 2026 to meet CEO Jensen Huang in person, alongside SK hynix's exhibit of custom HBM structures and HBM3E/HBM4 hardware.

So Where Does SK hynix Actually Stand on HBM4 Right Now?

According to the company's own second-quarter 2026 earnings disclosure, SK hynix began mass-production shipments of HBM4 in Q2 2026 and has flagged a broader ramp-up in the second half of the year. The current production reality, per that same disclosure:

• 12-Hi HBM4 is already in mass production.
• 16-Hi HBM4 (48GB) is in the customer qualification stage — not yet shipping at volume, but being validated with customers.
• HBM4E samples were completed and delivered in the first half of 2026, with the company describing its process choice as balancing "technology maturity and manufacturing stability."
• SK hynix says HBM4 delivers the speed customers need while also hitting industry-leading power efficiency and cost competitiveness — the company's own characterization, not an independently verified benchmark.

The push toward 16-Hi stacks isn't just SK hynix's initiative — Nvidia has been pressing for early adoption of 16-layer HBM4 across its accelerator roadmap, which has put Samsung, SK hynix, and Micron all in a race to qualify 16-Hi production.

On supply share, this is where the record shifts from company statements to outside estimates: UBS projected SK hynix would capture roughly 70% of the HBM4 supply for Nvidia's Rubin platform in 2026, while TrendForce separately estimated SK hynix would supply about two-thirds of Nvidia's HBM4 demand.

Counterpoint Research, using a different measure, put SK hynix's overall HBM market share at 58% in Q1 2026 based on shipment volume. These are three different research firms using different methodologies and time windows, not a single confirmed number — but they point in the same direction.

SK hynix itself has framed 2026 as a "memory supercycle led by HBM" in its own market outlook materials, and on its Q2 earnings call the company said HBM4 volume expansion combined with increased shipments of 1c-nanometer conventional DRAM should push second-half bit growth above first-half levels.

What's Actually Public About HBM5, Then?

This is the part where the distinction between "company roadmap" and "analyst forecast" matters most, because the two get blended together in a lot of coverage.

So What Has SK hynix Actually Confirmed About Its Own Roadmap?

At SK AI Summit 2025 in November 2025, the company laid out its "Full Stack AI Memory Creator" vision and a next-generation memory roadmap built around custom HBM, AI DRAM, and AI NAND. The official newsroom writeup itself doesn't name HBM5 or HBM5E explicitly. However, outlets covering the event's presentation slides directly — including TrendForce — reported that SK hynix's roadmap slide showed a two-phase structure: 2026–2028, followed by 2029–2031, with standard and custom HBM5/HBM5E development placed in that second, later window.

Then Where Does That "2029" Date Actually Come From?

So the claim "HBM5 sits in SK hynix's 2029–2031 roadmap window" is grounded in the company's own presented slide. A specific, singular "HBM5 launches in 2029" statement from the company itself is a different claim — and that one hasn't turned up in the company's own materials.

Where that specific year comes from: a widely circulated headline framed it as "SK Hynix to Launch HBM5 in 2029," but the substance of that reporting traces back to a forecast from Counterpoint Research, an analyst firm — not a company announcement. The roadmap window (2029–2031) and the analyst's point estimate (around 2029–2030) point in a similar direction, but they're not the same kind of source, and conflating them risks treating a forecast as a confirmed launch date.

What Has the Company Said About the Packaging Technology Itself?

At Hot Chips 2026 in August 2026, Jaesik Lee, VP of package engineering at SK hynix America, said directly that hybrid bonding — a technique that joins copper layers directly, without solder bumps — won't be ready in time for HBM4E, but is targeted to arrive starting with HBM5. Multiple outlets covering the talk cited Lee's comment that logic wafer thickness (775 microns) is currently the practical ceiling the company can build to, and noted that JEDEC recently raised its maximum HBM stack-height standard from 720 to 775 microns. Separately, Counterpoint Research has forecast that hybrid bonding will see meaningful production use around the same 2029–2030 window it associates with HBM5 — again, an analyst projection layered on top of the company's own technology statement.

What Are Samsung and Micron Saying About Their Own Next Generation?

SK hynix isn't the only company talking about HBM5 in public. Samsung Electronics unveiled an HBM5 mock-up — its 8th-generation HBM — at Computex in June 2026. Samsung DS Division president and CTO Jae-hyuk Song described a "chimney-style" structure adding extra thermal pathways between chips to reduce thermal resistance and improve operating stability.

Samsung is reportedly planning to use its own 2nm foundry process for the HBM5 base die, preparing 12-, 16-, and 20-layer configurations, with mass production targeted around 2028 — a year that, notably, comes in earlier than the 2029–2031 window associated with SK hynix's roadmap. Samsung followed up at Hot Chips 2026 with an even further-out concept called "zHBM," which stacks memory directly on top of the GPU, claiming performance and efficiency gains well beyond standard HBM5.

Micron, for its part, hasn't published a specific HBM5 roadmap or launch timeline in the period covered by this research. At Hot Chips 2026, Micron fellow Raghu Sreeramaneni instead spoke about a different problem: the silicon (wafer) area penalty HBM carries relative to DDR5 keeps growing with each generation, currently sitting at roughly triple DDR5's usage for the same capacity — a comment about HBM's underlying cost structure industry-wide, not a Micron HBM5 announcement.

So Is There an Official HBM5 Standard Timeline?

JEDEC's own website wasn't directly accessible for this research (it returned an access error), so the organization's official schedule for a formal HBM5 standard couldn't be independently confirmed. What's known is limited to secondary reporting: JESD270-4 (the HBM4 standard) remains JEDEC's most recent published HBM standard, and industry discussion references bandwidth roadmaps extending toward HBM5 through HBM8, but a confirmed publication date for an HBM5 standard doesn't appear in what's currently public.

But What About the Earnings Backdrop?

None of this technology roadmap unfolds in a vacuum — it sits against a backdrop of extraordinary, and extraordinarily scrutinized, financial results.

SK hynix reported second-quarter 2026 revenue of 79.3187 trillion won and operating profit of 60.5426 trillion won — a 76% operating margin — in what the company called its best quarterly results ever. That's up 257% in revenue and 557% in operating profit year-over-year, and up 51% and 61% respectively from the prior quarter. First-half cumulative revenue crossed the 100-trillion-won mark for the first time.

The company described its plan going forward in fairly general terms: continuing to prepare for mid- to long-term growth opportunities while maintaining capex discipline to strengthen both production capacity and financial health — direction-setting language rather than a specific numeric guidance.

Here's the part that looks contradictory at first glance: despite posting record results, SK hynix's stock fell, because the numbers still came in below what the market had priced in, according to CNBC's coverage of the July 29, 2026 earnings release.

That gap between "record results" and "stock falls anyway" becomes easier to understand once you look at how far analyst expectations had already run up over the course of 2026. A sample of brokerage forecasts for full-year 2026 operating profit, tracked by announcement date:

Date	Firm	2026 Operating Profit Estimate	Price Target
Jan 28, 2026	KB Securities	132 trillion won	1.2 million won
Mar 3, 2026	Kiwoom Securities	170 trillion won (revenue: 230 trillion won)	1.3 million won
Mar 13, 2026	KB Securities (revised up)	177 trillion won	1.7 million won
May 20, 2026	Korea Investment & Securities	Not disclosed	3.8 million won
May 22, 2026	NH Investment & Securities	Not disclosed	3.1 million won
May 23, 2026	Nomura Securities	Not disclosed	4.0 million won (reported as the highest of any domestic or foreign brokerage at the time)
Jun 22, 2026	iM Securities	261 trillion won (assumes a 10% employee bonus payout; +452% year-over-year)	3.5 million won (revised up from 2.76 million won)

The spread here is large — full-year operating profit estimates ranging from 132 trillion to 261 trillion won, and price targets from 1.2 million to 4.0 million won — and it's worth being precise about why: this isn't one "market consensus" number, it's a series of individual forecasts published at different points in the year, with the range shifting upward as the year progressed and as individual firms adjusted their assumptions (bonus payouts, DRAM/NAND pricing trajectories, and so on).

iM Securities' own single-quarter estimate for Q2 2026 operating profit, published in June, was 63.7 trillion won — versus the actual reported figure of 60.5426 trillion won, illustrating just how much even a well-resourced forecast can miss by. On its earnings call, SK hynix itself pointed to rising DRAM shipment volumes and improving product mix pushing average selling prices higher as reasons it expects second-half results to strengthen further from the first half.

Put together, the picture that emerges from public earnings data and public analyst notes is less "the company disappointed" and more "the bar kept rising faster than even record results could clear."

So Where Is This Actually Heading?

Watching this space for a while, what stands out isn't any single number — it's the shape of the transition. A jump from 1024-bit to 2048-bit interfaces, a base die that moves onto logic process nodes, and now a packaging method (hybrid bonding) that's being pushed out to the following generation because the physical stack has run into a literal thickness ceiling — that's not routine product cycling. It's the kind of shift where the underlying manufacturing approach changes, not just the spec sheet.

What also stands out is how much of the HBM5 conversation right now is still roadmap and lab-talk rather than shipping product. SK hynix, Samsung, and Micron have each said something about the generation after HBM4, but what they've said differs in kind: a roadmap slide with a multi-year window, a packaging technique confirmed for a future generation without a hard date, a mock-up at a trade show, and in Micron's case, commentary about industry-wide cost structure rather than a roadmap at all. None of that is a criticism — it's simply where things are in 2026, this many years out from an actual HBM5 launch. The gap between "this is on our roadmap" and "this ships next quarter" is one worth keeping in mind whenever a headline compresses the two into a single date.

Wrapping Up

HBM4 is the generation where the base die moved to logic process, the interface doubled in width, and Nvidia built its next GPU platform around it as the default. SK hynix has confirmed mass-production shipments started in Q2 2026, with 12-Hi in production and 16-Hi in customer qualification, while outside estimates put its share of Nvidia's HBM4 supply somewhere around two-thirds. HBM5, by contrast, is still mostly a roadmap item: SK hynix has publicly placed it in a 2029–2031 window and confirmed hybrid bonding is targeted to begin there, but the specific "2029" year attached to HBM5 in a lot of coverage traces back to analyst estimates, not a company announcement — and Samsung's own HBM5 mock-up points to an earlier 2028 production target from a different manufacturer entirely. On the earnings side, SK hynix just posted its best quarter ever and still saw its stock fall, a reminder that in this cycle, the bar for "good enough" has been rising just about as fast as the results themselves.