HBM Yield and Supply-Chain Bottlenecks Explained

A single HBM stack can pile 12 memory dies on top of each other. A defect in one layer, one connection or the final package can throw away the whole high-value unit.

Watch it play outOne defect anywhere can scrap the whole stack

An HBM stack can pile 12 memory dies on a base die, joined in a final package.

Tiny vertical wires, through-silicon vias, run through every layer.

A defect in one layer, one connection or the final package...

...can throw away the whole high-value unit, every good die included.

Replay

That is why HBM supply isn't set by DRAM wafer starts. Usable output is capped by the weakest link in a chain that runs from the memory fab to a customer's sign-off, so more announced capacity doesn't immediately mean more memory shipping inside AI chips. A short calculation below shows why each extra layer makes the job harder, and why testing dies before stacking changes the answer so much.

Watch it play outUsable HBM output is capped by the weakest link

HBM supply is often counted from DRAM wafer starts.

But every wafer has to pass a chain that runs from the memory fab to a customer's sign-off.

Usable output is capped by the weakest link in that chain.

So more announced capacity doesn't immediately mean more memory shipping inside AI chips.

Replay

The short version

• "Yield" for HBM is at least five different numbers, from raw die yield to customer-qualified yield. Companies rarely disclose all of them.
• Samsung describes 4-high, 8-high and 12-high stacks. Taller stacks add capacity and bandwidth but raise heat and manufacturing difficulty.

Watch it play outTaller stacks add capacity, and heat and difficulty

Samsung describes common 4-high, 8-high and 12-high stacks.

Each extra layer adds capacity and bandwidth...

...and raises heat and manufacturing difficulty. Every extra layer is another chance of failure.

Replay

• Micron shipped its HBM4 12-high product in volume while still sampling the 16-high version, and said it expected mature HBM4 yields faster than it reached them for HBM3E.
• "Sold out" describes contracted demand, often for future periods. It is not the same as memory shipping today.

If you're new to what HBM does inside an AI accelerator, start with our HBM explainer. This piece covers why supply is hard to grow.

Check the claimThree claims to test as you read

ClaimHBM supply grows in step with DRAM wafer starts.Tap to checkMythUsable output is capped by the weakest link, from the memory fab to a customer's sign-off.Flip backClaimMemory makers publish their HBM yield.Tap to checkMythThey rarely disclose a full yield waterfall; investors piece it together from commentary.Flip backClaimA taller stack is simply a better product.Tap to checkPartly trueIt adds capacity and bandwidth, but every extra layer adds heat and another chance of failure.Flip back

What does HBM yield mean?

Yield is the share of manufactured units that meet the required electrical, thermal, performance and reliability specs. For HBM, first ask which yield someone is talking about:

• DRAM die yield: usable memory dies coming off the wafer
• Stacking yield: good output after several dies are aligned and bonded
• Base-die and interface yield: usable control logic and connections
• Package yield: good output after the accelerator, HBM stacks and interconnect are joined
• Qualified yield: finished output that passes a customer's platform tests

A supplier can improve one stage while another stays stuck. Treat any comment about "mature yield" in its exact context.

HBM yield is at least five different numbers

1. Die yield (Off the wafer) to Stacking yield (Aligned and bonded)
2. Stacking yield (Aligned and bonded) to Base die (Logic and interface)
3. Base die (Logic and interface) to Package yield (Joined to the chip)
4. Package yield (Joined to the chip) to Qualified yield (Passes customer tests)

Why stacking makes the math harder

HBM stacks DRAM dies vertically and links them with through-silicon vias, or TSVs: tiny vertical wires drilled through each die. Samsung describes common 4-high, 8-high and 12-high configurations and notes that extra layers raise capacity and bandwidth along with thermal and manufacturing challenges. See Samsung's HBM technical overview.

Every component that has to pass adds another chance of failure. The real process is gentler than multiplying die yields together: manufacturers test at several stages, repair some defects, pick known-good dies and improve over time. The intuition still holds. A valuable stack can be lost to a flaw in one layer, one interconnect or the later package assembly.

Worked example: how stack height eats yield

Hypothetical example with round numbers: each DRAM die has a 95% chance of being good, and each bonding step that adds a layer succeeds 99% of the time. When every part must work, you multiply the chances together.

Make a guess

Each untested DRAM die has a 95% chance of being good. What share of 12-high stacks built from them work?

Tap your guess

0%Your guess: 0%10%Your guess: 10%20%Your guess: 20%30%Your guess: 30%40%Your guess: 40%50%Your guess: 50%60%Your guess: 60%70%Your guess: 70%80%Your guess: 80%90%Your guess: 90%100%Your guess: 100%

01234567890123456789%Actual

Spot on.Close, a little low.Close, a little high.Way off, too low.Way off, too high. It was 54%.

Every die must work, so the chances multiply (0.95¹²). Testing dies before stacking lifts the result to about 89%.

Actual54%

With tested known-good dies89%

Height compounds losses; testing claws them back

Pick a stack height and whether dies are tested first

Stack height

4-high8-high12-high

Dies

UntestedKnown-good

Stacks that work81%96%66%92%54%89%

54%96%

95% per die99% per bond

Now carry the tested 12-high stack through two more hypothetical stages: a base die that passes 98% of the time and package assembly at 97%.

A tested 12-high stack still loses 4 points after stacking

After stacking88.6%

Base die (98%)−1.7 pts= 86.9%

Packaging (97%)−2.6 pts= 84.3%

Stacks still good84.3%

Three lessons for investors:

• Height compounds losses. Going from 4-high to 12-high without testing drops yield from about 81% to about 54% in this example.
• Testing is a big lever. Screening dies before stacking lifts the 12-high result from about 54% to about 89%. That's why "known-good die" work and repair matter so much.
• Late failures are expensive. A stack that fails at packaging has already used 12 good dies and every step before it.

The five stages of the chain

A few details matter for investors:

• HBM can use more wafer capacity per useful bit than ordinary DRAM, so a mix shift toward HBM tightens the wider memory market.
• The HBM4 base die adds another technology and another supplier dependency.
• Memory sitting in inventory can't become accelerator revenue while packaging capacity is unavailable.

Samples, production and shipments are different things

Sampling, qualification, volume production and volume shipments are separate milestones. A supplier can make HBM well before it contributes meaningful qualified volume.

Micron's fiscal second-quarter 2026 remarks show why. It described volume shipments of its HBM4 12-high product, separately described sampling its 16-high version, and said it expected to reach mature HBM4 yields faster than it had for HBM3E. Generation, stack height, yield stage and customer status all have to be tracked on their own. See Micron's prepared remarks.

Read closely

Which phrase describes HBM shipping in revenue volume today?

Micron, fiscal Q2 2026 prepared remarks

Micron described volume shipments of its HBM4 12-high product, separately described sampling its 16-high version, and said it expected to reach mature HBM4 yields faster than for HBM3E.

Tap the phrase you think is the tell.

Spotted it.Not that phrase. Generation, stack height, yield stage and customer status all have to be tracked on their own.

• “Micron described volume shipments of its HBM4 12-high product” Volume shipments are the milestone that turns into revenue.
• “separately described sampling its 16-high version” Samples reached customers; qualification and volume come later.
• “and said it expected to reach mature HBM4 yields faster than for HBM3E” An expectation about process progress, not usable volume.

Explain every phraseHide the other explanations

Member research: The same chain-of-yields logic governs the optical parts that link AI clusters together. Issue 06 works through yields, reliability and cash conversion at five photonics companies.

Why "sold out" doesn't mean shipping now

Supply agreements and pricing commitments improve visibility, but they don't remove execution risk. Contracted demand can refer to future periods, specific products, or capacity that still has to ramp and qualify.

Contracted is not shipped

"Sold out"Contracted demand

• Often for future periods
• Specific products
• Capacity that still has to ramp and qualify

Visibility, not volume

vs

Shipping nowQualified supply

• Passed customer qualification
• Packaged beside an accelerator
• High-volume shipments

Revenue today

Micron has described multi-year supply commitments that can include HBM, and it has separately reported HBM4 high-volume shipments. Both are useful signals of demand and execution. Neither should be collapsed into a single current-volume number. See its fiscal third-quarter 2026 results.

What each signal proves

Watch inventory and capital spending next to these milestones. A fast ramp can tie up working capital before revenue arrives, and poor yields can soak up capacity without producing matching saleable output.

The supply-side risks

• Generation transitions. A supplier that executes HBM3E well can stumble on HBM4. New interfaces, base dies, stack heights, thermals and customer platforms reset part of the learning curve.
• Customer concentration. Qualification is often tied to a few accelerator platforms. One platform delay can hit memory, packaging, networking and server supply chains at once.
• Over-ordering. Customers facing scarcity may reserve more than they end up using. Contract terms, deposits and cancellation clauses matter.
• Margin normalization. Scarcity supports pricing until qualified supply catches up or demand shifts. HBM can stay strategically important while supplier returns become less exceptional. Memory has long been a cyclical business for exactly this reason.

Scarcity pricing invites the supply that ends it

1. Scarce qualified HBM to Exceptional margins
2. Exceptional margins to Qualified supply catches up
3. Qualified supply catches up to Margins normalize

Try it yourself

A 10-minute exercise using the primary sources above:

1. Open Micron's fiscal second-quarter 2026 prepared remarks and search for "HBM".
2. For each statement, write down the generation (HBM3E or HBM4), the stack height, and the milestone: sample, qualification, production or shipment.
3. Do the same with the fiscal third-quarter 2026 results, and mark anything that is a supply commitment rather than a shipment.
4. Match each statement to a row in the "What each signal proves" table, and note what it can't prove.
5. Open the HBM demand calculator and see how many HBM stacks one accelerator deployment implies.

You should now be able to answer: how much of what the company has said describes qualified HBM shipping today, and how much is still a sample, a plan or a commitment?

Our read

HBM yield is a chain, not one percentage. The output that counts is qualified, packaged memory attached to an accelerator a customer can deploy. Investors who track wafer starts without stacking, packaging and qualification will mistake announced capacity for usable supply.

Here is the sequence we use:

1. Pin down the exact HBM generation and stack height.
2. Separate samples, qualification, production and shipments.
3. Check whether packaging capacity is growing alongside memory capacity.
4. Compare supply commitments with the delivery periods disclosed.
5. Track inventory, gross margin, capital spending and customer concentration.
6. Stress-test accelerator demand and efficiency assumptions.

The bottleneck is easing when packaging grows in step with memory and HBM4 qualifications land without delays. It is getting worse when inventory and capex climb faster than qualified shipments. To translate an accelerator fleet into memory capacity and bandwidth, use the HBM demand calculator.

Questions readers ask

Why don't companies just publish their HBM yield?

Public companies rarely disclose a full yield waterfall. Investors have to piece it together from commentary such as "mature yield," which is only useful once you know which stage, generation and stack height it refers to.

Is a 16-high stack better than a 12-high one?

It holds more memory and can deliver more bandwidth, but every extra layer adds heat and another chance of failure, as the worked example shows. Micron was shipping its 12-high HBM4 in volume while still sampling the 16-high version, so the taller product was at an earlier milestone.

If HBM is sold out, how can supply still disappoint?

"Sold out" usually describes contracted demand, often for future periods. That capacity still has to ramp, pass qualification and get packaged. Customers facing scarcity may also reserve more than they end up using.

Does better yield mean higher profit?

Usually, for the company that achieves it. Better yield means more saleable stacks from the same capacity, which lowers the cost of each good unit. Across the industry, though, better yields add supply, and more supply is what eventually pulls scarcity pricing down.

Go deeper with member research

Yield is where memory profits are won or lost. Member case studies apply this to the companies making and equipping HBM:

• The HBM Capacity Trade: Micron’s HBM capacity thesis as the evidence stood on December 18, 2025, with timing, valuation and later results kept separate.
• Lam Research and the economics of harder-to-build memory: NAND conversions, manufacturing complexity and cash conversion as of January 30, 2026.

Limits of this data

• Public companies rarely disclose a full yield waterfall, so stage-by-stage yields have to be inferred from commentary.
• Supplier roadmaps, qualifications, yields and customer commitments change. Check them against current filings and technical disclosures.