Tight. 3 of 4 warning signs are red, 1 is unknown. Confidence: medium.
Many leading data-center AI accelerators depend on HBM, and only three companies make it at scale. Micron has already contracted most of its 2027 HBM supply at significantly higher prices, and SK hynix is negotiating 2027 volumes and prices now. TrendForce, a memory market tracker, expects supply to stay tight through 2027.
Outlook: tight through 2027 on TrendForce's forecast. New fabs begin output from mid-2027, but Micron says new-fab output only "gradually becomes more meaningful starting a few quarters after initial output." How long the tightness lasts depends on qualified HBM output and on demand, not just fab openings.
New here? Skip to Background below for the basics. Coming back? Start with What changed.
This month
October 2026 · Updated Oct 4 · Next update after Samsung and SK hynix Q3 results
What changed since last month
First reading. From next month this section leads with what moved and why.
Scorecard
Warning sign | What we saw | Reading | Source |
Long waits | Micron: "We have completed agreements for the vast majority of our calendar 2027 HBM bit supply." Its 2027 HBM is effectively sold a year ahead. SK hynix says talks on 2027 HBM volume and pricing are under way. | 🔴 Yes | |
Orders above supply | TrendForce expects memory supply to stay tight through 2027. It says GPU and custom-chip makers are considering using less HBM per chip, moving from 12-high to 8-high stacks, to manage cost and limited supply. | 🔴 Yes | |
Prices going up | Micron signed the vast majority of its 2027 HBM supply "with significant price increases year over year." That covers most of a major maker's book, not a single deal. TrendForce separately forecasts the 2027 blended average HBM price up 121%. That's a forecast, and a blended average also rises when buyers shift to pricier generations like HBM4, so it isn't the same product doubling in price. | 🔴 Yes | |
New capacity not enough | Not established for HBM specifically. Micron expects memory and storage overall to be "much tighter in fiscal 2027 and 2028 than they were in 2026," but that covers all memory, not HBM after planned expansion. No maker or tracker gives an HBM-specific view of whether new capacity will fall short. | ⚪ Unknown |
Score: 3 red, 0 green, 0 mixed, 1 unknown = 🟠 Tight. Severe needs all 4 confirmed red. One unknown, so confidence is medium. HBM-specific evidence that supply will still fall short after expansion would turn the fourth sign red.
Dates to watch
Oct 8: Samsung preliminary Q3 results. A first look at memory revenue and profit.
Oct 27 (expected): SK hynix Q3 results. Listen for 2027 HBM contract terms and HBM4 volumes.
Oct 29 (expected): Samsung full Q3 results and call. Listen for HBM4 progress with major customers and new capacity.
Dec 23 (expected): Micron fiscal Q1 2027 results. Listen for HBM output from new capacity and 2027 pricing.
"Expected" dates come from earnings calendars. The company hasn't announced them yet.
Status history
- Oct 2026: 🟠 Tight (3 red, 1 unknown), medium confidence. First reading.
Background
Reference. Updated when the facts change, not every month.
What it is
HBM (High Bandwidth Memory) is a tower of memory chips stacked on top of each other and wired straight through with tiny vertical connections. It sits right next to the AI chip so data can move fast. A modern AI chip carries several of these towers.
Where it sits: Memory chips made on wafers (SK hynix, Samsung, Micron) → stacked into an HBM tower → bonded next to the AI chip by TSMC (CoWoS) → server → data center
Many leading data-center AI accelerators, such as NVIDIA's and AMD's top GPUs, depend on HBM. Some smaller AI chips use other memory instead; NVIDIA's L4, for example, uses GDDR6. For the chips that need it, the chip is only as fast as the memory feeding it, so an HBM shortage holds back the most expensive chips.
When something is short, two things happen. The seller can charge more and stays fully booked. The buyer has orders it can't fill, or pays more. Same shortage, opposite effect on the stock.
Why can't they just build more?
- It eats normal memory capacity. Micron says HBM uses roughly three times the wafer area of regular DDR5 memory for the same capacity, and the gap widens with each generation. Every HBM wafer means less memory for phones, PCs and servers.
- The towers keep getting taller. The HBM4 standard allows stacks of up to 16 chips. Taller stacks are harder to build.
- New memory fabs take years. Micron's Idaho fabs start wafer output in mid-2027 and late 2028, Japan in late 2028, and New York in 2030. SK hynix's Yongin cleanroom opens in early 2027.
- Only three companies make it at scale: SK hynix, Samsung and Micron.
How big is it, and who makes it?
HBM revenue share, Q2 2026 (Counterpoint Research). Shares add to 101% because of rounding.
Maker | Share | Note |
SK hynix | 50% | The market leader. Began HBM4 mass production shipments in Q2 2026. |
Samsung | 33% | Gaining fast, up from 21% in Q1 2026. |
Micron | 18% | The only US maker. Most of its 2027 HBM supply is already under contract. |
What it means: with only three suppliers, each one holds pricing power. Samsung's rise adds supply but hasn't loosened the market yet.
Who wins, who waits
The effects below are our interpretation, not company statements.
Company | Likely effect | Why |
Wins | Half the market by revenue. 2027 prices are still being negotiated, but in a tight market. | |
Wins | Has locked in most of its 2027 HBM at significantly higher prices. Separately, its 26 strategic customer agreements carry $32B of customer financial commitments, mostly cash deposits. That figure covers all products, not HBM alone. | |
Samsung (005930.KS) | Wins | Taking share in HBM, and also benefits as regular memory gets scarcer. |
Pays more | A major HBM buyer. Its supply and capacity commitments rose from $119B to $279B in one quarter, "primarily related to the procurement of memory." That's all supply commitments, not an HBM-only total. Higher memory prices raise its costs. | |
Pays more | Their AI chips need HBM too, and they compete for the same limited supply. | |
SMCI, Dell, HPE | Pays more | Server builders. Regular server memory is getting tighter as capacity shifts to HBM. |
Phone and PC makers | Pays more | Less memory capacity is left for consumer devices, which points to higher costs or thinner margins. |
The takeaway: HBM supply and contract pricing are tight. How long that lasts depends on qualified HBM output and on demand, not simply on announced fab openings.
The companies behind HBM
- Stacking machines: Hanmi Semiconductor (042700.KS) makes the bonders that press the memory chips into a tower. SK hynix has ordered them for its HBM4 ramp.
- The base chip: for HBM4, the bottom chip of each tower is a logic chip. SK hynix partnered with TSM to make its HBM4 base die on TSMC's logic process. Samsung is reported to make its own.
- Testing and grinding: Advantest (6857.T) tests memory chips. Disco (6146.T) makes the machines that thin wafers so they can be stacked.
When does it end, and what could replace it?
- Early 2027: Micron's Singapore HBM packaging facility starts initial output. SK hynix's Yongin cleanroom opens. A cleanroom opening isn't output: tools still have to be installed and the line ramped.
- Through 2027: TrendForce expects supply to stay tight. Chip designers are considering shorter 8-high stacks to stretch supply.
- Mid-2027 to 2028: Micron's Idaho fabs start wafer output (mid-2027 and late 2028), and its Japan fab in late 2028. Micron says new-fab output becomes meaningful only a few quarters after it starts. Finished HBM then still has to be stacked, tested and qualified by customers before it eases the shortage.
- The risk after that: if new capacity lands while demand cools, prices could fall. Memory has historically been cyclical, with booms followed by price drops.
- No close substitute. Nothing else delivers the same speed next to the chip, so the fix is more supply, not a different product.
How to read a change
Good easing
Prices level off because new plants came online, while AI chip orders stay strong. Chip designers get relief and memory makers still sell more.
Bad easing
Prices fall because AI customers ordered less. In past memory downturns, memory stocks have usually fallen early and hard.
What would change our view
- Gets tighter: a maker reports HBM4 yield problems, or next-generation AI chips need more HBM per unit.
- Gets easier (good): Samsung adds major HBM4 volume with large customers, while orders keep growing.
- Gets easier (bad): memory makers mention order cuts or rising stockpiles at customers.
Past scorecards
Each month the previous scorecard moves here, newest first.
- None yet. October 2026 is the first reading.
Evidence and sources
Company statements
- Micron fiscal Q4 2026 prepared remarks (Sep 30, 2026): "We have completed agreements for the vast majority of our calendar 2027 HBM bit supply with significant price increases year over year"; supply-demand "much tighter in fiscal 2027 and 2028 than they were in 2026"; Singapore HBM advanced packaging initial output early 2027; ID1 wafer output mid-2027, ID2 late 2028, Japan late 2028, New York 2030; customer financial commitments $32B, mostly cash deposits. Prepared remarks
- SK hynix Q2 2026 call (Jul 28, 2026): "Discussion is underway for 2027 HBM supply volume and pricing with our key customers"; HBM4 mass production shipments began in Q2; Yongin Phase 1 cleanroom early 2027. Transcript
- NVIDIA CFO commentary, Q2 fiscal 2027: commitments rose from $119B to $279B, "primarily related to the procurement of memory." CFO commentary
- Micron at Hot Chips (Aug 2026): HBM uses roughly three times the wafer area of DDR5 for the same capacity. Tom's Hardware
Market data and forecasts
- TrendForce (Sep 29, 2026): 2027 blended HBM average price forecast +121% year over year; memory supply tight through 2027; GPU and custom-chip makers considering less HBM per device (12-high to 8-high). TrendForce
- Counterpoint Q2 2026 HBM revenue shares: SK hynix 50%, Samsung 33% (from 21% in Q1), Micron 18%; total 101% due to rounding. Counterpoint · Seoul Economic Daily, Sep 3
Background
- HBM4 standard up to 16-high stacks. JEDEC via Tom's Hardware
- SK hynix HBM4 TC bonder orders from Hanmi. TrendForce, Jun 9
- SK hynix and TSMC partnership (Apr 19, 2024): HBM4 base die on TSMC's advanced logic process. SK hynix. Samsung in-house base die (reported). BigGo Finance
- NVIDIA L4 uses GDDR6 memory, an example of an AI chip without HBM. NVIDIA L4 product brief
- Micron on new fabs: production "gradually becomes more meaningful starting a few quarters after initial output"; the $32B covers its 26 signed strategic customer agreements. Prepared remarks
- Where sources disagree: the bull case says AI has made memory structurally scarce. The bear case says new fabs arriving in 2027–2028 could meet demand just as orders cool, as in past memory cycles.
Words used on this page
- HBM (High Bandwidth Memory): a tower of stacked memory chips placed right next to an AI chip.
- HBM4: the newest generation, with taller towers and faster speeds.
- DRAM (Dynamic Random-Access Memory): the main type of working memory in computers. HBM is made from DRAM chips.
- DDR5: the current standard type of regular computer memory.
- GDDR6: a graphics memory type used on some smaller AI and graphics chips instead of HBM.
- Qualified: approved by the customer after testing, so the part can go into its products.
- 8-high / 12-high / 16-high: how many memory chips are in one tower.
- TSV (Through-Silicon Via): tiny vertical wires drilled through each chip to connect the tower.
- Bonder (TC bonder, thermo-compression bonder): the machine that presses the chips together into a tower.
- Base die: the logic chip at the bottom of an HBM4 tower that controls it.
- Bit supply: the total amount of memory capacity a maker ships, measured in bits.
- Blended price (ASP, Average Selling Price): the average price across all products and generations sold.
- Wafer: the round silicon disc chips are made on.
- Cleanroom: the dust-free factory space where chips are made.
Invest Deeper · Educational research, not investment advice.
Made with Bullet