The UBS HBM Chart Says More Than "Shortage"
A UBS chart titled "Adjusted HBM supply vs demand" is making the rounds with the tagline that 2027 will be a brutal year for HBM. Six bars, three years. Read against the gridlines, the chart says something more specific than "shortage": 2027 is the first year in which end-consumption alone exceeds bit shipments, before any inventory is counted.
This post converts the bar heights into numbers, works out what the red inventory box means, and fills in from public sources why consumption jumps and supply does not keep pace. Every figure from the chart is a read-off, and the post says so where it matters.
KEY TAKEAWAYS
1. Read from the chart, HBM bit shipments run 21.4 / 36.6 / 53.6 billion Gb for 2025-2027E and end-consumption 17.6 / 26.1 / 59.7. In 2027 consumption alone is 6.1 billion Gb above shipments.
2. Adding the demand-side inventory box takes 2027 total demand to 76.4 billion Gb, a 43% shortfall against shipments. Through 2026 supply exceeded total demand (+2.9 billion Gb in 2026).
3. In growth terms, 2027 consumption rises 129% against shipments of 46%. Kiwoom estimates HBM wafer capacity grows 49% and CoWoS 25% the same year. Capacity, not demand, sets the ceiling.
Turning the chart into numbers
UBS labels the axis in "Mil Gb", millions of gigabits. Reading the bar tops against the 0-90,000 gridlines and converting to billions of Gb gives the table below. These are read-offs, not UBS's published figures, so allow a few percent either way.
| Year | Shipments | End-consumption | Inventory (demand side) | Total demand | Supply minus total | Supply minus consumption |
|---|---|---|---|---|---|---|
| 2025 | 21.4 | 17.6 | 4.8 | 22.4 | -1.0 | +3.8 |
| 2026E | 36.6 | 26.1 | 7.6 | 33.7 | +2.9 | +10.5 |
| 2027E | 53.6 | 59.7 | 16.7 | 76.4 | -22.8 | -6.1 |
Billion Gb. Read from UBS Figure 4. Negative = shortfall.
The circulating summary says supply ran ahead of demand in 2025-26, inventory built up, and in 2027 consumption overtakes shipments so hard that drawing down inventory does not cover it. The direction is right. Two points need sharpening.
First, 2025 was not balanced. Shipments exceeded consumption by about 3.8 billion Gb, and that excess is nearly identical to the inventory box (4.8). The chart is saying the 2025 surplus went into customer stock. Second, 2027 is short even if you delete the inventory box entirely: consumption of 59.7 is above shipments of 53.6. Inventory does not create the 2027 shortfall. It multiplies it more than threefold.
What the red dashed box is
The legend reads "Inventories" and the box sits on top of the consumption bar, on the demand side. Supplier inventory would sit on the supply bar. Stacked on demand, the natural reading is inventory customers want to hold, in other words inventory-build demand. "Adjusted" in the title then means end-consumption adjusted upward for that stocking.
On that reading, a 2027 inventory box of 16.7 billion Gb, 2.2x the 2026 box, is consistent with customers wanting deeper safety stock going into a year they expect to be short. I could not confirm UBS's definition of the line from the report itself. If it means something else, such as channel inventory on hand, the 76.4 total changes meaning. That is why this post leans on the consumption-versus-shipments comparison (59.7 vs 53.6), which holds under any definition.
Can the 2025-26 cushion cover 2027?
Taken literally, shipments above consumption in 2025 and 2026 total about 14.3 billion Gb (3.8 + 10.5), of which 12.4 shows up as inventory boxes. The 2027 consumption overshoot of 6.1 is smaller than that, so on paper stock could cover it. UBS instead adds another 16.7 of inventory demand in 2027, assuming customers do not release stock but keep building it.
There is a reason that assumption is defensible. HBM is not a commodity that trades through a channel. It is qualified against a specific GPU or ASIC and bought on annual contracts. Korea Investment & Securities notes that 2026 HBM contracts are already fixed, so suppliers cannot reprice mid-year the way they can in commodity DRAM, but can pull the 2027 contract cycle forward. Stock of HBM3E does not substitute for HBM4 demand in 2027. If the inventory is the wrong generation, holding it does not close the gap.
Where 129% consumption growth comes from
In 2026 shipments (+71%) outgrew consumption (+48%). In 2027 the order flips: consumption +129%, shipments +46%. Public sources point to three drivers for consumption more than doubling in a year.
Content per accelerator. In May, UBS cut its Rubin Ultra HBM4E assumption from 16-high 1TB to 12-high 768GB, suggesting Nvidia may be pre-empting supply constraints at the DRAM die level (Investing.com). The lowered assumption is still 768GB per package. Unit volume. Nvidia guided fiscal 2028 revenue growth of about 70%, explicitly supply-constrained, and lifted supply commitments to $279B (covered in the memory spot-market post). Non-Nvidia demand. BofA reported that the ASIC and TPU camp is bidding for HBM at prices above Nvidia's.
One illustrative conversion: at 288GB of HBM per accelerator, 2027 consumption of 59.7 billion Gb (7.46 billion GB) equals roughly 26 million accelerator-equivalents, against about 7.6 million for 2025. Illustrative calculation. Not actual company figures.
Why shipments stop at 46%
Kiwoom Securities estimates HBM wafer capacity rising from 530K to 790K wafers per month in 2027 (+49%) and CoWoS from 120K to 150K (+25%), per TopStarNews. The UBS shipment growth of 46% sitting next to a 49% capacity figure is not a coincidence. Output follows wafer starts.
Capacity cannot grow as fast as consumption because of wafer intensity. At the KeyBanc forum in August, Micron said each HBM3E wafer costs three wafers' worth of conventional DDR output. Every wafer moved to HBM shrinks commodity DRAM supply, which is what is driving the spot-price spike, and suppliers pace the conversion with that trade-off in view. Greenfield fabs arrive late: Micron's Idaho fab targets mid-2027, SK hynix's Yongin phase 1 end-2027, and Samsung's new P4 capacity is allocated to HBM first (BofA roadshow notes). None of that contributes fully to 2027 shipments.
On top of this framework, UBS's SK hynix ADR initiation put 2027 HBM bit share at Samsung 41%, SK hynix 39%, Micron 20% (NewsSpace), and in May UBS raised its 2027 HBM4 and HBM3E ASP assumption to +30% year over year from flat.
What I actually watch
| When | What | Why |
|---|---|---|
| Q4 2026 | 2027 HBM annual contract prices and volumes | First test of the +30% ASP assumption and the 16.7 billion Gb inventory line |
| Q4 2026 | Samsung HBM4 qualification at Nvidia | A large part of the 53.6 billion Gb shipment figure rides on Samsung's ramp |
| H1 2027 | Final Rubin Ultra HBM4E content | If 768GB moves again, the consumption curve moves with it |
| Quarterly | HBM bit shipment commentary from the three suppliers | Replaces chart read-offs with reported numbers |
Value chain read-through
| Segment | Direction | Evidence |
|---|---|---|
| Three HBM suppliers | Pricing and allocation leverage into 2027 | Consumption above shipments; ASP +30% assumption; 41/39/20 share |
| Front-end equipment | Direct beneficiary of +49% wafer capacity | Korea Investment: the current shortage is front-end cleanroom, not back-end |
| CoWoS, back-end | +25% expansion, bottleneck persists | Kiwoom capacity estimate |
| GPU and ASIC buyers | HBM cost burden, competition for stock | Nvidia margin guidance; BofA on ASIC camp bidding |
| Commodity DRAM | Supply shrinks by the HBM conversion | 1:3 wafer ratio (Micron) |
Risks to this view
- Read-off error and definitions. Every supply-demand figure here is read from gridlines, and the inventory line is not defined in the material available. Under a different definition the "43% short" figure changes; "consumption above shipments" does not.
- Demand slope. Consumption +129% is steeper than Nvidia's ~70% FY28 guide plus ASIC demand would imply on its own. It requires higher content per accelerator, an assumption UBS itself trimmed in May.
- Over-ordering. TrendForce flagged in 2023 that customers expecting shortage inflate orders. A 16.7 billion Gb inventory line may be a measure of that risk.
- Samsung execution. If the 41% 2027 share does not materialize, the 53.6 billion Gb shipment figure itself comes down.
Bottom line
The number that matters in the UBS chart is 59.7, not 76.4. Whatever the inventory line means, 2027 is drawn as the first year end-consumption exceeds shipments. The Q4 2026 contract round and Samsung's HBM4 qualification are what will turn these read-offs into reported figures.
Sources: UBS estimates, Figure 4 "Adjusted HBM supply vs demand" (published chart image; values are read-offs) / Investing.com on UBS report, May 13, 2026 (HBM4E 768GB, ASP +30%) / NewsSpace, July 2026, citing UBS SK hynix ADR initiation (share) / TopStarNews, Aug 2026, citing Kiwoom Securities (HBM and CoWoS capacity) / Korea Investment & Securities, "2026-2027 HBM Update" / Micron at KeyBanc Technology Leadership Forum, Aug 2026 (via ad-hoc-news) / Nvidia Q2 FY2027 results, Aug 26, 2026. Everything here is from public sources.
Comments
Post a Comment