A forecast that NVIDIA will account for 37.3% of global HBM demand in 2027 shows how fierce the competition for AI memory has become.
Morgan Stanley's estimate, reported by ChosunBiz on August 5, puts NVIDIA, Google and AMD together at roughly 85% of HBM demand for GPUs and ASICs. Meanwhile, NVIDIA's commitments to secure future supply and production capacity reached $279 billion as of July 26.
On September 29, TrendForce also forecast that the average selling price of HBM overall will rise 121% year over year in 2027.
While large customers try to lock in future supply, the HBM shortage and rising prices are prompting designs that reduce the memory capacity per AI chip so that more products can be shipped.
However, demand forecasts, the amounts committed to secure supply, and the volume of HBM that can actually be shipped are separate measures. Signing huge contracts does not by itself guarantee that the necessary computing resources will be secured on schedule.
NVIDIA's 37.3% is a demand forecast, not a supply allocation
According to ChosunBiz's report, Morgan Stanley estimates 2027 HBM demand for GPUs and ASICs in bits, a measure of total memory capacity.
Gb stands for gigabit, and here it is used as a unit for comparing total HBM capacity.
| Buyer | Estimated 2027 HBM demand (billion Gb) | Share of global demand |
|---|---|---|
| NVIDIA | 18.125 | 37.3% |
| 17.507 | 36.0% | |
| AMD | 5.9 | 12.1% |
| Global total | 48.618 | — |
Adding the published shares, NVIDIA, Google and AMD together come to 85.4%.
What stands out in this estimate is that demand is not concentrated in NVIDIA alone.
Google is forecast at 36.0%, close behind NVIDIA, and AMD at 12.1%. If cloud providers expand their in-house AI accelerators, HBM makers will need to accommodate the product plans of several large customers, not just NVIDIA.
Still, the 37.3% figure is only a forecast of 2027 demand.
Morgan Stanley's original report is not publicly available, and the figure does not mean NVIDIA has contracted 37.3% of global HBM production capacity.
Saying that "NVIDIA has secured 37.3% of global HBM supply" would change what the estimate actually measures.
The amount of HBM per AI chip also matters when considering the scale of HBM demand.
According to Google's official technical overview, the training-focused TPU 8t carries 216GB of HBM and the inference-focused TPU 8i carries 288GB.
NVIDIA's official specifications list HBM4 capacity of 288GB per Rubin GPU.
HBM stacks multiple DRAM dies and places them close to a GPU or AI accelerator to feed it large amounts of data at high speed. It is not used by NVIDIA GPUs alone.
Capacity per chip also differs from product to product.
As a result, similar shares of HBM demand do not mean similar shares of AI chip shipments or computing performance. Nor can 37.3% be read as NVIDIA's share of the AI market.
The $279 billion is not just HBM purchases
According to Note 10 of NVIDIA's quarterly report, its commitments to secure future supply and capacity rose from $119 billion in the previous quarter to $279 billion as of July 26, 2026.
The CFO's commentary says the increase mainly relates to memory procurement.
Official documents confirm that NVIDIA is securing future components on a large scale to prepare for demand for its AI products.
However, not all of the $279 billion is for HBM purchases.
NVIDIA explains that the supply and capacity needed for its data center products relate mainly to memory and manufacturing equipment. The commitments cover not only current products but also future product generations.
Nor does the figure mean the full $279 billion has already been paid.
The 37.3% is a forecast of 2027 HBM demand, whereas the $279 billion is NVIDIA's multiyear supply and capacity commitments as of July 26, 2026.
These two numbers cannot be used to calculate how much NVIDIA will pay for HBM in 2027.
| Figure | Scope and period | Source | What it tells us |
|---|---|---|---|
| NVIDIA's 37.3% | Calendar 2027 HBM demand for GPUs and ASICs, based on total bit capacity | Morgan Stanley estimate reported by ChosunBiz | A forecast that NVIDIA will account for a large share of HBM demand; not secured supply |
| $279 billion | Multiyear supply and capacity commitments as of July 26, 2026 | NVIDIA Form 10-Q | The scale of future component and manufacturing capacity secured; not HBM purchases alone |
| 121% year-over-year increase | Average selling price of HBM overall in calendar 2027 | TrendForce forecast | A change in average price reflecting supply shortages and a rising mix of higher-priced products such as HBM4 and HBM4E |
These three figures differ in both units and time periods.
Multiplying them together will not yield the size of the overall HBM market or NVIDIA's purchase amount.
According to NVIDIA's disclosure, of the $279 billion, $92 billion in spending is scheduled for the remainder of fiscal 2027, $87 billion for fiscal 2028 and $88 billion for fiscal 2029.
After that, the amounts are $6 billion in fiscal 2030, $5 billion in fiscal 2031 and $1 billion in fiscal 2032 and beyond.
Because NVIDIA's fiscal year does not match the calendar year, the "remainder of fiscal 2027" cannot simply be treated as spending in calendar 2027.
NVIDIA also explains that for some contracts, cancellation, rescheduling or quantity adjustments may be possible before formal purchase orders are placed, though changes can bring additional costs.
The $279 billion is therefore a large commitment to secure future supply, but it is not a figure that directly indicates future volumes of good products shipped.
The "121% rise" in HBM prices also reflects the shift to HBM4
According to TrendForce's September 29 announcement, growing AI server demand means HBM and conventional DRAM continue to compete for limited advanced process and wafer production capacity.
Because raising yields and output of new-generation HBM also takes time, the firm expects memory supply to stay tight in 2027.
On that basis, it forecasts a 121% year-over-year increase in the average selling price of HBM overall in 2027.
However, it would not be appropriate to read the 121% as meaning the same HBM products will uniformly more than double in price.
The rise in average price reflects not only supply shortages but also a higher share of the more expensive HBM4 and increased HBM4E shipments from the second half of 2027.
Even if prices of existing products do not change, the average selling price rises when higher-priced new-generation products make up a larger portion of total sales.
A price increase on the same product and a shift to higher-priced products need to be considered separately.
HBM and conventional DRAM share some wafer production capacity.
Increasing production for HBM could therefore affect the supply of general-purpose DRAM.
However, the forecast of a 121% rise in 2027 HBM average selling prices cannot be applied directly to retail prices of DDR5 memory for PCs.
HBM and consumer DRAM differ in product specifications, customers and distribution channels.
8-high HBM allows more GPUs to be built, but costs more per unit of capacity
The HBM shortage and price increases are beginning to affect how much memory goes into next-generation AI chips.
According to TrendForce, GPU and ASIC makers are increasingly considering 8-high configurations as a priority, in addition to the 12-high HBM originally planned.
The background is not weaker demand for AI computing but limited HBM supply and rising system-wide costs.
Because an 8-high stack uses fewer DRAM dies than a 12-high one, it reduces both capacity and component cost per HBM stack.
Even when the same amount of DRAM die is used, reducing the amount mounted on each AI chip allows HBM to be allocated to more GPUs and ASICs.
On the other hand, the price per unit of capacity does not necessarily fall.
Each HBM stack needs a base die separate from the DRAM dies, and its cost does not fall proportionally when the number of layers is reduced.
With an 8-high stack, each base die carries less DRAM capacity than in a 12-high one, so the base die's cost burden per Gb is larger.
TrendForce forecasts that in 2027, the price per Gb of 8-high products could be 10–20% higher than that of 12-high products.
In other words, the adjustment here is a trade-off between how many AI chips can be shipped and how much HBM is mounted on each one.
Reducing memory capacity per chip allows more AI chips to be built from the same HBM supply. At the same time, the memory available to each GPU could be smaller than originally planned.
TrendForce also reported in August that, for NVIDIA's Rubin Ultra, the company is considering multiple configurations, including 8-high HBM4E and HBM4, in addition to the originally planned 12-high HBM4E.
However, it says the final specifications have not been decided, so it cannot be said that NVIDIA has decided to switch to 8-high HBM.
Will huge supply commitments translate into actual shipments?
If demand from Google and others grows as Morgan Stanley forecasts, multiple large customers will be competing for limited supply capacity in the 2027 HBM market.
NVIDIA's $279 billion in supply and capacity commitments can also be seen as a move to secure future components and manufacturing capacity early in such an environment.
However, a large contract size is not the same as receiving enough HBM of the required specification at the required time.
In assessing AI infrastructure investment in 2027, it will be necessary to check not only GPU and ASIC shipment volumes but also the HBM capacity actually mounted per chip and the volume of good products memory makers were actually able to ship.
For example, even if 8-high stacks raise AI chip shipments, the HBM shortage itself has not been resolved if the total HBM capacity supplied has barely increased.
Conversely, if HBM makers improve output and yields so that the required number of units can be supplied without cutting GPU memory capacity, it becomes easier to judge that supply constraints have eased.
To see whether NVIDIA's huge commitments translate into more actual computing resources, it is necessary to track not just contract amounts but also HBM output, yields, mounted capacity and, ultimately, shipments of finished AI systems.