Nvidia has raised its supply commitments to $279 billion for the quarter ended July 26, 2026, more than double the $119 billion reported in the prior period. According to Morgan Stanley estimates cited by The Inference, Nvidia, Alphabet, and AMD could together account for 85% of the world’s projected high-bandwidth memory (HBM) capacity for 2027, a situation that is adding pressure on the availability and price of other types of memory.
AI memory, the key points in 30 seconds
- Nvidia reported supply commitments of $279 billion, up from $119 billion in the prior quarter. The company attributed the increase mainly to memory purchases.
- Morgan Stanley estimates Nvidia will absorb 37.3% of the world’s projected HBM capacity for 2027. Alphabet would reach 36%, and AMD 12.1%.
- HBM demand is competing for manufacturing resources with other memory types, including the DRAM used in servers and PCs.
- Rising DDR5 prices are already hitting the retail market, though forecasts differ on how long the shortage will last.
- Nvidia’s earnings, expected in November, will show how its supply commitments, cash generation, and margins are evolving.
The $279 billion figure doesn’t represent spending already executed, nor does it correspond exclusively to memory already delivered. It reflects supply commitments disclosed by Nvidia, which include future obligations. According to the disclosed information, roughly $267 billion would come due before the end of fiscal year 2029. The scale of these commitments reflects how far in advance major buyers of artificial intelligence (AI) accelerators are locking in components for their next systems.
The move lines up with Nvidia’s own growth forecasts. Its CEO, Jensen Huang, said on September 17 that the company would sell twice as many chips the following year as it does this year. If that forecast holds, demand for the memory tied to its accelerators will stay elevated. However, how many chips actually get sold will depend on how orders, manufacturing capacity, and data center deployments evolve.
Three buyers could account for 85% of HBM in 2027
Morgan Stanley’s estimates, built from industry checks rather than direct company disclosures, put Nvidia at 37.3% of the world’s projected HBM capacity for 2027. Alphabet would reach 36%, and AMD 12.1%. Combined, the three companies approach 85.4%, while the rest of buyers would share roughly 14.6%.
These figures describe a capacity forecast attributed to buyers, not an official breakdown of all production or a guarantee that each company will receive exactly that share. The distinction matters because projected allocations can shift with orders, commercial agreements, factory yields, and the availability of new production lines.
HBM is used in AI accelerators and other high-performance processors because it can move large amounts of data at high bandwidth. Manufacturing it requires specialized processes and industrial resources that are also relevant to other memory families. When large customers reserve production in advance, manufacturers can prioritize those orders over other market segments.
Nvidia isn’t the only company trying to lock in supply. Alphabet develops its own AI processors, and AMD competes in the accelerator market. Their demand forecasts add further pressure to a supply chain with a limited number of memory manufacturers and specialized equipment suppliers.
The estimated concentration also raises a question for companies that don’t have the purchasing volume of the big tech groups. Server makers, computer vendors, and companies that buy memory for other uses may find themselves with less availability or higher prices. Even so, the projected HBM split alone doesn’t determine how much conventional memory will be available for each sector.
HBM pressure is spilling into the DDR5 market
Competition for manufacturing capacity isn’t limited to memory destined for AI accelerators. According to The Inference’s report, HBM production can consume more than three times the wafer capacity of conventional DRAM. In addition, some of the resources and processes used for this memory can’t be shifted immediately to manufacturing other products.
Samsung, SK hynix, and Micron account for more than 90% of the world’s DRAM wafers, according to figures cited in the report. The article states that their HBM production reservations for 2027 are already full and that new capacity won’t start arriving until late that year. It also notes that Samsung has allocated more than half of its 4-nanometer capacity to manufacturing base dies for HBM4, the memory’s next generation.
The effect on DDR5 prices offers a sense of the strain running through the market. The report notes that German retail prices were 544% above their July 2025 level. In the United States, 32GB kits that had sold for around $72 were going for roughly $409.
These are figures from specific markets and reference points, not a global price average. The final cost also depends on each module’s capacity, brand, distribution channel, and purchase timing. Even so, they illustrate why memory availability has become a relevant variable for computer and server manufacturers.
Top-tier server builders are said to have cut their inventories down to between eight and eleven weeks of supply, according to data cited by The Inference. With less stock on hand, shifts in component deliveries or prices can pass through to equipment purchases more quickly.
Manufacturers disagree on how long the shortage will last
Not every forecast points to restrictions lasting several years. Jason Chen, CEO of Acer, argues that memory makers are overstating supply constraints to protect their margins, according to statements reported by Notebookcheck. Chen expects prices to start falling after mid-2027, as DRAM output increases at China’s CXMT.
That forecast doesn’t mean an immediate drop. Chen’s own outlook includes computer price increases of between 5% and 20% during the fourth quarter, according to the same report. How things play out will depend on how much additional capacity reaches the market, AI-linked demand, and production decisions from the major suppliers.
On the supply side, SK hynix holds a leading position in HBM supply for Nvidia’s accelerators. The article attributes the company an estimated 60% to 70% share of the HBM4 volume destined for Vera Rubin, Nvidia’s upcoming platform. That share should be read as an estimate tied to one specific platform, not as SK hynix’s share of the entire global HBM market.
New capacity coming online and the development of alternatives could ease the constraints, but manufacturing lead times limit how fast supply can respond to rising demand. That’s why forecasts about when the shortage will end still depend on factors that can still change.
November’s earnings will test Nvidia’s commitments
The market’s next reference point will be Nvidia’s quarterly results, expected in late November. How its supply commitments, cash flow, and margins evolve will show how these advance investments are showing up in the company’s books.
The report notes that Nvidia’s operating cash flow fell from $50.3 billion to $24.1 billion between the two most recent fiscal quarters cited. The publication ties the decline to advance payments linked to its supply commitments. That connection doesn’t mean every payment corresponds exclusively to memory, although Nvidia did attribute the increase in its obligations mainly to memory purchases.
The company had also forecast a gross margin of 71% to 72% for its fiscal fourth quarter. That’s a company forecast, not a confirmed result. How it plays out will depend on prices, manufacturing costs, sales mix, and other operating factors.
Nvidia has also warned that growing supply obligations pose a risk if data center deployments slow down. Reserving capacity helps prepare for high demand, but it also exposes the company to commitments that could become harder to absorb if customers delay their investments.
The Vera Rubin platform’s timeline adds another data point: volume shipments are expected in the first half of 2027. Whether those systems actually ship on time will depend on accelerator, memory, and other component production tracking the current plans.
For the rest of the market, the question isn’t just how much Nvidia will spend, but how much supply will remain available for buyers with less negotiating leverage. Morgan Stanley’s figures point to a heavily concentrated HBM market for 2027, but they don’t by themselves determine future DDR5 prices or confirm how long the shortage will last. Upcoming earnings and how industrial capacity evolves will help show how far today’s forecasts hold up.
Frequently asked questions
How much has Nvidia committed in supply deals?
Nvidia reported supply commitments of $279 billion for the quarter ended July 26, 2026, up from $119 billion in the prior quarter. The company attributed the increase mainly to memory purchases.
What share of global HBM could Nvidia reserve in 2027?
Morgan Stanley estimates Nvidia will account for 37.3% of the world’s projected HBM capacity for 2027. The estimate comes from industry analysis, not an official allocation published by manufacturers.
Why could HBM demand push up DDR5 prices?
HBM manufacturing consumes more wafer capacity than conventional DRAM and relies on specialized industrial resources. That competition can limit the availability of other memory types, though the final price also depends on demand and production capacity.
When could the memory shortage ease?
Forecasts differ. Acer expects prices to start falling after mid-2027, while estimates cited in the report point to meaningful new capacity arriving closer to the end of that year. Neither timeline guarantees a broad price drop.
via: theinference.org
