Nvidia, Alphabet, AMD secure 85% of 2027 HBM supply

The AI arms race has a new chokepoint, and it’s not the chips themselves. It’s the memory stacked on top of them.

Nvidia, Alphabet, and AMD are projected to consume roughly 85% of global high-bandwidth memory supply by 2027, according to Morgan Stanley estimates. That leaves every other company building AI hardware, from startups to established hyperscalers, fighting over what amounts to table scraps in one of the most consequential supply chains in tech.

The numbers behind the squeeze

Morgan Stanley’s breakdown puts Nvidia at 37.3% of global HBM demand by 2027. Alphabet, fueled by its custom TPU chip program, comes in at 36%. AMD rounds out the trio at 12.1%.

HBM is projected to shift from roughly 20% to over 50% of the total bill-of-materials cost in upcoming GPU designs. Memory used to be a line item. Now it’s the line item.

The three primary HBM manufacturers, SK Hynix, Samsung, and Micron, have reportedly already filled their 2027 production capacity. Significant new manufacturing capacity isn’t expected to come online until late 2027 or potentially 2028.

Nvidia alone has disclosed procurement commitments that have ballooned to $279 billion, predominantly targeting memory-related supplies.

Why HBM matters more than you think

HBM achieves this by stacking multiple layers of DRAM dies vertically and connecting them with thousands of tiny wires called through-silicon vias. The result is memory that sits physically closer to the processor and moves data at dramatically higher rates than conventional DRAM.

HBM has leapfrogged both logic wafers and advanced packaging to become the single most limiting factor in AI chip output.

SK Hynix maintains the strongest position in the HBM market, with particularly deep ties to Nvidia. Samsung has been gaining ground with AMD as a customer. Micron rounds out the trio but all three are capacity-constrained through the forecast window.

What this means for the broader market

The ripple effects extend into consumer electronics as well. Advanced memory technology shares manufacturing infrastructure with the DRAM that goes into smartphones, PCs, and gaming consoles. When HBM commands premium pricing and guaranteed allocation, consumer-grade memory can get squeezed on both availability and cost.

Alphabet’s 36% share of projected demand is particularly notable because it reflects the scale of Google’s custom silicon ambitions. While Nvidia’s share is driven by its position as the dominant merchant GPU supplier, Alphabet’s allocation signals that its TPU program has grown into one of the largest consumers of advanced semiconductors on the planet.

AMD’s 12.1% share represents a meaningful validation of its Instinct accelerator lineup as a credible alternative in the data center AI market.

With new HBM capacity additions not expected until late 2027 at the earliest, the supply-demand imbalance could persist for another two to three years.