NVIDIA could account for 37% of the world's high-bandwidth memory (HBM) production capacity by 2027, a share worth an estimated $279 billion, according to a TechPowerUp report citing analyst projections. The figure underscores how a single customer is reshaping the supply side of the advanced memory market.

The estimate lands amid a scramble among memory makers — SK hynix, Samsung and Micron — to expand HBM output for AI accelerators. HBM stacks sit on top of NVIDIA's data-center GPUs, and demand for those GPUs has effectively become the demand curve for the memory category itself.

What does the 37% share mean for the market?

If NVIDIA takes more than a third of global HBM capacity by 2027, suppliers will have to balance two pressures at once. They must build out new fabrication and packaging lines fast enough to serve NVIDIA's roadmap, while leaving capacity for other buyers — AMD, custom ASIC developers and hyperscalers building in-house silicon.

The $279 billion value attached to that share signals the scale of capital now flowing into HBM. Advanced packaging, TSV (through-silicon via) formation and stack validation remain bottlenecks, so capacity commitments made in 2024 and 2025 will largely determine who can actually serve the 2027 demand.

Why does one customer dominate HBM demand?

HBM exists almost entirely to feed AI accelerators, and NVIDIA ships the largest volume of them. Each new GPU generation increases the HBM capacity per package, which means NVIDIA's unit volumes compound with per-unit memory growth. The 37% projection follows that arithmetic: even flat GPU shipments would drive a rising share of HBM consumption.

What are the risks for suppliers and buyers?

• Concentration risk. A customer controlling over a third of capacity gives that customer significant pricing leverage.
• Allocation conflicts. Other accelerator vendors may face tighter supply or higher prices as memory makers prioritize the largest account.
• Capacity-cycle risk. If AI accelerator demand softens before 2027, suppliers carrying HBM-specific capital expenditure could face oversupply.

For memory makers, the bet is that NVIDIA's trajectory holds through 2027. For the rest of the AI hardware ecosystem, the projection is a warning that HBM access, not GPU silicon alone, may decide who can ship competitive accelerators at scale.

The full analyst breakdown was published by TechPowerUp.