TechGlobal.news

Nvidia discloses new multi‑year HBM and foundry supply pacts, locks in 2027 AI capacity

Enzo Reno ·

Nvidia said Wednesday it signed new multi year supply agreements and made additional prepayments to secure high bandwidth memory and advanced packaging capacity through 2027. The company disclosed the commitments in an SEC filing, citing tight supply for AI accelerators used by cloud and enterprise customers.

The move targets what analysts have called the primary choke point in AI server production. TrendForce reported SK Hynix held roughly 53 percent HBM market share in the second quarter of 2024, with Samsung and Micron racing to ramp HBM3E, and that packaging throughput for CoWoS remained a limiting factor for shipments of Nvidia systems.

Nvidia did not detail vendors or dollar figures in the filing. People familiar with the agreements said they include prepayments for HBM3E and CoWoS capacity with SK Hynix, Micron, and Taiwan Semiconductor Manufacturing Co, describing the total as spread over several years and tied to volume milestones.

TSMC told investors it planned to at least triple CoWoS capacity by the end of 2024 and continue expansions into 2025 to meet AI demand. Samsung has been pushing its SF3 foundry node and advanced HBM3E stacks, while Micron announced volume production of HBM3E in 2024, moves that collectively aim to loosen constraints on GPU availability.

Nvidia’s data center segment became its dominant business after 2023, driven by hyperscaler orders. Meta guided 2024 capital expenditure of 35 billion dollars to 40 billion dollars, Microsoft signaled sustained elevated capex for AI infrastructure, and Google parent Alphabet raised outlays for TPU and GPU capacity, figures that have continued to underpin long term GPU contracts.

Export controls remain a swing factor for allocations. The U.S. Commerce Department tightened rules in October 2023 and again in 2024 that restricted sales of certain Nvidia accelerators to China, prompting compliant product variants and reallocation of supply to other regions.

If sustained, the prepayment model could further entrench long horizon supply agreements between chip designers, memory makers, and foundries. That would favor incumbents with balance sheet capacity, reshape pricing power in HBM and packaging, and set a higher bar for new entrants attempting to scale AI silicon into the largest cloud buyers.