NVIDIA Plans to Expand Substrate Capacity 2.5x Within Three Years
nashnova research
Nvidia announced a sweeping capacity expansion with six supply-chain partners, targeting a 2.5× increase in substrate output over three years and a far more aggressive 15× ramp in optical-engine capacity — its most explicit commitment yet to locking down AI hardware supply.
Why pull six partners into one announcement?
AI chip demand keeps growing, but the bottleneck is not chip design — it is packaging, substrates, testing, and optical interconnects, the steps that turn a designed chip into a shippable product.
This means → Nvidia must scale every link at once; a shortfall in any single one stalls the entire line.
The six partners span substrates (IBIDEN, Samsung Electro-Mechanics, Unimicron), advanced packaging (Amkor), testing (King Yuan Electronics), and optical engines (ASE / SPIL) — a full-chain simultaneous buildout.
Substrate capacity at 2.5× — what does that number tell us?
Nvidia plans to raise substrate capacity to 2.5× current levels within three years — one of its most specific public commitments on expansion scale.
Substrates — the circuit boards that physically carry and connect chips — are a critical input for high-end GPUs; a substrate shortage directly caps shipments.
In plain terms = no matter how advanced the chip design, without enough substrates to mount it on, the product cannot ship.
Samsung Electro-Mechanics joins the supply chain as a new vendor; Unimicron is set to ramp from Q2 2027, reaching full output by mid-2028.
How big are the packaging and testing moves?
On advanced packaging, Amkor plans to boost interposer wafer — the "bridge" layer that stitches multiple chips together — capacity by over 300% by the end of 2028.
On testing, Nvidia expects total test capacity in 2027 to double versus 2026.
The standout signal: Nvidia prepaid $1.1 billion to King Yuan Electronics as a capacity reservation, funding a new U.S. test facility targeting volume production by April 2027.
This means → Nvidia is not buying capacity on the open market; it is locking it in upfront with capital — the sheer size of the prepayment signals urgency.
Optical engines at 15× — why is this the most aggressive target?
Nvidia and ASE / SPIL plan to scale co-packaged optics (CPO) — embedding optical-communication modules directly inside the chip package — optical-engine capacity to 15× current levels by mid-2027, then double again by the end of 2028.
That makes this the single largest expansion in the announcement, dwarfing the 2.5× for substrates and 3× for packaging.
This reflects a heavy bet on the CPO roadmap — future AI super-clusters need data-transfer speeds that conventional electrical signals can no longer deliver; the path forward runs through light.
Everything converges on 2027–2028 — where is the risk?
Every expansion timeline clusters in 2027 to 2028: packaging by end-2028, testing by April 2027, substrates by mid-2028, optical engines by mid-2027.
In plain terms = Nvidia has placed its bets on a single two-year window; a delay in any one link could drag down the whole delivery schedule.
This means → whether these timelines hold will be the key test of Nvidia's AI supply-chain resilience.
市场有风险，内容仅供研究参考，不构成投资建议。