Nvidia scales AI capacity to multi-gigawatt with Oracle after locking six supplier deals, including a $1.1B U.S. testing advance

The details

• On Oct. 21 at GTC Berlin, Nvidia announced capacity agreements with six supply chain partners, per its press release: substrate makers IBIDEN, Samsung Electro-Mechanics and Unimicron (2.5x aggregate Nvidia substrate capacity over three years); Amkor (interposer wafer capacity up more than 300% at its South Korea plants by end-2028); King Yuan Electronics (new U.S. testing facilities funded by a $1.1 billion Nvidia capacity advance, high-volume production targeted for April 2027); and Siliconware Precision Industries (15x optical-engine capacity for co-packaged optics by mid-2027, with another doubling targeted by end-2028).
• Total semiconductor testing capacity is targeted to double from calendar 2026 to 2027, with buildouts in the U.S., Taiwan and Singapore. Tom’s Hardware notes KYEC currently handles nearly all of Nvidia’s chip testing and is building a $1.4 billion U.S. facility; by 2028 a significant portion of Samsung Electro-Mechanics’ annual package-substrate output is projected to be allocated to Nvidia.
• CEO Jensen Huang framed the stakes: “Demand is extraordinary. Capacity is the constraint.”
• On Oct. 27, at Oracle AI World in Las Vegas, Oracle and Nvidia announced plans to expand OCI Superclusters with Vera Rubin NVL72 systems to multi-gigawatt scale, with Vera Rubin systems and Vera CPUs now generally available on OCI and a smaller HGX Rubin NVL8 configuration planned — two 64-core Intel Xeon 6 CPUs, 4 TB of DDR5 memory and 6.4 Tbps RDMA networking per the Oracle release. OpenAI’s GPT-6 Astra was trained and served on an OCI Supercluster with more than 100,000 Blackwell GPUs in Abilene, Texas, Oracle said.
• Supply Chain Dive summarized the supplier deals as scaling certain component production by as much as 15-fold; press coverage of Nvidia’s latest filing puts its total supplier commitments at $279 billion, including roughly $160 billion added in the last three months.

What buyers should do

• Treat AI compute as allocation-priced into 2027-2028: if you are buying GPU cloud capacity or servers, negotiate multi-year terms and reserved capacity now, while providers compete for anchor customers.
• Check your electronics BOM against the constrained steps — substrates, advanced packaging, test capacity and DDR5 — and ask suppliers for allocation status, not just forecasts; the DRAM squeeze showed what reserved capacity does to merchant pricing.
• Watch supplier concentration in the other direction: test houses and substrate makers prioritizing Nvidia volume may lengthen lead times for automotive, industrial and consumer electronics buyers at the same suppliers.
• Where US-based supplier capacity is being added — KYEC’s testing facilities are part of the October U.S. factory investment wave — build it into dual-sourcing and tariff-risk maps for 2027.

The bigger picture

Nvidia is running a wartime-style playbook: prepay suppliers for dedicated capacity and anchor its infrastructure pipeline to a handful of cloud partners, in exchange for exposure to the same constraint everyone else faces — power, memory and packaging. For procurement, the takeaway is not that AI capacity is growing (it is, dramatically) but that its growth is being sequenced by one buyer’s reservation schedule. Every category touching silicon — from laptops to building automation — now inherits Nvidia’s supply chain decisions, and the “capacity is the constraint” posture predicts elevated electronics pricing and long commit leads well into 2028.