Celestica
Celestica designs AI systems for hyperscalers from the silicon up, but its compute design work is custom-ASIC (Google TPU rack integration, Meta's Minerva, AMD Helios switching), not NVIDIA GPU racks: its own partner list names AMD, Broadcom and Intel. Listed for its hyperscale design model and its component role in the GB200 supply chain, with that distinction stated plainly.
- Joint-design-manufacturing model for hyperscalers, designing from silicon level up rather than build-to-spec only
- In-house networking switch design: DS6000/DS6001 (102.4 Tbps), DS5000 (51.2 Tbps) on Broadcom Tomahawk silicon
- Awarded design and manufacturing of a co-packaged-optics 1.6T Ethernet switch for a hyperscaler, production 2027
- In-house liquid cooling: CDUs, cold plates and manifolds, single and dual-phase
- Leads R&D and production of the scale-up switches for AMD's Helios rack-scale AI platform; JDM partner for Meta's Minerva ASIC servers and reported primary partner for Google TPU rack integration
| Model | Platform | Form factor | Cooling |
|---|---|---|---|
| DS6000 1.6T switch for AI back-end networks | Ethernet switch (non-NVIDIA) | 3U, 64-port 102.4 Tbps | air |
| DS6001 Hybrid-cooled variant for OCP racks | Ethernet switch (non-NVIDIA) | 2 OU ORv3 | DLC |
| GB200 NVL72 NVLink switch fabric boards Reported supplier alongside Amphenol; component role, not full rack builder | GB200 NVL72 | board-level component | DLC |
- No confirmed NVIDIA HGX/MGX/NVL72 builder status: its technology-partner page lists AMD, Broadcom and Intel, and it is absent from NVIDIA's Vera Rubin builder list
- Reported component-level role in the GB200 NVL72 supply chain (NVLink switch fabric boards), per secondary supply-chain research
- Hyperscale AI compute business is custom-ASIC-centric (Google TPU, Meta MTIA), not NVIDIA GPU platforms
- Fort Worth, Texas (new campus, $876M+ investment tied to AI datacenter electronics)
- Monterrey, Mexico; Kulim and Johor, Malaysia; Thailand campus
- ~$1B new investment across Mexico, Japan, Thailand and the US for 2026-2027
- Hyperscaler-concentrated: top 10 customers were 79% of FY2025 revenue
- Buyers cannot procure Celestica-designed hyperscaler racks off the shelf
- Listed here for its hyperscale design-manufacturing model; its AI compute design work is custom-ASIC (Google TPU, Meta), not NVIDIA racks
Is Celestica an OEM, ODM or integrator for NVIDIA AI servers?
Celestica sits in the "Hyperscale design-manufacturers" tier of our index. Joint-design-manufacturing model for hyperscalers, designing from silicon level up rather than build-to-spec only
Does Celestica design its NVIDIA servers in-house?
Celestica designs AI systems for hyperscalers from the silicon up, but its compute design work is custom-ASIC (Google TPU rack integration, Meta's Minerva, AMD Helios switching), not NVIDIA GPU racks: its own partner list names AMD, Broadcom and Intel. Listed for its hyperscale design model and its component role in the GB200 supply chain, with that distinction stated plainly.
Which NVIDIA platforms does Celestica ship?
Celestica ships systems on Ethernet switch (non-NVIDIA), GB200 NVL72. The systems table on this page lists the specific models.
Where does Celestica manufacture its AI servers?
Fort Worth, Texas (new campus, $876M+ investment tied to AI datacenter electronics). Monterrey, Mexico; Kulim and Johor, Malaysia; Thailand campus. ~$1B new investment across Mexico, Japan, Thailand and the US for 2026-2027