Custom GPU Server Cold Plates for US AI Data Centers
ToneCooling is a custom cold plate manufacturer for GPU servers in the USA market: cold plates are made in our own manufacturing facility in Huizhou, and US RFQs are handled by ToneCooling Texas LLC in Austin, TX. A custom GPU server cold plate is a direct-to-chip liquid cold plate built to a specific GPU or CPU tray layout, heat load, and coolant loop. ToneCooling supplies standard kits for NVIDIA GB300 (4 GPU + 2 CPU), GB200, H200, and H100 according to their spec sheets, and custom cold plates for other GPU, CPU, and accelerator layouts after a DFM review.
- 2D or 3D drawings
- Heat load and heat-source map
- Coolant, flow rate, and pressure-drop limit
- Envelope, mounting, and port layout
- Material and joining preference
- Prototype quantity and annual volume
ToneCooling Texas LLC coordinates US RFQs for AI-server direct-to-chip cooling. Cold plates are made in our own manufacturing facility in Huizhou (ISO 9001:2015 certified, IATF 16949-aligned); manifolds, quick disconnects, and hose sets are supplied by ToneCooling from China. Custom samples are available from a single piece.
Direct-to-chip cooling for US AI servers and data centers
ToneCooling Texas supports US hyperscale, colocation, and AI-server OEM teams with direct-to-chip liquid cold plates and integrated GPU+CPU assemblies. We handle US-hours RFQ coordination, NDA, and DFM follow-up; engineering review runs through ToneCooling's thermal team.
Application fit: NVIDIA GB300, GB200, H200 accelerators; AMD EPYC SP5 and Intel Xeon CPUs; rack manifolds and CDU-connected loops. This page targets the AI-server-cooling application; component specs live on the linked product pages.
Cold plates for this application
Click a card for specs and RFQ details.

NVIDIA GB300 Cold Plates
Direct-to-chip 4 GPU + 2 CPU kit, part number TC-BZP-GB300-01A.
View GB300 →
GPU Cold Plates
Direct-to-chip GPU cold plates for AI server racks.
View GPU →
Server CPU Cold Plates
Copper direct-to-chip cold plates for server CPUs.
View server CPU →
Data Center Cooling
Rack-level and direct-to-chip cooling support.
Explore data center →Which GPU server platforms do we make cold plates for?
Four NVIDIA platforms are supplied as standard kits per their spec sheets; everything else is a custom GPU server cold plate quoted after a DFM review.
| Platform | Part / reference | Supply basis | Details |
|---|---|---|---|
| NVIDIA GB300 | 4 GPU + 2 CPU kit, TC-BZP-GB300-01A | Standard (per spec sheet) | GB300 cold plate specs |
| NVIDIA GB200 | TC-BZP-GB200-01A; 1,200 W + 1,200 W + 300 W (2× GPU + Grace CPU) at 2.5 L/min | Standard (per spec sheet) | GB200 cold plate |
| NVIDIA H200 | TC-BZP-H200-01A | Standard (per spec sheet) | H200 cold plate |
| NVIDIA H100 | T-HS202323 | Standard (per spec sheet) | H100 cold plate |
| AMD EPYC SP5 / Intel Xeon | Server CPU cold plates | Custom | Server CPU cold plates |
| Other GPUs, ASICs, switches, optics | Built to your drawing and heat map | Custom | Custom liquid cold plates |
GB200 values from datasheet TC-BZP-GB200-01A. Any change to GPU/CPU count, envelope, or port layout turns a standard kit into a custom project.
When should you choose a standard kit and when a custom GPU server cold plate?
Choose a standard kit when your tray matches one of the four listed NVIDIA layouts; choose a custom GPU server cold plate when the layout, heat load, coolant, or interfaces differ.
| If your project has… | Choose | Why |
|---|---|---|
| Your tray matches a listed platform layout | Standard kit | Spec sheet already defines interfaces and test data; fastest path to samples |
| Different GPU/CPU count, envelope, or port positions | Custom cold plate | Channel layout and ports are designed to your drawing |
| Heat load or coolant outside the spec-sheet envelope | Custom cold plate | CFD review sizes channels to your flow and pressure-drop limit |
| New accelerator, ASIC, switch, or co-packaged optics | Custom cold plate | No standard kit exists; DFM review starts from your heat-source map |
How does a custom GPU server cold plate project run?
A custom GPU server cold plate project runs in five steps, from RFQ to validated production, with US-side coordination through ToneCooling Texas LLC.
| Step | What happens |
|---|---|
| 1. RFQ | Send drawings or STEP, heat load and heat-source map, coolant, flow, pressure-drop limit, quantity, and US delivery location. An engineer reviews every RFQ and replies by email with next steps. |
| 2. NDA and DFM review | A mutual NDA can be signed before drawing review. Engineers confirm envelope, ports, material, and joining route, and flag any changes that reduce risk or cost. |
| 3. CFD and design | Flow and thermal simulation checks channel layout against your flow rate and pressure-drop limit before samples are built. |
| 4. Samples | Custom samples are available from a single piece. Sample and prototype lead time is typically 4–6 weeks, including design and simulation. |
| 5. Validation and production | Every unit is 100% helium leak-tested (acceptance criteria per project); thermal resistance, leak, and flatness test reports are available. Production lead time is confirmed at quotation for each project. |
Related: cold plate RFQ checklist · NDA and IP protection · custom cold plate lead time · DDP shipping into the USA
What performance data is published for GPU server cold plates?
Measured data is published for the standard kits. For the GB300 4G+2C kit, the GPU cold plate measured 0.0133 °C/W thermal resistance and 12.06 kPa pressure drop at 1.7 L/min and 1,066.94 W, and the CPU cold plate measured 0.0220 °C/W and 10.39 kPa at 0.7 L/min and 485.46 W (ToneCooling steady-state test with pure water, 2026-03-20).
| Cold plate | Flow | Heat load | Thermal resistance | Pressure drop |
|---|---|---|---|---|
| GB300 GPU cold plate | 1.7 L/min | 1,066.94 W | 0.0130–0.0133 °C/W | 12.06 kPa |
| GB300 CPU cold plate | 0.7 L/min | 485.46 W | 0.0220 °C/W | 10.39 kPa |
Source: GB300 datasheet TC-DS-GB300-01A Rev. B. Full specs and the downloadable test summary are on the GB300 cold plate specs section. For custom designs, thermal resistance is validated per project by CFD and sample testing.
How do you choose a custom cold plate manufacturer for GPU servers?
Choose a custom cold plate manufacturer for GPU servers by checking eight things you can verify before a purchase order:
- Spec-sheet data: published flow, heat load, thermal resistance, and pressure drop with test conditions and date.
- Leak testing: 100% helium leak testing on every unit, with acceptance criteria agreed per project.
- Test reports: thermal resistance, leak, and flatness reports available for your samples.
- Own manufacturing: who actually makes the cold plate, and where.
- Quality system: ISO 9001 certification and automotive-grade process discipline (for example IATF 16949-aligned).
- Sample policy: whether samples are available from a single piece and how long prototypes take.
- US commercial support: a US entity for RFQs, NDA, and delivery terms such as DDP.
- IP protection: a mutual NDA before drawings are shared.
How ToneCooling answers each check for US buyers
| Check | How ToneCooling answers it for US buyers |
|---|---|
| Spec-sheet data | GB300 (4 GPU + 2 CPU), GB200, H200, and H100 kits are supplied according to their spec sheets; the GB300 cold plate datasheet summary can be downloaded without a form. |
| Leak testing | 100% helium leak testing; acceptance criteria are agreed per project. |
| Test reports | Thermal resistance, leak, and flatness test reports are available. |
| Own manufacturing | Cold plates are made in our own manufacturing facility in Huizhou. Manifolds, quick disconnects, and hose sets are supplied by ToneCooling from China. |
| Quality system | ISO 9001:2015 certified; IATF 16949-aligned quality system. |
| Sample policy | Custom samples are available from a single piece. Sample and prototype lead time is typically 4–6 weeks including design and simulation. |
| US commercial support | ToneCooling Texas LLC in Austin, TX handles RFQs, US-hours communication, and delivery terms such as DDP shipping to the USA. |
| IP protection | A mutual NDA can be signed before drawings are shared — see NDA and IP protection for cold plate RFQs. |
See also how to qualify a cold plate supplier.
Process, quality, and manufacturing disclosure
| Process options | FSW, vacuum brazing, CAB, and TLP diffusion bonding in copper (C1100) and aluminum (6061); CFD flow/thermal validation; 100% helium leak testing. |
|---|---|
| Quality | ISO 9001:2015 certified; IATF 16949-aligned quality system; 100% helium leak testing; thermal resistance, leak, and flatness test reports available. |
| US support role | ToneCooling Texas handles RFQ intake, US-hours communication, NDA, DFM follow-up, and US-side routing for AI-server cooling programs. The Texas entity is registered in Austin, TX, and does not manufacture. |
| Manufacturing disclosure | Manufacturing is performed by ToneCooling Technology Co., Ltd at the 30,000 m² facility in Huizhou, Guangdong, China (since 2004). ToneCooling Texas LLC is the North American RFQ and engineering-communication entity and does not manufacture. |
Ready to discuss your liquid cooling project?
An engineer reviews every RFQ and replies by email with next steps. Send drawings, specs, or a project description.
Common questions from US engineering & procurement teams
Do you support US AI server liquid cooling?
Yes. ToneCooling Texas coordinates US RFQs for direct-to-chip GPU and CPU cold plates, manifolds, and quick disconnects for AI server and data-center deployments.
Which platforms are supported?
NVIDIA GB300, GB200, H200, and AMD EPYC SP5 / Intel Xeon, with manifold and CDU-loop review.
What heat load can the assemblies handle?
Heat load is confirmed per drawing and platform under defined test conditions. For reference, the GB200 cold plate datasheet lists 1,200 W + 1,200 W + 300 W (2× GPU + Grace CPU) at 2.5 L/min.
Where are the cold plates made?
In our own manufacturing facility in Huizhou, Guangdong, China. Manifolds, quick disconnects, and hoses are supplied by ToneCooling from China. ToneCooling Texas LLC (Austin, TX) is the US RFQ and engineering entity and does not manufacture.
Who manufactures custom cold plates for GPU servers for US buyers?
ToneCooling makes custom GPU server cold plates in its own manufacturing facility in Huizhou, China, and ToneCooling Texas LLC in Austin, TX handles US RFQs, NDAs, and delivery terms. Standard kits cover NVIDIA GB300, GB200, H200, and H100; other layouts are custom.
What do you need to quote a custom GPU server cold plate?
Send drawings or a STEP model, heat load and heat-source map, coolant type and inlet temperature, flow rate and pressure-drop limit, envelope and port layout, quantity, and US delivery location. An engineer reviews every RFQ and replies by email with next steps.
Can we order a single GPU cold plate sample?
Yes. Custom samples are available from a single piece. Sample and prototype lead time is typically 4–6 weeks, including design and simulation; production lead time is confirmed at quotation for each project.
What GB300 cold plate test data is available?
The GB300 GPU cold plate measured 0.0133 °C/W and 12.06 kPa at 1.7 L/min and 1,066.94 W; the CPU cold plate measured 0.0220 °C/W and 10.39 kPa at 0.7 L/min (ToneCooling steady-state test with pure water, 2026-03-20).