They’re Putting Diamonds in GPU Cold Plates. Who’s Watching the Water?
AI infrastructure / Liquid cooling / Independent engineering
They’re putting diamonds in GPU cold plates. Good. Now show me the coolant report.
That is not an argument against better technology. It is a question about what happens after the impressive hardware leaves the test bench.
Before approving the upgrade, I would want to know who is responsible for the fluid, what proves the installed loop is clean, and how the team will recognize a loss of cooling performance before it becomes a computing problem.
Those questions will not compete with diamonds for attention. They belong in the same purchasing decision.
Diamond cold plates are real. The operating questions are, too.
On October 7, 2026, Frore Systems announced LiquidJet Diamond, a liquid-cooling cold plate incorporating a diamond heat spreader positioned at processor hotspots. The company reports an additional 10°C reduction in hotspot temperature compared with its base LiquidJet design, and up to 35% more AI tokens per watt compared with traditional skived cold plates.
Those are Frore’s reported performance figures, not independently verified results from this article or a guaranteed outcome for every facility. Tokens per watt relates AI output to power use; it does not, by itself, establish facility profit.
The diamond is part of the heat-spreading assembly. This is not a proposal to circulate diamond particles through the coolant.
The engineering deserves a serious look. So does the assumption that the installed system will preserve the conditions behind the advertised improvement.
The distinction: A better cold plate can improve one part of the heat-transfer path. It does not certify the condition of everything connected to it.
What are you going to do with the extra thermal margin?
Suppose a new cooling design gives you more room between the processor’s operating temperature and its limit. You now have a decision to make.
Will you preserve that margin as operating reserve? Use some of it to support a warmer coolant supply? Or pursue more computing output?
Depending on the equipment and its controls, those may be legitimate options. But the same temperature margin cannot be counted in full toward every benefit at once.
This is the discussion I would have before converting a cold-plate performance claim into an energy-saving or capacity forecast. What is held constant in the comparison? Processor power? Coolant temperature? Flow? Pumping energy? Workload?
My concern is not that cold plates are becoming too sophisticated. It is that an attractive component result can be mistaken for a complete operating plan.
Nothing here establishes a defect in Frore’s product. This is a question about evaluating and operating any advanced cooling technology.
First, which water are we talking about?
“The water tested fine” is incomplete information. Which circuit was sampled?
In a liquid-to-liquid coolant distribution unit, or CDU, a heat exchanger transfers heat between the building-side fluid and a separate technology-side coolant circuit. Under normal operation, those fluids do not mix. The technology-side coolant circulates through the cold plates.
That separation matters. A satisfactory building-water report does not establish the condition of the fluid reaching the processors. Equally, dirty building water does not automatically mean those contaminants are flowing through the servers.
The circuit arrangement must be established before interpreting the sample. KMC²’s explanation of data center closed-loop cooling follows that distinction through the thermal chain.
The Open Compute Project also treats the technology cooling system as an interconnected scope, covering cold plates, tubing, manifolds, connectors and CDUs. Its work includes dedicated coolant-fluid and wetted-materials workstreams.
That is the right scale for the conversation. The question is not whether someone is “watching the water.” It is whether someone owns the condition of the correct circuit.
Clear coolant is not a condition report.
A clear sample is useful information. It is not proof that the coolant chemistry, material compatibility and contamination levels meet the equipment requirements.
Eaton’s liquid-cooling design guidance describes two distinct corrosion consequences: damage that can lead to leaks, and corrosion products that can accumulate elsewhere, restricting passages or filters. Deposits on heat-transfer surfaces can also add thermal resistance.
In other words, there does not have to be a puddle on the floor for coolant condition to matter.
Nor is “use purer water” a complete answer. The approved fluid and additive package must be compatible with the materials throughout the circuit. Eaton specifically notes that incompatible materials elsewhere in a system can undermine corrosion inhibitors.
I would not borrow a pH range, conductivity limit or filter rating from another installation and assume it protects these cold plates. The limits need to match the approved equipment, coolant formulation and system design.
A smaller number on a water-quality meter is not automatically a better engineering outcome. First establish what the number means for the fluid actually specified.
A small resistance change can cost real temperature margin.
Here is a deliberately simplified example, not a prediction about LiquidJet Diamond or any operating facility.
Assume a processor transfers 2,000 watts through a cooling path. At otherwise unchanged conditions, an additional thermal resistance of just 0.0015°C per watt produces:
Illustrative calculation — not measured product data
Temperature penalty = heat load × added thermal resistance
2,000 W × 0.0015°C/W = 3°C
Three degrees of additional temperature rise across that path.
Assumes steady heat flow through the modeled path. The added resistance is hypothetical, not an estimate of fouling or deterioration in a particular product.
That calculation does not tell us how quickly a system will deteriorate, whether a processor will throttle, or how much computing output would change. It shows why relatively small changes deserve attention when evaluating thermal margin.
It also explains why one temperature reading is not enough to assign a cause. A warmer processor could reflect a different workload, warmer inlet coolant, changed flow, a mounting issue or another change in the heat-transfer path.
The useful question is: What changed under comparable operating conditions?
Establishing that comparison is more valuable than collecting an impressive number of unrelated readings.
Five questions before buying advanced cold plates
These are the five questions I would put beside the performance specification. They apply whether the cold plates contain diamond, copper or another engineered material.
What, exactly, is approved to circulate?
Identify the coolant product, required concentration where applicable, approved additives and compatible wetted materials. Then identify who controls substitutions and top-ups.
“It is the same type of glycol” would not close that question for me. I would want documented compatibility, not a description on a service ticket.
What proves the installed circuit is clean?
Ask for the project’s acceptance criteria and the evidence that the assembled loop met them before serving the cold plates.
A completed flushing procedure documents an activity. The acceptance results need to establish the outcome. The same distinction matters after modifications or service that opens the circuit.
What will you compare six months from now?
Record a usable starting condition: coolant identity and condition, inlet temperature, relevant flow and pressure readings, processor load and temperature, and control settings.
Capture enough context to make a later comparison meaningful. A temperature trend without workload or coolant context can raise questions without answering them.
Can the loop be maintained without improvising?
Review access for sampling, filtration service, isolation, fluid handling and return-to-service checks. Establish the required operating state for each task.
The practical question is whether the approved procedure can actually be performed in the installed space, with the tools and outage arrangements available to the team.
Who acts when the cold plates are still cooling, but performance is drifting?
Name the person responsible for reviewing the evidence and the process for escalation. Establish how facilities, IT, the equipment supplier and the coolant specialist will compare their information.
A report sitting in an inbox is not a response. Neither is an alarm that everyone assumes belongs to somebody else. Decide who can investigate, who can authorize corrective work, and what evidence is required before the system returns to service.
The cold plates deserve a better question than “How much cooler?”
“How much cooler?” is a fair opening question. It should not be the last one.
“What will keep that advantage available after installation, service work and months of operation?”
That question puts the hardware, fluid, controls, maintenance access and operating responsibility into the same discussion. It does not diminish a breakthrough. It makes the purchase decision more complete.
I want to see better cold plates. I also want to see an operating team that can explain what is flowing through them, what has changed, and what happens next.
Put the diamonds in the cold plate. Put the coolant in the engineering review.
Two questions worth clearing up
Do diamond-enhanced cold plates require a special coolant?
The presence of diamond does not answer that question. Follow the complete assembly’s approved coolant and material-compatibility requirements. Do not infer a fluid specification from the heat-spreader material or from this article.
Does a closed cooling loop eliminate water-quality concerns?
No. Closed describes the circuit arrangement, not proof of cleanliness or chemical stability. Assess the specified fluid, compatible materials, service history and condition evidence for that particular loop.