⏱️ Lectura: 10 min
UC San Diego announced that it will test an experimental power technology designed specifically for AI data centers, according to UC San Diego Today. The university thus joins a small group of academic institutions using their own campuses as testing grounds to solve the energy bottleneck facing the AI industry.
📑 En este artículo
TL;DR
- UC San Diego confirmed that it will test an experimental power technology for AI data centers (UC San Diego Today).
- The project aims to meet the electricity demand required by large-scale AI training and inference.
- The test uses the campus as an energy testing ground before scaling the technology to the industry.
- PUE (Power Usage Effectiveness) remains the standard metric for measuring a data center’s efficiency.
- Redfish, the open DMTF standard, allows monitoring server power consumption via a REST API.
- Electricity, not the chip, is now the main limit to continuing to scale AI infrastructure.
- The initial announcement doesn’t detail a results date or the pilot’s scale in megawatts.
What happened: UC San Diego will test a new power technology for AI data centers
The announcement came through a statement from UC San Diego Today, the university’s institutional media outlet, which confirmed that the La Jolla campus will test a power technology described as innovative to supply artificial intelligence data centers. The statement doesn’t yet detail which vendor or what specific type of technology will be installed, but places the test within the university’s applied research programs.
That a public university would announce a pilot test of this kind is no coincidence. University campuses have functioned as real-world laboratories for energy technology for years: they have their own microgrids, specialized technical staff, and, above all, computing infrastructure (GPU clusters for AI research) that reproduces, on a smaller scale, the same problems faced by large commercial data centers.
UC San Diego’s case also fits its profile: the university has a solid track record in energy research applied to high-performance computing, and its supercomputing infrastructure grew alongside the demand for AI computing within the University of California system.
Context and Background
The underlying reason behind this and other similar tests is simple to state and complex to solve: training and running large-scale AI models consumes far more electricity than traditional data centers needed a decade ago. A modern data center designed for AI concentrates GPU racks that multiply electrical density per square meter compared to a generic computing data center, forcing a rethink of both power supply and cooling.
For years, the industry’s response was simply to request more connection to the public power grid. But interconnection queues (the regulatory process for connecting a large facility to the grid) grew longer in several regions of the United States, pushing data center operators to seek their own generation: on-site gas turbines, fuel cells, large-scale batteries, and, in some cases, direct agreements with existing nuclear plants.
Universities like UC San Diego enter this conversation from a different angle than major cloud providers: they’re not trying to solve their own scaling problem (no university campus operates data centers the size of those built by Microsoft, Google, or Amazon), but rather to validate technology in a controlled environment and publish the results, something private companies rarely do with the same level of technical detail.
💭 Key takeaway: The bottleneck for scaling AI is no longer just the chip: today it’s the ability to generate and deliver electricity at the pace GPU clusters are growing.
Technical Details and Performance
To understand what testing a power technology in a data center means, we need to start with the metric the industry uses to measure efficiency: PUE (Power Usage Effectiveness). PUE is calculated by dividing the facility’s total energy consumption by the energy that actually reaches the computing equipment. A PUE of 1.0 would mean that all the electricity entering the building is used for computing, with no losses in cooling, conversion, or distribution; in practice no data center reaches that number, but the closer it gets, the more efficient the facility is.
The innovative power technologies typically tested in this kind of pilot aren’t a single concept, but several families of solutions with different goals:
| Technology | When to Use It | Advantage | Limitation |
|---|---|---|---|
| Fuel cells | When clean, constant generation is needed next to the data center | Low acoustic footprint and reduced local emissions | Cost of hydrogen or natural gas infrastructure |
| On-site gas turbines | When public grid connection has a long wait queue | Relatively fast deployment compared to new grid capacity | Carbon emissions and dependence on gas supply |
| Battery storage (BESS) | To buffer demand peaks from GPU clusters | Responds in milliseconds to load variations | Doesn’t generate new energy, only redistributes it over time |
| Small modular reactors (SMR) | For long-term baseload supply on large campuses | Constant generation with no direct emissions | Regulatory approval timelines of several years |
The typical flow of a pilot test like this follows a sequence similar to this one:
flowchart TD
A["Public power grid"] --> B["Test microgrid"]
B --> C["Conversion and storage system"]
C --> D[("AI data center")]
D --> E["GPU clusters"]Data center operators monitor power consumption in real time using open standards. One of the most widespread is Redfish, the DMTF specification that exposes power metrics for a chassis or server through a standard REST API.
curl -k -u admin:password \
-H "Content-Type: application/json" \
https://bmc-servidor-gpu-01.local/redfish/v1/Chassis/1/PowerThis call returns, among other fields, PowerConsumedWatts: the instantaneous consumption in watts reported by the server’s management controller (BMC). It’s the same type of reading a pilot like UC San Diego’s needs to capture at each point in the facility to compare the new technology against the baseline.
(
sum(facility_power_watts)
/
sum(it_equipment_power_watts)
)On a Grafana dashboard powered by node_exporter and ipmi_exporter, this PromQL query calculates PUE by dividing the facility’s total power (which includes cooling, distribution, and losses) by the power that reaches the computing equipment directly. It’s the standard way to verify, with data rather than marketing promises, whether a new power technology actually improves efficiency compared to the infrastructure it replaces.
How to Start Testing It
A pilot like this isn’t exclusive to universities with research budgets. Any infrastructure team that wants to evaluate whether it’s worth investing in its own generation can replicate the same methodology on a smaller scale, starting by measuring its baseline.
The first step is to instrument the current facility before changing anything, with node_exporter on the servers and an IPMI or Redfish exporter pointing to each chassis’s BMC:
# Install node_exporter to expose host power metrics
curl -LO https://github.com/prometheus/node_exporter/releases/latest/download/node_exporter-linux-amd64.tar.gz
tar xvfz node_exporter-linux-amd64.tar.gz
cd node_exporter-linux-amd64 && ./node_exporter --web.listen-address=:9100With that baseline established, any test of a new technology (a battery, a fuel cell, a turbine) is compared against the same indicator, under the same load conditions. It’s the only honest way to know if something actually improves, beyond what the vendor’s spec sheet says.
💡 Tip: Before evaluating any new power technology, measure your current PUE for at least one full week of real load, not just during peak hours.
Impact and Analysis
The impact of this initiative isn’t in UC San Diego’s size (a campus, compared to a large-scale cloud provider’s data center, is small), but in the role it plays as an independent testing ground. Unlike a cloud provider, which has a commercial incentive to present its own pilots favorably, a public university can publish negative results without it affecting a business line.
This matters especially at a moment when the conversation about energy and AI data centers has become political as well as technical: states like Wyoming already specifically regulate AI data centers because of their impact on the local power grid, and countries like Switzerland and Belgium have revisited long-standing nuclear policies given the growth in electricity demand driven by AI. In that context, having independent data on which generation or storage technologies actually improve a facility’s PUE has value beyond the La Jolla campus.
The honest limitation is that, for now, UC San Diego Today’s announcement doesn’t include a public date for publishing results or the exact scale, in megawatts, of the test facility. Until the university publishes concrete data, any conclusion about the technology’s real performance is premature: the value of this news today is the commitment to test it in an open academic environment, not a measurable result yet.
What’s Next
The logical next step is for UC San Diego, or another institutional outlet within the University of California system, to publish technical details of the pilot: what specific technology is being installed, in which campus facility, and under what computing load it will be measured. If you want to follow the project, the primary source to monitor is UC San Diego Today‘s own page, which tends to expand on initial statements with follow-up notes.
It’s also worth watching whether other universities with supercomputing infrastructure announce similar pilots in the coming months: if UC San Diego gets public, favorable results, it’s reasonable to expect the model of campus as energy testing ground to be replicated at other AI data centers within the US university system.
📖 Summary on Telegram: View summary
Try it yourself: if you manage your own infrastructure, install node_exporter and ipmi_exporter today and calculate your real PUE before evaluating any new energy provider.
Frequently Asked Questions
What exactly did UC San Diego announce?
According to UC San Diego Today, the university will test a power technology described as innovative, aimed at supplying artificial intelligence data centers within the campus.
Why do AI data centers consume so much electricity?
GPU clusters used to train and run inference for AI models concentrate far more electrical density per rack than a generic computing server, requiring more power and more cooling capacity per square meter.
What is PUE and why is it the benchmark metric?
Power Usage Effectiveness divides a facility’s total energy by the energy that reaches the computing equipment directly. The closer to 1.0, the more efficient the data center.
Will this technology be used in other AI data centers?
There’s no public confirmation of that yet. The current announcement is limited to a pilot test within UC San Diego; eventual adoption by commercial operators would depend on the results the university publishes.
How can I measure my own infrastructure’s PUE?
By instrumenting servers with node_exporter and BMCs with an IPMI or Redfish exporter, and calculating the ratio between the facility’s total power and the power that reaches the computing equipment, as shown in this article’s PromQL query.
When will the results of UC San Diego’s test be known?
The initial statement doesn’t specify a date. In this kind of academic pilot, results are typically published several months after the test begins, once enough real load cycles have been measured.
References
- UC San Diego Today: official statement about the power technology test for AI data centers.
- Wikipedia: Power Usage Effectiveness: definition and calculation of the standard data center energy efficiency metric.
- Wikipedia: Data center: general context on data center infrastructure and electrical density.
- DMTF: Redfish specification: open standard for monitoring server power and hardware via REST API.
📱 Enjoying this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day.
Imagen destacada: Foto de American Public Power Association en Unsplash
0 Comments