Green Data Centre Options for Hardware

Why Hardware Choices Define Green Data Centres

When I first joined ORIGINALGO TECH CO., LIMITED, I assumed that building a greener data centre was mostly about switching to renewable energy contracts. It took just one quarter of tracking power usage effectiveness across our financial AI clusters to realize how wrong that assumption was. The hardware we choose — servers, storage arrays, cooling systems, networking gear — determines not only how much electricity we consume, but also how much water we evaporate, how much e-waste we generate, and ultimately whether our sustainability claims hold up under scrutiny. In the financial data strategy team, we run risk models, fraud detection engines, and large-scale AI inference workloads. These are not lightweight tasks. Every watt matters, and every component decision ripples through our total cost of ownership and our carbon ledger.

The global data centre industry consumes roughly 1–2% of the world's electricity, and with AI workloads exploding, that share is projected to rise. According to the International Energy Agency, data centres and data transmission networks accounted for nearly 300 TWh in 2022, with AI-specific demand growing at double-digit rates. For financial institutions like ours, the pressure is twofold: regulators increasingly demand climate risk disclosure, and clients want to know that their transaction data is processed responsibly. Hardware-level green options are therefore not a nice-to-have; they are a strategic imperative.

This article explores the most practical green hardware choices available today. I will draw on our own experience at ORIGINALGO TECH, where we have deployed liquid-cooled racks, high-efficiency power supplies, and extended-life storage systems across our Asia-Pacific nodes. I will also share a few candid mistakes — like the time we underestimated the airflow complexity of a "green" server room — because those lessons matter as much as the success stories. Green hardware is not just about buying the most efficient chip; it is about matching the component to the workload, the climate, and the lifecycle. Let's dig in.

Low-Power Processors for AI Workloads

The processor is the heart of any server, and its power draw often dominates the rack's energy profile. For years, the industry chased raw performance, but modern green data centres prioritize performance-per-watt. At ORIGINALGO TECH, we tested several ARM-based server processors against traditional x86 chips for our financial risk simulations. The results were surprising: for parallel Monte Carlo simulations, the ARM-based solution delivered about 70% of the throughput while consuming only 45% of the power. That changed our procurement strategy overnight.

But low-power does not mean low-capability. Companies like Ampere, AWS with its Graviton line, and Apple with its M-series chips have demonstrated that RISC-based architectures can handle heavy AI inference when paired with the right accelerators. The key is workload matching. Not every financial AI task needs a 400-watt GPU; many inference jobs run perfectly well on a 100-watt CPU with vector extensions. We learned this the hard way when we initially over-provisioned GPUs for a natural language processing pipeline that was mostly I/O-bound. After moving it to a cluster of low-power CPUs, we cut energy use by 60% with no latency penalty.

Another angle is chiplet design. AMD's EPYC processors use multiple small dies rather than one large monolithic die, which improves manufacturing yield and reduces wasted silicon. That translates into lower embodied carbon per chip. For a financial data centre that refreshes hardware every four to five years, embodied carbon can be 20–30% of the total lifecycle emissions. So choosing a chiplet-based processor is a quiet but powerful green move.

Of course, there is a catch: software compatibility. Many legacy financial applications are compiled for x86 instruction sets. Migrating to ARM or RISC-V requires recompilation, testing, and sometimes rewriting optimized libraries. At ORIGINALGO TECH, we spent three months porting our stochastic volatility models to ARM. It was painful, but the long-term energy savings justified the effort. My advice: start with containerized workloads that are already architecture-agnostic, then expand.

What about the future? I expect to see more heterogeneous computing — CPUs, GPUs, FPGAs, and specialized AI ASICs working together under a power-aware scheduler. The scheduler, not the human, will decide which chip handles which task based on real-time electricity prices and carbon intensity. That is already happening in some research labs. For now, the practical step is to audit your processor fleet and replace the oldest, least efficient nodes first.

Efficient Power Supplies and Voltage Regulation

Power supplies rarely get attention, yet they waste a surprising amount of electricity as heat. A typical server power supply unit (PSU) with 80 PLUS Bronze certification is about 85% efficient at 50% load; an 80 PLUS Titanium unit reaches 96%. That 11% difference may sound small, but across 5,000 servers drawing 400 watts each, it adds up to hundreds of megawatt-hours per year. At ORIGINALGO TECH, we standardized on Titanium-rated PSUs for all new financial AI nodes. The payback period was under 14 months, even without factoring in the reduced cooling load.

Voltage regulation is the next frontier. Traditional voltage regulator modules (VRMs) on motherboards convert 12V down to the 1V or so that processors need. These conversions lose 5–10% of energy as heat. New digital VRMs with adaptive phase shedding can improve efficiency by another 3–5%, especially at low loads. Financial AI workloads are bursty — they spike during market open and idle overnight — so low-load efficiency matters enormously. We measured a 7% reduction in idle power after switching to digital VRMs, which translated directly into lower cooling demand.

There is also the question of power distribution. In a traditional data centre, AC power comes in, gets converted to DC for UPS batteries, then back to AC for server PSUs, then back to DC inside the server. Each conversion loses energy. A 48V DC distribution architecture eliminates two conversions. Companies like Google and Facebook have deployed 48V DC racks, and the savings are real — up to 8% overall. For a financial firm, the barrier is legacy equipment. Most servers expect 120V/230V AC input. We solved this by using rack-level DC-to-AC inverters only for older gear, while new gear runs native 48V DC. It is a hybrid approach, but it works.

One personal reflection: I once approved a purchase of cheap PSUs to save $20,000 upfront. Within a year, we spent $34,000 extra on electricity and had three times the failure rate. That mistake taught me to calculate total cost of ownership over five years, not just sticker price. Green hardware is almost always cheaper over its lifetime if you do the math correctly.

Liquid Cooling and Heat Reuse

Air cooling has served data centres well, but it hits physical limits around 20–25 kW per rack. Modern AI servers with multiple GPUs can easily exceed 40 kW per rack. At that density, air cannot carry away heat fast enough without enormous fans and exaggerated airflow. Liquid cooling — either direct-to-chip or immersion — solves this problem with far less energy. A direct-to-chip system uses cold plates on processors and memory, circulating water or dielectric fluid. Immersion cooling submerges entire servers in a non-conductive liquid.

At ORIGINALGO TECH, we piloted a single-phase immersion tank for our most power-hungry AI inference cluster. The results: a 40% reduction in cooling energy, a 90% reduction in fan power, and a 15% increase in server density per square meter. The servers also ran cooler, which extended their expected lifespan. The biggest surprise was noise — or rather, the lack of it. Our engineers could finally have a conversation in the server room. That is a small thing, but it matters for morale.

Heat reuse is the next step. Instead of rejecting waste heat into the atmosphere, you can pipe it to district heating systems, greenhouses, or even office buildings. A data centre in Stockholm heats 20,000 apartments. In our case, we partnered with a nearby commercial laundry that needed hot water. The laundry saved on natural gas, and we reduced our cooling tower load. It was a win-win that also improved our community relations. The technical challenge is temperature: liquid cooling loops typically run at 40–50°C, which is too cool for many industrial processes but perfect for space heating or pre-heating.

Cost is the main barrier. A direct-to-chip system costs roughly 20–30% more than air cooling upfront. However, when you factor in the reduced cooling energy, higher density (which saves floor space), and potential heat reuse revenue, the payback is often 2–3 years. For financial AI workloads that run 24/7, the payback is even faster. My recommendation: start with a small pilot on your highest-density racks, measure everything, then scale.

There is also a maintenance consideration. Liquid cooling introduces new failure modes — leaks, pump failures, fluid degradation. We trained our facilities team on leak detection and added redundant pumps. The first year had two minor leaks, both caught by sensors before any damage. That is an acceptable risk given the energy savings. And frankly, I sleep better knowing we are not wasting half our electricity on fans.

Storage: The Forgotten Energy Consumer

When people talk about green data centres, they focus on compute. But storage can consume 20–30% of total IT power, especially in financial services where we retain years of transaction data for compliance. Traditional hard disk drives (HDDs) spin constantly, drawing 6–10 watts each. Solid-state drives (SSDs) draw less than 2 watts at idle, but they have higher embodied carbon due to semiconductor manufacturing. The greenest option depends on access patterns.

At ORIGINALGO TECH, we implemented a tiered storage strategy. Hot data — recent trades, real-time risk positions — lives on NVMe SSDs. Warm data — monthly reports, historical prices — lives on high-capacity SSDs that can power down. Cold data — seven-year-old audit logs — lives on tape or optical media. Tape is surprisingly green: a tape cartridge draws zero power when idle and has a 30-year lifespan. We moved 400 terabytes of cold financial data to tape and cut storage energy by 65%. The trade-off is slower access, but for compliance data that is accessed once a year, that is fine.

Software-defined storage (SDS) also helps. By pooling storage across nodes, SDS can keep only the necessary disks spinning and put the rest into standby. We use erasure coding instead of triple replication, which reduces the number of copies from three to 1.5, saving both energy and raw capacity. The math: triple replication uses 3x the disks; erasure coding with 6+3 uses 1.5x. That directly halves the storage power footprint.

Another angle is drive longevity. Financial data centres often replace drives every three years due to warranty, not failure. But modern SSDs can last 5–7 years in write-light workloads. By extending replacement cycles, we reduce embodied carbon and e-waste. We now run drives until they hit 80% of their rated write endurance, not until the calendar says so. This simple policy saved us $200,000 over two years and kept 1,200 drives out of the recycling stream.

One challenge: performance variability. Some financial AI models require consistent low-latency storage. Tape and spun-down disks introduce latency spikes. We solved this by keeping a small NVMe cache for metadata and using machine learning to predict which cold data will be accessed next. It is not perfect, but it works for 95% of our queries. The remaining 5% get a latency warning, which is acceptable for batch compliance jobs.

Network Gear and Optical Interconnects

Networking equipment — switches, routers, transceivers — accounts for 10–15% of data centre power. In a financial AI environment, we move enormous datasets between compute nodes, so the network is always busy. Traditional copper cables are cheap but power-hungry at high speeds. Optical interconnects use light instead of electricity, which means lower loss and lower power per bit. A 400G optical transceiver draws about 15 watts; the copper equivalent draws 25 watts or more.

At ORIGINALGO TECH, we replaced our top-of-rack copper switches with optical ones for inter-rack connections. The energy savings were modest — about 8% of network power — but the reliability improved dramatically. Optical cables are immune to electromagnetic interference, which matters in a room full of power supplies and motors. We also reduced cable clutter, which improved airflow and allowed us to raise the cold aisle temperature by 2°C, saving additional cooling energy.

Another green networking option is time-synchronized sleep. Many switches run at full power even when traffic is low. Newer switches support Energy Efficient Ethernet (EEE) and low-power idle (LPI). During night hours when our AI training jobs are paused, the switches drop to 20% power. We schedule our heaviest training runs during the day when solar power is abundant, and let the network sleep at night. It is a small operational change with a big impact.

What about software-defined networking (SDN)? SDN allows centralized control of traffic flows. By routing traffic along the shortest path and avoiding oversubscribed links, SDN reduces the number of active switch ports. We implemented SDN across our three Asia-Pacific nodes. The result: 12% lower network power and 20% lower latency. Latency matters for financial AI — every microsecond counts in arbitrage detection — so this was a double win.

Looking ahead, silicon photonics promises to integrate optical transceivers directly onto switch chips, eliminating the power-hungry electrical-to-optical conversion. That technology is still 2–3 years away from mass deployment, but it could cut network power by another 50%. For now, the practical step is to audit your network for idle ports and replace copper with optical where distance exceeds 3 meters.

Lifecycle Management and Circular Hardware

Green hardware is not just about operational energy; it is also about embodied carbon, e-waste, and recycling. A server's manufacturing accounts for 30–50% of its total lifecycle carbon. If you replace servers every three years, you are essentially buying a new carbon debt each time. Extending refresh cycles to five years — possible with modern hardware — cuts embodied carbon by 40% per year of service.

At ORIGINALGO TECH, we adopted a "repair first, replace last" policy. When a server fails, our technicians attempt component-level repair — replacing a failed DIMM, PSU, or fan — rather than swapping the whole unit. This reduced our server replacement rate by 35%. We also partner with a certified e-waste recycler that recovers 95% of materials by mass, including gold, copper, and rare earth elements. That is not just green; it is good business. Recovered metals offset about 10% of our new hardware procurement costs.

Another angle is refurbished hardware. For non-critical workloads — development, testing, batch reporting — refurbished servers from reputable vendors offer 70–80% of new performance at 40% of the cost and 20% of the embodied carbon. We use refurbished gear for our sandbox environments. The only failure we had was a fan, which cost $15 to replace. I know some financial firms are nervous about refurbished hardware for compliance reasons, but for non-production workloads, it is a no-brainer.

Circular economy principles also apply to cooling fluids, cables, and racks. We clean and reuse dielectric fluid from immersion tanks, we re-terminate fiber cables instead of replacing them, and we buy racks made from recycled steel. None of these actions are glamorous, but together they reduced our Scope 3 emissions by 18% in one year. My advice: measure your embodied carbon first. You cannot manage what you do not measure.

One challenge is data security. When decommissioning storage drives, you must ensure data is unrecoverable. We use cryptographic erasure — destroying the encryption key — which is faster and more reliable than overwriting. After cryptographic erasure, drives can be reused or recycled safely. This process is audited by our compliance team, and we have never had a data breach from decommissioned hardware.

Monitoring, Metrics, and Continuous Improvement

You cannot be green without data. At ORIGINALGO TECH, we installed per-outlet power meters on every rack, plus temperature and humidity sensors at three heights per rack. The data feeds into a dashboard that shows real-time power usage effectiveness (PUE), carbon intensity per workload, and water usage effectiveness (WUE). PUE is the ratio of total facility power to IT power; a PUE of 1.1 means only 10% overhead, while 2.0 means 100% overhead. We started at 1.8 and are now at 1.25.

The dashboard also flags anomalies. One day, it showed a 15% power spike in a rack that was supposed to be idle. We investigated and found a stuck fan controller that was running at 100% speed. Replacing the controller saved 8 kW per hour. Without monitoring, that waste would have continued for months. I now tell my team: "If you cannot see it, you cannot fix it."

We also use machine learning to predict cooling demand. By feeding weather forecasts, workload schedules, and historical sensor data into a model, we can pre-cool the data centre before a heatwave or reduce cooling before a cold night. This predictive approach cut our cooling energy by 22%. The model is not perfect — it once over-cooled during an unexpected cool spell — but the overall savings far outweigh the occasional error.

Benchmarking against peers is also useful. We participate in the EU Code of Conduct for Data Centres, which provides best practice guidelines and a maturity model. Going through the assessment forced us to document our processes and identify gaps. I recommend any financial data centre do the same, even if you are not in Europe. The framework is globally applicable.

Finally, do not forget the human factor. Our greenest hardware is useless if operators leave doors open or set thermostats too low. We run quarterly training sessions and publish a simple one-page "green checklist" for the facilities team. The checklist includes things like "close cold aisle doors" and "report any leaks immediately." Small actions, repeated daily, add up.

Conclusion: Hardware as a Sustainability Lever

Green data centre hardware is not a single product you buy; it is a portfolio of decisions across processors, power supplies, cooling, storage, networking, and lifecycle management. At ORIGINALGO TECH CO., LIMITED, we have learned that the biggest gains come from matching hardware to workload, measuring everything, and being willing to pilot new technologies even when they are not yet mainstream. Our PUE dropped from 1.8 to 1.25, our cooling energy fell by 40%, and our embodied carbon per compute unit dropped by 35% — all through hardware choices.

I will not pretend it was easy. We made mistakes: over-provisioning GPUs, buying cheap PSUs, underestimating liquid cooling maintenance. But each mistake taught us something. The financial industry is conservative for good reason — we handle other people's money — but sustainability is now a fiduciary issue. Climate risk is financial risk. Hardware efficiency is operational efficiency. The greenest data centre is also the most cost-effective one over a five-year horizon.

Looking forward, I expect to see more integration between hardware and carbon-aware scheduling. Imagine a server that refuses to run a non-urgent batch job when the local grid is powered by coal, but runs it two hours later when the sun is shining. That is technically possible today; the barrier is software and policy, not hardware. I also expect to see standardized carbon labels for servers, similar to energy labels on refrigerators. That would make procurement decisions much easier.

For now, my recommendation is simple: start with an energy audit, replace the oldest and least efficient hardware first, pilot liquid cooling on your densest racks, and extend your refresh cycles. You do not need to do everything at once. But do something. The planet — and your CFO — will thank you.

ORIGINALGO TECH CO., LIMITED's Insights

At ORIGINALGO TECH CO., LIMITED, our journey toward green data centre hardware has been shaped by the unique demands of financial AI workloads. We cannot sacrifice latency or reliability, yet we must reduce energy and carbon. Our key insight is that green hardware decisions must be workload-aware, not one-size-fits-all. A low-power ARM processor is perfect for batch risk simulations but wrong for real-time fraud detection. Liquid cooling is ideal for dense GPU clusters but overkill for storage racks. We have also learned that the biggest barrier is not technology but organizational inertia — procurement policies, compliance fears, and "we've always done it this way" attitudes. Overcoming those requires data. Our per-rack power meters and carbon dashboards gave us the evidence to convince even the most skeptical stakeholders. Finally, we believe that circular hardware — repair, refurbish, recycle — is the next frontier. Financial firms hoard old servers for compliance; we need secure, auditable ways to extend hardware life. We are now piloting a blockchain-based chain-of-custody system for decommissioned drives. It is early days, but we are optimistic. Green hardware is not a cost center; it is a competitive advantage in a carbon-constrained world.