Battery Precision Is Non‑Negotiable for AI Power Stability

upsengineering

21 Sept

Part 1: IPS Operations and the Full Technology Landscape

When a GPU cluster leaps from idle to peak power in milliseconds, the electrical system of a modern AI data centre faces an unprecedented stress test. Traditional UPS batteries were never designed for this kind of high‑intensity, high‑frequency cycling - and the latest operational data confirms that the old approaches are no longer fit for purpose.

Why AI Workloads Are Forcing a Rethink

AI workloads are inherently bursty. Large‑scale model training and real‑time inference produce repeated, rapid ramp‑up and ramp‑down cycles, with constant transitions between idle and peak draw. The exact magnitude and duration vary across GPU generations and applications, but the pattern is universal.

The traditional "brute‑force" response has been to oversize upstream infrastructure – bigger generators, transformers, and switchgear – to absorb worst‑case spikes. Others impose software power caps on GPUs, artificially limiting peak draw at the cost of compute utilisation and longer job completion times. Some facilities accept the volatility and manage generator instability reactively through operations teams.

Input Power Smoothing (IPS) offers a smarter path. When AI load fluctuates, the UPS actively charges and discharges its battery cabinet to compensate, keeping input power stable. But this mode places extreme demands on the accuracy and precision of battery state information - and that's where the choice of storage technology becomes absolutely critical.

Understanding the Battery States That Keep Your UPS Running

Before we explore the technologies themselves, it's essential to understand the three fundamental battery metrics that determine whether your UPS will perform reliably – or fail when you need it most.

State of Charge (SOC): The Fuel Gauge

State of Charge (SOC) represents the percentage of usable energy currently available in a battery relative to its total capacity. Think of it as the battery's equivalent of a car's fuel gauge – it tells you how much "fuel" is in the tank at any given moment.

SOC = 100% means the battery is fully charged and ready to deliver its maximum rated energy.
SOC = 0% means the battery is completely depleted and cannot deliver any further energy without risking damage.

In an input power smoothing architecture, SOC is not a passive status indicator. It is an active input to UPS control logic that governs when the battery charges, discharges, or holds idle. The UPS uses SOC to make real‑time decisions about:

Whether to continue IPS operations or stop to preserve backup reserve
When to initiate charging to replenish used energy
How aggressively to discharge when load spikes occur

Why accuracy matters: If the Battery Management System (BMS) reports SOC with a 5% error, the UPS may continue smoothing after the battery is actually depleted (risking over‑discharge), prematurely withdraw smoothing (exposing upstream gear to load swings), or oscillate between charge and discharge states unnecessarily – each action increasing stress and reducing service life.

State of Health (SOH): The Ageing Indicator

State of Health (SOH) represents a battery's actual usable capacity relative to its original nameplate rating – a measure of how much the battery has aged.

SOH = 100% means the battery delivers its full rated energy, just like new.
SOH = 80% means the battery has degraded to the point where it can only deliver 80% of its original capacity.

As batteries degrade through cycling, SOH declines, and both available runtime and smoothing capacity shrink. In AI power smoothing applications where batteries cycle far more frequently than in conventional backup roles, tracking SOH is essential to understanding what the system can actually deliver at any point in its service life.

Why accuracy matters: Inaccurate SOH hides degradation until it becomes operationally visible. This increases the risk of unplanned maintenance, unexpected cabinet replacements, and operational disruption. Accurate SOH supports maintenance strategies with service actions that can be planned rather than reactive, improving reliability while reducing lifecycle cost.

Depth of Discharge (DoD): The Usage Metric

Depth of Discharge (DoD) is the percentage of a battery's total nameplate capacity that has been discharged during a cycle. For a battery with a rated capacity of 10 kWh, a depth of discharge of 80% means 8 kWh has been withdrawn; the remaining 2 kWh stays in reserve.

DoD relates directly to SOC through a simple identity: DoD = (1 − SOC) × 100% .

DoD is arguably the single most important operational parameter for sizing energy storage because it directly determines:

How much energy the owner can actually use
How long the battery will last
What each stored kilowatt‑hour truly costs over the system's life

The DoD‑Cycle Life Trade‑off

There is no free lunch when it comes to DoD. Higher DoD levels accelerate degradation and increase stress on battery components. Every battery chemistry has a characteristic relationship between depth of discharge and cycle life – and understanding this trade‑off is essential for UPS specification in AI environments.

Key principle: A battery cycled daily at 90% DoD will degrade faster than an identical unit held at 80% DoD. However, drop the limit to 50% and calendar ageing starts to erode the economics from the other direction - the unused capacity sits on the wall, degrading quietly, while you still paid for every kilowatt‑hour of nameplate rating.

For UPS applications under AI loads, where batteries cycle far more frequently than in traditional backup roles, the DoD‑cycle life relationship becomes a primary design consideration. Going beyond 80% depth of discharge reduces the number of cycles exponentially.

How Input Power Smoothing Actually Works

With these definitions in mind, we can now understand what the UPS is doing during IPS – and why these battery states matter so much.

Traditional UPS role (standby/backup): The battery sits fully charged, doing nothing, for weeks or months. When a mains outage occurs, the UPS switches to battery, drains it steadily over several minutes, and that's that. The battery might cycle a handful of times per year.

IPS role (active power buffering): The battery becomes a continuously active participant in everyday operation. Here's how it plays out in practice:

Continuous load monitoring – The UPS constantly measures the real‑time power draw from the IT load (the GPU cluster) at microsecond intervals. It also monitors the upstream input power flow from the utility or generator.
Comparing against a target reference – The UPS control logic holds a defined input‑power setpoint (a moving average or a fixed ceiling). This setpoint represents the stable power level that the upstream infrastructure should see – typically the average AI load rather than the spiky peak.
Calculating the delta – The system subtracts the actual instantaneous load from the setpoint. If the load suddenly spikes above the setpoint, you have a positive delta (excess demand). If the load suddenly drops below the setpoint, you have a negative delta (excess supply).
Commanding the battery – This delta becomes the command signal for the battery converter:
- Positive delta (load spike) → The bi‑directional DC‑DC converter switches to discharge mode, pulling stored energy from the battery and injecting it onto the DC bus to supplement the rectifier. The result: input power from the utility/generator stays flat, even though the IT load just jumped.
- Negative delta (load drop) → The converter switches to charge mode, diverting the surplus energy back into the battery at a controlled rate. Again, input power stays flat.
Millisecond‑level response – The entire detect‑compute‑act loop completes in milliseconds – fast enough to catch those abrupt GPU transitions before they propagate upstream and stress transformers, switchgear, or generators.
Operating within defined SOC windows – The UPS will only allow IPS to operate between configurable SOC upper and lower bounds. It also reserves a protected SOC floor for genuine backup emergencies. Once the battery hits that lower threshold, the UPS gracefully tapers or stops the smoothing function, even if the AI load is still fluctuating – hence the absolute requirement for accurate SOC reporting from the BMS.

The critical operational difference: Under IPS, the battery isn't just a "last resort" - it's a working component that charges and discharges thousands of times over its service life, often in shallow but rapid micro‑cycles. This is why cycle life, thermal management, and state‑estimation precision are far more important here than they are in conventional UPS setups.

The DoD Imperative for AI Workloads

For AI data centres running IPS, the DoD‑cycle life relationship is not an academic exercise – it's a daily operational reality.

Frequent, shallow cycling vs. occasional, deep cycling: Traditional UPS batteries spend virtually their entire lives on float and seldom undergo a full discharge after commissioning. Under IPS, however, batteries experience thousands of shallow micro‑cycles over their service life – often at 20–50% DoD – as they continuously buffer GPU load fluctuations.

The critical insight: A battery rated for 6,000 cycles at 80% DoD may deliver 10,000+ cycles at 50% DoD, but only 3,500 cycles at 100% DoD. This means the operating DoD window chosen by the BMS and UPS control logic directly determines whether the battery lasts 5 years or 15 years.

DoD and SOC Precision: Two Sides of the Same Coin

Remember the SOC‑DoD identity: DoD = (1 − SOC) × 100% . If the BMS reports SOC with a 5% error, the UPS may inadvertently operate the battery at 85% DoD when it thinks it's at 80% DoD – potentially halving the cycle life of some chemistries. This is why accurate and precise SOC communication from the BMS is not just a monitoring nicety – it's a direct driver of battery longevity under IPS duty.

Chemistry‑Specific DoD Strategies

Different chemistries demand different DoD strategies:

Lithium LFP: Tolerates 80-100% DoD with minimal penalty. This makes it exceptionally well‑suited for IPS, where deep discharges may occasionally be required.
Lithium NMC: Performs best at 60-80% DoD. Going deeper significantly accelerates degradation.
VRLA: Should rarely exceed 50% DoD. This severely limits usable capacity in IPS applications.
Nickel‑Zinc: Delivers 600-1,500+ cycles at 80% DoD. At shallow DoD (10%), cycle life extends beyond 8,000 cycles.
Sodium‑Ion: Supports 80-95% DoD with 5,000–8,000+ cycles. Smaller depths of discharge significantly reduce degradation.

Energy Storage Technologies Compared

With that technical foundation in place, here is how the main UPS energy storage technologies stack up across key operational and environmental dimensions:

  
    











    Aspect
    VRLA
    LFP
    NMC
    NiZn
    Sodium‑Ion
    Flywheel
    Supercapacitor
  


    Service life
    3–7 years
    10–15 years
    3–5 years
    15 years
    15–20 years
    20 years
    15–20 years
  

    Cycle life (at design DoD)
    300–500 cycles
    3,000–6,000+ cycles
    800–2,000 cycles
    600–1,500+ cycles
    5,000–8,000+ cycles
    Virtually unlimited
    ~1,000,000 cycles
  

    Recommended max DoD
    50%
    80–100%
    60–80%
    80%
    80–95%
    N/A
    N/A
  

    Cycle life at 100% DoD
    ~200–300 cycles
    ~2,000 cycles
    ~440 cycles
    ~600 cycles
    2,500+ cycles
    N/A
    N/A
  

    Cycle life at 50% DoD
    ~600 cycles
    6,000–10,000+ cycles
    ~2,670 cycles
    ~8,000+ cycles (10% DoD)
    Extended significantly
    N/A
    N/A
  

    Energy density
    30–50 Wh/kg
    120–160 Wh/kg
    150–250 Wh/kg
    50–100 Wh/kg
    Moderate
    N/A
    Very high
  

    Footprint & weight
    Large, heavy
    Up to 60% smaller
    Up to 60% smaller
    50% smaller, 66% lighter
    Similar to LFP
    Smaller than VRLA
    Extremely compact
  

    Typical runtime
    5 min – 8 hours
    5 min – 8 hours
    5 min – 8 hours
    Short‑duration high power
    5 min – 8 hours
    1 sec – 1 min
    Seconds
  

    Response time
    Milliseconds
    Milliseconds
    Milliseconds
    Milliseconds
    Milliseconds
    < 0.5 seconds
    1–50 ms
  

    Thermal runaway risk
    Moderate (hydrogen evolution)
    Low (~270°C onset)
    High (150–200°C onset)
    None (intrinsically safe)
    None
    None
    None
  

    Recyclability
    >98% (lead)
    High
    Moderate
    >90%
    High
    100% (metals)
    ~80%
  

    Greenhouse gas emissions
    Baseline
    Lower than NMC
    Moderate
    25–50% lower than VRLA and Li‑ion
    Low
    Higher operational
    Very low
  

  
  

Nickel‑Zinc (NiZn) – The Rising Star in AI Data Centres

Nickel‑zinc is rapidly gaining traction as a compelling alternative for modern UPS deployments. ZincFive has launched the BC 2 AI battery cabinet specifically engineered for AI workloads, featuring 90Ah NiZn cells designed for high‑intensity discharge cycling, delivering 280–325 kWh capacity and 800A maximum discharge current. Major UPS vendors – including Vertiv, ABB, and Kohler – have already added NiZn to their product portfolios.

Key advantages for operators:

Intrinsically safe: The aqueous potassium hydroxide electrolyte means NiZn will not combust. Even under extreme abuse conditions – nail penetration, crushing, overcharging, or direct flame exposure – nickel‑zinc will not ignite or explode. It carries UL 9540A battery‑level safety certification.
Superior power density and space savings: It delivers triple the power density of conventional solutions while occupying roughly half the floor space and one‑third the weight.
Long operational life: NiZn offers a 10‑year warranty and a 15‑year design life. ZincFive even offers drop‑in retrofit kits for existing VRLA cabinets, extending the battery life to 15 years without major cabinet re‑engineering.
Remarkable environmental credentials: Lifecycle greenhouse gas emissions are 25–50% lower than both lead‑acid and lithium‑ion. Nickel and zinc are abundant, with no conflict‑mineral concerns, and over 90% of the materials are recyclable.
Exceptional thermal tolerance: Even if the cooling system fails and temperatures climb to 50°C, NiZn continues to operate safely and reliably – and remains within warranty.

Beyond Electrochemistry: Flywheels and Supercapacitors

Flywheel Energy Storage

Flywheels store energy kinetically in a spinning rotor. When power dips, the rotor continues to spin, generating DC power almost instantly.

Response time: < 0.5 seconds
Duration: 15–90 seconds (short‑duration ride‑through)
Lifecycle: virtually unlimited (mechanical bearings are the primary wear component)
DoD consideration: Not applicable – flywheels are not subject to electrochemical degradation
Environmental footprint: fully recyclable (steel, copper, aluminium) but carries higher operational energy consumption, resulting in a larger lifecycle carbon footprint than VRLA or NiZn.

Flywheels are exceptional at bridging the gap between a power disturbance and generator start‑up, and they have achieved 18% global penetration in the UPS market. Some AI data centres – including a notable facility in Finland – have deployed fleets of 208 flywheel UPS units.

Supercapacitors (Ultracapacitors)

These are solid‑state devices with no moving parts. They can deliver massive power pulses in 1‑50 milliseconds and endure millions of charge‑discharge cycles.

Ideal for instantaneous compensation of AI‑driven power transients
Maintenance‑free and highly reliable
Very low carbon footprint (~12 g CO₂e per unit cradle‑to‑gate)
~80% material recoverability
DoD consideration: Supercapacitors can be discharged to 0% without degradation – a unique advantage
Limited energy capacity – they cannot sustain long outages

In AI server environments, supercapacitors are increasingly used to absorb the millisecond‑level transients that would otherwise stress batteries and upstream switchgear.

Hybrid Architectures - The Best of All Worlds

No single technology covers every scenario. That is why the most sophisticated data centres are moving toward hybrid energy‑storage architectures:

Nickel‑Zinc + Lithium‑ion: NiZn handles instantaneous high‑power shocks, while lithium provides long‑duration backup.
Nickel‑Zinc + Flywheel: NiZn takes on short‑term high‑rate discharge, while the flywheel manages sub‑second fluctuations.
LFP + Supercapacitor: LFP handles bulk energy storage; supercapacitors absorb millisecond‑level transients, protecting the LFP battery from unnecessary micro‑cycling.
NiZn retrofit kits: Directly replace VRLA cells in existing cabinets - with 30–50% cost savings compared to new‑build installations – breaking the traditional 5‑ to 7‑year VRLA replacement cycle.

This article is published in two parts. Part 2 will examine nickel‑zinc in greater detail and explore why sodium‑ion batteries – with their fundamental advantages in resource abundance, safety, and lifespan – represent the true "game changer" for data centre UPS applications.

Albert Wong