Nvidia's DSX Ready BESS Program: Three Suppliers, One New Standard for AI Factory Power

Nvidia has qualified exactly three battery energy storage suppliers — Tesla, LG Energy Solution and Hitachi Energy — under its new DSX Ready program, launched September 22, 2026. The program vets power and cooling hardware against Nvidia's DSX AI factory reference designs, and the BESS category is its first. For an industry accustomed to sizing batteries around tariff arbitrage and peak shaving, the significance is not the three names — it is that a compute company has just written a performance specification for grid-scale storage.

What Nvidia Actually Announced

The DSX Ready program, announced on Nvidia's corporate blog and reported by pv magazine USA, is a vendor qualification layer on top of Nvidia's DSX reference architecture. It launched with two categories: battery energy storage systems and cooling distribution units (CDUs). In cooling, LG Electronics, LiquidStack and Vertiv qualified. In storage, only three names appeared.

The two qualification pathways differ in a way that matters for interpreting the list. According to AI Weekly's analysis, BESS vendors must submit qualification test data to Nvidia for review, while CDU vendors use a self-qualification suite. Storage is the harder door — vendors are handing over measured performance data, not self-attestations.

ESS News reported the specific products: Hitachi Energy's WD4 PCS with battery storage, and LG Energy Solution's Vertech JF2 2HR/4HR AC Link, both rated 5 MW AC / 10 MWh. Tesla's qualification is tied to its Megapack platform, according to Tesla North. Hitachi Energy confirmed its own qualification in a September 2026 news release.

Why an AI Chip Company Is Specifying Batteries

This is the part the storage industry should read carefully. Nvidia is not entering the battery business. It is defining the power interface that its compute hardware assumes exists.

AI factories — Nvidia's term for purpose-built AI data centers — have a load profile unlike anything utilities have served before. Training clusters ramp hundreds of megawatts in seconds, swing between near-idle and full power on job-scheduling timescales, and cannot tolerate voltage or frequency excursions that conventional data centers ride through on UPS systems. The DSX reference architecture, introduced in July 2026, explicitly targets scalability from 100 MW to multi-gigawatt scale, with DSX Flex as the component that "enables renewable generation and adaptive grid balance."

What DSX Ready does is turn that architectural intent into a procurement filter. If you are building an AI factory against the DSX reference design, you now have a shortlist of storage vendors whose PCS behavior has been validated against the load profile Nvidia expects. EVShift's coverage noted that the qualification is "focused more on PCS behavior and AI-factory integration than" on cell chemistry or energy density. That is a deliberate choice: the differentiator in this application is the inverter, not the battery.

The Three Duty Cycles That Define the Spec

Three requirements appear consistently across coverage of the program, and together they explain why both headline products converge on 5 MW / 10 MWh.

Ride-through

AI factory BESS must ride through grid disturbances rather than disconnect. Where a conventional utility BESS might trip offline on a voltage excursion and wait for the grid to stabilize, an AI factory battery is expected to hold the bus while the disturbance passes. mgrid.org reported that suppliers are "graded" on ride-through and black-start — meaning these are scored attributes, not pass/fail gates.

Black-start

An AI factory designed around the DSX reference architecture can restart from a de-energized state using its battery as the cranking source. This is a meaningful escalation from data center practice, where generators typically provide the black-start capability and batteries cover the transition. Giving the BESS black-start duty shifts the requirement from power quality asset to grid-forming asset — which is why PCS behavior, not capacity, is the qualification axis.

Load buffering

This is the requirement that drives the 2-hour duration. AI training load swings are large, fast and frequent, and the grid connection is sized for an average, not a peak. The battery absorbs the delta between instantaneous compute load and contracted import capacity. A 5 MW / 10 MWh system buffers 5 MW of swing for up to 2 hours — enough to bridge between grid ramping events without cycling a generator.

The Grid Connection Queue Is the Real Constraint

The DSX Ready program exists in the context of a grid interconnection bottleneck that has become the binding constraint on AI infrastructure build-out. JLL's 2026 Global Data Center Outlook found that average grid connection wait times in primary data center markets exceed four years, and reported that operators "are expected to increase behind-the-meter power arrangements and explore colocated battery storage."

That wait time explains why storage is being pulled forward from an optimization asset to an enabling asset. Energy-Storage.news reported in April 2026 that with connection wait times "in the order of four years or more" in most major US wholesale markets, IPPs and hyperscalers are restructuring behind-the-meter versus front-of-the-meter contracts specifically to shorten time to power. A battery that lets a data center operate on a smaller interconnection than its peak load effectively buys years of schedule.

How AI Factory BESS Compares to Conventional Utility BESS

The table below contrasts the specification logic. The divergence is concentrated in what the battery is asked to do, not in the technology it uses — both are LFP-based, AC-coupled, containerized systems.

The cycling-profile row deserves emphasis. A load-buffering battery accumulates far more partial cycles than an arbitrage battery, but far less total energy throughput. That is a materially different stress pattern for degradation modeling — calendar aging and cycle counting diverge from the assumptions baked into conventional arbitrage warranties. Energy Optima models battery SOH and RTE as a three-dimensional interpolation across year, C-rate and cycles-per-day, which is the correct frame for a duty cycle like this one.

Simulating Data Center BESS in Energy Optima

Designing storage for an AI factory load is a different modeling problem than designing for a utility PPA, and the platform is built for both. The workflow for a data center application looks like this:

• Load profile setup: Import or construct an 8760-hour profile with the fast-ramp characteristics of AI compute load. The auto-design wizard accepts the profile directly and sizes against it.
• Dispatch strategy selection: Select ECONOMIC_DISPATCH for tariff-optimized operation, or MILP_HYBRID when the objective is a constraint — capping grid import at a contracted capacity while serving an unconstrained compute load.
• Capacity optimization: The LP-optimized sizing engine finds the minimum power and energy rating that satisfies the import cap across the full 8760-hour series — the same calculation that determines whether a 5 MW / 10 MWh block is sufficient for a given load swing.
• Degradation projection: The battery's SOH trajectory is projected from manufacturer cell data using 3D interpolation rather than a fixed annual fade rate, so a high-partial-cycle duty cycle is priced correctly.
• Financial projection: A 25-year NPV, IRR, LCOE and payback model, with augmentation triggers when SOH crosses a configured threshold.

The component database covers 112 batteries from 44 manufacturers, 165 PCS units from 23 manufacturers, and 200 inverters from 20 manufacturers — so the qualified-vendor shortlist can be modeled at product level rather than at generic round-number ratings.

The broader takeaway from DSX Ready is that storage specification is fragmenting by application. A data center battery, a solar-co-located battery and a peak-shaving battery are no longer the same product with different nameplate numbers. They are different duty cycles with different qualification criteria. Engineers specifying them need dispatch and degradation models that reflect the actual load, not a generic cycle count.

Sources