Nvidia published its first qualified list for AI data center battery storage on September 21 — three systems, from Hitachi Energy, LG Energy Solution and Tesla. The list matters less for who is on it than for what it asks: the certificate is about how an inverter behaves at the grid connection, not about how many cells a container holds. And the absence of Chinese suppliers, at the moment when China builds most of the world’s grid batteries, is the more interesting result.
What Nvidia Actually Qualified
Power and cooling are the two things that decide how much of the computing hardware can actually run, and they are why equipment such as solid-state transformers is already being ordered for US AI campuses ahead of the compute.
Nvidia calls the programme DSX Ready, and it sits under the DSX platform the company uses to describe the design, simulation and operation of an AI factory. The first two categories are battery energy storage systems and cooling distribution units — power and cooling, in other words, the two things that decide how much of the computing hardware can actually run.
On the battery side, three products qualified at launch:
| Supplier | Qualified product | AC output | Energy |
|---|---|---|---|
| Hitachi Energy | WD4 power conversion system with battery storage | 5 MW | 10 MWh |
| LG Energy Solution | Vertech JF2 2HR/4HR AC Link (LFP) | 5 MW | 10 MWh |
| Tesla | Megapack 2XL | 1.93 MW | 3.85 MWh |
On the cooling side, the first qualified units come from LG Electronics, LiquidStack and Vertiv. Nvidia’s wording is deliberately provisional — the programme “launches with two initial categories”, and more infrastructure and software categories are to follow.
Note the spread in what counts as a qualified unit. Hitachi Energy and LG Energy Solution are offering five-megawatt blocks; Tesla’s Megapack 2XL is a 1.93 MW unit, and it is the product Tesla has shipped by the thousand. Nvidia has not published a minimum size, and nothing in the launch material suggests one.
Why AI Data Center Battery Storage Is Not Grid Storage
The scope document is the useful part. Nvidia’s qualification sits at the AC terminals of the storage system and tests the power conversion behaviour there: dynamic active and reactive power response, current limiting, ride-through, islanded operation and black start capability. Suppliers must hand over the underlying evidence — raw time-series data, summary plots, test configurations and pass-or-fail statements tied to Nvidia’s requirements — along with the standards and regulatory certifications that apply, covering grid connection, fire safety and cybersecurity.
Read that list again and the emphasis is unmistakable. Nothing in it rewards a bigger cell, a denser pack or a cheaper cell chemistry. It rewards an inverter that can hold a voltage steady while the load underneath it moves in steps.
That is a different problem from the one grid storage solves. A utility battery is paid to shift energy between hours: charge when power is cheap, discharge when it is expensive, plus a growing list of ancillary services. An AI campus battery is paid to survive milliseconds. Inference and training loads do not ramp — they step, and they step in both directions, which is why operators describe the profile of a large GPU cluster in terms of near-instant load jumps rather than a curve. Traditional uninterruptible power supplies bridge an outage measured in seconds and diesel generators take longer still; neither is designed to absorb a load that falls away and comes back inside the same second.
Nvidia is also careful about what the certificate means. Passing qualification “does not replace site-level engineering or imply site-level stability”, the company states. Transformers, switchgear, generators and the rest of the site infrastructure sit outside the boundary. A qualified block is a starting point for a design, not a guarantee of a stable campus — a distinction worth remembering the next time a qualified list is described as a certification of reliability.
The Handbook Behind the Certificate
Nvidia did not invent these requirements in September. Its BESS Self-Qualification Guidelines, published in May, set out the criteria, and Chinese industry coverage of that document describes a test regime built around ten admission criteria: islanded self-stabilisation, adaptive buffering of AI load, tiered current limiting, high- and low-voltage fault ride-through, dynamic reactive support, seamless on-grid to off-grid transfer, black start, remote telemetry and control, suppression of control-loop distortion, and real-time execution of demand response.
The measurement tolerances reported for those tests are the part that separates the programme from a marketing badge: voltage and current measurement accuracy within 0.2%, frequency error no greater than 0.01 Hz, and channel-to-channel time synchronisation under one millisecond. Testing runs on two tracks — hardware measurement and electromagnetic transient simulation — across twelve operating conditions, with simulation at SCR 2, the weak-grid end of the range, mandatory. A 24-hour test must hold net state-of-charge drift to 5% or less.
Then there is the commercial screen, which is where a battery maker rather than a system integrator would come unstuck: disclosure of real power conversion system shipments over the previous twelve months, the largest single deployment, a workable plan to raise capacity tenfold within 24 months, full disclosure of single-source components and their alternates, a mean time between failures of at least 150,000 hours, 72 hours of full-load burn-in before shipment, and a repair response within four hours.
Why No Chinese Supplier Made the First List
China builds most of the world’s grid-scale batteries, and several of its largest suppliers are already working on the AI-campus case. Sungrow has set up a dedicated AIDC division and is developing high-voltage direct-current and grid-forming storage. Huawei Digital Power’s full-DC source-grid-load-storage system, built around an 800 V architecture, has completed the twelve test conditions. Zhiguang Electric has filed technical documentation for a solid-state-transformer-integrated battery system. None of them is on Nvidia’s launch list.
Three explanations hold up. The first is the standard base: Nvidia’s requirements are written against North American practice, including IEEE 2800, ANSI C84.1 and NERC reliability guidance, and validation at SCR 2 calls for simulation work and field data that most Chinese suppliers have not accumulated in those frameworks. Grid operators are moving in the same direction: the IEA argued in a September report that existing networks could be used considerably harder before any of them is reinforced. The second is manufacturing and service geography — AI campus projects in the United States increasingly expect local assembly and a service organisation on the ground, which is precisely what LG Energy Solution has been building in Michigan and what Tesla has in Texas. The third is the tenfold capacity commitment, which is a supply-chain promise rather than an engineering one, and hard for anyone to sign without a qualifying reference project already running.
It is worth being precise about what a rejected supplier has actually lost. Nvidia is not a buyer, and DSX Ready is not a procurement ban. What it does is shape what hyperscalers and their engineering contractors put in reference designs and tender documents. Being off the list does not stop anyone from selling batteries into a data centre; it removes them from the shortlist that a design team starts from. In a market where the premium for AIDC-grade, liquid-cooled, high-voltage DC storage runs well above ordinary commercial and industrial products — with gross margins in the range operators describe as 30% to 40% — that shortlist is where the money is. It also sits against a cell market that has started pushing list prices up again while contract prices lag.
The Co-Designer of Nvidia’s Reference Architecture Is Not on the List
One absence deserves its own paragraph, because it cuts across the logic above. On June 3, Siemens published a reference electrical and power architecture for Nvidia DSX Vera Rubin NVL72 facilities — a 136 MW campus with 100 MW of IT load, running from a 34.5 kV utility connection through medium-voltage distribution and modular low-voltage blocks to the rack. It was developed with Nvidia and with Fluence, whose Smartstack battery platform was written into the design to handle voltage and frequency ride-through, black start, demand response and load smoothing for AI workloads. nVent-aligned design considerations covered thermal management.
Fluence is not among the three qualified battery systems. Nvidia has not explained the omission, and the programme is product-level: a company can be a design partner and still not have put a specific product through the test, or not yet have passed it. What can be said is that Fluence spent the same month dealing with a different problem. In its September 16 guidance update, the company cut its fiscal 2026 revenue outlook to about $2.4bn from about $3bn and its adjusted EBITDA outlook to roughly negative $200m from roughly negative $10m, and pushed about $400m of deliveries into fiscal 2027 — attributing the shortfall to the ramp at its Houston contract-manufacturing partner rather than to demand. Its chief executive said the international supply chain was working normally and the difficulty was concentrated in US manufacturing.
The same month Fluence was also the buyer in a deal that shows how much of this business now crosses the Pacific: in September it agreed a 206 GWh supply framework with EVE Energy, a volume larger than that company’s entire 2025 shipments across EVs and storage.
Set those two facts side by side with Nvidia’s list and a pattern emerges that is bigger than any single company. The qualification rewards exactly the combination Fluence has struggled to deliver this year in the United States: a qualified AC-side control system plus a domestic line that can actually produce and service it. That is the criterion LG Energy Solution, Hitachi Energy and Tesla each meet in a different way.
Author’s Take: The interesting thing about DSX Ready is not the three names, it is that a chip company has started writing the specification for grid-facing power electronics. Nvidia is not a regulator and this is not a standard, but it does not need to be: when the reference design for a 136 MW campus comes from one vendor, the list attached to it becomes the shortlist. For Chinese storage makers this is the first hard evidence that the barrier in the AI-campus market is not cells or cost — the two things they have spent a decade winning on — but grid-forming control algorithms validated against someone else’s grid code, plus a local factory and a four-hour service promise. Sungrow, Huawei and Zhiguang Electric are all working on the right technical problems; the question is whether the qualification route has a door open to them at all. And for buyers, the caveat Nvidia wrote into its own programme is the one to keep: a qualified block is not a stable campus.
The Bottom Line: Nvidia’s DSX Ready list turns AI-campus storage into a qualification market rather than a price market. Hitachi Energy, LG Energy Solution and Tesla are on it at launch; Fluence, which co-wrote the reference architecture, is not; and no Chinese supplier is. The technical bar sits on the AC side — ride-through, black start, weak-grid simulation and millisecond synchronisation — so the suppliers that win this segment will be the ones with power electronics engineers and a local factory, not the ones with the cheapest cell.
Sourcing note: The DSX Ready launch, the three qualified battery systems, the cooling suppliers and the statement that qualification “does not replace site-level engineering or imply site-level stability” are taken from Nvidia’s own announcement of September 21, 2026. The product nameplates, power and energy ratings are from Energy-Storage.news’ report of September 22, which also describes the AC-terminal scope of the test. The ten admission criteria, the measurement tolerances, the twelve test conditions, the SCR 2 requirement and the commercial requirements are as reported by Chinese industry media covering Nvidia’s BESS Self-Qualification Guidelines. The Siemens, Nvidia and Fluence reference architecture of June 3 and Fluence’s role in it are reported by Data Center Dynamics. Fluence’s September 16 guidance revision is taken from the company’s own announcement.
Accuracy note: We have read Nvidia’s launch announcement, the Energy-Storage.news report and the Data Center Dynamics report on the reference architecture, but we have not read Nvidia’s BESS Self-Qualification Guidelines themselves; the detailed tolerances and commercial conditions above rest on Chinese industry media and should be treated as reported rather than confirmed, and we have not independently verified the twelve test conditions. The judgement that Fluence’s absence reflects product-level qualification rather than any deficiency is our reading, not a statement by Nvidia or Fluence; Nvidia has not commented on the omission and has not said whether further battery systems are under review. The described load behaviour of large GPU clusters and the comparison with uninterruptible power supplies and diesel generation are our framing. The 30% to 40% margin range for AI-campus storage comes from Chinese industry coverage and is not a figure either company publishes. Dollar conversions use RMB 7.25 per USD where RMB figures appear.
Sources & Further Reading
- Nvidia — “NVIDIA Launches DSX Ready to Qualify Power and Cooling Products for AI Factories” (2026-09-21) — the programme launch, the qualified battery and cooling suppliers, and the limits of qualification.
- Energy-Storage.news — “Nvidia launches BESS qualification program for AI data centers, names just Tesla, LG, Hitachi” (2026-09-22) — product ratings, the AC-side scope, the evidence suppliers must submit, and statements from LG Energy Solution and Hitachi Energy.
- Data Center Dynamics — “Siemens, Nvidia, and Fluence develop reference electrical and power architecture for data centers running Vera Rubin NVL72 platform” (2026) — the 136 MW / 100 MW reference design and Fluence’s Smartstack role in it.
- Nvidia — DSX Ready qualification categories — the programme’s current scope.
- Fluence Energy — fiscal 2026 guidance update (2026-09-16) — the revenue and adjusted EBITDA revisions and the attribution to US manufacturing ramp-up, as announced by the company.
- Chinese industry coverage of Nvidia’s BESS Self-Qualification Guidelines (May 2026) — the ten admission criteria, measurement tolerances, twelve test conditions, SCR 2 requirement and commercial conditions; also the progress reported at Sungrow, Huawei Digital Power and Zhiguang Electric.
- EVsays — earlier coverage of the same demand: Jupiter Power’s $1.4bn of storage financing and Sunrun and Tesla’s 580 MW residential dispatch, which show the same load pattern being met from the other end of the wire.
