Sizing a BBU for a GB300-class rack

Start from the event, not the number of GPUs

Ride-through and bridging are different requirements and they are sized differently. Ride-through means the rack survives a transfer with no node state lost: the bus deviates, the battery takes over, and the load never sees a reset. On the VB-5125 shelf, bus deviation is detected in microseconds and solid-state transfer into the battery completes in under 1 ms; the VB-H400 is quoted at under 2 ms. That is a fixed property of the bridge electronics, and it does not change with rack size.

Bridging means holding the load for a defined window while something else takes over: a generator reaching speed, a UPS picking up the bus, a controlled migration of work off the rack, or a load shed. Ride-through is measured in microseconds. Bridging is measured in minutes.

The sizing target therefore has two parts: a duration in minutes and a load in kilowatts. Neither comes from a GPU count. NVIDIA does not publish an official rack power figure for GB300-class racks, so every single-number rack kilowatt figure in circulation is an industry estimate. Treat any such figure as a range that brackets the design, then replace it with a measured profile before committing to a shelf count. A shelf count derived from an estimate is the opening of a conversation, not a purchase order.

Measure the load, then shape it

Capture five things from the rack you intend to protect:

• Average draw across a representative training day, not across a benchmark run.
• Peak draw, and how long the peak lasts.
• The amplitude and the period of the swings between average and peak.
• Settle time after a step change, for example when a checkpoint write lands or a new job starts.
• The auxiliary load that must ride through with the IT load: rack management, switching gear, cooling controls. These are small in kilowatts and they are the loads that decide whether the rack returns cleanly.

GB300-class power systems already integrate energy storage for power smoothing, power capping and GPU burn, so the fastest swings are flattened before your BBU sees them. Futurum Group estimates that smoothing can cut peak grid demand by up to 30%. That is an analyst estimate, not an NVIDIA specification, and it should be read as a direction rather than a number to size against.

Two cautions. A nameplate figure over-states the bridge need, because it describes a configuration rather than the mixed duty cycle a production cluster actually runs. And the ramp after a transfer can be steeper than the average, because everything that was capped or throttled returns at once. Size against the post-transfer ramp, not the daily mean.

The events being covered are not hypothetical. A 54-day Llama 3 pre-training run logged 466 interruptions, 419 of them unexpected, with roughly 78% hardware-related and GPU issues making up 58.7% of the unexpected events (arXiv 2407.21783). On the supply side, US interconnection queues hold over 2,060 GW (LBNL, "Queued Up"), so a site can run at partial grid capacity for years, which turns the bridging window into a curtailment tool as well as a transfer tool.

Turn minutes into kilowatt-hours

Two formulas carry the exercise:

• Usable energy (kWh) = shelf nameplate capacity (kWh) × depth of discharge × conversion-path efficiency.
• Run time (min) = usable energy (kWh) ÷ bridge load (kW) × 60.

Example A — one VB-5125 shelf at 40 kW and 60 kW. The shelf is 5.12 kWh nameplate (51.2 V, 16S, 100 Ah, 21700 LFP). At 80% depth of discharge: 5.12 × 0.80 = 4.096 kWh. The 48V bridge path has no AC conversion stage, so at this step the full shelf energy is treated as delivered to the bus. Then 4.096 ÷ 40 × 60 = 6.1 minutes at a 40 kW bridge load, and 4.096 ÷ 60 × 60 = 4.1 minutes at 60 kW.

Example B — an 80 kW bridged load on a 10-minute window. Energy needed: 80 × (10 ÷ 60) = 13.3 kWh. Shelves by energy: 13.3 ÷ 4.096 = 3.3, rounded up to 4. Four shelves deliver 4 × 4.096 = 16.38 kWh, which is (16.38 − 13.33) ÷ 13.33 = 22.9%, so about 23% headroom over the window.

The rate check. Energy answers how many kilowatt-hours the window costs. It does not answer whether the shelves can deliver them fast enough. One VB-5125 is rated 100 A continuous at 1C, and 100 A × 51.2 V = 5.12 kW per shelf. An 80 kW load at 51.2 V draws 80,000 ÷ 51.2 = 1,563 A, so the group needs 1,563 ÷ 100 = 15.6, or 16 shelves — past the 15-shelf ceiling of one supervised bus. The energy method said 4. The rate check says 16. Both must pass, and the larger count governs. This mismatch is the most common error in rack BBU sizing, and it is why the reference band below carries two shelf-count columns.

The 1U in-node module. VB-E3 stores 1.5 kWh and is rated at 5 minutes or more at full load and 20 minutes or more at quarter load. The two figures are internally consistent: a 4:1 load ratio against a 1:4 time ratio moves the same energy. If the full 1.5 kWh is delivered, the full-load point is 1.5 ÷ (5 ÷ 60) = 18 kW and the quarter-load point is 4.5 kW. If the usable window is 80% instead, the same hold times imply 1.2 ÷ (5 ÷ 60) = 14.4 kW. Either way, the module is a per-node bridge sized to node consumption, not a rack-level substitute. It is 48V/54V on the server bus, bridges in under 1 ms, and takes PMBus telemetry to the host.

The HV shelf. VB-H400 carries 11.6 kWh at ±400VDC with 97% or better efficiency at the bus. Usable at 80% DOD is 11.6 × 0.80 = 9.28 kWh, and 9.28 × 0.97 = 9.0 kWh delivered. At 80 kW that is 9.0 ÷ 80 × 60 = 6.8 minutes per shelf; two shelves give 18.0 kWh, or about 13.5 minutes. The HV shelf needs far fewer units because of current, not energy: the same 80 kW at ±400 V is 80,000 ÷ 400 = 200 A instead of 1,563 A.

Depth of discharge: the setting that decides shelf count

The usable number is always smaller than the nameplate number, for three reasons. The BMS reserves capacity above and below the normal operating window to protect the string: headroom at the top for charge balance, a floor at the bottom so the pack can still hold its end-of-discharge voltage above the load's cut-off. The cycle-life rating of 6,000 cycles or more is quoted at 25 C and 80% DOD, so anything past 80% trades cycle life for minutes. And the bottom of the discharge curve is where voltage sags hardest under load, so the last few percent are the most expensive minutes available.

The -48V DC interface is specified as a -40.5 to -57.0 VDC window under ETSI EN 300 132-2, and the shelf has to stay inside that window for the whole event, not just at the start. That constraint, not the cell, is what fixes the floor of the operating window.

The 40 kW column is a yardstick for comparing DOD settings; 40 kW at 51.2 V is 781 A, a load one shelf cannot carry alone, which is the rate check from the previous section. The 1C column is the load a single shelf is actually rated for, so it is the column that maps to a real one-shelf bank.

End-of-life rule. Size against a state-of-health threshold, not new-cell capacity. If the bank must still hold the window at 80% state of health, usable energy per shelf becomes 5.12 × 0.80 × 0.80 = 3.28 kWh. A 40 kW / 10-minute window needs 40 × (10 ÷ 60) = 6.67 kWh, so 6.67 ÷ 3.28 = 2.03, which is 3 shelves by energy at end of life against 2 shelves on new-cell energy. The telemetry to enforce this already exists: state of health is trended against a per-cycle log, so the year-five derate is a number read from the same bus that reports state of charge.

A reference sizing band for GB300-class racks

The band below applies both checks at 80% depth of discharge. Loads are industry-estimated classes, not published rack figures. Shelves by energy use 4.096 kWh usable per shelf; shelves by rate use 100 A continuous, which is 5.12 kW per shelf.

Read three things out of the band. Every row is set by the rate check, not by energy, so a 10-minute window at 30 kW uses only about a fifth of the energy the specified shelves can deliver — short windows run at shallow depth of discharge, which is easier on the pack than the 80% basis. Granularity is coarse: a shelf is indivisible, so headroom arrives in steps of 4.096 kWh. And 15 shelves is the ceiling of one supervised parallel group, where 15 × 5.12 kW is roughly 77 kW of an 80kW-class bay. An 80 kW bridged load at 51.2 V wants 16 shelves, which is the point where the architecture changes rather than the shelf count: a ±400VDC bay using VB-H400 shelves at 11.6 kWh each covers an 80kW-class load with two or three units at 200 A instead of 1,563 A, and it is 800VDC ready for the next bus generation.

How to read a vendor's sizing table

Four assumptions hide inside a single runtime number.

• Nameplate versus usable. If a table quotes 5.12 kWh per shelf, it is quoting nameplate. Usable at 80% DOD is 4.096 kWh, 20% less, before any efficiency term. Ask which number the runtime was built from.
• New cells versus end of life. A runtime quoted on fresh cells shrinks as the pack ages. At 80% state of health the same shelf holds 3.28 kWh usable instead of 4.096 kWh, another 20% off. Ask the state of health at which the figure still holds.
• The discharge rate behind the quote. Run time falls as the discharge C-rate rises, so a runtime without a stated load is not a specification. The 6,000-cycle rating at 25 C and 80% DOD is qualified by both rate and temperature; a runtime should be quoted the same way.
• Temperature and scope. VB-5125 discharge is rated from -20 to 60 C and charge from 0 to 55 C; the 1U module is rated -5 to 45 C. Auxiliary and conversion losses are frequently left outside the quoted number, and the -40 C storage figure applies only to the self-heating variant.

Send this list back with any sizing table you receive:

• Is the capacity figure nameplate or usable, and what depth of discharge does the runtime assume?
• At what state of health is the runtime quoted — new cells or end of warranty?
• At what load was the runtime measured, stated in kW or in C-rate?
• At what ambient temperature, and are auxiliary and conversion losses inside the number?
• How many shelves may be paralleled on one supervised bus, and what current-sharing tolerance is specified?
• What happens to the runtime with one shelf removed from the group?
• Which documents ship with the batch: UN 38.3 Test Summary on request, MSDS, Class 9 transport documentation, drop test report, per-batch QC records retrievable by serial number, UL 9540A test data per product, and CE marking with a DoC available under LVD 2014/35/EU and EMC 2014/30/EU?

There is a commercial reason to interrogate the table rather than accept it. Uptime Institute's 2026 analysis found that 57% of operators' most recent significant outages cost over $100k and about one in five exceeded $1M. An optimistic runtime estimate is not paid for in shelf price.

Redundancy: N+1 on shelves, not on watts

Because the shelves are hot-swap, the practical redundancy unit is one shelf. A bank designed to hold the window with one shelf removed is a different number from a bank sized to the target alone, and the difference is not a percentage of watts — it is one physical unit.

Take the 40 kW / 10-minute case on energy alone. The window needs 40 × (10 ÷ 60) = 6.67 kWh. On new-cell usable energy that is 6.67 ÷ 4.096 = 1.63, so 2 shelves; sized against 80% state of health it is 6.67 ÷ 3.28 = 2.03, so 3 shelves; one more shelf for N+1 gives 4. That is the answer an energy-only sizing table produces.

The rate check moves both numbers. At 40 kW the group must pass 40,000 ÷ 51.2 = 781 A, and 781 ÷ 100 = 7.81, so 8 shelves carry the load and 9 shelves carry it with one shelf removed. With one shelf out, the remaining eight share 781 ÷ 8 = 97.6 A each, just inside the 100 A continuous rating — which is why N+1 is the honest claim here and N+2 would require buying more shelves rather than asserting more redundancy.

Two consequences follow. Nine 3U shelves occupy 27U, which is rack space a ±400VDC bay would not consume, since the same 40 kW is 100 A per pole at ±400 V. And the field-replaceable unit is a 3U shelf at 14.8 kg including modules, swapped on a live supervised bus, not the bay. The alarm ladder reports the degraded shelf from warning through to contactor drop, with time stamps, so the swap is scheduled work rather than an incident.

Validate before you commit

A sizing exercise ends with a test, not a spreadsheet. The sequence that holds up:

• Freeze the specification: rack count, measured load profile, bridging window in minutes, depth-of-discharge basis, state-of-health threshold, ambient envelope, and the DCIM interface you need.
• Order samples: 2 to 4 weeks after specification freeze, with engineering samples supported for qualified projects. Module MOQ runs 10 to 50 units depending on model.
• Replay the measured profile against the bank, including the post-transfer ramp, rather than testing against a resistive load at a single set point.
• Time the transfer and the hold: bus deviation to transfer under 1 ms on the 48V shelf and under 2 ms on the HV shelf, then hold measured to the minute at the specified load.
• Check current sharing across the parallel group at part load and at full load. Unequal sharing shows up as one shelf reaching its limit first, and it is the failure that a kilowatt-hour-only sizing sheet cannot predict.
• Confirm the telemetry path: state of charge and state of health into your DCIM over CAN 2.0B, RS485 Modbus or SNMP. A healthy shelf at rest reports a snapshot like BUS 51.18 V, CUR -46.2 A, cell delta-V 18 mV, delta-T 2.1 C, SOC 87.4%, SOH 99.2%, state BRIDGE-READY.
• Test at the temperature boundaries you will actually run: discharge -20 to 60 C, charge 0 to 55 C, and -5 to 45 C on the 1U module, with cold-plate liquid cooling where the rack provides it.
• Agree the paperwork before the first batch: cells UN 38.3 tested with the Test Summary on request, packs designed and tested to IEC 62619:2022, a UL 1973 certification path through accredited labs, UL 9540A test data per product, RoHS/REACH compliance statements, Class 9 transport documentation (IATA PI965/PI966 by air, IMDG by sea), and QC records from the six gates — cell IQC and grading, weld and assembly inspection, in-process test, formation and aging, end-of-line verification, OQC — retrievable per batch by serial number.

Mass production runs 4 to 8 weeks per batch after the sample is signed off. Build the validation into that calendar rather than after it: the EU Battery Regulation 2023/1542 battery passport rules for batteries over 2 kWh arrive in 2027, so the per-cycle log and the serial-number QC trail you validate now become a regulatory record later.

Sources

• UNECE Manual of Tests and Criteria, Section 38.3 — transport testing for lithium cells and batteries; in-house pre-compliance bench for T.1 to T.8 before accredited third-party labs
• IEC 62619:2022 — safety requirements for industrial lithium cells and batteries; packs designed and tested to this standard
• ANSI/CAN/UL 1973 3rd edition — stationary battery certification path through accredited labs
• UL 9540A — thermal runaway fire propagation test method; test data available per product
• EUR-Lex Regulation (EU) 2023/1542 — EU Battery Regulation; battery passport rules for batteries over 2 kWh from 2027
• ETSI EN 300 132-2 — -48V DC interface, -40.5 to -57.0 VDC operating window
• NVIDIA developer blog on the 800 VDC architecture and GB300 NVL72 power — integrated energy storage for power smoothing, power capping and GPU burn
• Futurum Group — estimate that power smoothing can cut peak grid demand by up to 30% (analyst estimate, not an NVIDIA specification)
• Uptime Institute Annual Outage Analysis 2026 — 57% of operators' most recent significant outages cost over $100k; about one in five exceeded $1M
• arXiv 2407.21783 — Llama 3 infrastructure: 466 interruptions over 54 days of pre-training, 419 unexpected
• LBNL "Queued Up" — US interconnection queue data, over 2,060 GW waiting for capacity