A battery energy storage system (BESS) is a bank of batteries, inverters and controls that stores electricity and returns it later — and until very recently, a data center's battery was sized for minutes, not hours. The uninterruptible power supply (UPS) that every conventional facility carries holds just enough energy — typically 5 to 15 minutes at full load — to bridge the seconds until diesel generators start. That is the entire job the battery was hired for, and it shaped everything: lead-acid or lithium chemistries chosen for discharge rate rather than duration, a battery room rather than a battery farm, and a capital line nobody outside the facilities team ever read.
What appeared in Memphis in 2025–2026 is a different object wearing the same word. Canary Media, the clean-energy trade outlet, counted 720 Tesla Megapack battery containers visible on satellite imagery of xAI's Colossus 2 campus dated 11 July 2026 — a configuration it estimates at about 2.8 GWh of storage and between 720 MW and 1,400 MW of discharge power, depending on which Megapack generation is installed. The story broke not through a press release but through a utility board meeting: on 20 August 2026, Riley Trettel, an energy and data-center developer working with SpaceXAI (the combined entity after SpaceX acquired xAI in February 2026), told the Tennessee Valley Authority's board that the battery holds 3.3 GWh — "enough to power all of Memphis for two hours" — and the board approved a direct hookup to the TVA grid that same day. Two weeks later, Doug McGowan, chief executive of the municipal utility Memphis Light, Gas & Water (MLGW), went further at a storage industry conference: "It's the world's largest grid-connected battery system … I think it's 2,000 MW of batteries behind the meter that they have to help as a backup when they have to curtail."
Three statements, three numbers, one direction. We are not going to pretend they reconcile — 2.8 GWh from counting containers, 3.3 GWh from the operator's own developer, and 2,000 MW of power capacity from the utility CEO, a figure Canary Media notes "is not exactly consistent" with the satellite count. What matters is that all three describe the same category: a gigawatt-hour-class battery attached to a compute campus, built without an announced utility contract, large enough to compete with the biggest batteries the US power industry has ever energised.
How big is 2.8–3.3 GWh next to the rest of the grid?
For scale, the benchmark is Edwards & Sanborn, the solar-plus-storage plant in Kern County, California — per Canary Media, the largest grid battery currently operating in the United States at more than 1,066 MW and 3,287 MWh. A utility-scale project of that class takes years of siting, permitting, community process and a utility contracting cycle that guarantees the finished plant revenue. The Memphis battery went from invisible to the national record contender in roughly a year, behind the meter, on private land, with its capacity first disclosed at a board meeting of the federal utility that serves it.
Now the arithmetic on duration, which is where the category change hides. Discharge the visible 2.8 GWh at the satellite-derived 720 MW and the battery runs about 3.9 hours; at the utility CEO's 2,000 MW figure, about 1.4 hours. Either way the design point is hours. That single fact separates it from every battery a data center has ever hosted, and it is worth being precise about why.
Why does an AI campus want hours of storage when 15 minutes used to be the standard?
Because the battery's job changed from ride-through to load-shifting. Run the conventional number first, since it is the baseline everyone's intuition was trained on. A 100 MW data hall that wants the standard 15 minutes of UPS ride-through needs, at full load, 100 MW × 0.25 h = 25 MWh. Colossus 2's visible battery is roughly 2.8–3.3 GWh — on the order of 100 times the ride-through storage of a 100 MW hall. No outage lasts four hours before the diesels are warm. Storage sized like this is not backing up the servers; it is backing up the grid connection.
The use case is right there in the MLGW chief executive's phrasing: batteries "to help as a backup when they have to curtail." An AI training campus is a giant interruptible-industrial load with a razor-thin tolerance for interruption. When the utility is capacity-constrained — a hot evening, a generation shortfall — the campus can drop its grid draw and run its margin from the battery instead, freeing gigawatts for everyone else. It charges when the grid is loose and discharges when the grid is tight. That is a peak-shaving generator that charges from the grid, and the reason a company builds one is exactly the reason xAI's history suggests: before the battery, the campus's gap-filler was a fleet of temporary gas turbines — dozens of them, operated without air permits, now the subject of a lawsuit by the Southern Environmental Law Center over emissions in adjacent neighbourhoods. Batteries are the legally survivable version of the same function: instant, dispatchable capacity with no stack emissions, no fuel logistics and no permit category that varies by county.
There is a second function that gets less attention because nobody markets it: the battery as a shock absorber for the grid connection itself. A training cluster that ramps from idle to full draw in minutes is exactly the kind of load that stresses a utility feeder and its protection equipment. A battery sitting between the campus and the TVA grid can smooth those ramps, which makes the utility more willing to connect the load at all. The TVA board's willingness to approve a direct transmission-level hookup in a single meeting — for a customer whose battery had not yet been energised — is evidence that this trade is being understood at the institutional level.
What did Tesla actually ship this year?
The Memphis battery was built from Megapack 2-class units. The product that supersedes it started production this quarter, and its specs define what the next campus-scale battery looks like. Tesla began manufacturing Megapack 3 on 6 August 2026 at a new Megafactory in Brookshire, Texas (outside Houston), designed for 50 GWh per year dedicated to this one product. Each Megapack 3 holds about 5 MWh and delivers 2.5 MW — a 28% energy increase over the 3.9 MWh Megapack 2 XL, in the same 28-foot footprint, via new 2.8-litre LFP cells. Tesla reports cutting thermal-system connection points by 78% and collapsing 24 cable connections to 3 busbar connections — the failure-point arithmetic we care about, because every connection in a liquid-cooled battery loop is a future field leak. Round-trip efficiency is quoted at 91% at medium voltage including auxiliary loads, with more than 10,000 cycles and a 25-year design life.
The system-level product is the more interesting one. A Megablock bundles four Megapack 3 units (20 MWh) with a factory-installed medium-voltage transformer and switchgear in one deployable assembly — the on-site electrical work that used to take weeks of electrician hours becomes a crane lift. Tesla claims 23% faster installation measured in labor hours, up to 40% lower construction cost, 248 MWh per acre of site density, and a crew able to deploy 1 GWh in 20 business days. We label all four numbers as vendor figures with no independent build to check against: the first Megablock sites will not energise until 2027, and every claim above originates with the seller. What is independently verifiable is the demand side — Tesla's Q2 2026 filing shows 13.5 GWh deployed in the quarter, up 41% year over year, within an energy generation and storage business that booked $3.14 billion of revenue. A second 50 GWh/year factory is roughly four times Tesla's entire global quarterly deployment, which tells you how the demand curve is being read in Fremont.
The customer side is documented too, in filings rather than press releases. Tesla's January 2026 SEC filing acknowledged $430 million of Megapack sales to xAI in 2025; after the SpaceX acquisition, the combined company bought another $405 million in the first half of 2026. Call it $835 million in eighteen months from two related customers, against a visible installed base of 720 containers and growing. At typical Megapack pricing that spend brackets the visible installation with room for what has arrived since the July satellite pass — which is the honest way to say the 2.8 GWh figure is a floor, not a ceiling.
What does this change about the AI power bill?
Two things, one visible and one structural. The visible one: storage is now a procurement decision, not a facilities afterthought. Yesterday's 800 VDC piece ended with NVIDIA's observation that energy storage is part of the new power architecture — at the time, a forward-looking line with no quantified example. Memphis is the quantified example. The battery farm is now a line in the campus bill of materials, sitting between the transformers and the generators, sized in gigawatt-hours, bought by the hundred-count container.
The structural one is about queue positions. The binding constraint on every AI campus build is time-to-power — the interconnection study, the substation, the transmission upgrade, measured in years (our time-to-power piece runs that timeline). A behind-the-meter battery changes the negotiation: a campus that can trim its own draw on command is a better customer to the utility, and the TVA's one-meeting approval of a direct hookup is what that looks like in practice. The battery is not just an energy asset; it is an instrument for buying grid access faster. That is a use case the traditional storage industry never priced, because the buyer was never this rich and never this impatient.
What it does not change: the diesel generators are still there, and the gas turbines came first for a reason. Batteries have energy, not endurance — 3.3 GWh against a multi-gigawatt campus load is hours, and multi-day grid shortfalls remain the domain of fuel-based generation. The honest reading of Memphis is not "batteries replace generators" but "batteries replace the everyday need for generators" — the curtailment events and ramp events that are frequent, short and politically expensive — while the diesels retreat to true emergencies. In facilities language: the battery moved from the UPS room to the substation yard, and the generators moved from the front line to the last resort.
For everyone building smaller than Colossus, the transferable arithmetic is the ratio, not the gigawatt-hours. Ride-through needs 0.25 MWh per MW of IT load (15 minutes). Load-shifting starts being interesting at 1–4 MWh per MW (1–4 hours) — ten to sixteen times the ride-through battery, at Tesla's published system prices a seven-figure line item per 10 MW of campus, and recoverable through demand-charge reduction and curtailment participation rather than avoided downtime. That is the calculation every campus operator in the queue is now running, which is why the Memphis installation matters even to people who will never own 720 containers: it is the first proof, at record scale, that the calculation closes.
Comments