AI data centers have a power problem, and it’s not the one most people think about. It’s not “is there enough electricity” — it’s “can the grid survive how erratically AI actually draws it.” TeraFlow, an OEM building vanadium redox flow batteries for the data center market, thinks the answer requires rethinking battery chemistry, campus architecture, and how these facilities relate to the grid entirely.
The core pitch: LDUPS
TeraFlow’s product is something they call an LDUPS — Long Duration Uninterruptible Power Supply. Think of a standard UPS (the small battery that keeps your desktop alive during a power blip) scaled up to grid size and stretched from a 10-minute buffer to a 10-hour, gigawatt-scale battery built roughly one-to-one with the data center’s capacity. The goal: turn the data center itself into a grid asset — something that can absorb shocks, ride through faults, and even feed power back — rather than a liability that trips offline and takes chunks of the grid with it.
Why “off-grid” rarely means fully islanded
Most so-called off-grid data centers aren’t really off-grid. They build behind-the-meter generation (solar, wind, gas, diesel, and — soon, per this conversation — small modular reactors) because it’s faster to get power than waiting on a grid interconnection. Most still apply for a grid connection eventually; it just comes later. True islanding is rare, and usually temporary.
The “EKG” problem
Traditional data center and cloud loads are flat and predictable. AI is not. Training workloads create a heartbeat-like power curve — servers train, sync, share, and repeat — producing violent swings that can trip circuit breakers, generator protection systems, and even grid substations. Inference is calmer, but still far from steady. A recent real-world example: a fault in Virginia’s “Data Center Alley” cascaded from a 200-megawatt trip into a multi-town event.
The uncomfortable truth, per this conversation: the volatility is a design problem, not an inevitability. Data centers built to properly ride through grid faults (a standard called low-voltage ride-through) can be genuine grid assets. Ones that aren’t can take entire regions down — and ratepayers absorb the cost either way.
Why vanadium over lithium-ion
Lithium-ion is the default because it’s bankable and familiar, not because it’s well-suited to this job. The case against it here:
- It’s built for roughly one cycle per day; running it hot and continuously accelerates degradation and raises fire risk.
- Cell balancing and accurate state-of-charge readings are harder than people assume — batteries can be off by a meaningful margin without anyone knowing.
- Even top-tier systems reportedly solve only 60–70% of the volatility problem, and heavy cycling can shrink a 7–10 year battery life down to a matter of months for some cells — turning a capital expense into a recurring operating cost.
Vanadium flow batteries, by contrast, use a single chemistry, which means degradation (crossover) is recoverable rather than permanent. That, plus no thermal runaway risk, made it the pick for a battery meant to run continuously for a decade-plus.
The real innovation: where the battery sits
Most battery and generation assets are wired into a campus’s medium-voltage layer. TeraFlow instead places its battery on the low-voltage side, in line between the step-down transformer and the circuit breaker panel — meaning all power runs through the battery continuously. That lets it act as a shock absorber: the generator and the grid only ever see a flat, stable draw, while the battery quietly soaks up the spikes and dips. It also strips out a lot of otherwise-necessary equipment (UPS units, harmonic filters, extra backup generators) and dovetails naturally with the industry’s shift toward 800-volt DC rack architecture.
Insurance, communities, and NIMBYism
No thermal runaway risk reportedly lowers insurance costs relative to lithium-ion. It also removes one of the bigger community objections to hosting a data center. Combined with closed-loop cooling (no water draw) and behind-the-meter generation lowering local power costs, three of the four classic objections to new data centers — water, cost of power, fire risk — can largely disappear. Noise remains the holdout.
The compliance angle: low-voltage ride-through
NERC and ERCOT are reportedly tightening scrutiny on data centers tripping offline during routine grid faults. The fix is low-voltage ride-through capability — staying connected and coming back online instantly when grid voltage sags rather than dropping off. Because power runs through the battery continuously, the architecture claims to bring load back up instantaneously when the grid returns, rather than the few-second lag typical of other approaches.
Skepticism about the “50 gigawatts off-grid” narrative
Asked directly whether the off-grid AI power narrative is real engineering or press-release capacity, the answer was blunt skepticism. Chip supply, transformers, switchgear, and generation equipment are all bottlenecked — no single manufacturer, battery or otherwise, could meet the demand implied by the gigawatts of projects currently sitting in interconnection queues. The expectation is that only a fraction of announced projects — maybe a tenth to a quarter — actually get built, and that even those get built in stages (25 MW, then 200 MW, then more) rather than all at once.
The one-year prediction
The forecast for where this all lands: less appetite for giant purpose-built gigawatt campuses (which risk being obsolete by the time they’re finished, given how fast GPU generations turn over), and more movement toward smaller, modular, distributed sites — 25 to 100 megawatts — especially as workloads shift from training to latency-sensitive inference. Done well, a distributed fleet of grid-friendly, storage-backed sites could ease grid stability concerns rather than worsen them. The core hope: that data centers get built well, so the current wave of local moratoriums gives way to genuine buy-in rather than backlash.
