Flex to buy EPC Power for $4.4 billion as AI data centres move toward 800V DC power
Power conversion is becoming a strategic layer of AI infrastructure, and the physics of voltage, current and conversion losses explains why.
Flex has agreed to acquire EPC Power, a maker of power-conversion equipment, for $4.4 billion, Data Center Knowledge reported on 4 September. The deal is expected to close in the fourth quarter of 2026 and will be financed with a combination of debt and equity. EPC Power will join Flex's Cloud and Power Infrastructure segment, which Flex plans to spin off as an independent public company in the first quarter of 2027.
EPC Power builds power-conversion systems for data centres, utility-scale energy storage and microgrids, including grid-forming technology, digital rectifiers, DC-DC converters and its M and MRACK series inverters, with solid-state transformers on its roadmap. The company is projected to generate about $800 million of revenue in 2026, with organic growth of roughly 40 percent expected in 2027 and an EBITDA margin near 30 percent. In July it opened a 167,000-square-foot factory in Fountain Inn, South Carolina, with initial capacity of 27 gigawatts a year that can scale to 40 gigawatts; its US capacity is expected to exceed 30 gigawatts in 2027.
The article ties the deal to rising rack densities in AI facilities and a shift toward 800-volt direct current distribution. Higher voltage delivers the same power at lower current, which reduces conductor size and electrical losses and removes conversion stages between the grid and the rack. EPC Power's Agile Grid-Forming platform is described as responding to demand changes within milliseconds while integrating energy storage and on-site generation and supporting both grid-connected and microgrid operation.
Neil Osnato of Persistence Analytics Group told the publication that the price signals power conversion moving from a supporting component to a strategic control point, and that AI infrastructure is becoming a power-systems problem as much as a compute problem.
Why it matters
Accelerator racks now draw enough power that the electrical path from substation to chip is a design constraint rather than a commodity. The move to 800V DC, now backed by a multibillion-dollar manufacturing deal, suggests the industry expects rack densities to keep climbing. Grid-forming converters and on-site storage point to data centres that can ride through grid instability and even support the grid. Manufacturing capacity measured in tens of gigawatts shows how large the power side of the AI build-out has become.
Accelerator racks now draw enough power that the electrical path from substation to chip is a design constraint rather than a commodity.
Conventional low-voltage AC path
Utility AC stepped down, converted to DC for UPS batteries, back to AC, then rectified in every server power supply; more stages, higher current
800V DC distribution
Higher voltage carries the same power at lower current, so thinner conductors, lower resistive loss and fewer conversion stages between grid and rack
What you can learn from this
Power equals voltage times current, and resistive losses scale with current squared. Delivering 1 megawatt at 400 volts requires 2,500 amps, while at 800 volts it requires 1,250 amps. Loss in a conductor is proportional to the square of the current, so halving the current cuts that loss to a quarter and allows thinner, cheaper busbars. This is the whole physical argument for higher-voltage distribution as rack power climbs into the hundreds of kilowatts. The same logic explains why long-distance transmission lines run at hundreds of kilovolts.
Every conversion stage costs efficiency and space. A traditional data centre path takes medium-voltage AC from the utility, steps it down, converts to DC for batteries in the UPS, converts back to AC for distribution, and finally rectifies to DC inside each server power supply. Each conversion is perhaps 95 to 98 percent efficient, and the losses multiply along the chain. Distributing DC at high voltage lets operators remove stages and place a single conversion close to the rack.
Grid-forming inverters set the frequency rather than follow it. Most inverters attached to solar or batteries are grid-following: they synchronise to the voltage waveform already present and inject current. A grid-forming inverter generates its own stable voltage and frequency reference, which lets it start a microgrid from black, absorb sudden load swings and stabilise a weak connection. AI clusters produce sharp power steps when training jobs start and stop, so millisecond response is a functional requirement, not a luxury.
Energy storage on site turns a data centre into a controllable load. Batteries behind a grid-forming converter can smooth those power steps so the utility sees a flatter demand curve, ride through short outages, and in some markets earn revenue by reducing draw during peaks. This is why power-conversion vendors describe storage integration as part of the data centre product. Interconnection capacity is scarce, and a site that can shape its own load is easier for a utility to approve. Utilities increasingly ask for this flexibility as a condition of connection.
Manufacturing capacity is quoted in gigawatts because that is what customers order. A factory rated at 27 gigawatts a year is describing the total electrical output of the converters it can ship, not its own consumption. Comparing that with a single large AI campus, which can be planned at a gigawatt or more, shows the scale at which the supply chain is now sized. It also shows why acquirers value proven production lines and not only designs.
We teach this
How to use this in practice
Calculate the current for a rack at three voltages. Take a 120 kilowatt rack and work out the amps required at 208V AC, 415V AC and 800V DC, then estimate resistive loss for a 20 metre run using a published copper resistance table. Done is a small table you could explain to a colleague showing why cable sizes shrink as voltage rises. Then note which of the three could use a standard cable gauge you already recognise.
Draw the power path from utility to GPU for a conventional facility. Sketch each transformer, switchgear, UPS, PDU and server power supply as a box, and label each with an assumed efficiency. Multiply them together to get end-to-end efficiency, then draw a second version with an 800V DC bus and count the stages you removed. Done is two diagrams side by side with an efficiency figure under each. The difference between the two figures, multiplied by a 100 megawatt load, is the loss an operator is paying for.
Read one grid-forming inverter datasheet. Download a public datasheet from any vendor and find the response time, the operating modes and whether black-start is supported, then note which of those figures the marketing page omitted. Done is a half-page summary in your own words of what the device does when the grid disappears. Compare its response time with the millisecond figure the article attributes to EPC Power's platform, and note whether the datasheet quotes it under the same conditions.
Model the power step of a training job. In a spreadsheet, plot a cluster ramping from idle to full load in one second and back, and calculate how much energy a battery would need to supply to smooth that step over ten seconds. Done is a chart plus the kilowatt-hours figure, which gives you an intuition for why storage sizing sits alongside converter sizing. Double the ramp rate and watch how the required energy changes, which shows why converters need headroom.
Sources
- Flex Pays $4.4B for EPC Power as AI Data Centers Push 800V Architecture — Data Center Knowledge
Our reporting is an original summary; full coverage is at the links above.
Don't just read about it — build it.
Square 1 teaches the skills behind the headlines, with every line of your work graded by AI. Find your starting point in 3 minutes.
Get your free skill report