A contract manufacturer bets big on power electronics
Flex announced on September 3 that it will acquire EPC Power, a California-founded power conversion specialist, for 4.4 billion dollars. EPC Power CEO Jim Fusaro framed the rationale plainly: as demand for AI infrastructure accelerates, customers need power systems that are more intelligent, efficient and resilient. Flex CEO Revathi Advaithi went further, calling it evidence of a generational shift in power architecture driven by rising power density and the changing demands of digital infrastructure.
The price tag works out to roughly 5.5 times EPC Power's expected 2026 revenue of about 800 million dollars, a rich multiple for a hardware manufacturer, but one that reflects where the real bottleneck in AI infrastructure now sits. EPC Power has deployed more than 15 gigawatts of power conversion capacity across 62 countries and is scaling its Fountain Inn, South Carolina facility to 40 gigawatts of annual capacity, adding roughly 275 jobs there. The deal is expected to close in the fourth quarter of 2026, financed through a mix of debt and equity with committed financing from Citi and Bank of America.
Why power conversion, and why now
Every gigawatt of GPU compute that gets announced needs an equivalent amount of electrical infrastructure to actually convert grid power into something a rack of accelerators can use, and to do it reliably as AI training and inference loads spike and drop in ways traditional data center power systems were never designed for. EPC Power specializes in exactly that: power conversion and grid-forming technology built for the 800 volt architectures that next-generation AI data centers are adopting to handle higher power density per rack.
That is a narrower, more specialized problem than it sounds. Traditional data centers ran on relatively stable, predictable loads, sized once and rarely revisited. AI training clusters create violent power swings as thousands of GPUs synchronize compute cycles and then idle together during checkpointing, and grid operators have started pushing back hard on facilities that cannot smooth those swings before they hit the transmission network. Grid-forming power electronics are one of the few technologies that can absorb that volatility locally rather than exporting it to the grid, which is exactly the capability utilities are increasingly requiring as a condition of interconnection, and why demand for it has outpaced almost every other category of data center hardware this year.
The spinoff signals where Flex thinks the real value sits
Flex is not just bolting EPC Power onto its existing contract manufacturing business. It plans to fold the acquisition into its Cloud and Power Infrastructure segment and then spin that unit off as an independent, publicly traded company in the first quarter of 2027. Flex is projecting roughly 40 percent organic revenue growth for that segment in 2027 and is targeting EBITDA margins near 30 percent, well above what a typical contract manufacturing business commands.
That spinoff plan tells you how Flex's own leadership is pricing this market. A standalone public company focused purely on cloud and power infrastructure can command a growth-stock multiple that a diversified contract manufacturer never could, and separating it lets investors bet directly on AI power infrastructure demand without diluting that thesis with Flex's lower-margin legacy electronics assembly business. It is the same logic driving data center operators toward their own IPOs this year: pure-play AI infrastructure exposure is what public markets currently want to buy.
The real supply chain bottleneck is shifting away from chips
For the past two years, most enterprise conversations about AI infrastructure constraints have centered on GPU availability, and understandably so given how tightly Nvidia rationed allocation through 2024 and 2025. That constraint is easing as Nvidia, AMD and the hyperscalers' own custom silicon programs ramp production in parallel. What has not eased, and what deals like this one are a direct response to, is the availability of transformers, switchgear, and power conversion hardware capable of handling AI-scale density. Lead times on grid transformers alone have stretched past two years at some utilities, and power electronics manufacturers report backlogs that now extend further out than GPU allocation windows, a reversal few infrastructure planners saw coming even a year ago.
That shift changes what CTOs and infrastructure leaders should actually be tracking. A cloud provider or colocation partner's ability to deliver new capacity on the timeline they promise now depends more on their power electronics supply chain than on their chip allocation from Nvidia. Vendors who have locked in dedicated power conversion capacity, through deals like Flex's or through direct long-term supply agreements, are better positioned to hit stated delivery dates than vendors still shopping the open market for transformers and switchgear.
What this means for capacity planning
If you are negotiating a colocation or wholesale capacity agreement for AI workloads over the next 18 months, add a specific question to your due diligence: who supplies your provider's power conversion equipment, and do they have committed capacity or are they buying on the spot market. A provider whose power electronics come from a vertically integrated supplier like the newly combined Flex and EPC Power has a materially different risk profile than one relying on fragmented, oversubscribed vendors.
This deal also validates a broader thesis worth building into your infrastructure roadmap: power delivery capability, not compute availability, is becoming the true differentiator between cloud vendors who can actually deliver committed capacity on schedule and those who cannot. Expect more consolidation in this specific corner of the supply chain over the next year, as other contract manufacturers and industrial conglomerates follow Flex's lead into power electronics, and price that consolidation into your longer-term vendor negotiations now rather than after your next capacity request gets delayed. The companies that lock up power conversion supply early will be the ones still hitting their delivery dates when everyone else is explaining a slip.
