Flex’s $4.4B EPC Power Buy Is an AI Power Tollbooth Bet
AI is not being held back by chips. It is being held back by electricity — and Flex has paid $4.4 billion to own a tollbooth between the grid and the GPU.
Most AI companies are fighting over chips while the real choke point is the electricity getting into the building. Flex has just paid $4.4 billion in cash for EPC Power because selling the shovel is good; owning the only road to the mine is better.
On September 3, Flex announced a definitive agreement to buy EPC Power, a California-based power-conversion specialist. The deal is expected to close in the fourth quarter of 2026, after which EPC will sit inside Flex’s Cloud and Power Infrastructure business — the very division Flex plans to spin into a separately listed company in the first quarter of 2027. ([investors.flex.com](https://investors.flex.com/news/news-details/2026/Flex-to-Acquire-EPC-Power-Adding-Leading-Power-Conversion-Capabilities-for-AI-Data-Centers-and-Grid-Applications/default.aspx))
That is not a routine bolt-on. It is a deliberate attempt to turn a giant manufacturer into the owner of a far more valuable piece of the AI infrastructure stack.
The $4.4 billion bet in plain English
EPC Power does not make glamorous consumer tech. It makes the equipment that converts, regulates and stabilises power for data centres, utility-scale battery storage and microgrids. Think rectifiers, DC-to-DC conversion, controls that help stabilise the grid, and the early foundations of solid-state transformers. These are the systems that make a grid connection usable for a warehouse full of power-hungry AI hardware. ([investors.flex.com](https://investors.flex.com/news/news-details/2026/Flex-to-Acquire-EPC-Power-Adding-Leading-Power-Conversion-Capabilities-for-AI-Data-Centers-and-Grid-Applications/default.aspx))
That matters because modern AI data centres are changing the electrical design brief. Flex says EPC’s systems are built for next-generation 800-volt DC architectures — designed to reduce conversion steps and deliver power more efficiently to increasingly dense compute loads. EPC also brings grid-forming technology intended to smooth large swings in demand and support grid stability. ([sec.gov](https://www.sec.gov/Archives/edgar/data/866374/000119312526382492/d108491dex992.htm))
The price tells you where the confidence sits. Flex says EPC is expected to generate about $800 million of revenue in calendar 2026, which puts the headline purchase price at roughly 5.5 times revenue. Flex is also forecasting about 40% organic revenue growth in 2027 and EBITDA margin expansion to about 30%. Those are forecasts, not cash in the bank, so don’t get carried away. But the arithmetic explains the cheque: Flex is buying a business it believes could produce roughly $1.12 billion in 2027 revenue if that growth rate lands. ([sec.gov](https://www.sec.gov/Archives/edgar/data/866374/000119312526382492/d108491dex992.htm))
The seller matters too. Goldman Sachs Alternatives and Cleanhill Partners are selling EPC after backing its expansion. EPC says its domestic manufacturing footprint grew nearly tenfold during that period. Flex says the company has more than 15 gigawatts deployed across 62 countries, with annual U.S. manufacturing capacity expected to exceed 30 gigawatts in 2027. ([epcpower.com](https://www.epcpower.com/news/epc-power-announces-sale-to-flex-for-4-4b))
That is what Flex is buying: not just clever hardware, but manufacturing capacity, customer relationships, engineering depth and the right to be in the conversation before a data-centre operator has finalised its electrical design.
Flex is buying position, not merely product
Here is the uncomfortable truth for anyone who thinks the AI boom is simply an Nvidia story: the GPU is only useful if the building can feed it, cool it and keep it alive when the grid has a wobble.
Flex already sells power, cooling, compute infrastructure and manufacturing services. EPC fills a strategic gap between the utility connection and the rack. Its platform is designed to connect 800V DC loads to grid voltages and combine work traditionally handled by uninterruptible-power systems and AC distribution equipment. That is why Flex keeps using the phrase “grid to chip.” ([sec.gov](https://www.sec.gov/Archives/edgar/data/866374/000119312526382492/d108491dex992.htm))
It is also why this deal is smarter than it first appears.
A commodity manufacturer gets paid for assembling what somebody else designed. An infrastructure platform with proprietary controls, power electronics and long customer qualification cycles can earn a much fatter return. The latter gets specified earlier, becomes painful to replace and can capture more dollars per megawatt.
That last bit is the real game. Flex’s investor material explicitly points to expanding “content per megawatt” as AI infrastructure shifts toward integrated architectures. In normal-person language: rather than selling one box, Flex wants to sell more of the system around every megawatt of data-centre capacity. ([sec.gov](https://www.sec.gov/Archives/edgar/data/866374/000119312526382492/d108491dex992.htm))
I like that logic. The best businesses do not hunt endlessly for new customers. They increase the amount they can reliably sell to an existing customer at the moment the customer has no appetite for failure.
A data-centre developer can save money on office chairs. They will not happily save money on the equipment that stops a multi-hundred-million-dollar facility from going dark.
Why the spin-off changes the deal
The transaction gets more interesting because Flex is not keeping the asset buried inside the parent forever.
EPC is slated to join Cloud and Power Infrastructure in Q4 2026. Then Flex intends to separate that division into a standalone public company in Q1 2027. The obvious ambition is to hand public investors a cleaner AI-infrastructure story: power conversion, cooling, compute integration and related services, rather than a broad manufacturing group with a collection of unrelated moving parts. ([investors.flex.com](https://investors.flex.com/news/news-details/2026/Flex-to-Acquire-EPC-Power-Adding-Leading-Power-Conversion-Capabilities-for-AI-Data-Centers-and-Grid-Applications/default.aspx))
That can work. Markets often value a focused growth platform more generously than a conglomerate where the good business is hard to see.
But it also makes the execution risk very real. Flex is trying to acquire, integrate, finance and then separate a $4.4 billion asset in short order. Anyone who has actually run a business knows these are not PowerPoint verbs. Every one creates distractions: systems, sales incentives, reporting lines, leadership retention, customer communication, financing documents and regulatory approvals.
Flex has secured a 364-day bridge facility of up to $4.4 billion from Citi and Bank of America as a backstop, while saying it expects to replace it with a combination of debt and equity financing. The purchase agreement includes customary regulatory conditions, including the U.S. Hart-Scott-Rodino waiting period. If the deal has not closed by December 31, 2026, either side can terminate it, subject to two automatic three-month extensions in certain circumstances. ([sec.gov](https://www.sec.gov/Archives/edgar/data/866374/000119312526382492/d108491d8k.htm))
None of that means the deal is dodgy. It means it is a deal, not a completed fact. There is a difference, and investors forget it every time a shiny press release lands.
The overlooked angle: this is a grid deal wearing an AI hat
The AI label will get the headlines. The sturdier thesis is the power grid.
EPC serves data centres, yes, but also utility-scale storage and microgrids. Its grid-forming technology is meant to respond quickly to load swings, while its systems can support backup, storage and onsite generation configurations. ([epcpower.com](https://www.epcpower.com/news/epc-power-announces-sale-to-flex-for-4-4b))
That diversification matters. AI spending will be lumpy because every capital boom is lumpy. Some projects will get delayed. Some customers will overbuild. Some data-centre dreams will turn out to be spreadsheets with a car park attached.
A business that only sells to frontier-model builders is exposed to that mood swing. A business that also sells into grid modernisation, energy storage and reliability has more legs under the table.
The contrarian view, then, is this: Flex may not be paying up for an AI hardware supplier at all. It may be paying up for an electrical-infrastructure business that happens to have AI as its most urgent new customer.
That is a more durable idea — provided the technology really is differentiated and the margins arrive.
And there is the catch. A 5.5-times-revenue deal leaves very little room for average execution. Flex needs the 2027 growth, the expected margin expansion and the integrated-system cross-sell to show up in actual results. If EPC becomes another expensive box inside a sprawling supply chain, this purchase will look bloody ambitious in hindsight.
What this means for you
For founders: stop chasing fashionable markets and look for the constraint that everybody else is pretending does not exist. AI demand is sexy. Getting power, cooling, permits, components and reliable operations in place is not. The unsexy constraint is often where the pricing power lives.
For operators: map your business like a data centre. Where does work back up? Where does a customer face a delay, risk or failure they cannot tolerate? Do not ask only, “What product can we sell?” Ask, “At what point are we expensive to replace?” Build there.
For investors: separate the headline from the mechanism. The headline is Flex buying an AI-power company for $4.4 billion. The mechanism is a planned standalone infrastructure platform trying to own more of the grid-to-chip bill of materials. Watch the financing, closing timetable, 2027 revenue growth and margin delivery — not the conference-call adjectives.
And for anyone building a company: learn this lesson properly. You do not need to own the whole market. You need to own the part of it that becomes indispensable when the stakes are high. Flex has just spent $4.4 billion betting that power conversion is one of those parts. Time will tell whether it bought a tollbooth — or paid tollbooth money for a very nice transformer.