Amit Mathrani: “These companies didn’t secure their AI futures by buying model companies or chip designers. They went and bought power.”
A C-Suite Thought Leader Interview
A leading energy sector strategist from Rabobank on why power has crossed from procurement to strategy and what that means for every executive who isn’t a hyperscaler.
Amit Mathrani is an Executive Director of Energy Transition Research at Rabobank North America, where he leads sector research on power markets, renewables, and energy infrastructure. Before joining Rabobank’s RaboResearch team, he led corporate strategy at Consolidated Edison and National Grid, focusing on electric business growth across New York and the Northeast. Prior to that, he worked in management consulting across capital-intensive industries including mining, oil and gas, and industrials. His recent research has reframed the data center power crunch as a strategic and financial challenge for senior executives, not merely an infrastructure management problem.
C-Suite: Your report describes a structural mismatch at the heart of the AI build-out: data centers go up in 12 to 24 months, but grid connections take 36 to 84 months. How big is this problem?
Bigger than most executives have absorbed. US data centers consumed roughly 176 terawatt-hours of electricity in 2025. By 2030, that figure is forecast to reach somewhere between 420 and 728 terawatt-hours. The incremental demand alone is equivalent to the combined annual electricity consumption of New York and California. The five largest hyperscalers – the cloud and AI infrastructure giants building data centers at a scale no other companies can match – plan to spend between $745 billion and $775 billion on capital expenditure in 2026 alone, and every one of them reported on their most recent earnings call that AI compute demand continues to outpace available capacity.
That scale is the context for the mismatch. In California, interconnection processing times now exceed seven years. AEP Ohio told PJM’s Load Analysis Subcommittee that none of the 13 gigawatts of new data center load in its territory can be reliably served until Q4 2031. Companies setting the pace for AI innovation have concluded they cannot wait. They are taking power procurement into their own hands, and in doing so constructing what is effectively a parallel energy system.
C-Suite: Why does that make it a boardroom issue rather than a facilities management problem?
Because of where the real cost lives. Power is only 10 to 15 percent of a data center’s total cost of ownership. But it is the single binding constraint on whether a facility generates revenue at all. A data center’s semiconductor assets cost $10 to $15 billion and depreciate at $3 to $5 billion per year. The moment those chips arrive, the facility needs to be operational. You cannot wait four to seven years for a grid connection. That is why companies are paying $100 to $165 per megawatt-hour for behind-the-meter power versus $90 to $95 for grid alternatives. The cost of the power is not the issue. The cost of idle hardware is.
And then there is the price signal in the capacity markets. PJM’s capacity auction cleared at $333 a megawatt-day in December 2025, against a price of $29 two years earlier. PJM estimated the clearing price would have approached $530 without the regulatory cap. That is not a procurement fluctuation. That is a signal that belongs in the boardroom.
C-Suite: Can a non-hyperscaler realistically build its own power solution?
My honest answer, if I were sitting on the board of a Fortune 500 industrial today, is: no. And I would say that directly. The risk is asymmetric in a way that makes self-supply a dangerous idea for most companies.
A hyperscaler can build or even overbuild a behind-the-meter solution and either sell the excess into the wholesale market or absorb the write-down. They have trillion-dollar balance sheets. A non-hyperscaler has none of that. We have seen what happens when companies get that wrong: the merchant gas turbine cycle of 1997 to 2002 ended in a wave of stranded asset write-downs. Enron is the most famous name, but there were many others.
The framework I would offer has three variables. Scale: unless you are thinking about at least 500 megawatts to a gigawatt, the economics do not work. Below 100 megawatts, they definitely do not work. Duration: can you credibly commit to a 20- or 25-year contract? Most companies will overestimate that certainty, and the CFO signing that contract will probably have left the company before it expires. Balance sheet: can you absorb that asset if your demand curve flattens? If the answer to any of those three is no, you should not be owning your own power plant.
C-Suite: So what are the realistic options for a company below that threshold?
The most underappreciated tool in this conversation is the virtual power purchase agreement, or VPPA. The logic is straightforward: a VPPA means the renewable asset or power plant sits in an entirely different electricity market from where your facility is located. You never touch the plant. The developer builds it, operates it, finances it, and carries the construction and permitting risk. What you get is long-term price certainty and the financial benefit when market prices rise above your contract strike, which, given the trajectory of capacity markets, they increasingly will. If prices fall, you owe a modest difference. It is not a perfect hedge, but it gives a company genuine optionality without a physical asset on its balance sheet.
Mars recently signed a wind VPPA in Europe on exactly this logic. The company gets the economic benefit of renewable ownership, reduces its carbon exposure, and manages long-term price volatility, without building a wind farm. That is capital-light exposure to an infrastructure project that keeps balance sheet flexibility intact. It is the right model for most companies sitting below the self-supply threshold – with one important caveat: a VPPA is a financial instrument, not a physical one. It does not help you connect to the grid or power new facilities.
Beyond VPPAs, there are smaller-scale options: rooftop solar on an industrial facility, battery storage to offset peak hours, and fractional capacity arrangements with merchant developers. The electric grid remains one of the greatest inventions of the past hundred years. The advice for most companies is: stay with the grid, use financial instruments to manage your exposure, and treat power as the strategic variable it has become.
C-Suite: How serious is the price volatility risk for companies that cannot self-insure?
Very serious, and not yet fully reflected in how most boards think about risk. Last year we saw a 5 percent increase in national retail rates. This year, 6 percent is expected. I think the real numbers are probably higher than the forecasts suggest, given what we are seeing in capacity markets.
The uncomfortable implication: a company that misjudges its energy position today will not just feel it in next year’s earnings. It will feel it in 2031, when it discovers it cannot expand because it did not lock in the megawatts in time. The cost of getting this wrong compounds. It is not a question of what power costs per kilowatt-hour. It is a question of what it means for your production capacity and your growth options not to have secured it.
A CFO getting this right understands that this is no longer a procurement question. It requires someone on the management team, or on the board itself, who can read a power purchase agreement the way someone reads a credit commitment. Long-duration physical assets behave very differently from the asset-light businesses most boards know how to oversee. The risk profile of a balance sheet changes materially when you start taking positions in energy infrastructure, even through financial instruments. Boards that do not have that expertise in the room are taking a risk they may not fully understand.
C-Suite: Will the data center build-out slow, and what are the real constraints?
It will be constrained, not stopped. Of the 130-plus gigawatts of co-located energy resources announced across the US, roughly a quarter is expected to actually deploy by 2030. The rest runs into a cascade of overlapping constraints. Renewable tax credits, the Income Tax Credit and Production Tax Credit that made solar and wind behind-the-meter solutions economically attractive, had their qualification timelines compressed to mid-2026 under the One Big Beautiful Bill Act, removing a major cost advantage from the renewable pipeline beyond that date. Turbine backlogs at GE Vernova, Siemens, and Mitsubishi are running four to five years out. Transformer lead times are 36 to 48 months. One constraint after another.
The technology choices being made today reflect that reality directly. Fuel cells can be deployed in as little as 90 days, the only near-term option for immediate energization. Gas engines and aeroderivatives deploy in 12 to 24 months. Heavy-duty turbines and combined-cycle plants, which are the most efficient at scale, take 36 to 84 months and face wait times of up to seven years for new orders. Gas accounts for over 80 percent of the announced behind-the-meter pipeline, not because of any preference for gas, but because it is the fastest workable option on the timeline AI infrastructure demands. The market is paying a premium for speed.
C-Suite: For companies that depend on third-party compute for their AI strategies, what should they be doing?
Treating compute access as a strategic resource, not a commodity input. The announced pipeline overstates what will actually materialize. That supply shortfall is real, and it will affect pricing and availability. Companies that wait for the market to normalize may find themselves renting their growth capability from someone else at terms they did not negotiate.
Locking in capacity now, even at a premium, may well be the more defensible decision, depending on how central AI compute is to your competitive position over the next three to five years. The question to ask is not what compute costs today. It is what it costs your business not to have it when you need it.
C-Suite: Power availability is reshaping where companies can locate and grow. What should executives be factoring into expansion decisions?
The US industrial geography is shifting faster than most corporate real estate and site selection teams have absorbed. ERCOT in Texas can get a data center interconnection in about three years, the shortest timeline anywhere in the country. Land is available, cheap, and largely without community resistance in rural Texas. PJM and MISO – the regional transmission grids stretching from New Jersey to Minnesota – account for more than 60 percent of projected US data center capacity growth but face the most severe interconnection backlogs. Demand is clustering faster than infrastructure can expand, turning what should be a national growth story into a set of regional capacity crises.
The belt pulling ahead runs from Texas through Ohio and Pennsylvania. These states are actively making it easier: the Texas Energy Fund is providing capital for gas plant developers; PJM is working on cluster connections and process reform. The coasts keep the talent advantage. They are losing the build advantage.
The uncomfortable corollary is this: if you are not hyperscaler scale, you are effectively renting your growth capability from someone else. That should be a conscious strategic choice, not something that happens by default. Where you can locate, where you can grow, whether you can execute on the strategy you are developing, all of that is being reshaped by power access. Don’t find out too late that you didn’t lock in the megawatts.
C-Suite: If you had to leave boards and senior executives with one thought, what would it be?
Power has crossed from procurement to strategy, and most boards have not fully absorbed what that means yet.
Tech companies are becoming infrastructure companies. Industrials are becoming compute buyers. Utilities are becoming financiers. The boundary between these worlds is dissolving faster than the organizational charts can keep up. Look at where the most sophisticated capital allocators in the world are putting their money: Google acquired Intersect Power to own the power pipeline. BlackRock GIP and its partners acquired a utility company. Microsoft secured Three-Mile Island. Amazon signed a contract for an entire campus. These companies did not secure their AI futures by buying model companies or chip designers. They went and bought power.
When the smartest balance sheets in the world are acquiring power plants and utility companies, that tells you where the moat is actually living.
The harder question, whether behind-the-meter power is a temporary bridge to the grid or the foundation of a permanent parallel energy system, remains open. That answer will determine the economics of this decade.
The leaders and boards that can recognize that, and reorient their capital allocation toward it, will outperform. The ones who keep defending the old boundaries won’t. The gap will be measured in enterprise value and total shareholder return. Not in basis points on the electric bill.