Why is power the bottleneck for AI data centers?

Because an AI data center is a factory whose only significant input at volume is electricity. K.R. Sridhar, founder, chairman and CEO of Bloom Energy, describes it that way directly: the input is electricity and data, data is everywhere, the chips are the machines, and the output is intelligence. That leaves electricity as the one scarce input, which is why he concludes that clearly electricity becomes the key factor. The constraint is also a scheduling problem rather than only a supply one. Chips arrive on a purchase order, but delivered power depends on interconnection queues, permits and local fuel supply, none of which compress when you raise more money. The same point shows up from people with nothing to sell in energy: on Sequoia's stage Nvidia's Jensen Huang places energy at the bottom of the AI buildout, saying the first layer of the cake is energy.

How long does it take to build a data center and get power to it?

Sridhar's figure for a greenfield facility at the scale of roughly 2 gigawatts is that it takes 12 to 18 months minimum, and he is describing the building rather than the power. The gating items he lists are the construction trades, the copper inside the facility, the cooling, and the availability of people who can build at that scale, all of which he says are stressed at once. His claim about generation is that on-site power can be delivered faster than the shell goes up, with the remaining constraints being whether the customer finishes on time, whether the permits come through, and whether gas can reach the site. Bloom's own proof point is a contract to deliver more than 50 megawatts within 90 days of signing for a delayed Utah project, which he says was energized in 55 days.

What is a solid oxide fuel cell and why do data centers use them?

A solid oxide fuel cell converts fuel, usually natural gas, into electricity through an electrochemical reaction rather than by burning it to spin a turbine. Bloom Energy's product, the Energy Server, is built from modules of roughly 50 kilowatts each. Sridhar's pitch rests on three properties that he argues matter more than the generation method. They can be sited on the customer's premises, so they bypass part of the grid interconnection queue. They are modular, so any one unit can be hot swapped without taking the facility down. Sridhar's contrast is that a single large turbine averages annual availability somewhere in the low 90s as a percentage over a 20-year life and therefore needs a second turbine as backup. And because they are solid state, he says they ramp up and down in milliseconds, which suits the spiky load profile of AI inference without a bank of batteries to absorb the ramp.

Does the cost of data center power affect what I pay per token?

Yes, indirectly, and the chain is shorter than it looks. Sridhar calls electricity the only real cost input into manufacturing intelligence, so your price per token is an electricity price with several layers of abstraction stacked on it. The useful part of his framing is that he refuses to treat it as a cost question at all: "it's neither price nor cost, it's value." His reasoning is that the token has a value, and the question he asks is what that entire stream looks like end to end. The practical implication for a startup is that the lever is rarely negotiating a better rate. It is the shape and efficiency of what you run: how much of your traffic needs a frontier model, how spiky your load is, and whether your margin survives a flat token price rather than a falling one.

Data Center Power Is the Binding AI Constraint

Ask most founders what gates their AI roadmap and the answer is chips, money or hiring. Ask the people who build the physical layer and you get a different answer, and it is not on that list.

K.R. Sridhar co-founded the company that became Bloom Energy in 2001, under its original name Ion America. Before that he was a professor of aerospace and mechanical engineering at the University of Arizona and director of its Space Technologies Laboratory, where he led NASA-funded research on converting gases in the Martian atmosphere into oxygen for life support and propulsion. Run that process in reverse, turning fuel into electricity instead, and you have the solid oxide fuel cell Bloom sells today.

In a long conversation with Harry Stebbings on 20VC, he made a claim worth taking seriously even if you will never buy a megawatt: what rations AI right now is not intelligence and not silicon. It is data center power, in the literal sense of electricity energized at the rack, and it runs on a clock that funding does not compress.

Why data center power became the binding constraint

Sridhar's first move is to separate two things founders routinely merge, the market's mood and the physical demand curve.

"People confuse markets, stock prices, and infrastructure, and secular changes."

His case for the demand being secular rather than cyclical is a rhetorical question rather than a forecast: he asks when any person in any civilization last said there was too much intelligence and decided to stop. He expects bumps and pauses along the way, and says so plainly, while arguing the trajectory survives them.

That part is a vendor's view of his own market, so it deserves the discount you would apply to any vendor. What makes it worth more than that is how often the same conclusion arrives from people with no fuel cells to sell. On Sequoia's stage, Nvidia founder and CEO Jensen Huang lays the AI buildout out as a five-layer cake and puts energy underneath everything else: "The first layer of the cake is energy." His own accounting prices each gigawatt of AI factory at roughly 50 billion dollars. An a16z panel introducing the firm's Machine Age fund is blunter about present supply: "power, cooling, memory, GPUs, like you name it, we're out of it."

Where Sridhar is more specific than the usual permitting complaint is in naming what the bottleneck actually is. Not technology, and not capital:

"The single biggest bottleneck that I would say we face in terms of introducing a new concept into society is the friction associated with new ideas."

Large infrastructure was deliberately designed and regulated to move slowly, for reasons he accepts were good ones. His argument is that deliberate friction becomes an impediment once the thing on the other side of it moves on an AI clock, and that the friction is unevenly distributed, so a region that throttles while others do not pays for it. He still declines to bet against American capacity to work around that: "don't short the US, don't short Silicon Valley."

Electricity is the only real input to intelligence

This is the reframe worth taking. Sridhar describes an AI data center as a factory and then points out how strange its bill of materials is. Unlike a chemical plant, where ores and feedstocks arrive at the gate, here:

"the input into this factory is simply electricity and data"

Data is everywhere, so it is not the scarce one. "The machines are the chips. The output is intelligence." That leaves a single input arriving at volume, which is why he concludes that "clearly electricity becomes the key factor." His framing for the whole shift: "for the first time in human history, we are manufacturing intelligence."

Follow the logic to your own P&L and it says something uncomfortable. Your cost per token is an electricity price with several layers of abstraction stacked on top of it. You do not see a kilowatt hour on your invoice, but the person who sells you tokens does, and their ability to sell you more of them next year is set by how much power they can get energized this year.

He is also realistic about how early this is. "Electricity in the world is a $5.5 trillion market," he says, and against that, "last year our revenues were $2 billion." The companies positioned to relieve the constraint are small relative to it, which is the part of the story that should make you cautious about anyone promising the shortage ends soon.

The lead time you cannot engineer around

The most operationally useful thing in the interview is a number about buildings, not generators.

"it takes 12 to 18 months minimum to build something like that"

He is describing a greenfield facility at roughly 2 gigawatts, and in the same answer he stretches it to a year and a half or two, so 12 to 18 months is his most favourable version of the number. The gating items he lists are not exotic. The construction trades, the copper that goes inside the building, the cooling, and the supply of people who can build at that scale, all stressed simultaneously. Against that baseline, his claim about generation is the inversion: "We can stand up that power faster than they can build a data center." Bloom's own remaining constraints, by his account, are whether the customer finishes the shell on time, whether the permits come through, and whether gas supply can physically reach the site. The promise he makes to customers who put down a firm deposit is deliberately narrow: "We will not be the bottleneck."

The proof point he offers is a contract rather than a projection. A data center in Utah being built for an Oracle customer had its power systems delayed, so Bloom signed to deliver more than 50 megawatts within 90 days: "we would do 50 plus megawatts, which is an unbelievable speed." He says they energized it in 55 days. That relationship has since been formalized at a different order of magnitude, with Bloom and Oracle signing a master agreement under which Oracle intends to procure up to 2.8 gigawatts of fuel cell systems, of which an initial 1.2 gigawatts is contracted.

Worth noting that none of this started with AI. By Sridhar's account, Bloom's first mission-critical data center deployment was in 2013, at the eBay facility in Utah that also ran PayPal, and he frames it as a site that needed power and did not want to relocate. eBay's own announcement at the time framed the same project differently: 30 Bloom servers totalling 6 megawatts, designed in as the primary energy source with the utility grid kept only as backup, and presented as a clean energy decision. Either way, ten to twelve years of 4 to 15 megawatt data center installations is what made the 55-day job credible to sign.

On capacity, Stebbings puts current manufacturing at a gigawatt a year, which Sridhar does not dispute, and Sridhar's own number for the end of this year is "We'll be over 2 gigawatts. We said that publicly." That matches Bloom's stated plan to reach 2 gigawatts of annual production capacity by the end of 2026, which its FY2025 annual report describes as doubling the Fremont run rate from about 1 gigawatt. He declines to give a number past that, citing forward guidance rules, which is the right answer and also a reminder that nobody has promised you the ramp you are modelling.

Three things to take from this if you are planning capacity:

Ask for energization dates, not availability dates. A provider telling you capacity is available in Q3 may mean the racks, the shell or the power, and the three can be a year apart. The date that constrains you is when electrons reach the rack.
Treat the permit as a dependency with an owner. Sridhar lists permits alongside gas supply as a real constraint on his own delivery. If your roadmap depends on a facility that depends on a permit, that permit belongs in your risk register, not in your vendor's.
Assume the lead time is the price. This is the same conclusion the capex cycle produces from the other direction, which the AI capex buildout works through: in a supply constrained market the binding term is when you can get capacity, not what it costs per hour.

Modular beats monolithic, in power and in your own stack

The best transferable idea in the interview is an architecture argument, and it generalizes well past energy.

Power a 500 megawatt data center with one large turbine and you inherit the turbine's duty cycle. Sridhar's figure is that such a machine's availability, averaged over a 20-year life, lands "somewhere in the low 90s" as a percentage, because like a jet engine or a car it has to stop for maintenance. His question is what you do with the remaining tenth of the year, given that "The data center has to be always on." When data centers were small the grid absorbed that, but at gigawatt scale there is no grid standing by that is large enough to back you up. So you buy two turbines to run one.

Bloom's answer, in the Energy Server product line that eBay's 2013 announcement already called Bloom servers, is granularity. The system is built from modules of roughly 50 kilowatts, and "So you can hot swap any one when you're working with it." He draws the analogy to the thing it powers: a data center is server blades in racks precisely so that no single blade or rack failing takes the facility with it, and he claims the same topology for his own product, which means the overbuild needed for redundancy is a small percentage rather than a second machine.

Take that one level up and it is an argument about how you consume compute, not how you generate power. A single provider, in a single region, serving a single model, is the 500 megawatt turbine. It is cheaper per unit right up to the maintenance window. The modular version costs a little more and fails in pieces: a second inference path you have actually tested, an eval suite good enough to tell you what switching would cost, and a router that can degrade to a smaller model instead of returning an error. What that trade looks like when the decision is renting versus owning hardware is the subject of the GPU cluster decision.

Spiky load is a cost structure, not a footnote

The second technical argument matters for anyone who runs inference, because it is about load shape rather than load volume.

"AI loads are like the brain." They spike into heavy activity, draw a great deal of power, then fall away. A mechanical generator cannot follow that; Sridhar's comparison is that even a fast car takes a few seconds from zero to 60 because something physical has to spin up. A solid state device has no such limit: "we can ramp up and ramp down in power in milliseconds, which is what the AI loads need." The commercial consequence he claims is that the batteries otherwise installed to absorb the ramp stop being necessary.

The same a16z panel fills in what this is doing inside the building. Compute density is climbing by something like 70x, cooling is "moving from air to liquid as a requirement," and at the resulting power per rack "AC power doesn't work anymore," which pushes the facility to high voltage DC distribution. One panelist asserts, as a figure we could not trace to any published statistic, that only about 2 percent of electricians in the US are certified on DC power. Treat the number as unverified and the constraint as real: the same panel notes that Meta runs a free training program to close exactly that gap.

For a startup, the thing to extract is that your load shape is a cost input in its own right. Batching tolerant work, scheduling it off your own peaks, and smoothing a jagged inference profile are not micro-optimizations at the margin; they are the shape of the bill that your provider is paying and passing on. The arithmetic on where that money actually goes is in what AI inference really costs.

Value per token, not cost per kilowatt hour

Asked when customers stop caring about speed and start caring about cost, Sridhar rejects the frame:

"it's neither price nor cost, it's value"

His reasoning is that you are generating tokens, the token has a value, and the question he poses is what that entire stream looks like end to end. His term for the product is "What Bloom makes is designer electricity." Grid power, in his telling, is one-size-fits-all, so you buy it and then build the gear that adapts it to your actual need, and you own, operate and maintain that gear, and it fails, and the failures are yours.

That is a vendor's argument for vertical integration and should be read as one. The transferable part is the pricing instinct underneath it, which is the same mistake founders make one layer up: pricing a product on the cost of the tokens it consumes rather than on the value of the work it completes. Charging for inference by the unit of inference hands your margin to whoever sets the token price. How to avoid that is the whole subject of AI pricing for startups, and the wider frame for running a company on these inputs is our pillar on AI for startups.

Power at the edge is a resilience argument

The part of this that should interest anyone responsible for security is not generation at all. It is topology.

Sridhar's case for distributed power starts from what a fully digital world cannot tolerate: "having power outages for any reason is unacceptable" once surgery, logistics and transport depend on compute responding in time. Then he describes the delivery mechanism we currently rely on, a line of poles and wire running hundreds of miles, as "capable of disruption by mother nature, capable of destruction by bad actors," whether that is a cyber intrusion or a physical attack. He also makes the latency argument: if the decision has to be made now, a large facility far away is not an answer.

He extends it to sovereignty, and this is where he is most direct. Asked whether energy sovereignty matters more than the model sovereignty everyone discusses, his answer is that after food, "next after food, it's energy sovereignty, period," on the grounds that wars have historically been fought over water, food and energy, and that this one is avoidable. His framing of why distribution is the fix: "when power is democratized, access is not restricted by people who are in power."

Two honest caveats, and these are my reading rather than his. On-site generation relocates the dependency rather than removing it: a fuel cell farm needs a fuel supply, which is a pipeline, which is its own single point of failure with its own threat model. And concentrating your generation at the same site as your compute removes the geographic diversity that made grid failures survivable. The right conclusion is not that distributed is safe and centralized is not. It is that power belongs in the dependency map you already keep for providers and regions, with the same question asked of it: what happens to us when this one thing stops, and have we ever tested that. The broader pattern of treating physical constraints as strategy rather than as someone else's problem runs through deep tech startups and the clock that is their real moat, and the most extreme response to the terrestrial energy constraint, moving the compute off the planet entirely, is the subject of the case for data centers in space.

What to do this week

Get an energization date in writing. Ask your cloud, colo or inference provider one question: for the capacity step we expect to need in 12 months, when does the power actually energize, and what is the permit status. Vague answers are information.
Write down your load shape. Not your monthly token volume, the profile: peak to median ratio, by hour. You cannot smooth a curve you have not plotted, and the curve is what your provider is paying for.
Find the turbine in your own architecture. Name the single provider, region or model whose maintenance window takes your product down. Then decide whether you are accepting that or fixing it, deliberately, in writing.
Re-run your unit economics on a flat token price. Not rising, just flat for 24 months. If the model only works on a falling price curve, you are betting your margin on somebody else's power procurement.
Add power to your risk register. One line, one owner, one question: what is our exposure if our primary region loses capacity for a week. If nobody owns that line today, the answer is that you will find out during the outage.
Separate the vendor's thesis from the physics. Sridhar is selling the solution to the problem he is describing. The 12 to 18 month build time, the interconnection queue and the permitting friction are real regardless of who relieves them. Plan around the constraint, not around any one company's promise to remove it.

The course is the structured version of this: AI Operating System for Startups covers how to budget compute, instrument the spend and build the seams that keep a provider or model swap cheap.

Sources

Bloom Energy CEO: Why We Aren't in an AI Capex Bubble (20VC with Harry Stebbings), the interview with K.R. Sridhar that this article distills.
NVIDIA's Jensen Huang on Building the Dynamo of the Intelligence Age (Sequoia Capital), for the five-layer framing that puts energy at the bottom of the AI buildout and the figure of roughly 50 billion dollars per gigawatt.
Why Top Founders Are Racing Into AI Infrastructure (a16z), for the state of supply across power, cooling and memory, and for rack density, liquid cooling and the DC power skills gap.
Bloom Energy and Oracle Expand Strategic Partnership to Deploy up to 2.8 GW (Bloom Energy investor relations), for the scale of the Oracle relationship beyond the Utah project.
Bloom Energy says it's on track for 2 GW annual production capacity (Utility Dive), for the 2 gigawatt annual production capacity target. Bloom's FY2025 annual report, credited without a link because the SEC blocks automated readers, is the source for that target being a doubling of the Fremont run rate.
Profiles: K.R. Sridhar, founder, chairman and CEO of Bloom Energy and previously a professor of aerospace and mechanical engineering at the University of Arizona, and Jensen Huang, founder and CEO of Nvidia.