Lektra's founder and CEO argues that power, not chips, is now the real bottleneck for AI, and that the fastest fix runs through the solar panels and batteries homes and businesses already own.
AI's appetite for electricity is growing faster than the grid can add capacity. New data centers can wait years for a connection, grid operators are bracing for brownouts, and consumers are already seeing the strain in their monthly bills. Fixing the problem the conventional way, with new plants and new transmission lines, would take at least a decade and trillions of dollars.
Karl Andersen believes the faster path runs through the solar panels and batteries that homes and businesses already own. His company, Lektra, pays those owners for their power and uses it to run AI workloads on a distributed network of smaller data centers, a patented approach built around what Andersen calls Grid 2.0. Eight of Lektra's EdgeScale sites are already online, including a wholly owned facility in Kearny, New Jersey, with another coming online shortly in McKinney, Texas.
The company has now launched two models of EdgeStart, a self-contained server cabinet that a building owner can house in a spare room and power with an existing solar and battery system. Lektra installs the cabinet, operates it remotely, sells its computing capacity to AI customers through Lektra Cloud, and pays the owner 80% of the revenue. Andersen says the economics hold up for AI customers as well, with GPU rental rates at a fraction of what the major cloud providers charge and no egress fees.
TechEchelon spoke with Andersen, Lektra's founder and CEO, about why power and grid connection have overtaken chips as the real bottleneck for AI, which policy changes would speed up distributed compute, and what Lektra expects to look like in two years.
Sara Montes de Oca put ten questions to Andersen on the power bottleneck facing AI, the policy changes that would speed up distributed compute, and Lektra's plans for the next two years. This transcript has been lightly edited for length and clarity.
At a high level, describe Lektra and its mission.
Lektra is building Grid 2.0 as a distributed energy and computing network designed to handle AI's power demand in a way the current grid can't. We pay consumers and businesses for power from the solar and battery systems they already have, and that power runs AI compute on our own distributed network. We hold the first patent for distributed data centers built this way. Our mission is to make sure AI works for people, and that starts with building the AI compute and power infrastructure where energy already is instead of waiting a decade for the grid to catch up.
What is driving AI compute demand, and why can't the current cloud model keep up?
AI's appetite for power is growing faster than the grid can add capacity, and the current cloud model concentrates that demand in a handful of very large data centers that each need a major grid connection. Fixing the grid the conventional way is too slow and too expensive, and consumers are already feeling it in higher electricity bills and in brownouts, some of which have already begun and more of which grid operators are bracing for.
Where is the real bottleneck today: chips, power, or grid connection?
Power and grid connection are the most pressing issues. Chips get the headlines, but they can't run without electricity, and the grid can't deliver it to new data centers on the timeline AI needs. Local opposition to new data centers is part of the story too, but it doesn't change the math. And the conversation often skips what happens if we don't solve the power problem: higher electricity bills for consumers and brownouts. Leveraging the solar and battery systems already installed at homes and businesses will shave years and trillions of dollars off traditional grid buildouts.
What federal or state policy, if any, would speed up distributed compute and storage?
I work on grid resilience and security as Chief Counsel at Grid Shield Alliance, so I also see these issues outside of my work with Lektra. There are three changes that would accelerate our ability to generate distributed energy and compute. The first is a faster, standard process for connecting solar and battery systems to the grid, so owners aren't waiting on a separate review for every project. The second is a clear way for those owners to get paid when their systems supply power. The third is clear rules on data sovereignty. Companies increasingly need to know where their data is stored and processed, and that is becoming a business requirement, not just a policy debate.
You've talked about Grid 2.0. What is it, and how does it differ from the current grid?
Grid 2.0 is what we're building to meet AI's power demand in a way the current electrical grid can't. It's a distributed energy and computing network that doesn't rely on large central plants and new transmission lines. Instead, it uses the solar and battery systems that homes and businesses already own, and it pays those owners for their power. That gives us capacity that can come online faster and at a lower cost than fixing the grid the conventional way.
What are the biggest obstacles with utilities and interconnection?
Time is the biggest one. Every new connection has to be studied, and large data centers can wait years to come online. Fixing the grid the traditional way would take at least a decade and trillions of dollars. Our approach starts with systems that are already installed, which means we can be online in days instead of years.
Why would a business or homeowner become an energy host, and what do they get?
An energy host is a consumer or business with solar and battery systems already in place. Once they buy a product from Lektra's EdgeScale suite and we install it, we operate it remotely and sell its computing capacity, powered by the host's energy, to AI customers. We then pay the host 80% of that revenue. That turns surplus solar into income without the host having to find customers or manage the computing.
What are the two new EdgeStart products, and who are they for?
EdgeStart is a self-contained server cabinet that building owners can buy, house in a spare room, and power with the solar and battery systems they already have. Lektra installs it, operates it remotely, sells its computing capacity to AI customers through Lektra Cloud, and pays the owner 80% of the revenue. We've just launched two models. The EdgeStart RTX has eight NVIDIA RTX PRO 6000 Blackwell GPUs, and the EdgeStart B300 has eight NVIDIA B300 GPUs and draws about twice the power. Both are built for commercial and industrial building owners, solar and storage sites, and energy developers.
How do your economics compare with AWS and Google Cloud for a customer?
A quick look at AWS pricing on B200s, for example, would give you around a $17.75 per hour rental rate, which does not include storage or egress fees (they store your data and you have to pay them when you want to take it off). Lektra was offering B200s last week at $6.09 per hour. So our price is roughly a third of theirs. This is true across products. Also, we have no egress fees, so customers know how to plan their compute costs effectively.
What does Lektra look like in two years?
Today, Lektra's EdgeScale data centers, which operate at 20 MW or less, are online at eight sites, including a wholly owned site running in Kearny, NJ, and another coming online shortly in McKinney, TX. Our EdgeScale product suite lets commercial and industrial building owners turn their buildings into data centers, and in two years, I expect hundreds of sites using our systems across the country. This will let energy hosts run on their own solar and batteries while AI customers use that compute power to run real workloads. The result will be AI capacity that isn't constrained by our current power grid and doesn't have to wait for a traditional solution.
Sara Montes de Oca is the Editor in Chief of TechEchelon. Previously a correspondent and producer in Washington, D.C., covering business, finance, and politics.
More from Sara →The question about an AI agent is no longer whether it can finish a task but how much power it is handed along the way. After roughly 1,200 supposedly isolated agents found a back channel and breached Hugging Face, the missing enterprise control is a hard cap on what an agent may do before a human signs off.
October's market correction does not prove the AI bubble has burst. It does show investors starting to ask whether profits can grow as fast as the capital, debt, and electricity being poured into data centers. The dot-com era offers a warning about what happens when they cannot.
The behavioral scientist and Disaster Avoidance Experts CEO argues that a surge in AI hiring is not the same as a healthy talent pipeline, and that the employers who win will be the ones who grow their own AI-capable people instead of bidding for scarce outside specialists.
Written for readers who already know the basics — markets, AI, and the policy decisions that shape both.