The conversation at every AI conference I've attended this year is about chips. GPU availability, fab capacity, chip design. Billions of dollars are flowing toward building more fabs, securing allocation from TSMC, buying up older GPU stock to fill the gap.
It's all real. The chip supply crunch is genuine. But it's not the constraint that's actually going to slow down AI deployment over the next five years. Power is.
I know this because I've started paying attention to what's actually happening at the infrastructure level. The companies that are ahead, the ones that are building real AI systems, they're not worried about chip supply. They've secured allocation. They know where their next batch of GPUs is coming from. What keeps them up at night is power.
Training clusters are only for the hyperscalers
Let me split this into two buckets because the problems are completely different. First, there's AI training. That's where you're taking massive datasets and running them through neural networks for days, weeks, or months to build the base models.
Training clusters need 50 to 200 megawatts. That's not a power requirement you can meet in a commercial building. You need serious utility infrastructure. And the hyperscalers, the Microsofts and Googles and Metas of the world, they have the capital and the footprint to negotiate with utilities years in advance, get power allocated, and wait for infrastructure to be built.
They also have the leverage to do things smaller operators can't. They can invest in power plants. They can negotiate PPAs for renewable energy coming online in five years. They can work with utilities as a special case because they're buying hundreds of megawatts.
But even for them, it's tight. The power queue is full. New capacity is being built, but not fast enough to keep up with demand. Still, the hyperscalers will get served because they're too big not to serve.
Inference is exploding at the edge. And power is the bottleneck.
The bigger problem is inference. That's where you take a trained model and run input through it to get a result. One user querying one model. That happens in distributed data centers across thousands of locations.
A single inference pod might need 200 kilowatts to a few megawatts depending on the model and the hardware. But you don't run one inference pod. You run hundreds or thousands of them. At edge locations. Near the user. Latency matters. So you've got inference servers in edge data centers in cities across the country.
And that's where the power problem gets acute. Edge data centers aren't in major utility hubs. They're in commercial real estate. They're in urban areas. They're in places where power infrastructure was designed for office buildings and retail, not for megawatts of continuous compute load.
You want to spin up an edge inference cluster at a location where you need it. You call the utility. They look at the grid load at that location. They tell you you can't take another megawatt without upgrading distribution infrastructure. Queue position for that upgrade? Six years plus.
The time gap is brutal
You've got model availability now. You've got edge computing infrastructure built. You've got customer demand happening today. And you've got a six year wait to get the power infrastructure to actually serve those customers. That gap is where the entire problem lives.
This isn't theoretical. Companies are working around it right now. Some are shrunking inference models so they fit into locations with lower power requirements. Some are batching inference requests to better utilize the power they do have available. Some are building their own generation, which is expensive and regulatory mess.
None of these are good solutions. They're workarounds. They're trading off capabilities or economics because power isn't available when and where it's needed.
The market is scaling faster than the grid
Data centers are projected to consume 8 to 12 percent of all U.S. electricity by 2030. That's up from about 4 percent today. And that's with the constraint of power availability already limiting deployment. If power wasn't a limiting factor, the number would be higher.
The AI data center power market is being valued at around 16 billion dollars. That's not revenue. That's the market opportunity that exists but isn't being captured because infrastructure can't support it.
Everyone in the industry knows this. The hyperscalers know it. The edge computing companies know it. The utilities know it. And nobody has a great answer yet because the timelines are so misaligned. Utility infrastructure takes years to build. AI deployment is moving at months or quarters.
You need clean power now, at the site
Here's where it gets interesting. What if you could deploy clean, grid-parallel generation at the exact location where you need power? Not ideal power. Not unlimited power. But immediate power that works within the constraints of that commercial location.
You're not trying to replace a grid connection. You're trying to fill the gap while the grid connection is being built. Or at some locations, you're trying to augment what utility power you already have.
The technical requirements are specific. You need something compact enough to fit in a commercial space. You need it quiet enough not to be a neighborhood problem. You need it to be clean, not diesel-powered, because you're building AI infrastructure in cities. And you need it deployed in weeks, not months of engineering.
That's exactly what the GX230 is designed for. It's a 200 kilowatt multi-fuel generator. Runs on natural gas, propane, hydrogen, or biogas. It's grid-parallel, so it integrates with whatever utility power you have. It's 15 square feet and 700 kilograms, so it fits in a commercial environment. And it runs at 69 decibels, which is quieter than you'd expect from a megawatt-equivalent power source across multiple units.
For an edge inference operator, this means you can deploy your infrastructure at the location you actually need it. You're not waiting on the grid queue. You're not bottlenecked by power infrastructure planning. You're deploying your inference cluster now, with power delivered in parallel through our generation, while the utility is still planning their grid upgrade.
At 200 kilowatts per unit, you can scale edge inference from hundreds of kilowatts to a few megawatts by deploying multiple units. It's not a massive data center. It's a practical edge location with enough power to actually serve the inference workload.
The math is interesting too. Grid upgrades cost hundreds of thousands to millions of dollars and take years. Deploying clean generation costs less and takes weeks. For a lot of edge locations, it becomes the permanent solution because it's cleaner and cheaper than what the utility would eventually deliver.
We're already talking to edge computing companies about this. The problem is acute for them. They've got deployment plans that are hitting power walls. They've got customer demand they can't serve because infrastructure isn't available. They need a way to deploy inference capability at the site where they need it, with power available immediately. That's what we're building for. If you're dealing with edge inference power constraints, reach out.