Conviction in AI

if you can’t have a strong view on what data centers look like in 10 years, you shouldn’t be deploying capital into the category today. that applies to investors, GCs, and executives deciding whether to stand up a mission critical division. the reason is that data center construction has a payback period measured in decades, the technology it serves is evolving on a timeline measured in months, and the gap between those two creates a sorting mechanism for who actually has conviction vs. who is chasing backlog because it’s there.

i’ll walk through my perspective across three time horizons as the conviction required at each one is different.

something i don’t think most people in this industry realize: all of the gains we’ve seen from AI to date have come from existing data centers. every model, every app you’ve vibecoded, every product experience, all of it was trained on infrastructure that was built years ago when most of the construction industry wasn’t really paying attention to data centers. the announcements and starts dominating backlog reports right now won’t produce a trained model for years. openAI’s stargate campus in abilene broke ground early 2025, they won’t train there until at least 2029. the michigan facility broke ground in february. play that forward and that’s beyond 2030 before it hosts a completed training run.

construction alone takes 18 to 36 months for a hyperscale facility, then commissioning, network config, etc., then the training run itself which for a frontier model takes months and each generation takes longer as models get bigger. the supply everyone is calling a bubble hasn’t even entered the production function yet. meanwhile the demand side is absolutely on fire. construction and manufacturing firms that can support mission critical are turning away work because it’s hard to justify taking the incremental school project at 2.5% fee for 12 months when you can have guaranteed work for three years at a higher absolute dollar revenue. electrical subs are booked 18+ months out. so even if every hyperscaler paused tomorrow, work under contract sustains elevated activity in this sector through 2028. the labor constraint is the limiting factor before the demand constraint. in the near term, 1 to 3 years, i don’t believe there is a bust.

before getting to the medium term though, some context. people throw around “data center” like it’s singular thing with exact characteristics. it’s not. there are generally two buckets: storage and compute. and then within compute you have training and inference. each of these have different demand drivers, different power profiles, and different risk, and so lumping them together is like calling every building with a loading dock a warehouse.

training facilities are the headline grabbers. massive campuses, thousands of GPUs in parallel for months. you can’t pause a training run because the grid dipped so facilities these need nameplate, baseload power with zero interruption, which in practice means co-located natural gas generation running behind the meter. hence why those projects are always co-announced with power generation.

inference facilities handle the actual use of trained models. the load is demand responsive, much like electricity does, it spikes during the day and drops at night, so these tolerate hybrid power: gas plus BESS, grid connected with peaker backup, renewables in the right geography. inference is also very latency sensitive so these typically sit closer to population centers.

storage is the least discussed and the most durable. every piece of digital content ever created has to live somewhere and storage demand isn’t conditional on AI scaling laws. it’s conditional on people or things using the internet.

the profile for storage varies wildly and the easiest way to explain it is your recycling habits. if you throw a bunch of boxes, cardboard, cans into the bin without breaking anything down, it takes up a lot of space. that’s warm storage. now imagine you break down the boxes before putting them in the bin. same bin, more capacity. that’s cool. your apartment complex gathering everyone’s recycling and compacting it before shipping it away, that’s cold. the recycling plant compressing everything further into bales, that’s frozen. you see this every time you pull up Instagram and someone’s recent post loads instantly but when you scroll back to 2014 it takes a couple seconds. that photo was created, sorted by tier, and stored in a facility optimized for exactly that retrieval pattern. given the explosion of digital content from current AI capabilities alone, we’re probably still UNDERinvesting in storage and inference datacenters.

the 4 to 6 year horizon is where it gets murky and where the framing above actually matters.

scaling laws are the idea/concept where model performance improves predictably as you increase compute. double the training compute, and you get a measurable improvement. if scaling laws keep holding, training demand is functionally unbounded and the boom times roll on. if scaling laws break, training demand will change very rapidly because labs won’t spend billions on clusters that produce marginal gains and so then training data center (mega campus) construction slows. but remember, that’s only one type of data center.

if we see the headline that scaling laws are breaking, there will be a massive overcorrection. capital markets will try to halt all data center construction even though training is just one bucket. the market won’t distinguish between training datacenters and inference datacenters, let alone storage. everything data center related will sell off. construction firms, power developers, equipment manufacturers. that overcorrection, for anyone who understands the taxonomy, is latent opportunity.

inference and storage don’t depend on scaling laws. they depend on adoption. and most of society hasn’t really seen AI diffuse through everything yet. it’ll follow a similar curve to the internet spreading through devices, geos, industries over a decade. as that happens, it dramatically increases the need for inference and storage facilities. remember the earlier point: the capabilities we see today are based on existing, prior data centers. the new facilities outside of training will provide so much more capacity for inference and storage that the demand from current AI capabilities alone already exceeds what’s being built. i’m confident in non-training DC spend over the next 4 to 6 years regardless of what happens with scaling laws.

now where all of this gets built matters because every data center thesis bottlenecks at energy. training needs baseload nameplate power, consistent megawatts 24/7 for months. natural gas is the default because it’s dispatchable, dense, and can sit behind the meter. inference works with a broader mix but heavy inference still needs a gas backbone. storage cares about cooling efficiency and density more than raw compute.

what most people outside the Midwest don’t realize is how much dormant infrastructure exists in places like Ohio, Michigan, Pennsylvania. these are states where heavy industrial loads, steel mills, auto plants, uranium enrichment facilities, once pulled enormous power from a grid sized to serve them. that industrial base contracted over decades but the grid and infrastructure didn’t. the substations, transmission corridors, gas pipelines, right of way easements, all still in the ground. two weeks ago, the DOE announced a partnership to convert a former uranium plant in ohio into a 10gw data center campus. bechtel and kiewit are building it with softbank as the private partner.

the 7 to 10 year view is almost too conditional on scaling laws to predict with precision but the direction holds. by the mid-2030s i expect everything to be AI-enabled, AI-connected, much like the internet is everywhere now. autonomous vehicles will be running local inference, robots in warehouses and factories and eventually jobsites, phones running on-device models for real time tasks while routing complex reasoning back to a data center. all of that needs compute infrastructure. the question is where inference runs and that’s genuinely hard to predict. but the physical footprint of compute grows in every scenario, it just grows differently. however, given the power constraints for the various types of data centers, it begins to get easier to plan for where these will be located across the country.

betting on the future requires conviction in the present. i'm confident in the near term work. i believe the taxonomy separates who's positioned for the medium term from who gets caught in the overcorrection. and i have a view on what gets built when AI is everywhere, which is more than most people can say. i'm betting on American infrastructure.