The AI Boom Is an Infrastructure Supercycle
AI is about infrastructure more than it is about software
Welcome to the hundreds of new subscribers who have joined over the past few weeks. In today’s free post, I look at AI as infrastructure. The durable value, I argue, accumulates in physical infrastructure, not software written on top of LLMs.
And if you like what you read here and you’re not yet subscribed, consider subscribing. Most of my posts are free though deeper dives tend to be paid.
For years, Silicon Valley has sold a familiar story: that the next great fortunes will come from software. AI, we were told, would be no different. The prevailing narrative framed AI as the latest software wave: a Cambrian explosion of startups wrapping foundational models with slick UX, novel workflows, and bespoke vertical integration. The vision was clear: LLMs as platforms, wrappers as the next unicorns.
For the most part, this vision is wrong. Most of the real action is in the physical world.
What we’re witnessing is not a software revolution. It’s an infrastructure revolution: capital-intensive, geopolitcally sensitive, physically rooted. Durable value in AI is aggregating not in apps, but in the substrate of energy, compute, cooling, water, and land. GPUs are the tip of the spear, but they’re drawing us into a full-stack industrial realignment. One where venture capital is not the dominant allocator. One where private equity, sovereign wealth, real estate developers, and energy firms lead.
This isn’t SaaS with a twist. This is the early-stage buildout of AI’s physical empire.
The Ephemerality of the Software Layer
Let’s start with what the venture crowd is betting on: chatbots, copilots, agents, wrappers, workflows. These products are typically thin veneers over foundational models. At best, they provide UX convenience; at worst, they’re glorified prompt templates. And as model capabilities increase, the surface area for differentiation shrinks.
The problem isn’t that these companies don’t work. It’s that they’re trivial to copy and impossible to defend. There’s no real moat, just temportary distribution edges or branding games. Foundational models themselves will continue to absorb functionality, compressing the value capture of app-layer plays.
In short: the better the models get, the worse the wrapper economics become.
There is a thin layer of sticky software, where the switching cost is regulatory rather than technical. For example, FDA-validated radiology workflows or export-controlled ML Ops pipelines. Flag it, but don’t build an allocation thesis around it.
This is why the center of gravity is moving down the AI stack, to the hardware, to the energy grid, and to the land beneath the server racks.
Stranded Asset Risk: Betting on the Wrong Physics Curve
In spite of this, all is not without risk in the world of physical infrastructure. Today’s AI datacenters are optimized for Nvidia GPUs: hot, power-hungry chips with massive cooling needs. But what happens when the compute substrate changes?
Post-GPU alternatives are no longer fringe ideas. There are a few different potential post-GPU chips in development: analog, photonic, and neuromorphic. It’s unclear at present which of these will make it to market, but only one has to succeed to disrupt the GPU market. These chips run much cooler than GPUs, require different rack configurations, and depend on new energy envelopes. This means datacenters built for today’s thermal and electrical assumptions may be structurally incompatible with tomorrow’s compute.
This introduces a stranded asset risk to the AI infrastructure boom. Like coal plants or DSL lines, today’s hyperscale facilities may become obsolete faster than expected, especially if their layout or cooling systems can’t accommodate next-gen chips.
Who eats that risk? Cloud providers? REITs? Sovereigns? The answer isn’t clear yet. But the capital allocators who understand compute as a moving target, not a fixed substrate, will have an edge.
Securitization of Power: Compute as an Energy Derivative
AI’s energy needs are rewriting the relationship between compute and electricity. Power is no longer a utility bill. It’s a first-class financial asset. Frontier model labs are now locking in power through long-duration PPAs, inking deals with wind farms, nuclear operators, and hydro stations to guarantee training capacity.
But it goes deeper. These agreements behave like financial instruments. We’re watching the birth of synthetic energy markets, where inference futures, GPU-hour options, and energy-backed compute credits start to resemble commodity derivatives. Don Wilson, who founded famed prop trading firm DRW, seeded Compute Exchange, an innovative GPU auction site. GPU futures, akin to oil futures, seem possible.
The implication: infra investors are no longer passive landlords. They’re becoming liquidity providers to the AI economy.
AI-native securitization models could include:
Tradable inference capacity futures
Compute-backed energy tokens
Sovereign-guaranteed PPAs indexed to model throughput
The entire capital stack around power is being reconfigured to reflect the volatility and strategic nature of AI workloads.
State-Backed AI Ecosystems: From Startup to Strategic Asset
AI is no longer a sector. It’s a sovereign priority.
Consider the UAE, where G42 is vertically integrating everything from chip imports to data hosting to model development. Or China, where AI labs are national champions. Or the US, where DARPA and DoE are stepping in to fund compute, control safety, and absorb risk.
In this world, AI companies behave more like defense contractors than startups. They require sovereign guarantees, export licenses, and embedded regulators. They live or die not by their LTV:CAC ratio, but by energy access, geopolitical alignment, and compliance with national security laws.
This might look like:
UAE & KSA: vertical integration from chip import corridors to Falcon-class model weights, underwritten by gigawatt hydrogen and SMR projects.
China: national champions tethered to Five Year Plan capital sequencing.
United States: DARPA, DoE and Commerce jointly funding secure fabs + export-controlled model checkpoints.
The result is a capital stack inversion:
Top equity: Sovereign equity or strategic procurement
Middle: Infra funds, PE credit, and energy-backed lenders
Bottom: Token equity from venture investors, useful only for signaling
This is the return of industrial policy, applied to intelligence infrastructure.
AI-Native Urbanization: Cities for Machines
As AI inference becomes large-scale and persistent, we may see AI-native urbanization: regions and cities built around compute.
These aren’t “smart cities.” They’re cities for smart machines.
Inference-heavy industries, including autonomous logistics, AI-driven healthcare, robotic agriculture, and more, require extremely low-latency, high-throughput compute. That creates locational gravity: firms will colocate near inference nodes to reduce latency, cut power costs, and access privileged infrastructure.
This creates:
Boomtowns near stranded hydro, solar, and gas assets
Zoning arbitrage in low-cost real estate markets
Colony urbanism: enclave cities designed around compute cores
Microgrids with embedded data centers as civic anchor points
Think less like Silicon Valley, more like Company Town 2.0, where the company is a sovereign-scale AI stack.
Case in point: Quincy, Washington. Fifteen years ago, it was alfalfa fields; today Sabey runs a >600 MW datacenter for Microsoft powered by Columbia River hydro.
Land-bank now, rezone later.
Final Frame: AI as the Birth of a New Asset Class
If you’re still viewing AI through a startup lens, you’re already behind. This isn’t about apps. It’s not even about models. It’s about industrial realignment: the reconfiguration of energy, compute, land, and capital flows to support silicon cognition at planetary scale.
AI is not a software play. It is a sovereign-industrial play.
The winners will be those who understand how to own and operate the physical choke points. Not the APIs but the megawatts, water rights, and geopolitical permissions that underwrite them.
This is the new oil. And the scramble has only just begun.
Investor Playbook
Timeline for Post-GPU Compute
2025-2027: H200/B100 era; incremental thermal gains; safe to fund air-liquid retrofits.
2028-2030: Early photonics/analog pilots; finance shells with modular utility corridors.
2031-2034: Volume photonics; only Tier-A retrofits survive without gut renovation.
2035+: Neuromorphic wildcard; treat any 2025-vintage air-cooled build as fully depreciated by now.
Coda
If you enjoy this newsletter, consider sharing it with a colleague.
Most posts are public. Some are paywalled.
I’m always happy to receive comments, questions, and pushback. If you want to connect with me directly, you can:
Inference at the edge will be a thing
Great to see what I was thinking written down. You are on the leading edge here.