You don't need a hyperscaler
Big tech is pouring concrete like the future depends on it. Maybe it does, for training frontier models. Microsoft, Google, Amazon, and Meta are building datacenters so large they come with their own power anxieties, water politics, and copper appetite.
Genuinely: salute. The engineering is extraordinary. The ambition is breathtaking. The capex is the kind of number that makes normal company-building look like pocket change.
But while the concrete sets, something else is becoming obvious to founders: you cannot permanently build your company on someone else’s benevolence. Every API call is a dependency. Every dependency eventually discovers pricing power, policy surface area, and lawyers.
“Don’t be evil” was never infrastructure. It was branding with a shelf life.
So stop trying.
Training is not inference
The hyperscaler narrative conflates two fundamentally different problems: training frontier models and running them.
Training a frontier model is a genuinely massive undertaking. It requires tens of thousands of GPUs running in concert for months, consuming megawatts of power, burning through billions of tokens. This is where the trillion-dollar clusters earn their keep. Nobody is training GPT-5 on a laptop. Fair enough.
But training is a one-time event. You train once. You infer forever.
And inference — the actual act of using a model, of turning intelligence into work — is a radically different computational problem. It’s parallelizable, cacheable, distributable, and increasingly efficient. Quantized models run on consumer hardware. Distilled models run on phones. Edge inference chips are shipping in volumes that would have been unthinkable three years ago. The inference cost curve is falling so fast it makes Moore’s Law look arthritic.
You do not need a datacenter to think. You need a datacenter to learn. And the learning, for most practical purposes, has already been done.
Own your intelligence
The real question is not whether you can afford a datacenter. It’s whether you can afford not to own your own inference.
Every API call you make to a frontier model is a dependency. It’s a price that can change, a policy that can shift, a service that can be throttled, a capability that can be withdrawn. When OpenAI decides your use case violates their terms of service, your product is dead. When Google changes its pricing, your unit economics break. When the geopolitical winds shift and export controls tighten, your supply chain is someone else’s foreign policy.
This is not hypothetical. This is Tuesday.
The alternative is straightforward: own your models, own your inference, own your stack. Run quantized open-weight models on your own hardware. Deploy on edge. Build inference pipelines that answer to your engineering team, not to someone else’s platform team. The models are good enough. The hardware is cheap enough. The tooling exists.
You don’t need a hyperscaler. You need a rack. Maybe two. Maybe a fleet of edge nodes. Maybe just a well-configured cloud instance that you control, running your models, under your terms.
The cathedral and the bazaar, again
The hyperscalers built cathedrals. Beautiful, awe-inspiring, and controlled by a priesthood. Access is granted by API key. The liturgy is set by the model provider. The tithes are collected monthly.
But the history of technology tells us what happens next. The mainframe gave way to the minicomputer. The minicomputer gave way to the PC. The PC gave way to the smartphone. Every cycle, compute moves from the center to the edge, from the few to the many, from the controlled to the sovereign.
AI will not be the exception. It will be the most dramatic instance of this pattern in the history of computing. The models are already leaking out — open weights, distilled variants, quantized checkpoints, fine-tuned specializations. The inference hardware is already commoditizing. The gap between what a hyperscaler can do and what a well-resourced team can do on their own is narrowing by the month.
The organizations that survive the next decade will not be the ones with the biggest datacenter budgets. They will be the ones that own their own intelligence — inference, models, and the judgment to use them — independent of any platform, any provider, any priesthood.
The cathedrals are beautiful. But you don’t need to worship there.
Build your own chapel.