Most buyers have never been told how AI infrastructure actually works.

Everyone throws around the word hyperscaler. AWS. Google. Microsoft. But what actually is one?

A hyperscaler is a landlord. A very large landlord, with very expensive buildings, renting you a very small piece of shared space by the second.

That's it? A landlord?

That's it. They own real estate, power contracts, cooling systems, thousands of servers. They slice all of it into small pieces and rent those pieces to millions of customers at once. That slicing is the whole business.

So when I pay eleven cents an hour for a server, I'm actually sharing that machine with strangers?

Correct. The same physical hardware is being rented to a hundred other tenants at the same moment. It's called multi-tenancy. It's the only way the eleven-cent number works.

Why do they charge by the second? By the gigabyte? By the API call?

Because their cost is capital. They spent tens of billions building the buildings. To make that back, every second of every server has to be earning revenue from someone. The meter running on you is what pays down their mortgage.

That model worked fine for hosting a website. Does it work for AI?

No. A modern AI workload is not a small slice of a small server. It's a full rack of GPUs drawing a quarter megawatt, running for weeks. You're paying multi-tenancy overhead on a workload that fills the whole machine.

And my data?

Your model weights, your customer data, your trade secrets — all of it moves through networks and storage that the landlord operates. That was fine when you were renting a web server. It matters enormously when you're running an AI trained on your intellectual property.

So the summary is: hyperscaler equals landlord with a meter on every wall.

That's not a criticism. That's the accurate description of the business. Next episode, we look at LLMs — and why they charge you by the syllable.

When I use ChatGPT or Claude, I'm being charged by the token. What is a token?

A token is roughly three-quarters of an English word. You're paying by the syllable.

By the syllable. And each syllable has a different price?

Yes. Three or four categories. Input tokens — what you send. Cached input — cheaper, because the model has seen that context before. Output tokens — the most expensive, because generating text takes real compute.

Give me a number.

A reference provider set runs about ninety-four cents per million input tokens, fourteen cents per million cached, and three dollars fifteen per million output. Those numbers change every quarter. The number isn't the point.

What is the point?

The structure. You're paying different rates for different kinds of syllables. And none of those rates map to a finished business task.

Meaning?

Meaning: read this contract, flag the indemnity clauses. That one task might use ten thousand input tokens, thirty thousand output tokens, three tool calls, two retrievals, a self-check pass. The bill is the sum of all of that. Not the sticker price.

So the same question, asked twice, could cost me different amounts.

Correct. Same document, uploaded twice, chunked into different token counts by different providers. Your bill varies. That's not a bug. The meter is behaving normally. It's just a meter designed for the provider, not for you.

And every token I send crosses the provider's boundary.

Every one. Your prompts. Your documents. Your customer data. You're trusting the terms of service.

So the token is a real meter, but it's not the product I'm actually trying to buy.

The token is the flour. The bakery is where the margin lives. Next episode: training versus inference — the two very different bills.

People say AI like it's one thing. But you keep telling me there are two.

Training and inference. Two completely different activities. Two completely different bills.

Start with training.

Training is building the brain. You feed a model trillions of examples and adjust billions of internal weights until it can predict the next word well. Monstrously expensive. Hundreds of GPUs running for weeks, drawing megawatts, generating heat you have to liquid-cool.

And inference?

Inference is the brain thinking. You send the finished model a question. It runs forward through those billions of weights and produces an answer. Much cheaper per event. But you do it billions of times.

Every ChatGPT reply is an inference.

Every one. Every AI-generated image. Every AI agent decision. The economics of a company running AI are dominated by inference volume, not training.

So training is a capital event. Inference is an operating expense.

Exactly. Confusing the two is the most common mistake in AI budgeting.

Is there a third thing?

Fine-tuning. Between training a base model and running inference, sophisticated operators customize the model to their business. Has its own compute cost. Its own governance risk — did your proprietary data get baked into the weights? And its own operational discipline.

And the infrastructure looks different for each?

Very different. Training loves dense clusters of GPUs with ultra-fast interconnects — the fabric costs more than the GPUs. Inference loves distributed capacity close to the user, with predictable latency. Building a facility that does both well is a different problem than building one that does one.

And each has a different data-exposure profile.

Where training happens determines who owns the model. Where inference happens determines who sees the live data. Train on a hyperscaler, run inference on an LLM provider, and you've handed both training data and live production traffic to third parties. Defensible decision. Should be a decision, not an accident.

Three different activities. Three different bills. Three different exposures.

Next episode: GPU-as-a-service — the model that pretends they're all the same thing.

GPU-as-a-service. Sounds efficient. What is it?

A provider buys thousands of NVIDIA H100s, H200s, or Blackwell chips. Racks them. Powers them. Cools them. Rents them to you by the hour. Two to eight dollars an hour depending on the silicon.

That sounds like a good deal.

For some workloads it is. Bursty training runs. Research spikes. One-off experiments. But for production AI, there's a trap on the pricing page that nobody points at.

Show me the trap.

Utilization. You pay for the hour whether you're using the GPU or not. Real production AI runs at maybe forty percent utilization. So your effective rate is two-and-a-half times the sticker price. GPUaaS isn't three dollars an hour. It's three dollars an hour divided by your actual utilization — a number you don't know until you're already spending.

What else?

The data path. Your model weights, your training data, your customer traffic all move onto and off the provider's storage and network. Egress is metered. Storage is metered. And the network between GPUs — the interconnect that determines whether your training run even works — is often oversubscribed with other tenants.

So the GPUs are yours, but the network between them is shared.

Correct. And there's a third problem. When capacity is tight — and for the newest GPUs it's always tight — you can be preempted, deprioritized, or asked to wait. Your workload is subject to the provider's queue.

Are all GPUaaS providers the same?

Spectrum. Hyperscalers renting GPU instances the same way they rent CPU instances — full multi-tenancy, full markup. Specialist neoclouds — CoreWeave, Lambda, others — better price-performance because they skip the multi-tenant tax. Bare-metal GPU rental — you get a whole physical machine and the provider stays out of your way.

But even the best of those still has the same underlying problem.

You don't own the outcome. You own the hours. Finish your training run in thirty percent of the reserved time? No refund. Model doesn't converge? No refund. Their product is the hour. Your product is the finished work. Those aren't the same thing.

So GPUaaS is a car rental.

Rent it if you know the trip. Don't rent it if the trip is your company. Next episode: sovereign data — where your AI physically runs and why that turns out to be one of the most consequential decisions your business makes.

You keep bringing up data. Where does my AI actually run?

That is the most important question in the entire category, and almost nobody asks it. Not where the model was trained. Not where the company that made it is headquartered. Where does your actual prompt — your document, your customer record, your trade secret — physically live for the seconds it takes the AI to answer?

And whoever owns that place—

—owns your data.

Walk me through the plumbing.

You use a public AI service. Your data leaves your walls. Travels over shared infrastructure. Lands in a provider's memory long enough to be processed. Depending on the terms, it may be logged. Cached. Retained. In some cases, used to improve future models. That's not a criticism of any specific provider. It's how the plumbing works.

And that plumbing creates real risks.

Three of them. IP leakage — once your prompt crosses the boundary, a third party possesses a copy of it. Even if they promise not to train on it. Subpoena exposure — data held by a third party is subject to legal process directed at that third party. Your data can be swept into a case that has nothing to do with you.

And the third?

Hyperscaler dependency. If your entire AI capability lives on someone else's infrastructure, they can raise prices, change terms, deprecate services, or refuse to serve you. Your business continuity is a footnote in their business plan.

So sovereign data means what, exactly?

The AI runs inside your boundary. Not theirs. Compute on your campus, or in a facility you control. Private network. Your keys. Your logs. Your ability to prove nothing left.

That sounds like a marketing word.

It's a technical one. Dedicated bare-metal, not virtualized slices. Private optical fabric, not public internet. Model weights hash-verified at boot. Runtime local-first, deny-by-default — external contact blocked unless explicitly approved, and every approved call recorded.

What about air-gapped?

Stronger claim. The facility has no live network connection during operation. Common request from defense, intelligence, regulated finance. And a claim you don't make without a completed acceptance package. Disconnected operation is an engineering exercise. Not a slogan.

The trade-off?

Sovereign infrastructure costs more upfront than renting from a hyperscaler. Because you actually own something. In exchange: predictable pricing. Provable data custody. Independence from a vendor whose priorities are not aligned with yours. For any organization whose data is a competitive asset — universities, banks, hospitals, defense, energy, governments — the trade is not close.

The boundary is the product.

The boundary is the product. Next episode: the real cost of an AI factory. Five buckets. Most vendors show you two.

Someone quotes me a price for AI infrastructure. How do I know what's actually in the number?

Ask which of the five buckets are in it. Because there are five, and most vendors show you two.

Walk me through the five.

One: data center infrastructure. The powered shell. Building, raised floor, cooling loops, fire suppression, security. Reference range for a modern high-density AI build: fourteen to sixteen million dollars per megawatt of IT capacity.

Two?

IT hardware. GPUs, servers, storage, top-of-rack networking. About twenty-nine million per megawatt for current-generation direct-to-chip liquid-cooled builds. This is what most people mean when they say AI compute.

Three?

Electrical infrastructure. Transformers, switchgear, uninterruptible power, busway, rack-level distribution. Not part of the shell. Not part of the IT. Separate line item. Rarely in the sticker price you see quoted.

Four.

Onsite power generation. If you don't want to be at the mercy of the grid, or if the grid can't give you the megawatts you need on your timeline, you generate onsite. Fuel cells, solar plus storage, biomass, natural gas. Hydrogen fuel cell plus reformer runs about one and a half million per megawatt gross, about one million per megawatt after federal investment tax credits.

Five.

Workforce infrastructure. Training programs. Apprenticeships. Community relationships. Local hiring pipelines. Not a cost. A moat. But a moat you fund upfront.

Add the first three.

Forty-three to forty-five million per megawatt, all-in, before power generation and workforce. Working benchmark. Every site is different. But that's the range for a modern direct-to-chip AI factory.

And tax incentives?

Where sophisticated developers make money that unsophisticated developers leave on the table. Section 48E investment tax credit offsets a substantial portion of power generation capex. Low-income community adders can reach up to sixty percent of eligible capex — with allocation caps, a five-megawatt-per-facility limit, site qualification, and tax counsel review. Section 6418 transferability lets you sell those credits for cash. Bonus depreciation accelerates the shield.

And state incentives.

Another layer. Virginia's distressed-locality data center program has a seventy-million-dollar investment and twenty-five-jobs threshold. Machinery-and-tools tax abatement. Workforce credits. Sales and use exemptions.

So when a hyperscaler quotes me eleven cents an hour for a GPU—

Embedded in that number is their recovery of all five buckets, plus their margin, plus multi-tenant overhead, plus risk premium. When a GPUaaS provider quotes three dollars an hour, they're recovering bucket two and part of one. Their model outsources the rest.

And a SAVRN AI factory prices what?

Finished work directly. Lets the cost of the buckets fall out to its own capital structure. The only honest comparison across all these vendors is dollars per finished task at real utilization. Not tokens. Not GPU hours. Not rack space. Finished work.

Five buckets. Ask which ones are in your price.

The ones they leave out are the ones you'll pay for twice. Next episode: why SAVRN is not a hyperscaler.