Netflix, Uber, or Your Own Car: The Three Ways to Pay for AI

Netflix, Uber, or Your Own Car: The Three Ways to Pay for AI

Published: January 28, 2026 - 15 min read

This is Part 5 of the Tokenomics for Humans series. If you haven't read Part 4 on Why AI Costs What It Costs, I recommend starting there.

In Part 4, I gave you homework.

I asked you to think about how you currently access AI. Do you use a subscription? Do you access it through another product? Have you ever paid per-token through an API?

Today, we're going to break down those options in detail.

Because here's the thing: the way you pay for AI affects how much you pay for AI.

Same usage. Different pricing models. Wildly different costs.

Understanding your options is the first step to making smart decisions about AI spending.

Let's go.

The Three Models: An Analogy

Before we dive into the technical details, let me give you an analogy that makes this intuitive.

Think about transportation. You have three basic options:

Netflix model (SaaS): You pay a flat monthly fee. Watch as much as you want (within limits). Simple, predictable, no surprises. But if you barely watch anything, you're overpaying. And if you're a power user, you might hit limits.

Uber model (API): You pay per ride. Only pay for what you use. Great visibility into costs. But if you use it a lot, those per-ride fees add up. And costs can be unpredictable.

Own a car (Self-Hosted): Big upfront investment. Ongoing maintenance costs. But once you own it, your per-trip cost is low. Makes sense if you drive a lot. Doesn't make sense if you rarely need a car.

AI works the same way.

THE THREE WAYS TO ACCESS AI
================================================================

SAAS (Subscription)         API (Pay-Per-Use)         SELF-HOSTED
     Netflix                     Uber                   Own a Car
================================================================

  $20/month flat            $0.003 per 1K tokens       $500K+ upfront
  Simple billing            Pay for what you use       You own it
  Limited control           Full visibility            Full control
  May hit limits            Costs can spike            Requires expertise

================================================================

Let's look at each in detail.

Option 1: SaaS (Software as a Service)

SaaS = You pay a subscription fee and use AI through a website or app.

This is how most people access AI today.

Examples

• ChatGPT Plus ($20/month) - Access to GPT-4, faster responses, priority access
• Claude Pro ($20/month) - Access to Claude, higher usage limits
• Microsoft Copilot (bundled with Microsoft 365) - AI features in Office apps
• AI features in other products - Notion AI, Grammarly, etc.

How the Economics Work

SAAS MODEL: HOW IT WORKS
================================================================

    YOU                    VENDOR                 AI INFRASTRUCTURE
     |                        |                          |
     |----[$20/month]-------->|                          |
     |                        |                          |
     |                        |------[manages]---------->|
     |                        |                          |
     |<---[AI access]---------|                          |
     |                        |                          |

================================================================
   What you pay: Fixed subscription fee
   What you see: Features and usage limits
   Token costs: Hidden from you (vendor handles it)
================================================================

You pay a flat fee. The vendor manages everything behind the scenes. You don't see tokens, you don't see infrastructure. You just use the product.

Pros

Predictable costs. You know exactly what you'll pay each month. No surprises on your credit card.

Simple to use. No technical setup. Just log in and start using it. Your grandmother could do it.

No infrastructure management. Someone else handles the GPUs, the data centers, the cooling systems. Not your problem.

Regular updates. New features, new models, improvements. All included.

Cons

Usage limits. Most SaaS plans have caps. Hit the limit, wait until next month (or pay for a higher tier).

No visibility. You can't see how many tokens you're using. Hard to optimize what you can't measure.

Potential overpaying. If you're a light user, you're subsidizing heavy users. The flat fee doesn't reflect your actual usage.

Limited customization. You use the product as designed. Can't fine-tune the model, can't integrate deeply with your systems.

Who Should Use SaaS?

• Individuals who want simple, predictable AI access
• Small teams without technical resources
• Anyone who values simplicity over optimization
• Explorers still figuring out their AI needs

Option 2: API Access (Pay-Per-Use)

API = You connect directly to an AI company's system and pay per token.

This is how developers and businesses build AI-powered applications. This is how the French Writing Playground was built.

Examples

• OpenAI API - Access to GPT-4, GPT-3.5, DALL-E, etc.
• Anthropic API - Access to Claude models
• Google AI API - Access to Gemini models
• Other providers - Cohere, AI21, Mistral (hosted), etc.

How the Economics Work

API MODEL: HOW IT WORKS
================================================================

    YOU                      API                   AI INFRASTRUCTURE
     |                        |                          |
     |----[request]---------->|                          |
     |                        |------[process]---------->|
     |                        |                          |
     |<---[response]----------|<-----[output]------------|
     |                        |                          |
     |----[$0.003/1K tokens]->|                          |
     |                        |                          |

================================================================
   What you pay: Per token used (input + output)
   What you see: Exact consumption for every request
   Token costs: Fully transparent
================================================================

You send requests, you receive responses, you pay for exactly what you use.

Real API Pricing (Early 2026)

Note: Prices change frequently. Always check current rates.

Pros

Pay for what you use. Light usage = low cost. No minimum fees (usually).

Full visibility. You see exactly how many tokens each request uses. You can optimize.

Flexibility. Build custom applications. Integrate with your systems. Use multiple providers.

Scale efficiently. Cost scales linearly with usage. No step-function pricing jumps.

Cons

Unpredictable costs. Hard to forecast monthly spend. One viral feature could spike your bill.

Technical complexity. Requires development skills. Not plug-and-play.

No bundled features. You get the raw AI capability. Everything else (UI, user management, etc.) you build yourself.

Potential for runaway costs. A bug or unexpected usage pattern can get expensive fast.

Who Should Use API?

• Developers building AI-powered products
• Businesses with technical teams who need custom solutions
• Heavy users who want cost visibility and optimization potential
• Anyone integrating AI into their own software

Option 3: Self-Hosted (Your Own Infrastructure)

Self-Hosted = You buy your own hardware and run AI models yourself.

This is the most complex option, but offers the most control.

Examples

• A large bank running AI on their own servers for security/compliance
• A tech company with their own GPU clusters
• A hospital keeping patient data on-premise
• A government agency with data sovereignty requirements

How the Economics Work

SELF-HOSTED MODEL: HOW IT WORKS
================================================================

    +-----------------------------------------------+
    |          YOUR INFRASTRUCTURE                  |
    |                                               |
    |   +-------+  +-------+  +-------+            |
    |   |  GPU  |  |  GPU  |  |  GPU  |  ...       |
    |   +-------+  +-------+  +-------+            |
    |        |          |          |               |
    |        +----------+----------+               |
    |                   |                          |
    |           +-------v-------+                  |
    |           |   AI MODEL    |                  |
    |           | (Open-Source) |                  |
    |           +---------------+                  |
    |                                               |
    +-----------------------------------------------+

================================================================
   What you pay: Hardware + electricity + staff + maintenance
   What you see: Everything (you own it)
   Token costs: Essentially zero per token (after initial investment)
================================================================

You buy the GPUs, set up the infrastructure, download an open-source model, and run it yourself.

The Cost Structure

Upfront costs:

• GPUs: $30,000-$40,000 each (AI-grade), need multiple
• Servers: $10,000-$50,000 each
• Networking: $50,000-$200,000+
• Setup and configuration: Engineering time

Ongoing costs:

• Electricity: Thousands per month
• Cooling: Additional thousands per month
• Staff: Engineers to maintain the system
• Hardware refresh: GPUs need replacement every 3-5 years

Rough estimate for a minimal self-hosted AI setup:

• Initial investment: $500,000 - $2,000,000
• Annual operating costs: $100,000 - $500,000

Pros

Complete control. Your data never leaves your infrastructure. Customize everything.

No per-token costs. Once you've paid for the hardware, tokens are essentially free (just electricity).

No usage limits. Run as much as your hardware can handle.

Data privacy. Critical for healthcare, finance, government, and other regulated industries.

Cons

Massive upfront investment. Half a million dollars minimum for a serious setup.

Technical complexity. You need skilled engineers to build and maintain everything.

Slower model updates. When new models come out, you have to set them up yourself.

Hardware obsolescence. GPUs get outdated. Today's $40,000 GPU might be outperformed by a $5,000 chip in 3 years.

Who Should Self-Host?

• Large enterprises with regulatory/compliance requirements
• Organizations with massive, predictable AI usage (billions of tokens monthly)
• Companies where data cannot leave their infrastructure
• Anyone who has the capital and engineering resources

The Hybrid Reality: Cloud-Hosted Self-Managed

There's actually a fourth option that sits between API and Self-Hosted:

Cloud-Hosted Self-Managed = You rent GPU infrastructure from cloud providers and run your own models on it.

The Hyperscalers

"Hyperscaler" is the industry term for the giant cloud companies:

These companies have massive data centers worldwide. Instead of buying your own servers, you rent computing power from them by the hour.

How It Works

You rent GPU instances from AWS, Azure, or Google Cloud. You deploy an open-source model (like Llama or Mistral) on those GPUs. You pay for the compute time, not per token.

CLOUD-HOSTED SELF-MANAGED
================================================================

    +-----------------------------------------------+
    |     CLOUD PROVIDER'S DATA CENTER              |
    |          (AWS, Azure, or GCP)                 |
    |                                               |
    |   +-------+  +-------+  +-------+            |
    |   | GPU   |  | GPU   |  | GPU   |  (rented)  |
    |   | $3/hr |  | $3/hr |  | $3/hr |            |
    |   +-------+  +-------+  +-------+            |
    |        |          |          |               |
    |        +----------+----------+               |
    |                   |                          |
    |           +-------v-------+                  |
    |           |   YOUR MODEL  |                  |
    |           | (Open-Source) |                  |
    |           +---------------+                  |
    |                                               |
    +-----------------------------------------------+

================================================================
   What you pay: Hourly GPU rental + data transfer
   What you see: Compute usage, your own token consumption
   Token costs: Determined by your efficiency
================================================================

Pros of Cloud-Hosted

• Lower upfront cost than self-hosted (no hardware purchase)
• More control than pure API (you run the model)
• Scale up/down as needed (rent more GPUs during busy periods)
• Use open-source models (no per-token fees to AI companies)

Cons of Cloud-Hosted

• More expensive than self-hosted at very large scale
• Still requires technical expertise
• Hourly costs add up (a GPU running 24/7 = ~$2,000+/month)
• Data leaves your premises (unless you use private cloud)

Open-Source vs. Proprietary Models

There's another dimension to this decision: which models can you actually use?

Before we go further, let me explain what "open-source" actually means.

Open-source = The creators make the underlying code (the "source code") publicly available for anyone to see, use, and modify. Think of it like a recipe that's been shared publicly. Anyone can follow it, tweak it, or improve it. The "source" is "open" for everyone.

Proprietary (also called "closed-source") = The creators keep the code private. You can use the product, but you can't see how it works under the hood. Like a secret recipe that stays in the vault.

Now let's see how this applies to AI models.

Proprietary Models (Closed-Source)

Some AI models are only available through the company that created them:

You cannot download these models. You cannot run them on your own hardware. You must pay the company for access.

Open-Source Models

Other models are freely available for anyone to download and run:

You can download these models. You can run them on your own hardware. You don't pay per-token fees.

PROPRIETARY VS OPEN-SOURCE
================================================================

PROPRIETARY (Claude, GPT-4, Gemini)
+------------------------------------------+
|  Locked inside company servers           |
|                                          |
|  - Pay to access via API or subscription |
|  - Cannot download                       |
|  - Cannot modify                         |
|  - Cannot see how it works               |
|                                          |
|  Usually more capable (for now)          |
+------------------------------------------+

OPEN-SOURCE (Llama, Mistral, Falcon)
+------------------------------------------+
|  Available for anyone                    |
|                                          |
|  - Download for FREE                     |
|  - Run on YOUR hardware                  |
|  - Modify and customize                  |
|  - See exactly how it works              |
|                                          |
|  Catching up in capability rapidly       |
+------------------------------------------+

================================================================

The Trade-Off

Proprietary models (Claude, GPT-4) are generally more capable. But you pay per token and can't self-host.

Open-source models (Llama, Mistral) are free to use. But you need your own infrastructure to run them, and they may be slightly less capable for some tasks.

For many business use cases, open-source models are now "good enough." The gap is closing rapidly.

Decision Framework: Which Option is Right for You?

Here's a practical guide to choosing:

DECISION FLOWCHART
================================================================

START: What's your monthly AI usage?
            |
            v
    +---------------+
    |  < $100/mo    |-----> SaaS subscription
    |  equivalent   |       (ChatGPT Plus, Claude Pro)
    +---------------+
            |
            | $100 - $10,000/mo
            v
    +---------------+
    | Need custom   |-----> API access
    | integration?  |       (Build your own solution)
    +---------------+
            |
            | No custom needs
            v
    +---------------+
    | OK with       |-----> SaaS subscription
    | usage limits? |       (Upgrade to higher tier)
    +---------------+
            |
            | No, need more
            v
    +---------------+
    | > $10,000/mo  |
    | equivalent?   |
    +---------------+
            |
            v
    +---------------+
    | Data privacy  |-----> Self-hosted or
    | requirements? |       Cloud-hosted self-managed
    +---------------+
            |
            | No strict requirements
            v
    +---------------+
    | Have GPU/ML   |-----> Cloud-hosted self-managed
    | expertise?    |       (Cheaper at scale)
    +---------------+
            |
            | No expertise
            v
        API access
        (Pay for convenience)

================================================================

Quick Reference

What This Means For You

If You're a Freelancer or Solopreneur

You're probably best served by SaaS subscriptions for now.

Practical takeaways:

• $20/month for Claude Pro or ChatGPT Plus is a bargain for most use cases
• If you're hitting limits frequently, evaluate whether API access would be cheaper
• Track your usage patterns (from Parts 2 and 3) to know if you're getting value

If You're a Tech-Forward Manager

You likely need a hybrid approach: SaaS for general use, API for custom integrations.

Practical takeaways:

• SaaS subscriptions for team members' daily AI use
• API access for any AI features in your products or internal tools
• Consider cloud-hosted self-managed if you have ML engineers and high volume
• Build visibility into AI costs across all access methods

If You're a CFO or Finance Lead

You need to understand the cost dynamics of each option.

Practical takeaways:

• SaaS: Predictable monthly costs, easy to budget
• API: Variable costs, need monitoring and potentially rate limiting
• Self-hosted: High capex, lower opex at scale, 3-year payback typical
• The right choice depends on volume, compliance needs, and technical resources
• More detail on TCO calculations in Part 6

The Bigger Picture

The way you access AI affects everything: cost, control, flexibility, and risk.

There's no universally "best" option. It depends on:

• Your usage volume
• Your technical capabilities
• Your compliance requirements
• Your budget structure (capex vs opex)

Most organizations end up with a hybrid approach:

• SaaS for easy access and exploration
• API for custom integrations
• Self-hosted for high-volume or sensitive workloads

Understanding all three options gives you the knowledge to make strategic decisions, not just default to whatever's easiest.

Quick Reference: Access Methods Cheat Sheet

Coming Up Next

Part 6: TCO, ROI, and the Money Nobody Talks About

Now that you understand how to access AI, let's talk about the real costs. Spoiler: it's more than just the subscription fee or API bill.

In Part 6, we'll cover:

• Total Cost of Ownership (TCO): What AI really costs when you add everything up
• The ROI problem: Why measuring AI value is so hard
• Capex vs Opex: How the accounting treatment affects decisions

Your Homework for Part 6

Think about your organization's (or your own) AI usage:

1. What are the obvious costs? (Subscriptions, API bills)
2. What are the hidden costs? (Staff time, training, fixing mistakes)
3. How do you measure the value AI provides?

These questions will help Part 6 make more sense.

See you in Part 6.

As always, thanks for reading!