The model is the cheapest part of this.

There is a category error that runs through almost every conversation about AI cost. Business owners assume the expensive part is the model: the GPT-4o subscription, the Claude API access, the enterprise license. So when they think about whether they can afford AI, they're asking whether they can afford the model.

This is the wrong question. The model is, by a wide margin, the cheapest component in a serious AI architecture. Usage-based API pricing means that for most small and mid-market businesses, monthly model costs run from tens of dollars to a few hundred, depending on volume. The expensive parts are the decisions that come before the model: data infrastructure, integration design, workflow orchestration, observability, governance. Those are architecture decisions. And architecture decisions don't appear on any vendor's pricing page.

The average SMB already spends $18,000 a year on AI and gets experimental results

Eighty-two percent of SMB employers have made investments in AI tools (SBE Council, 2026). The median small business runs five AI tools simultaneously. Annualized, the average SMB spends approximately $18,000 per year on AI-related software: subscriptions to ChatGPT Plus, Jasper, Midjourney, Zapier AI, HubSpot's AI features, and whatever else the team adopted after seeing a LinkedIn post.

Approximately 70% of those businesses describe their AI adoption as "experimental" or "opportunistic" (ArticSledge, 2026). They have the tools. They don't have a system. None of the five tools talk to each other by design. There is no shared data layer, no unified observability, no governance model that controls what the AI can access across the business. The $18,000 per year is being spent on five isolated products rather than one coherent architecture.

The expensive irony is that the scattered-product approach usually costs more than a structured architecture would, both in licensing and in the opportunity cost of tools that don't compound. Five disconnected AI tools running in parallel generate five times the integration debt, five times the context-switching overhead, and zero times the compounding value that comes from a system where outputs from one workflow feed the next.

The Krastor reference stack and what it actually costs

Here is the actual infrastructure stack we build on, with the actual monthly costs at a typical single-client deployment. These are not illustrative figures. This is what the stack costs.

Workflow orchestration: a self-hosted, open-source orchestration layer. Approximately $5 to $15 per month depending on instance size. It handles the workflow logic that routes data between systems, triggers actions, and manages the handoffs between human and AI processes. It is not a SaaS subscription with a vendor that can raise prices or deprecate features. It is a codebase running on a server you control.

Data layer: a managed database (~$25 per month). The data layer provides the Postgres database, authentication, and real-time subscriptions that most business workflows need. Open-source at the core and self-hostable if the client outgrows the managed tier. The data lives in a standard SQL database: exportable, portable, not trapped in a vendor's proprietary format.

Email infrastructure: transactional email (~$20 per month at the operational scale most SMBs need). Workflow notifications, automated client communications, and system alerts. Priced on volume, not seats: you pay for what you send, not for the right to send.

Model API: Claude via Anthropic API (or equivalent), passed through at cost. No markup. Usage-based pricing means the client pays for tokens consumed, not for access rights. At typical SMB workflow volumes, monthly model costs run $30 to $150 depending on complexity.

Observability: an observability dashboard, either open-source self-hosted or managed at approximately $49 per month. Every prompt, response, cost, latency, and trace is logged and searchable. Without observability, you're running blind: you don't know what the model is doing, at what cost, or when it's degrading. The observability layer is the instrument panel for the system.

Add the token-governance layer (a governance proxy for model routing and cost allocation, deployed with a Postgres backend) and you're under $50 per month for infrastructure that gives you vendor-agnostic model routing, per-client cost attribution, rate limiting, and fallback logic. The entire governance layer that most enterprise AI teams spend six figures building costs $50 per month to run.

Client tool subscriptions (HubSpot, Notion, Slack, whatever the client uses) are paid directly by the client to each vendor. Krastor doesn't mark up tools. The invoice covers architecture, design, build, and the ongoing retainer for maintenance and evolution. The $200 is the infrastructure. The architecture decisions are the product.

"The model is the cheapest part of this."

Ownership vs. rental: the on-premise math

For businesses with sufficient query volume, the ownership case becomes even more compelling when you move inference on-premise. NVIDIA's DGX Spark, announced in early 2026 and shipping to SMB and professional buyers, is a personal AI supercomputer at $4,699. It runs open-weight models (Llama, Mistral, Qwen, and others) locally, with no API costs, no per-token billing, no data leaving the building.

The running cost of a DGX Spark is approximately $50 per month in electricity. At 10,000 queries per day at 500 tokens each, cloud API costs range from $450 to $2,250 per month depending on the model tier. The hardware pays for itself in three to twelve months. After that, every query is free. Open-weight models carry no licensing fees. Llama 3, Mistral Large, and Qwen are released under licenses that permit commercial use without royalties.

"The DGX Spark costs $4,699. After it pays for itself, every query is free."

This is not a recommendation for every client. Cloud inference has advantages in flexibility, redundancy, and access to the frontier models. The point is that the cost structure of AI infrastructure has fundamentally changed. You are no longer choosing between "expensive and capable" and "cheap and limited." The open-weight models running on affordable hardware are capable of performing enterprise-grade tasks. The architecture decisions about which tasks to route to which model tier (cloud frontier for complex reasoning, local inference for high-volume structured tasks) are what create efficiency. That is an architecture decision. It does not appear on any vendor's pricing page.

What you're actually paying Krastor for

The question clients ask most often: if the tools are this cheap, what am I paying for?

The answer is architecture decisions. Specifically: which tools, connected how, with what data flowing where, at what governance level, observed how, and evolved in what sequence. Those decisions take years of production experience to make well. Making them wrong costs multiples of what making them right costs, both in direct rework and in the compounding value of a year spent on the wrong foundation.

The build itself: integration design, workflow orchestration, data pipeline construction, the MCP-native control plane that connects AI to the client's business systems, the observability layer, the governance configuration. That is the construction work.

The ongoing retainer: monitoring, model updates as the underlying landscape evolves, integration maintenance as the surrounding systems change, prompt iteration as the use cases mature, and the strategic evolution of the architecture as the client's operation grows. The retainer is not IT support. It is architecture stewardship: the ongoing decisions about what to build next, what to upgrade, and what to retire.

Context switching is the hidden cost nobody is measuring

The average company runs 130 SaaS applications (BetterCloud, 2025). Context switching between applications costs the US economy an estimated $450 billion per year in lost productivity (Speakwise, 2025). The average knowledge worker loses approximately 200 hours per year to application switching: time spent moving context from one tool to another, re-establishing state, translating information between systems that don't share a data model.

Five disconnected AI tools are five new context-switching vectors added to the existing 130. Every time a team member moves from the AI writing tool to the AI CRM assistant to the AI email helper, they're paying the context-switching tax again. The architecture problem and the productivity problem are the same problem. A coherent AI architecture eliminates the switching: the data flows between systems automatically, the AI operates on a unified data model, and the human interacts with one surface rather than five.

The pricing principle: consumption pricing is a warning sign

In June 2026, GitHub moved all Copilot plans to consumption-based pricing: pay per AI interaction, not per seat (TechCrunch, June 2026). The reaction from the developer community was immediate and negative, captured succinctly in the widely shared TechCrunch headline paraphrasing the community response. The shift to consumption pricing is not a user-friendly innovation. It is a mechanism for revenue expansion as AI usage increases, creating a situation where the more successfully you adopt the tool, the higher your bill becomes.

The Krastor pricing model is structurally different. The build is fixed price: scoped, quoted, and delivered for a defined cost with no consumption component. The retainer is a flat monthly fee for architecture stewardship, not a usage meter that accelerates as the system handles more work. As usage grows, the per-transaction cost of the architecture declines. The architecture scales; the price doesn't.

This is not a marketing claim. It is what happens when the architecture is owned by the client rather than rented from a vendor. The infrastructure costs $200 per month whether the system handles 1,000 queries or 100,000. Model API costs scale with usage, but usage-based pricing at cost (not marked up) means the scaling cost is the actual cost of compute, not the vendor's margin on top of it. The model costs what it costs. The architecture is yours.