Back to News
Article

The Case Against Vendor-Locked Infrastructure.

We keep hearing about large AI investments. When we ask what the money went to, the answer is usually platform licenses and consulting fees. In a well-implemented AI infrastructure, most of the cost comes from tokens and in-house talent. Not vendors. Not consultants.

March 2026 · 5 min read
01

Your Vendors Are the Bottleneck

The large platform vendors see AI the same way they see every technology shift: as a reason to sell more licenses. Microsoft repackages existing data products as Fabric. Copilot is $30 per user per month. Salesforce adds an AI label to features that existed last year. The pattern is the same. Take what you already have, attach AI to the name, charge a premium.

The problem is not just the cost. These vendors move slowly. Their release cycles are quarterly. Their AI capabilities lag months behind what is available in the open market. When you build on top of them, their pace becomes your pace. Their limitations become your limitations. You end up paying more for less, and the gap between what is possible and what your platform supports grows wider every month.

This applies to AWS, Google Cloud, Salesforce, Oracle, and every other major platform vendor. It also applies to AI model providers. The best model today may not be the best model in six months. Build so you can switch. The moment your infrastructure depends on a single vendor's roadmap, you have handed over control of your own pace of innovation.

02

Most Companies Do Not Need a Hyperscaler

A Postgres database on Supabase. A frontend on Vercel. Async jobs on Modal. That combination runs internal tools, customer applications, analytics, and automation for a fraction of what the equivalent costs on AWS or Azure. No VPC configuration. No IAM policy debugging. Simpler to manage, faster to ship.

Hyperscalers are the right choice when you need them. Petabyte-scale data, millions of concurrent users, globally distributed compute. Most mid-market companies are not there. They are paying for complexity they do not use.

Build infrastructure you can migrate away from. That single principle changes every technology decision that follows.

We are not suggesting anyone tear down working systems. But the cost of running production software has dropped, and the platforms available today are built for portability. The organisations that will navigate the next few years well are the ones that can adopt new tools in weeks, not years.

03

Tokens Are an Operating Cost

Every AI feature consumes tokens. Most organisations have limited visibility into how many, from which models, and whether those models are the right fit for the task. Claude Opus, GPT-4o, and Gemini Pro exist for complex work. Haiku, GPT-4o-mini, and Gemini Flash exist for simple work. Using a top-tier model for a classification task or a draft email is technically fine but economically wasteful at scale.

Token pricing will increase as enterprise adoption grows. The same dynamic played out with cloud compute. Organisations that build cost tracking and model routing early will manage this. Those running everything through a single expensive model with no monitoring will not.

In practice, this means an infrastructure layer with visibility into every model call, cost attribution per workflow and per team, and automatic routing based on task complexity. This is not something to add later. It is part of the foundation.

04

If It Is Not Working, Change the Approach

If you have implemented AI and the results are disappointing, do not invest more in the same direction. The learnings and problems from your first projects are not wasted. They are useful input. But the approach needs to change. Implementations done well are fast and focused. If yours was slow and unclear, the methodology was the problem.

For organisations starting from zero, the right first project is small. One manual workflow, well understood by the team doing it. Automate it. Measure the outcome. Low risk, visible value, low cost. If it works, expand from there. If it does not, you have spent weeks learning, not years.

Build infrastructure that
moves with you.

Speak with our team →