Software used to be written. Now it is described.

For forty years, getting custom software meant the same three things. You defined what you wanted. You found people who could build it. Then you paid them, waited, and hoped what came back matched what you meant. A simple business app could run tens of thousands of dollars before it did anything useful.

That equation has broken. Large language models can now write, test and fix working software from a plain description. The cost of producing code has fallen so far that it is no longer the hard part. What stays hard is everything around the code: knowing what to build, keeping it running, keeping data safe, and getting customers to use it.

To see where the opportunity is, it helps to understand the machinery. The guide below walks the AI stack from the bottom layer to the top. Each layer ends with a short note on what it means for a business owner.

A large language model (LLM) is a program trained to predict the next piece of text. It reads text as tokens, which are short chunks of words, and learns patterns across an enormous amount of writing, code and documentation. Predicting the next token sounds modest. At large enough scale it produces something that can summarize a contract, answer a customer, or write a working database query.

Models come in sizes. Frontier models from labs such as Anthropic, OpenAI and Google are the largest and most capable, and they handle hard reasoning and long, multi-step work. Small models are faster and far cheaper, and they handle simpler jobs like sorting an inbox or tagging a lead. Many models are also multimodal: they take in images, audio and documents, and some speak back in real time.

The money behind this layer is enormous. The 2026 Stanford AI Index puts U.S. AI investment in 2025 at about $286 billion, more than ten times Europe's figure. Almost none of that is spent by the businesses that end up using the models. It is spent so that they don't have to.

Pre-training is the expensive first stage. The model reads a vast corpus and learns language, facts and the structure of code. It takes huge clusters of specialized chips running for weeks or months, which is why only a handful of organizations do it from scratch.

Post-training turns that raw knowledge into a useful assistant. It usually combines three techniques: fine-tuning on high-quality examples, learning from human feedback where people rank answers, and reinforcement learning (RL), where the model attempts tasks and is rewarded when it succeeds.

RL works best where success can be checked automatically. Code is the clearest case: it either runs and passes its tests or it does not. That is why models have become dramatically better at software work, and why they can now plan, act and recover across many steps. The 2026 AI Index shows how fast this is moving on OSWorld, a test where AI agents must operate a real computer to finish tasks.

Inference means running a trained model to produce an answer. Training happens once. Inference happens every time anyone uses the model, so it is the cost that grows with your business. Providers price it per token, usually quoted per million tokens.

The price of a fixed level of capability has collapsed. The Stanford AI Index found that GPT-3.5-level performance fell from $20 to $0.07 per million tokens in under two years. Epoch AI's analysis across six benchmarks found prices for a given level of performance falling anywhere from 9 to 900 times per year, depending on the task.

Two forces push the other way. Reasoning models deliberately spend more tokens "thinking" before they answer, which buys better results on hard problems at a higher cost per answer. And latency matters as much as price. A chatbot can take a few seconds, but a live phone call cannot, so voice agents rely on fast, purpose-built speech models.

A model's weights are the billions of numbers it learned in training. Closed models keep their weights private, and you reach them only through the provider's service. Open-weight models, such as Meta's Llama, Mistral, Alibaba's Qwen and DeepSeek, publish their weights so anyone can run them on their own servers or in any cloud.

"Open weights" is not the same as open source. The training data and code are often not released, and licenses vary in what they allow. But the quality gap has nearly closed, and usage has followed. On OpenRouter, open-weight models handled roughly a third of all tokens by late 2025, and Chinese open models alone averaged about 13% of weekly volume over the year.

Because models differ in cost, speed and skill, a new layer has grown up to choose between them. OpenRouter, one of the largest, connects more than five million developers to over 300 models from more than 60 providers through one connection. Companies also run their own AI gateways between their software and the model providers.

A router decides, call by call, which model gets the job. If one provider is down, it fails over to another. The same layer can enforce rules, such as masking card numbers before a prompt leaves the building, and it can keep an exact ledger of what each call cost.

OpenRouter's data from 2025 also shows what people use AI for, and it changed fast. Programming went from a small slice of traffic to the majority, and reasoning models went from almost nothing to more than half of all tokens. Software is now the biggest single job AI does.

Every time an app sends a prompt to a model, some of your data travels with it: a customer's name, an invoice, a medical note, a pricing sheet. Data sovereignty is about controlling where that data is processed and stored, who can see it, and which laws apply. Rules such as Europe's GDPR, along with industry requirements in healthcare and finance, make this a legal question as well as a trust question.

Spending shows how seriously organizations now take it. Gartner forecasts worldwide spending on sovereign cloud infrastructure will reach about $80 billion in 2026, up 35.6% in a year, with Europe expected to pass North America in 2027. It also estimates that about a fifth of current workloads will move from global to local cloud providers.

The practical tools are well understood. Choose providers that commit to not training on or keeping your data. Keep processing in the right region. Mask sensitive fields before they leave. For the most sensitive work, run open-weight models in an environment you control. The routing layer is the natural place to enforce all of this, because every AI call already passes through it.

Add the layers up and the picture is clear. Models keep getting better, post-training keeps making them better at code, inference keeps getting cheaper, and open weights and routers keep prices competitive. The bottom of the stack is turning into a utility.

The builders have noticed. Lovable, one of the AI app builders, went from launch in November 2024 to a $200 million annual run rate by December 2025, and was tracking toward about $600 million by August 2026. Its users have created some 60 million projects, and the company says nearly eight in ten builders plan to make money from what they build.

Yet most small businesses are not using AI at all. The U.S. Census Bureau's survey found about one in five firms using AI in May 2026. Among the largest firms it was 37%. Among the smallest firms, with four or fewer employees, it was under 20%, and use at firms with fewer than 20 employees did not grow over the previous six months.

So there are two booming facts and one gap between them. Making software has never been cheaper, and the builders are growing fast. But a prototype is not a business, and the owners who could gain most are the least served. The questions they live with sit at the top of the stack:

• Does it run reliably, with updates, backups and security handled, a year from now?
• Is the AI inside it governed, with the right model for each job, sensible costs, and data kept where it belongs?
• Does it bring in customers, or does it sit on a domain nobody visits?

The last one is the biggest gap. AI made software cheap to build, but it did nothing to make customers easier to find.

sup3rapp was built by people who spent decades selling and building sales and marketing software, and who were asked over and over to find a team that could reliably build a customer's product. Our approach follows from everything above:

• Start from the job, not the code. You describe what your business needs. We turn it into a working app.
• The right model for every task. Each AI feature in your app is routed to the model that does that job best for the money, with automatic fallback.
• Deployed and maintained. Your app goes live on managed cloud hosting and stays updated. Nothing gets handed off and abandoned.
• Governed by default. Sensitive fields are masked, usage is logged, and costs are capped.
• Growth built in. Optional modules for landing pages, campaigns, an AI chat and phone assistant, and lead routing come with the Growth and Business plans.

Read our full approach →

sup3rapp is custom software for a segment of one: your business. Tell us what you need, such as a client portal, an intake and booking app, a quoting tool or a membership site. sup3rapp builds it, puts it live, keeps it running, and can help bring it customers through Kai, our AI assistant that answers chats and calls and books the work.

There's a free plan, and paid plans start at $24 a month.