188% of Organizations Now Use AI, and 40% of Large Companies Are Scaling Agents

The gap between the 88% and 44% figures is the most useful number on this page. Nearly every organization uses AI somewhere, but only about half have moved past pilots into enterprise-wide deployment. Using a chatbot is easy. Getting a model into production reliably is a software engineering problem, and it is the hard part of building AI products for real use.

McKinsey's measured jump in agent scaling, from 27% to 40% of large companies, points the same way as Gartner's forecast. Agents connect models to real systems such as calendars, CRMs and ticket queues. That is why the workflow automation tools that already link those systems are a common first step. For teams without in-house engineers, the 44% scaling figure also explains the demand for outside help, including nearshore agencies that build and maintain AI features.

2NVIDIA Booked $215.9 Billion in Fiscal 2026 as AI Compute Demand Surged

Data centers accounted for roughly 90% of NVIDIA's fiscal 2026 revenue ($193.7B of $215.9B, calculated from NVIDIA's reported figures). NVIDIA attributes that demand to AI compute, so its results work as a direct measure of what is being spent on the hardware that AI software is trained and served on.

The two model makers' numbers carry a caveat: both are privately held, so the figures come from press reports, not audited filings. Read them as directional. Even so, OpenAI near $70 billion and Anthropic above $65 billion show how much enterprise spending now flows through model APIs. Businesses reach those models through products and APIs, often as custom assistants built on top of them.

3Which Tools Do AI Developers Rely On?

Copilot's reach in the Fortune 100 means AI-assisted coding is now standard practice in large companies. For people who don't code, the same shift has made it more realistic to build working apps without writing every line by hand. Anyone still choosing which programming language to learn first will now learn it alongside an assistant.

PyTorch is an open-source framework built to shorten the path from research prototype to production. CUDA lets developers write kernels that run in parallel across GPU cores for training and inference. Together they connect the software stack to the NVIDIA revenue in the table above. They also explain why GPU memory and cores matter when choosing a laptop for data science and model experiments. Hugging Face, with more than 2 million public models, is where finished models are shared and downloaded.

4Inference Got 280x Cheaper: The Technologies Behind Modern AI Software

The 2017 paper "Attention Is All You Need" introduced the transformer. Its self-attention mechanism processes a whole sequence in parallel instead of one step at a time, which is why nearly every modern large language model uses it. A separate 2017 paper introduced mixture-of-experts layers, which send each input to only a few specialized sub-networks. That lets a model hold a very large number of parameters while keeping per-token compute low.

Three later techniques shape how companies customize models. Retrieval-augmented generation, published in 2020, connects a model to a document retriever so its answers are grounded in up-to-date knowledge without retraining. It is the pattern behind chatbots that answer from a company's own documents. LoRA, published in 2021, freezes the original weights and trains small matrices instead, cutting trainable parameters by orders of magnitude. That is the kind of saving that no-code AI and ML tools pass on to non-specialists. RLHF, shown at scale in OpenAI's 2022 InstructGPT paper, aligns models with human intent using a reward model trained on human preference rankings.

5What Are the Stages of AI Software Development?

Of the eight steps, only one is model training. Three come before it (extracting, analyzing and preparing data) and four come after (evaluating, validating, serving and monitoring). Seven of the eight steps are about data handling and operations, not modelling. That is why no-code machine learning platforms concentrate on automating the data and deployment steps.

Continuous training is the clearest difference from ordinary software delivery. A web app keeps working until someone changes its code. A model can get worse as the world it was trained on changes, so a level 1 pipeline retrains it automatically. Level 2 adds full CI/CD automation, so the pipeline itself is tested and deployed like any other code. The self-assessment below places a team on Google's scale.

6How Is AI Software Regulated in 2026?

For development teams, the Act turns risk classification into a design input. Teams have to decide which of the four risk levels a system falls into before building it, and some uses are banned outright. With broad applicability now in effect, compliance work belongs in the validation and monitoring stages of the workflow above. The rules cover legitimate builders. They do little about the attackers using AI to scale their attacks, so security remains a separate job.