IBM
Why it matters
IBM represents a distinct bet about where the value of AI lies: that most businesses don't need the biggest model, but a trustworthy, affordable one that they can customize with their own data and run where that data lives, with clear licensing and legal protection. Its Granite models are among the most permissively licensed and transparently documented from any major company.
Through Red Hat, IBM is also one of the biggest corporate backers of open-source AI infrastructure. It contributes heavily to projects that much of the industry runs on, including vLLM.
Deep Dive
IBM has been part of AI since the field began. An IBM researcher, Arthur Samuel, coined the term "machine learning" in 1959 while building a checkers program that improved through self-play. IBM's Deep Blue beat world chess champion Garry Kasparov in 1997, and Watson won the quiz show Jeopardy! in 2011. Both were global media events that shaped what the public thought AI was.
Today IBM is not competing to build the largest frontier models. Led by CEO Arvind Krishna, it positions itself as the AI supplier for large enterprises and governments: smaller, open and efficient models, tools for governance, and software to run AI across a company's own data centers and multiple clouds.
The Watson Lesson
After Jeopardy!, IBM tried to turn Watson into a business, above all in healthcare, with promises of transforming cancer treatment. The technology couldn't deliver on the marketing, flagship hospital projects were abandoned, and IBM sold the Watson Health assets in 2022. It's a widely cited cautionary tale about overpromising AI in high-stakes domains, and it helps explain the more sober tone of IBM's current strategy.
watsonx
Launched in 2023, watsonx is IBM's enterprise AI platform. It has three main parts: watsonx.ai, a studio for building with IBM's models as well as third-party and open models; watsonx.data, a data lakehouse; and watsonx.governance, for tracking models, risks and compliance. On top sits watsonx Orchestrate, for building and managing AI agents that automate business tasks. IBM's large consulting arm sells and implements all of it.
Granite
Granite is IBM's family of open models, released under the Apache 2.0 license. Instead of chasing scale, it targets small, efficient models for enterprise tasks such as RAG, summarization, extraction and tool calling.
- Granite started in 2023 with granite.13b on watsonx, and Granite 3.0 (October 2024) brought models of 2 and 8 billion parameters.
- Granite 4.0 (October 2025) introduced a hybrid architecture that combines Mamba state-space layers with transformer blocks, cutting memory use substantially for long inputs and many simultaneous sessions.
- Specialized models include Granite Guardian for safety checks, embedding and reranking models, tiny time-series forecasting models, vision models for document understanding and speech models.
IBM also created Docling, a popular open-source toolkit for converting PDFs and other documents into structured data for AI pipelines.
Red Hat and Open Source
IBM bought Red Hat in 2019 for about $34 billion. Red Hat's AI products include RHEL AI and OpenShift AI, for running and serving models on a company's own infrastructure. In early 2025 Red Hat completed its acquisition of Neural Magic, a leading commercial contributor to the vLLM inference server.
IBM also co-founded the AI Alliance with Meta in December 2023 to promote open AI development, and it created the Agent Communication Protocol (ACP), which merged into the A2A protocol under the Linux Foundation in 2025.
Beyond Models
IBM designs its own hardware, including AI accelerators for its mainframes, and it has one of the largest quantum computing programs in the world. It regularly argues that quantum computing and AI will reinforce each other, though a practical quantum advantage for AI remains an open research question.