What’s next in computing is generative and quantum
At Think 2025, IBM Research leaders explored how generative computing swaps AI prompting for programming, and outlined that quantum advantage is just around the corner.
At IBM’s flagship Think conference this week, the company demonstrated how businesses are putting its latest advances in AI, hybrid cloud, hardware, and quantum computing to work. IBM has been pushing the computer industry forward for as long as there has been one, and on Wednesday, IBM Research leaders took the stage to show off what’s next in computing and discuss the wide-reaching implications of two of the fastest moving areas of tech today — AI and quantum computing.
For AI, this means the debut of generative computing — a new way to interface with large language models. Generative computing will center the LLM as a compute element with a runtime built around it. For IBM’s clients, this development will make building AI agents and applications more secure, portable, maintainable, and efficient, said IBM Research VP of AI Sriram Raghavan. “It isn’t every day that a new computing element shows up in our industry,” he said. “Generative computing is a way to move away from prompting to real programming.”
And for quantum computing, the next two years will bring quantum advantage, meaning that IBM’s quantum computers will be able to perform calculations of practical, commercial, or scientific importance, more cost-effectively, faster, or with greater accuracy than a classical computer alone could achieve.
Taking the stage first, IBM Research Director of Science and Technology Zaira Nazario highlighted the founding of the “Mathematics of Computation” group within IBM Research. “I’m proud to be part of the rich history of theory at IBM,” she said, pointing out that IBM has been the world’s top institutional contributor of algorithms. “We’ve been pushing the computing industry forward as long as there’s been one,” said Nazario.
From the IBM Think stage, Raghavan outlined how LLMs need a better interface than prompt engineering. For a consumer using a chatbot to plan a vacation, prompting a chatbot to refine its output may be fine. But it’s not acceptable for enterprises using LLMs as sophisticated compute tools. An interface with LLMs must be secure, portable, maintainable, and efficient.
“It’s a recognition that an LLM is a fundamentally new computing environment,” said Raghavan. Along with the progression of application frameworks, from chatbots to assistants to agents, there must be new ways of interfacing with them. Now we’re asking models to do more: not just answer questions but also perform retrieval-augmented generation (RAG) and orchestration. Data, too, has ballooned from text documents to speech, tabular data, JSON files, and more.
“But has the fundamental mechanism for interacting with models changed?” Raghavan asked. “No. We continue to do prompt engineering.” Soon enough, he said, the industry will have books and books of prompts for AI, an untenable situation. Besides, even well-engineered prompts are brittle, and they’re meant for specific models. They aren’t portable.
When you interact with an LLM today, you use an API to interact with a model via tokens. With generative computing, IBM Research has replaced the API with a runtime that’s equipped with programming abstractions. These abstractions can create safety guardrails, outline structured requirements (including explicit calls that software at runtime can check and enforce), and use instructions to implement a generation strategy.
And this strategy will pay off for enterprises. Generative computing will make it possible to use much smaller models to achieve the same level of accuracy compared to the use of prompts, Raghavan said.
The runtime can also be used to detect hallucination, bias, and prompt injection, he explained. “As opposed to hoping that a piece of English will get correctly interpreted, you have well-trained adaptors that are going to implement security checks.”
Portability, too, is enabled by generative computing because the runtime’s structured abstractions aren’t programmed to meet the whims of a specific model.
Demonstrating the organization’s commitment to generative computing, Raghavan announced that IBM is releasing a Granite runtime, as well as the next generation of Granite 4.0 models (including one small enough to fit on a single GPU), which will come later this summer. Utilizing state-space models, transformer approaches, and a mixture-of-experts approach, early benchmarks suggest these models can perform inference two to five times more quickly than comparable models, Raghavan said.
“We’re really making cutting-edge technologies more performant and more consistent,” added Nazario, emphasizing what the new runtime and models will mean for enterprises. “We’re making them more cost effective, more flexible.”
The announcements were no less groundbreaking on the quantum side. “We believe quantum advantage will actually happen in 2026,” said IBM Research VP of Quantum Computing Jay Gambetta from the 2025 Think stage.
Reaching quantum advantage depends on cooperation between the quantum computing and high-performance computing communities, in the form of quantum-centric supercomputing, a compute paradigm that combines classical and quantum computing. “It’s not about classical versus quantum computing,” Gambetta said. “It’s about quantum plus classical.”
The belief that IBM will achieve quantum advantage by 2026 is rooted in IBM’s leadership and approach to quantum computing, treating it as an engineering problem rather than as a science project. Gambetta showed off IBM’s series of quantum processors and their relevant packaging technologies as evidence of its leadership in that space, backed up by IBM's long history of expertise in semiconductors.
In line with this, Gambetta’s team is working on new algorithms for quantum-centric supercomputers, such as sample-based quantum diagonalization (SQD), to shift what is possible with today’s pre-fault-tolerant quantum devices. Working with RIKEN in Japan, IBM Research was able to use the SQD technique to accurately simulate the ground state energy of [4Fe-4S], a 77-qubit problem beyond the scale of what's amenable to exact diagonalization methods on classical computers.
IBM Quantum expects to see advantage first in chemistry or materials science, then in optimization, and finally mathematical problems. With quantum advantage just around the corner, the IBM Quantum roadmap continues to push out toward fault-tolerant quantum computing with IBM Quantum Starling in 2029.