Beyond Experimentation: IBM’s Bold AI Vision at Think ‘25

Even lifelong IBMers admit they’ve never seen this level of energy and pace of development from the company. IBM is leveraging its core assets with renewed focus, pushing toward a leadership position in data and AI applications. At IBM Think 2025, the company once again underscored its strategic priorities: open standards, hybrid cloud, and AI. I’d argue there’s a fourth pillar — its unmatched mainframe expertise.

While previous decades were marked by a prolific wave of acquisitions, this era is defined by consolidation and strategic focus. That said, IBM’s recent acquisition of DataStax, announced just weeks before Think, is no exception — it’s a deliberate move to embed deeper AI capabilities across its stack.

In this blog, we unpack IBM’s key priorities and explore how Think 2025 showcased its evolving strategy to lead in the age of AI and hybrid data platforms.

AI Infrastructure and Models

The keynote at IBM Think 2025 centered on a powerful theme: productivity. As IBM Chairman and CEO Arvind Krishna put it, “Productivity is the heartbeat of every organization.” The message was clear — while generative AI has been around for a few years, the moment has come to move beyond experimentation and begin scaling AI to drive real business value, without requiring a massive expansion of IT infrastructure.

IBM’s focus this year is on the vast reservoir of enterprise data still untapped by AI: a staggering 99% of the world’s data, according to the company, much of it unstructured. Rather than competing in the race to build ever-larger foundation models, IBM is positioning itself to lead in context-aware small language models, specifically designed to effectively leverage rich enterprise datasets and purpose-built for specific enterprise use cases.”

Looking ahead, IBM envisions a future where digital workers outnumber human workers 10 to 1 — but where humans are still firmly in charge, directing and overseeing their digital counterparts. This blended model will even extend to IBM Consulting, which will be powered by a combination of human and digital talent working side by side.

IBM also announced preview versions of its upcoming Granite 4.0 models, including the lightweight yet powerful Granite 4.0 Tiny. Although still in training, these models are already demonstrating performance on par with Granite 3.3 2B Instruct, despite having fewer active parameters and requiring approximately 72% less memory. Granite 4.0 Tiny is particularly compelling — it’s designed to run efficiently on commercial-grade GPUs, offering a low-cost, low-footprint option for enterprise AI workloads.

A key innovation in the Granite 4.0 architecture is its departure from a pure Transformer design. Instead, IBM has adopted a hybrid Mamba-2/Transformer architecture, with a 9:1 ratio of Mamba to Transformer blocks. This blend combines the speed and efficiency of Mamba with the precision and contextual strength of Transformer-based self-attention.

The Granite 4.0 Tiny-Preview model is a fine-grained hybrid mixture-of-experts (MoE) architecture featuring 7 billion total parameters, of which only 1 billion are active during inference — enabling significant gains in performance and efficiency.

Mainframe and Linux

In the weeks leading to Think ’25, IBM announced the next generation of the Z series mainframe, the z17. If you thought the era of mainframes was long over, consider this: 70% of the world’s transactions run through mainframes and 95% of global financial institutions use them.

So, why the need for a refresh of the mainframe, you ask? If your guess is to accelerate AI workloads, then you are right. The latest mainframe comes with the Telum® II processor and the forthcoming Spyre AI Accelerator that have an incredible capability to perform 450B inferences per day.

With an ability to handle multiple models and data, z17 can run predictive AI workloads such as fraud detection faster than the previous mainframe generations and at a lower cost. Today’s mainframes already have an on-chip AI acceleration engine. But when workloads involve larger, more complex encoder-based Large Language Models, that’s where Spyre steps in, adding a new dimension of power, precision and possibility.

While a new operating system version, z/OS 3.2 is being readied to run z17, the icing on the cake is the Linux operating system as it opens up the entire ecosystem of applications to run on mainframes. The LinuxOne Emperor 5 is the fifth generation of IBM LinuxONE and comes three years after the release of the last generation. It has been updated for improved security, cost-efficiency, and AI acceleration of mission-critical enterprise workloads.

Not everyone has access to mainframes — we typically run analytics on more commercially available stacks. With that in mind, IBM launched Data Gate for watsonx at IBM Think last year, enabling the synchronization of mission-critical data from mainframe sources (including Db2 for z/OS, VSAM, and IMS) into Apache Iceberg tables managed by watsonx.data. This makes mainframe data readily accessible to distributed AI and analytics applications.

A key architectural detail: The Db2 log reader is optimized for maximum throughput and minimal resource impact. In fact, ingestion was so fast that Apache Iceberg became the bottleneck, prompting IBM Research to optimize high-speed ingestion into Iceberg.

At IBM Think this year, the company extended this vision further by enabling zero-copy integration between mainframe data and Salesforce Data Cloud — eliminating the need to move or duplicate data. This is all made possible through Data Gate and watsonx.data.

More on watsonx.data and Db2 later in this document.

AI Agents and Orchestration

IBM offers two options to build, deploy and manage gen AI applications — watsonx Orchestrate and watsonx.ai. Its goal is to:

1. Build agents in less than 5 minutes
2. Automate and orchestrate agents. IBM has launched over 150 pre-built agents.
3. Enrich agents with enterprise data
4. Deploy everywhere — edge, mainframes and in the cloud

While watsonx.ai has been a mainstay for data scientists and developers, IBM watsonx Orchestrate helps business users enhance productivity:

• Simplify daily routine task execution
• Streamline workflows and essential business processes
• Facilitate seamless team collaboration
• Orchestrate simple tasks — add rows to a spreadsheet, send emails
• Manage complex workflow — claims processing, manage sales, integrate CRM
• Improve employee experience by as AI understands intent
• Augment current investments, not disruptive overhaul of apps infrastructure

This no-code product is designed for non-technical users to build, deploy, and oversee intelligent AI assistants and agents through a natural language interface. In addition, as its name suggests, it enables business orchestration of agents, assistants to improve productivity of business operations, IT processes, financial planning & analysis, customer service, sales & marketing, etc.

IBM watsonx Orchestrate Key components include:

• Pre-built agents: for specific business domains and use-cases, with ready-to-use skills and integrations. Specialized areas include HR, coding, sales enablement (Salesforce, Box, Mailchimp), supplier management (Dun & Bradstreet), purchasing (ServiceNow), sourcing & contracts (Coupa, Ariba, SAP HANA)
• Build your own agents: Simplify the process of building, customizing and deploying agents; from no-code tools that allow you to build an agent in 5 minutes, to pro-code tools for developers.
• Agentic Orchestration: Integrate and automate agents to take on complex projects, enabling agents and assistants to work together across any tool, data source, or infrastructure.
• Agent Observability: Discover, monitor and optimize the use of agents across the enterprise in order to drive trust, performance and efficiency.

IBM watsonx Orchestrate has its own UI as well as can be used from 3rd party applications like Slack, WhatsApp, and Facebook Messenger. It runs on IBM’s own cloud or on AWS. It uses OpenAPI to connect to the applications mentioned above and is adding support for Anthropic’s Model Context Protocol (MCP).

While MCP establishes a standard for agent-to-tool communication, agent-to-agent communication is being addressed by a few emerging protocols — namely, Google’s A2A protocol and the Agent Communication Protocol (ACP) developed by IBM under the Linux Foundation.

watsonx.ai provides developers with an all-inclusive AI development studio, including AI Agent build, deployment and monitoring.

In addition, IBM Research has built the Agent Framework that uses IBM’s own open-source Bee framework or external ones like Crew AI, Microsoft Copilot or Salesforce Agentforce. It also integrates with a wider IBM ecosystem comprising Watsonx Code Assistant, IBM Business Automation Workflow (BAW), Granite LLMs, watsonx.data lakehouse, watsonx.governance, and 3rd party models.

Before we end this section, it is interesting to note that while IBM Think ’25 was going on, several other vendor user conferences were also in session, such as ServiceNow and Nutanix. Many weary analysts were criss-crossing the country trying to catch as much action as possible. At ServiceNow’s Knowledge 2025 user conference, the vendor announced their own automation approach called AI Control Tower for orchestration of agent workflows and listed IBM as one of its partners.

Lakehouse and Data Fabric

Launched at IBM Think ’23, IBM’s modern data lakehouse platform, watsonx.data, specifically engineered for AI and analytics workloads, has consistently evolved with notable new features over the past two years. Demonstrating IBM’s commitment to an open and interoperable data ecosystem within its watsonx platform, Apache Iceberg has been embraced as its default open table format.

Further underscoring IBM’s strategic focus on simplifying the development of data products by unifying disparate data sources are the recent releases of watsonx.data integration and watsonx.data intelligence. These additions highlight IBM’s ongoing investment in providing a streamlined and unified data foundation for modern data-driven initiatives.

watsonx.data integration is a SaaS product designed to unify various types of data integration approaches using a single control plane that supports ETL, ELT, real-time streaming, replication, data quality and observability. It also extends data integration space into unstructured data processing. Its no/low code drag & drop user interface is augmented with AI assistants and an SDK. Users specify tasks into the Assistant and it generates the flow diagram and the necessary Python code.

watsonx.data integration will be GA in June 2025. Its roadmap will support multiple deployment and runtime environments to optimize pipeline execution by pushing down processing to the multiple engines and will use Apache Arrow Flight for data exchange interoperability.

watsonx.data intelligence is a SaaS product (GA in June 2025) that becomes a single pane of glass to curate, manage, and process meaningful data through various products pertaining to data governance, lineage, data sharing, and data quality.

Figure 1 illustrates the diverse components of IBM’s rich offerings, all strategically integrated to deliver a comprehensive data fabric.

Figure 1: IBM’s AI-driven fabric offering is open by design, engineered to work in hybrid, multi-cloud environments and provides optimal price-performance.

These two new additions are geared to accelerate insights by providing automation and enhancing productivity. The watsonx.data offerings consolidate the experience of using individual products like IBM Knowledge Catalog, DataBand, Manta Data Lineage, Streamsets (more on this later), and Data Product Hub.

Third announcement in watsonx.data pertains to its integration with Meta’s Llama Stack. Now developers in Llama Stack can access watsonx.data via its APIs.

The final announcement in this space is for the Datastax acquisition which we will cover next.

Datastax and Unstructured Data

DataStax is a significant contributor to the NoSQL landscape, primarily recognized as the company behind the open-source Apache Cassandra database. This powerful database has gained widespread adoption among organizations like Apple and Uber that demand highly distributed, scalable, low-latency, and fault-tolerant data management solutions capable of operating at extreme scale. DataStax also offers Astra DB, a fully managed Database-as-a-Service (DBaaS) version of Cassandra, simplifying its deployment and management in the cloud.

In a notable move, IBM announced its intent to acquire DataStax in February 2025. While DataStax is well known for its leadership in open-source Cassandra, IBM’s interest goes beyond the database — it’s also about DataStax’s exceptional generative AI capabilities. A major theme at IBM Think ’25 was tackling the challenge of unstructured data, and this acquisition aligns squarely with that focus.

IBM’s ambitious strategy for its watsonx.data platform centers on a powerful synergy between key open-source technologies and strategic acquisitions, aiming to revolutionize how organizations manage and leverage both structured and, crucially, unstructured data for AI and analytics. This vision strategically combines Apache Iceberg as the foundational open table format with the open-source Milvus vector database for efficient similarity search and the capabilities of DataStax Astra DB.

Mirroring their acquisition of HashiCorp, intended to empower DevOps engineers with “shift-left” capabilities, IBM views the acquisition of DataStax as a strategic move to similarly empower developers. They plan to achieve this “shift-left” for data application development through two key DataStax assets:

• DataStax Langflow: This low-code/no-code platform simplifies the creation of complex AI pipelines, including RAG implementations. By abstracting away intricate coding requirements, Langflow enables developers to rapidly build and deploy AI-powered applications that can effectively reason over unstructured data.
• DataStax JSON Support: Native JSON document support within Astra provides developers with a flexible and familiar way to work with semi-structured and unstructured data directly within a highly scalable database. This eliminates the need for complex data transformations and simplifies the integration of diverse data types into AI and analytics workflows.

Databases

IBM has a rich portfolio of database technologies, including its flagship Db2, Informix, and Netezza. As previously discussed, it also offers a lakehouse solution in the data fabric space through watsonx.data, which includes the open-source Milvus vector database.

IBM announced several new enhancements to Db2, its trusted engine for both operational and analytical workloads, including:

• Db2 Intelligence Center: an AI-powered, unified database management platform designed to streamline Db2 administration across on-premises, hybrid, and cloud environments. At the heart of the offering is a new generative AI-powered database assistant that acts as a Db2 expert, provides recommendations on query optimizations, assists with root-cause analysis, and much more. The platform also provides deep monitoring, rapid troubleshooting, smarter query optimization, and advanced automation.
• Vector store and similarity search: Db2 12.1.2 introduces a new vector data type. Storing vector data alongside traditional scalar data makes it much simpler to do retrieval-augmented generation (RAG). Queries using familiar SQL syntax and built-in functions can now easily incorporate vector data and AI capabilities into existing apps and services with minimal code changes.
• Managed Db2 on Azure: Using the bring your own cloud (BYOC) deployment model, IBM Db2 and Db2 Warehouse SaaS will be GA on Azure in June 2025. In addition, Db2 data will be able to reside natively on Azure Blob Storage, making it easy for Db2 customers to run warehousing workloads on Azure at scale. Db2 will also be able to work with open formats, like Iceberg and Parquet on data residing in Azure Blob storage, accessed as an external table from Db2.
• Native Apache Iceberg: Iceberg tables now have full support for CRUD (create, read, update and delete) operations, making it easy for customers to unify their analytics across their Db2 Warehouse and lakehouse services.

Other Db2 enhancements pertain to Db2’s pureScale feature which ensures high availability and scalability.

IBM Netezza’s resurgence in the area of Data Analytics has been remarkable. Traditionally known for its blazing fast speed and built-in machine learning capabilities, it has transformed itself as a cloud agnostic massively parallel hybrid data warehouse engine embracing open standards, such as Apache Iceberg and Parquet, along with infused smart AI scaling and database assistant capabilities. Several new enhancements were announced recently:

• Netezza BYOC on native cloud object storage: The new BYOC (Bring Your Own Cloud) deployment model complements its existing Fully Managed Netezza SaaS offering available on AWS and Azure. Now, businesses can deploy the Netezza data warehouse engine directly within their own AWS/Azure/GCP Virtual Private Cloud (VPC) thereby enabling greater security and control, providing cost transparency and optimization, and the ability to draw down on their cloud provider credits. Netezza BYOC on AWS will be GA in June 2025 followed by Azure and GCP in the second half of 2025.
• Next Generation Netezza Appliance: This new offering is a follow-on to its on-premises IBM Netezza Performance Server and is designed to provide up to 2X performance improvement with seamless on-ramp to its public cloud fully managed and BYOC Netezza SaaS offerings. It is expected to be released in the second half of 2025.

Data Integration

We have covered data integration in this blog earlier in the section on lakehouse and data fabric. However, IBM’s cup runneth over in this section as there are two other products that also got a lot of attention at their user conference — webMethods and StreamSets. These two products were acquired from Software AG about nine months ago.

IBM webMethods Hybrid Integration addresses the enterprise integration needs by unifying diverse integration patterns (APIs, applications, B2B, files, events, and mainframe data) into a single platform, thereby reducing complexity. The goal is to enable AI-driven automation across hybrid and multicloud environments. It leverages agentic AI (via IBM watsonx) to automate complex integrations, supports mainframe data access, and offers a unified control plane for end-to-end monitoring across hybrid landscapes. Its consumption-based pricing and compatibility with hyperscalers like AWS and Azure further enhance flexibility.

IBM’s advancements in its broader data and AI portfolio includes StreamSets as a key component for real-time data integration. Its integration with IBM’s data fabric and AI platforms, such as watsonx.data, was showcased as part of IBM’s strategy to address data silos and enhance real-time decision-making. To understand the role it plays in delivering real-time data to AI agents, please watch this video.

Conclusion

IBM estimates that over one billion apps will emerge by 2028. Arvind Krishna also declared “The era of AI experimentation is over.” With this backdrop, IBM is planting its flag across the full stack — infrastructure to applications to become a leader in the rapidly emerging AI-powered brave new world that is much more integrated and inclusive of structured and unstructured data.

IBM is combining the prowess of its open and hybrid technology with deep industry expertise from IBM Consulting to get ready for the onslaught of new AI-driven apps, fueling them with AI-ready data. Its new announcements are geared to simplify a fragmented data infrastructure, orchestrate AI agents, and to drastically reduce the time needed to deliver insights.

Some key messages that emerged from the event are:

• IBM’s All In on AI: IBM is pushing hard to be a leader in AI and data, moving beyond just experimenting to actually using AI in businesses. Think ’25 really emphasized this big push. They are making bold bets on AI agents. IBM watsonx Orchestrate allows business users to create and manage these agents, even without coding. Over 150 pre-built agents are available!
• Smart Models: Instead of just creating huge AI models, IBM is focusing on smaller, more specific ones that can use a company’s data really well. For example, Granite 4.0 Tiny is a Big Deal. It’s meant to be efficient and cost-effective and blends Mamba and Transformer designs.
• Watsonx.data is the Hub: IBM’s watsonx.data platform is the place where all the data lives for AI and analytics. It uses Apache Iceberg as its default open table format and embeds Milvus open-source vector database.
• DataStax Acquisition is Pivotal: IBM bought DataStax, which is known for Cassandra and its AI capabilities. This helps IBM handle unstructured data better, like documents and images.
• Mainframes Are Still Key: Believe it or not, mainframes are still super important and they just launched the z17 that’s designed to handle AI tasks faster. It’s got new processors like Telum® II and the upcoming Spyre AI Accelerator. And a new version of Linux.
• Db2 Gets an AI Upgrade: IBM’s Db2 database is getting smarter with AI features, like an AI assistant for managing the database and better support for handling AI data. Db2 is also now coming to Azure.
• Netezza is Revamped: IBM’s Netezza is making a comeback, with faster speeds and better AI features. It’s also available in more ways, including “bring your own cloud” options.
• Data Integration Gets Easier: IBM is focusing on making it easier to connect different data systems together. webMethods and StreamSets are helping with real-time data.

Addendum: For a deeper dive straight from the event floor, watch my unscripted conversation on theCUBE at IBM Think ’25, where we discuss key announcements and strategic insights on cloud, mainframes, DataStax, RedHat, AI Agents, and much more. Watch the full video here: Sanjeev Mohan, SanjMo | IBM Think 2025