Part I: The Man Behind IBM's AI Resurrection

In October 2024, IBM released Granite 3.0, its latest family of enterprise AI models. The technical specifications were unremarkable by frontier model standards—8 billion and 2 billion parameter variants trained on 12 trillion tokens. No breathless claims of AGI capabilities. No promises to revolutionize how humans work.

What made Granite 3.0 significant was not what IBM promised, but what it did not. While OpenAI, Anthropic, and Google raced to build ever-larger foundation models capable of dazzling the public, IBM Research—under the leadership of Vice President Sriram Raghavan—was executing a contrarian strategy: building smaller, faster, auditable AI systems optimized for the unglamorous work of enterprise compliance.

Three weeks after the Granite 3.0 launch, IBM reported that its generative AI book of business had reached $9.5 billion inception-to-date, with approximately 80 percent coming from consulting engagements and 20 percent from software sales. The company projected that AI-enabled productivity savings would reach a $4.5 billion annual run-rate by the end of fiscal year 2025.

These numbers represented more than business growth. They marked the tentative rehabilitation of Watson, IBM's once-vaunted AI system that had spectacularly failed to deliver on its promises in healthcare, burned through billions in R&D spending, and left the company's AI credibility in tatters.

Sriram Raghavan, a 20-year IBM Research veteran who rose from staff researcher at the Almaden Research Center to leading a worldwide team of over 750 AI scientists and engineers, now carries the technical responsibility for IBM's AI future. His mandate: prove that in the race between frontier model capabilities and enterprise trust, trust can win.

Part II: The $4 Billion Humiliation

To understand IBM's current AI strategy, one must first understand the magnitude of Watson's failure.

In 2011, IBM's Watson AI system defeated human champions Ken Jennings and Brad Rutter on the quiz show Jeopardy, demonstrating what appeared to be human-level natural language understanding and reasoning. IBM CEO Virginia Rometty seized on Watson's victory to position IBM as the leader in cognitive computing, pouring billions into Watson Health with promises to revolutionize cancer treatment, drug discovery, and medical diagnosis.

By 2019, reality had set in. MD Anderson Cancer Center, one of Watson's flagship partners, had terminated its collaboration in 2015 after spending $62 million over two years with virtually nothing to show for it. Internal IBM documents revealed that Watson for Oncology provided "unsafe and incorrect treatment recommendations" in multiple cases.

A 2022 investigation documented the extent of the disaster. Around 50 partnerships between IBM Watson and healthcare organizations had been announced, including collaborations with the Mayo Clinic and national organizations for cancer, cardiology, and oncological research. None had produced usable clinical tools or apps. By 2018, more than a dozen IBM partners and clients had stopped or scaled back their oncology projects with Watson.

In 2022, IBM sold Watson Health to private equity firm Francisco Partners for approximately $1 billion—roughly one-quarter of the $5 billion IBM had spent on acquisitions alone for the division, not counting the additional billions in R&D expenditures.

The Watson Health failure left IBM with a credibility problem. In the emerging era of generative AI dominated by OpenAI's ChatGPT, Anthropic's Claude, and Google's Gemini, how could IBM convince enterprises that this time would be different?

Part III: The Scientist Who Built IBM Research India

Sriram Raghavan joined IBM Research's Almaden Center in San Jose, California, in 2004, fresh from completing his computer science PhD at Stanford University. His early work focused on natural language processing, data management, and distributed systems—technical domains that would prove directly relevant to the AI revolution that would unfold two decades later.

His most notable early contribution was SystemT, a declarative information extraction system that addressed fundamental limitations in classical NLP techniques. SystemT's approach—using declarative rules rather than purely statistical methods—foreshadowed IBM's eventual emphasis on explainability and auditability in AI systems.

In the mid-2010s, Raghavan transitioned from technical contributor to leadership roles. He was appointed Director of the IBM Research Lab in India and simultaneously served as CTO for IBM in India and South Asia. During this period, he established IBM Research India as a world-class center of competency for blockchain technology, serving as worldwide blockchain leader for IBM Research.

The blockchain work proved significant for unexpected reasons. Blockchain's core value proposition—transparency, auditability, and trustless verification—would later inform IBM's approach to enterprise AI governance. The technical parallels were obvious: both blockchain and enterprise AI required verifiable provenance, explainable decision-making, and compliance with regulatory frameworks.

By the time Raghavan was promoted to Vice President of IBM Research AI in the early 2020s, he led a global team of over 750 research scientists and engineers across all IBM Research locations. His research portfolio encompassed both foundational AI (advancing the state of the art in machine learning, natural language processing, and computer vision) and applied AI (integrating research innovations into IBM's commercial products).

Raghavan's leadership team was responsible for executing CEO Arvind Krishna's vision to rehabilitate IBM's AI brand through watsonx, a complete reboot of the Watson platform launched in May 2023. Unlike Watson's attempt to be an all-purpose AI solution, watsonx focused on providing enterprise-ready AI tools and infrastructure specifically optimized for business use cases.

Part IV: The Enterprise AI Divergence

While OpenAI, Anthropic, and Google DeepMind pursued frontier model development—building ever-larger language models trained on ever-more data to maximize general capabilities—IBM Research under Raghavan's technical leadership pursued a fundamentally different strategy.

The divergence centered on three core principles: explainability, governance, and efficiency.

Explainability: The Black Box Problem

Frontier models like GPT-4, Claude 3.5 Sonnet, and Gemini 1.5 Pro operate as black boxes. Even their creators cannot fully explain why these models produce specific outputs for given inputs. For consumer applications, this opacity is acceptable. For regulated industries—banking, healthcare, insurance, government—it is disqualifying.

A 2024 IBM Institute for Business Value study found that 80 percent of business leaders considered AI explainability, ethics, bias, or trust a major roadblock to generative AI adoption. Unpublished research revealed that enterprises were willing to accept lower raw performance in exchange for explainable AI systems that could survive regulatory audit.

IBM Research developed specific explainability capabilities integrated into watsonx.governance. The platform generates reason codes and feature importance rankings for individual predictions, showing visually why a model made specific decisions. For deeper analysis, IBM offers the AI Explainability 360 toolkit, providing techniques like SHAP values and counterfactual explanations.

This approach directly contradicted the frontier model paradigm. OpenAI and Anthropic focused on alignment—ensuring models behave as intended—but largely accepted interpretability limitations as a necessary trade-off for capability. IBM bet that for enterprise customers, interpretability constraints would prove more valuable than marginal capability gains.

Governance: The Compliance Infrastructure

Watsonx.governance, launched in December 2023, represented IBM's most direct challenge to hyperscaler AI strategies. The platform helps organizations monitor and govern the entire AI lifecycle, automating risk management, monitoring models for bias and drift, capturing model metadata, and facilitating organization-wide compliance.

IBM built a baseline governance framework mapping to multiple regulatory regimes: risk-based controls for the EU AI Act, NIST's trustworthy AI criteria, OECD's human rights focus, and sectoral laws like healthcare AI regulations and financial model risk management rules.

In January 2025, IBM collaborated with telecommunications conglomerate e& to deploy an end-to-end, multi-model AI and generative AI governance solution, announced at the World Economic Forum in Davos. The engagement demonstrated IBM's pitch: while AWS, Azure, and Google Cloud offered AI-as-a-service platforms, IBM provided the compliance infrastructure necessary to deploy AI in regulated contexts.

The economic logic was compelling. Under the EU AI Act, companies faced fines up to EUR 35 million or 7 percent of annual turnover for certain violations. For global financial institutions with hundreds of billions in revenue, a 7 percent fine could exceed $10 billion. Suddenly, paying IBM for governance infrastructure looked like cheap insurance.

Efficiency: The Inference Revolution

While competitors raced to build ever-larger models, IBM Research under Raghavan's direction focused on inference optimization—making existing models run faster and cheaper.

The strategy reflected harsh economic reality. OpenAI's GPT-4 reportedly cost over $100 million to train and required expensive NVIDIA H100 GPUs to run inference at scale. Anthropic's Claude models faced similar cost structures. For enterprises running millions or billions of AI inferences daily, these costs were prohibitive.

IBM's Granite 4.0 models, announced in October 2025, demonstrated the payoff from this efficiency focus. The hybrid architecture achieved speeds significantly faster than comparably-sized transformer models while using over 70 percent less memory. On retrieval-augmented generation tasks critical for enterprise applications, Granite 4.0 outperformed both similarly sized and larger open models.

For context, Granite 4.0's 1 billion parameter variant achieved a leading score of 68.3 percent across general knowledge, math, code, and safety benchmarks—performance competitive with models three to five times larger. The efficiency gains meant enterprises could deploy Granite models on cheaper hardware, reducing total cost of ownership.

This efficiency-first approach extended to edge deployment. Granite 4.0 Nano, IBM's smallest model family, delivered strong performance while remaining small enough to run directly in web browsers and on mobile devices. The 1-billion parameter Granite 4.0 Nano scored 78.5 on IFEval (instruction following), outperforming Qwen 3's 1.7-billion parameter model (73.1) and other models in the 1-2 billion parameter range.

The technical bet was clear: IBM believed the enterprise AI market would prioritize total cost of ownership and operational efficiency over raw capabilities. If a model could reliably perform specific business functions at one-third the cost of frontier alternatives, enterprises would choose cost savings over theoretical capability headroom they would never utilize.

Part V: The Granite Family—IBM's Model Strategy

IBM Research's Granite model family, developed under Raghavan's leadership, represented the technical culmination of IBM's enterprise-first AI philosophy.

The Granite roadmap from 2024 through 2025 revealed a coherent strategy focused on three market segments: general enterprise AI, edge computing, and multimodal understanding.

Granite 3.0: The Enterprise Foundation

Released in October 2024, Granite 3.0 featured 8-billion and 2-billion parameter models trained on over 12 trillion tokens of curated enterprise data. IBM's curation process filtered training data for relevance to business use cases, emphasizing financial documents, legal contracts, medical records, and technical manuals rather than general web content.

Benchmark results demonstrated Granite 3.0's enterprise focus. On academic benchmarks, Granite 3.0 8B matched or exceeded similarly-sized models like Llama 3.1 8B and Mistral 7B. But on enterprise-specific tasks—retrieval-augmented generation, classification, summarization, entity extraction, and tool use—Granite 3.0 outperformed larger competitors.

The safety metrics particularly stood out. Granite 3.0 achieved over 90 percent scores on safety benchmarks including SALAD and AttaQ, substantially higher than comparably-sized alternatives. For risk-averse enterprises, this safety performance justified accepting marginally lower performance on creative tasks or general knowledge questions.

Granite 3.2: Multimodal Enterprise Intelligence

In February 2025, IBM Research released Granite 3.2, adding vision-language capabilities for document understanding. The 8-billion parameter vision model delivered performance matching or exceeding significantly larger models—Llama 3.2 11B and Pixtral 12B—on essential enterprise benchmarks including DocVQA (document visual question answering), ChartQA (chart interpretation), AI2D (diagram understanding), and OCRBench (optical character recognition).

The document understanding capabilities targeted specific enterprise pain points. Financial services firms needed to extract structured data from unstructured loan documents, contracts, and regulatory filings. Healthcare organizations required automated extraction of information from medical charts, lab reports, and imaging studies. Insurance companies wanted to process claims documents, photos of damage, and policy contracts.

Granite 3.2's chain-of-thought reasoning capabilities represented a significant advance. Using novel inference scaling methods, the 8-billion parameter model achieved performance rivaling much larger models like Claude 3.5 Sonnet and GPT-4o on math reasoning benchmarks including AIME 2024 and MATH500. This performance breakthrough demonstrated that architectural innovations and inference optimization could compensate for smaller parameter counts.

Granite 4.0: The Hybrid Architecture Revolution

Released in October 2025, Granite 4.0 introduced IBM's hybrid model architecture—combining transformer layers with alternative architectures optimized for specific tasks. The hybrid approach delivered dramatic efficiency gains: Granite 4.0 models ran significantly faster than pure transformer models while using over 70 percent less memory.

The instruction-following performance particularly impressed enterprise users. Granite 4.0 demonstrated industry-leading instruction-following capabilities among open models, essential for agentic workflows where AI systems needed to reliably execute complex multi-step tasks.

On function calling benchmarks (BFCLv3), Granite 4.0 1B scored 54.8—the highest in its size class. Function calling capabilities enabled AI agents to interact with enterprise software systems, APIs, and databases, automating workflows that previously required human intervention.

All Granite models were released under the Apache 2.0 license, allowing enterprises to use, modify, and deploy them without restrictions or ongoing licensing fees. This open licensing strategy directly challenged proprietary model providers and aligned with enterprise preferences for avoiding vendor lock-in.

Part VI: The Model Gateway Strategy

In June 2025, IBM introduced Model Gateway as part of the watsonx platform—a capability that revealed the sophistication of IBM Research's competitive strategy.

Model Gateway provided seamless access to multiple foundation models through a single secure interface. Enterprises could integrate IBM's Granite models alongside frontier models from Anthropic, OpenAI, and other providers, switching between them based on task requirements, cost constraints, and compliance needs.

The strategic logic was counterintuitive. Rather than trying to convince enterprises that Granite models were superior to GPT-4 or Claude for all tasks, IBM acknowledged reality: frontier models were better at certain capabilities. But by providing the infrastructure to safely deploy any model while maintaining governance, data privacy, and auditability, IBM captured value regardless of which model customers ultimately chose.

The approach reflected Sriram Raghavan's technical realism. In interviews and conference presentations, he consistently emphasized that enterprise AI was not a winner-take-all market. Different workloads required different models. Cost-sensitive batch processing could use smaller, efficient models. Mission-critical decisions requiring frontier capabilities could route to larger models. Highly regulated tasks needing maximum explainability could use specialized interpretable models.

Model Gateway's economic model was clever. IBM charged for the governance layer, not the models themselves. Whether customers used Granite, GPT-4, or Claude, they paid IBM for the compliance infrastructure, security controls, and audit trails. This positioned IBM as the Switzerland of enterprise AI—neutral on model choice but indispensable for deployment.

The strategy also created optionality. If IBM's Granite models proved competitive with frontier alternatives for specific enterprise workloads, customers could switch to save costs. If frontier models maintained insurmountable capability advantages, customers could continue using them through IBM's infrastructure. Either way, IBM captured revenue.

Part VII: The Consulting Advantage

IBM's AI strategy diverged from pure software vendors in another crucial dimension: consulting integration.

Of IBM's $9.5 billion generative AI book of business as of Q3 2025, approximately 80 percent came from consulting engagements rather than software sales. This mix revealed the nature of enterprise AI adoption: most companies needed help figuring out how to deploy AI before they could benefit from AI tools.

IBM Consulting's AI practice employed thousands of consultants with deep industry expertise in financial services, healthcare, insurance, manufacturing, retail, and government. These consultants worked directly with clients to identify high-value AI use cases, architect AI-enabled business processes, integrate AI systems with existing enterprise software, and manage change management for AI adoption.

By Q3 2025, AI represented over 10 percent of IBM Consulting's total revenue. The firm had deployed over 200 consulting projects using "digital workers"—AI agents executing tasks at scale. Use cases ranged from automated financial report generation to AI-driven customer service to intelligent process automation in manufacturing.

The consulting model created a virtuous cycle. Consulting engagements generated insights into enterprise AI requirements, which fed back into IBM Research's product roadmap under Raghavan's direction. Real-world deployment challenges—data quality issues, integration complexity, compliance requirements, change management obstacles—informed Granite model development and watsonx platform enhancements.

This feedback loop gave IBM an advantage over pure model providers. OpenAI, Anthropic, and Google built excellent models but had limited visibility into enterprise deployment challenges. IBM's consultants encountered those challenges daily, enabling Research to build solutions for actual pain points rather than hypothetical use cases.

Part VIII: The Client Zero Initiative

In early 2023, CEO Arvind Krishna established "IBM as Client Zero"—a commitment to deploy IBM's own AI products internally before selling them to customers. The initiative gave Raghavan's research team invaluable real-world testing.

By 2024, IBM reported that AI had saved employees more than 3.9 million hours, with over 178,000 employees participating in internal watsonx application development challenges. The company projected $4.5 billion in annual productivity gains by the end of fiscal 2025.

These were not theoretical savings. IBM deployed AI across core business functions: automated code generation for software development, AI-driven contract review for legal operations, intelligent document processing for finance, and AI-enhanced customer support for client services.

The Client Zero initiative served multiple purposes. First, it validated watsonx capabilities at enterprise scale before customer deployments. Second, it generated detailed performance data and identified bugs, integration issues, and user experience problems. Third, it gave IBM salespeople credibility when pitching AI transformations—they could demonstrate real results from IBM's own operations.

For Raghavan's research team, Client Zero provided a unique testing environment. With 160,000-plus employees using watsonx internally, IBM Research could observe how AI systems performed across diverse use cases, identify failure modes, and iterate rapidly on improvements.

The feedback informed specific technical decisions. When internal users struggled with prompt engineering, IBM Research built prompt templates and few-shot learning capabilities. When compliance teams needed better audit trails, Research enhanced watsonx.governance's lineage tracking. When developers wanted faster inference, Research optimized Granite model architectures.

Part IX: The Regulated Industries Bet

IBM's enterprise AI strategy ultimately rested on a specific market hypothesis: regulated industries would adopt AI more slowly but eventually represent the largest and most profitable AI market segment.

Financial services, healthcare, insurance, pharmaceuticals, energy, and government sectors faced compliance requirements that dramatically constrained AI adoption. GDPR in Europe, HIPAA in US healthcare, SOC 2 compliance for cloud services, FDA regulations for medical AI, and sector-specific rules like financial services model risk management created complex regulatory environments.

For these industries, deploying black-box frontier models posed existential risks. A bank that could not explain why its AI rejected a loan application faced regulatory penalties and discrimination lawsuits. A hospital that could not audit its AI diagnostic system's decisions could not obtain FDA approval or defend against malpractice claims. An insurance company using biased AI for underwriting risked regulatory sanctions and reputational damage.

IBM's pitch to these sectors was straightforward: watsonx provided the governance infrastructure necessary to deploy AI safely. Explainability tools could generate audit-ready explanations for individual predictions. Model monitoring detected bias and drift before they caused compliance failures. Lineage tracking documented exactly which data influenced which decisions.

The strategy appeared to be working. In 2024, IBM and Thomson Reuters announced a joint offering combining Watson's AI capabilities with Thomson Reuters' regulatory intelligence covering thousands of financial services regulations across jurisdictions. The solution helped banks and investment firms automate regulatory reporting, fraud detection, and compliance monitoring.

In healthcare, IBM positioned watsonx as GDPR-ready and HIPAA-compliant, with built-in privacy controls and data residency options. Major health systems began deploying watsonx for clinical documentation automation, claims processing, and administrative workflow optimization—use cases where the value came from cost reduction rather than clinical breakthroughs.

The pharmaceutical industry represented another target. Drug development involved processing vast amounts of structured and unstructured data—research papers, clinical trial results, molecular databases, regulatory submissions. AI could accelerate literature reviews, identify drug candidates, optimize trial designs, and automate regulatory submissions. But pharma companies required explainable AI that could withstand FDA scrutiny. IBM positioned Granite models as purpose-built for these requirements.

Part X: The Competitive Reality Check

Despite IBM's progress, formidable competitive challenges remained.

OpenAI's ChatGPT Enterprise, launched in August 2023, rapidly gained traction among Fortune 500 companies. By mid-2025, OpenAI claimed millions of enterprise users across thousands of organizations. While OpenAI's governance capabilities lagged IBM's, its models delivered substantially better performance on complex reasoning tasks, creative work, and general-purpose capabilities.

Anthropic's Claude, particularly Claude 3.5 Sonnet released in June 2024, combined strong capabilities with a credible safety narrative that resonated with enterprises concerned about AI risks. Anthropic's Constitutional AI framework provided interpretability superior to OpenAI while maintaining near-frontier performance. Claude's context windows exceeding 200,000 tokens enabled processing entire codebases, legal briefs, and medical records in single inference calls.

Google Cloud's enterprise AI offerings leveraged the company's infrastructure advantages and technical AI leadership. Vertex AI provided access to Gemini models, PaLM 2, and numerous open-source alternatives, all integrated with Google Cloud's data analytics and machine learning tools. Google's enterprise relationships through Google Workspace created natural cross-selling opportunities.

Microsoft presented perhaps the most formidable challenge. The company's $13 billion OpenAI investment gave it exclusive access to GPT models for integration into Microsoft products. Copilot capabilities embedded in Office 365, Windows, GitHub, and Dynamics 365 created AI experiences directly within workflows millions of knowledge workers used daily. Azure AI Services provided infrastructure for custom model deployment, combining Microsoft's cloud capabilities with OpenAI's models.

AWS, while less aggressive on proprietary models, dominated cloud infrastructure for AI workloads. Amazon Bedrock's multi-model marketplace included Anthropic's Claude, Stability AI's models, Cohere's Command, AI21's Jurassic, and Amazon's own Titan models. AWS's neutral positioning—supporting all major model providers—resembled IBM's Model Gateway strategy but with substantially greater scale and cloud market share.

These competitive realities forced honest assessment. IBM's enterprise AI revenue of $9.5 billion was impressive in absolute terms but dwarfed by hyperscaler AI revenues. Microsoft's AI business was reportedly on track to exceed $20 billion annually by fiscal 2026. Google Cloud's AI revenue growth was similarly dramatic. AWS did not break out AI revenue separately but indicated AI was driving substantial cloud consumption growth.

Part XI: The Organizational Challenge

Beyond competitive threats, IBM Research under Raghavan faced significant organizational challenges.

IBM's workforce numbered approximately 280,000 globally, down from over 400,000 a decade earlier. The company had undergone multiple restructurings, spin-offs (Kyndryl in 2021), and workforce reductions. Employee morale and retention posed ongoing concerns, particularly for AI talent that could easily find higher compensation at hyperscalers, AI startups, or hedge funds.

Raghavan's IBM Research AI team of 750 scientists and engineers, while substantial, paled beside hyperscaler research organizations. Google DeepMind alone employed over 1,000 researchers. Meta's AI research organization (FAIR and applied teams) exceeded 1,500. Microsoft Research's AI groups spanned thousands of researchers when including OpenAI and product-focused teams.

Recruiting and retaining top AI talent required competing on dimensions beyond compensation. IBM Research emphasized publication freedom, intellectual property policies favoring researchers, opportunities to work on real-world enterprise problems, and the stability of an established company. But competing against OpenAI's mission to build AGI, Google's computational resources, or Microsoft's product reach remained challenging.

Cultural obstacles also persisted. IBM's reputation as a slow-moving enterprise incumbent conflicted with AI's rapid innovation pace. Stories of bureaucratic decision-making, missed opportunities, and strategic missteps (particularly Watson Health) created skepticism about whether IBM could truly compete in cutting-edge AI.

Raghavan's leadership approach emphasized pragmatic incrementalism rather than revolutionary claims. In interviews and conference presentations, he consistently positioned IBM's AI work as evolutionary progress toward enterprise-ready AI rather than breakthroughs toward artificial general intelligence. This measured tone built credibility with enterprise customers but lacked the inspirational vision that attracted top researchers to frontier AI labs.

Part XII: The Financial Inflection Point

IBM's financial results through 2024 and into 2025 suggested the AI strategy was beginning to generate meaningful revenue, though questions remained about profitability and sustainability.

In Q3 2025, IBM reported revenue of $16.3 billion, up 9 percent year-over-year, fueled by software and infrastructure growth. Software revenues improved to $7.21 billion from $6.52 billion, driven by Hybrid Cloud (up 12 percent), Automation (22 percent), and Data (7 percent). Consulting revenue reached $5.2 billion.

The company raised full-year guidance, projecting around $14 billion in free cash flow and revenue growth exceeding 5 percent. Management attributed improvements to "better mix and productivity gains"—code for AI-driven efficiencies and higher-margin AI software and consulting sales.

The generative AI book of business growth trajectory told the story of accelerating adoption. From "low hundreds of millions" in Q3 2023, the business had grown to over $2 billion by Q2 2024, then $5 billion by Q4 2024, $6 billion by Q1 2025, and $9.5 billion by Q3 2025. This growth rate—roughly doubling every two quarters—suggested strong enterprise demand.

But questions persisted about the quality and profitability of this revenue. The 80 percent consulting mix raised concerns. Consulting bookings often translated into revenue over multiple years as projects progressed. Actual revenue recognition lagged bookings by significant periods. And consulting gross margins typically ran 30-40 percent, substantially lower than software margins exceeding 80 percent.

AI software revenue—the 20 percent of AI bookings—remained relatively small compared to hyperscaler AI businesses. While IBM did not disclose exact watsonx software ARR (annual recurring revenue), analyst estimates suggested watsonx software revenue in 2024 was likely in the $200-300 million range, growing toward $500 million to $1 billion in 2025. These were respectable numbers for a two-year-old product but dwarfed by OpenAI's reported $4+ billion ARR, Anthropic's $1+ billion ARR, and Microsoft's tens of billions in AI-related revenue.

The path to substantial AI profitability required shifting mix from consulting to software. Consulting de-risked the strategy and built customer relationships, but software delivered the margins and scalability necessary for meaningful profit contribution.

Part XIII: The Technical Roadmap Ahead

IBM Research's technical roadmap for 2025 and beyond, developed under Raghavan's direction, revealed the company's AI ambitions and constraints.

Agentic AI represented the most significant near-term focus. IBM projected that by 2025, autonomous AI agents would handle substantial portions of enterprise workflows—customer service interactions, financial report generation, code review and testing, legal document analysis, and supply chain optimization. Granite models' strong instruction-following and function-calling capabilities positioned them well for agentic applications.

IBM Think 2025, the company's annual conference, showcased over 200 consulting client projects deploying "digital workers" at scale. These AI agents automated repetitive cognitive tasks previously requiring human knowledge workers. Early results suggested 30-50 percent productivity improvements for specific workflows, though implementation challenges around data quality, change management, and integration complexity remained substantial.

Multimodal capabilities represented another priority. Granite 3.2's vision-language models demonstrated IBM's commitment to document understanding, chart interpretation, and diagram analysis. Future releases would likely add video understanding, audio processing, and cross-modal reasoning—critical for enterprise use cases like manufacturing quality control, video meeting summarization, and customer service call analysis.

Inference optimization continued as a core focus. IBM Research partnered with Groq and other inference acceleration vendors to reduce latency and cost for Granite model deployment. The company's emphasis on hybrid architectures, quantization techniques, and efficient attention mechanisms aimed to maintain performance advantages over pure transformer models as competitors scaled up.

Vertical models for specific industries emerged as a strategic differentiator. Rather than building one general-purpose foundation model, IBM Research developed industry-specific Granite variants optimized for financial services, healthcare, manufacturing, and retail. These vertical models incorporated domain-specific training data, specialized vocabularies, and task-specific architectures.

The vertical model strategy reflected a broader philosophical difference from frontier labs. OpenAI and Anthropic believed general-purpose intelligence would eventually excel at all tasks. IBM bet that for many enterprise applications, purpose-built systems would prove more reliable, efficient, and compliant than general models.

Part XIV: The Trust Hypothesis

IBM's entire AI strategy ultimately rested on what might be called the trust hypothesis: that enterprise customers would prioritize transparency, explainability, governance, and compliance over raw model capabilities.

The hypothesis had both empirical support and significant uncertainties.

Supporting evidence came from IBM's own research. The 2024 IBM Institute for Business Value study found that 80 percent of business leaders saw AI explainability and trust as major adoption obstacles. Separate research indicated only 25 percent of enterprise AI initiatives achieved expected returns, often due to deployment failures rather than technology limitations.

Regulatory trends also supported the hypothesis. The EU AI Act, which came into force in 2024, imposed strict requirements for high-risk AI systems including transparency, human oversight, accuracy, and robustness. Penalties for violations could reach EUR 35 million or 7 percent of global turnover. Similar regulatory frameworks were emerging in China, Canada, and US states including California.

The financial sector provided a test case. After the 2008 financial crisis, regulators imposed model risk management requirements on banks, requiring documentation, validation, and ongoing monitoring of quantitative models used in lending, trading, and risk management. These same requirements now applied to AI models. Banks faced stark choices: deploy explainable AI systems that met regulatory standards, or risk sanctions from regulators increasingly scrutinizing algorithmic decision-making.

Healthcare presented similar dynamics. The FDA had signaled that AI diagnostic tools would require extensive validation, bias testing, and ongoing monitoring. Medical AI systems that could not explain their recommendations faced slim chances of regulatory approval. Hospitals and health systems, already operating under HIPAA constraints and malpractice liability concerns, showed strong preference for auditable AI.

But uncertainties remained. The trust hypothesis assumed enterprises would resist the temptation to deploy more capable but less explainable models for competitive advantage. It assumed regulators would enforce compliance requirements rather than tolerating violations. It assumed that explainability and governance overhead would prove manageable rather than prohibitive.

The hypothesis also faced a timing challenge. If frontier models achieved substantially superior capabilities—approaching or matching human expert performance across all enterprise tasks—customers might accept reduced explainability as a necessary trade-off. If GPT-5 or Claude 4 could reliably outperform Granite models by wide margins, enterprises might conclude that capability advantages outweighed governance concerns.

Part XV: The Alternative Scenarios

As of late 2025, multiple scenarios could unfold for IBM's enterprise AI strategy and Raghavan's research organization.

Scenario 1: The Trust Premium. Regulatory enforcement accelerates, high-profile AI failures generate public backlash, and enterprises face mounting pressure for explainable AI. IBM's governance infrastructure becomes table stakes for enterprise AI deployment. Watsonx captures 15-20 percent market share in regulated industries. AI software revenue exceeds $3 billion annually by 2027, with 60 percent gross margins. IBM Research expands to 1,000+ AI researchers. Raghavan is promoted to Senior Vice President with expanded responsibilities.

Scenario 2: The Hybrid Reality. Enterprises adopt a portfolio approach, using frontier models for some tasks and governance-focused models for others. IBM's Model Gateway strategy succeeds, with watsonx becoming the standard deployment platform for multi-model enterprise AI. Revenue grows steadily but IBM remains a second-tier player behind hyperscalers. AI software reaches $1-2 billion in annual revenue by 2027. Research budgets grow modestly. Raghavan continues in current role.

Scenario 3: The Capability Chasm. Frontier models achieve dramatic capability improvements that overwhelm governance concerns. GPT-5, Claude 4, and Gemini 2.0 deliver performance so superior that enterprises accept black-box operation. IBM's governance advantage proves insufficient to offset capability gaps. Watsonx growth stalls. IBM shifts strategy to reselling frontier models with thin governance wrappers. Research budgets face pressure. Raghavan's organization shrinks.

Scenario 4: The Acquisition Exit. IBM concludes it cannot compete effectively in AI software. The company doubles down on consulting and infrastructure, positioning watsonx as a channel for third-party models. IBM Research AI is restructured, with core teams potentially spun out or sold to a hyperscaler. Raghavan exits to join a frontier AI lab or starts his own company.

As of November 2025, Scenario 2—the hybrid reality—appeared most likely. Enterprise customers showed genuine interest in governance and explainability but also demanded competitive capabilities. IBM's Model Gateway strategy acknowledged this reality, allowing customers to use whichever models best fit specific tasks while IBM captured value through infrastructure and services.

Part XVI: The Unfinished Rehabilitation

Sriram Raghavan's IBM Research AI organization had unquestionably made progress rehabilitating Watson's failed legacy. Watsonx was a credible enterprise AI platform. Granite models delivered competitive performance for specific workloads. IBM's governance capabilities led the market. Revenue growth indicated real customer demand.

But the rehabilitation remained incomplete. IBM's AI business, while growing rapidly, represented a small fraction of overall enterprise AI spending. The company's 2025 AI revenue would likely reach $3-4 billion—impressive growth but dwarfed by Microsoft's $20+ billion, Google Cloud's tens of billions, and AWS's undisclosed but substantial AI revenue.

More fundamentally, IBM had not convinced the market that governance and explainability could compete with frontier capabilities. When enterprises needed maximum AI performance—whether for competitive advantage or genuine technical requirements—they still turned to OpenAI, Anthropic, or Google. IBM's positioning as the "safe" choice carried implicit admission that it was not the most powerful choice.

The organizational challenges persisted. IBM Research's 750-person AI team, while highly capable, could not match the resource intensity of hyperscaler research organizations numbering in the thousands. Recruiting top AI talent remained difficult when competing against companies offering higher compensation, more cutting-edge research, and better brand cachet.

The Watson Health shadow lingered. Despite watsonx's progress, skepticism remained about whether IBM could truly execute in fast-moving AI markets. Every product delay, every competitive loss, every consultant engagement that failed to convert to software revenue reinforced doubts about IBM's ability to compete with more agile rivals.

For Raghavan personally, the challenge was both technical and existential. Could a career IBM researcher, no matter how talented, lead an organization to compete with the concentrated genius and resources of frontier AI labs? Could incremental progress on enterprise-ready AI compete with revolutionary claims of approaching AGI? Could measured scientific pragmatism inspire the kind of missionary commitment that drove top researchers to work 80-hour weeks on alignment and capabilities?

Part XVII: The Broader Pattern

IBM's enterprise AI strategy under Raghavan's technical leadership represented a broader pattern in the AI industry: the divergence between frontier development and enterprise deployment.

Frontier AI labs—OpenAI, Anthropic, Google DeepMind, and to some extent Meta AI—focused on pushing the boundaries of what AI could do. Their goal was maximum capability: larger models, longer context windows, better reasoning, stronger multimodal understanding, and progress toward artificial general intelligence. Enterprise considerations like explainability, efficiency, and compliance were secondary concerns, to be addressed after achieving breakthrough capabilities.

Enterprise AI providers—IBM, Oracle, SAP, Salesforce, and to some extent Microsoft and Google Cloud—focused on making AI deployable within existing business processes and regulatory constraints. Their goal was reliable value delivery: automating workflows, reducing costs, improving decision-making, and generating measurable ROI. Breakthrough capabilities were valuable only insofar as they translated into business outcomes.

This divergence created both opportunities and risks for IBM. The opportunity was that enterprise requirements for governance, explainability, and efficiency represented genuine needs that frontier labs were not addressing. If IBM could own the "enterprise deployment" layer while frontier labs owned the "model capabilities" layer, a valuable market position existed.

The risk was that frontier capabilities would eventually obviate deployment concerns. If models became sufficiently reliable, inexpensive, and capable, enterprises might tolerate black-box operation. Or frontier labs might solve governance challenges themselves, developing explainability techniques and compliance frameworks that removed IBM's differentiation.

Early evidence suggested the divergence would persist. Frontier model development required massive capital investments—hundreds of millions to train GPT-4, likely billions for GPT-5. These economics favored a small number of hyperscale players with enormous computational resources. Enterprise deployment, conversely, required deep industry knowledge, integration expertise, and patient change management—capabilities that favored incumbent enterprise software vendors and consulting firms.

The market might therefore bifurcate: frontier labs selling model capabilities, enterprise vendors selling deployment infrastructure and services. IBM's Model Gateway strategy positioned the company to succeed in this bifurcated world, regardless of which models won the capabilities race.

Conclusion: The Long Game

In March 2025, Sriram Raghavan presented at MIT's AI Conference, discussing IBM Research's vision for enterprise AI. His talk notably avoided grandiose claims about artificial general intelligence or revolutionary breakthroughs. Instead, he focused on pragmatic progress: more efficient models, better governance tools, successful client deployments, and measurable business value.

The understated tone reflected IBM's strategic positioning. The company was not trying to win the race to AGI. It was not claiming to have the most powerful models. It was not promising to revolutionize how humans work.

IBM's bet was simpler and perhaps more realistic: that enterprises needed AI they could trust, explain, and control. That regulated industries would pay premiums for governance and compliance. That efficiency and reliability would ultimately matter more than maximum theoretical capability. That the path to $10+ billion in AI revenue ran through patient execution rather than viral consumer adoption.

Whether this bet would prove correct remained uncertain. The $9.5 billion AI book of business represented real traction, but converting bookings to profitable revenue at scale required sustained execution over years. Competitive pressures from hyperscalers would intensify as they built their own governance capabilities. Frontier model improvements might make explainability concerns obsolete.

But IBM had advantages. Twenty years of enterprise relationships. Deep industry expertise. A global consulting organization. Research capabilities spanning decades. And in Sriram Raghavan, a technically credible leader who understood both cutting-edge AI research and the prosaic realities of enterprise deployment.

The rehabilitation of Watson's legacy would be measured not in Jeopardy victories or breathless media coverage, but in steady revenue growth, expanding customer deployments, and demonstrated business value. By that unglamorous standard, IBM was making progress.

The question was whether progress would prove fast enough.