IBM Says Enterprise AI Has Moved Beyond Tokenmaxxing and Open Weights Could Decide Who Wins

The enterprise AI race has become less about intelligence and more about economics. Today, executives are asking a different set of questions in the boardroom. Can organizations afford to run frontier models at enterprise scale? Which workloads actually require them? And can businesses demonstrate that the returns justify the growing cost of inference?
The numbers behind those questions capture a defining contradiction. Executive confidence in AI has never been higher. Neither has the uncertainty surrounding how organizations will translate AI investment into measurable business value at scale.
New research from IBM’s Institute for Business Value, which surveyed more than 2,000 C-suite executives across 33 countries, puts figures on both halves of that contradiction. Nearly eight in 10 respondents expect AI to contribute significantly to revenue by 2030, almost double today’s figure. Yet only 24% say they understand where that revenue will actually come from. Executives expect AI investment to increase by roughly 150% over the next five years, while nearly seven in 10 worry their AI programs remain disconnected from core business operations.
IBM believes those findings signal a turning point. Cost optimization may have become enterprise AI’s immediate obsession, but the company argues the conversation is already evolving beyond tokenmaxxing. The next competitive advantage, it says, will come from combining efficient model selection with open-weight architectures, redesigned business processes, and measurable business outcomes rather than simply deploying the latest frontier model.
Most Fortune 500 companies no longer struggle to launch AI pilots. Instead, they struggle to move beyond them. Ritika Gunnar, IBM’s chief performance marketing officer and former general manager of Data and AI, notes that many organizations still approach AI as a technology deployment rather than an organizational redesign.
“A lot of what determines whether these projects scale has surprisingly little to do with the model itself,” she tells me. “Technology matters, but what separates pilots from production is how companies rethink their processes and more importantly, how they manage cultural transformation.”
For Gunnar, the missing layer sits above the models. “Many organizations are adopting AI and using agents in everything they do, but there’s a big gap between building these AI systems and actually governing them,” she says. “Can you understand why an agent made a particular decision? Can you manage first-party and third-party agents across your environment?”
That philosophy explains why IBM increasingly describes enterprise AI as an operating model rather than a software category. The company argues businesses need something resembling an “agent development lifecycle.”
Why ‘Modelmaxxing’ Is Replacing Frontier Model FOMO
If 2024 was defined by capability, 2026 is increasingly being defined by economics. Large reasoning models deliver impressive results, but many organizations are discovering they cannot afford to run every workflow through their most capable and most costly models.
Neil Dhar, senior vice president of IBM Consulting, describes that realization as the industry’s latest course correction. “When generative AI first arrived, there was enormous FOMO,” he says. “Business leaders felt pressure to deploy AI as quickly as possible instead of asking whether those deployments actually produced measurable business value.”
Dhar said many organizations initially assumed AI would automatically improve efficiency when layered onto existing workflows. In practice, he argued, failing to redesign those workflows can leave companies paying for both people and technology without realizing meaningful productivity gains.
That realization is fueling what IBM describes as the “tokenmaxxing” era—a growing emphasis on optimizing inference costs by using the right model for the right workload instead of defaulting to the largest one. One of its defining characteristics is modelmaxxing: routing each task to the model best suited for its cost, latency, and performance requirements.
“You don’t call a brain surgeon for a paper cut,” Dhar says. “You use the right model for the right workload.” IBM’s latest research found that 82% of executives expect organizations to operate multiple AI models by 2030. Nearly three-quarters believe smaller language models could outperform larger ones for many enterprise workloads.
Data May Decide Who Wins Enterprise AI
For Gunnar, the conversation about enterprise AI begins long before model selection. It starts with data. She argues that AI agents are only as effective as the information they can access, yet most enterprises still keep critical data fragmented across ERP systems, CRM platforms, operational databases, data warehouses, and lakehouses.
The company’s acquisition of Confluent and its investments around watsonx.data reflect a broader bet that enterprise AI increasingly depends on moving governed, contextual information to agents in real time rather than periodically copying data into centralized repositories. “By the time data lands in a lakehouse, it may already be stale for many agent-driven decisions,” she says. “You need governance and context at the point the data is created, not after it’s already been copied somewhere else.”
Competitors including Microsoft, ServiceNow, Snowflake, Databricks, and Salesforce increasingly make similar claims around governance, orchestration, and enterprise AI management. IBM’s challenge is convincing customers that its implementation stands apart.
The Real AI ROI Question Isn’t About Tokens
Dhar argues that enterprise AI is still obsessed with the wrong metric. Token savings may signal greater efficiency, but they remain an input rather than an outcome. The real measure, he said, is whether AI delivers the kind of business returns executives expect from any other strategic investment.
“The question isn’t how many tokens a workflow burned,” Dhar says. “It’s whether that workflow resolved customer issues faster, shortened the sales cycle, reduced operating costs, improved EBITDA, or generated new revenue. That’s what survives board scrutiny.”
IBM now encourages many clients to evaluate AI initiatives against a target return between 2.5x and 3x, measured through auditable business outcomes rather than productivity estimates alone. Dhar asserts that treating AI as primarily a technology initiative remains one of the biggest organizational mistakes companies make.
“The CFO brings financial discipline. The CTO determines the technology architecture. The CHRO prepares the workforce. The CEO establishes the strategic direction. If those functions optimize independently, even technically successful AI projects can fail,” he says.
Open vs. Closed AI Models: Why Enterprises Should Stop Picking Sides
On July 24, Nvidia CEO Jensen Huang used his first-ever X post to publish the “Open Weights and American AI Leadership” letter urging Washington against restricting downloadable models—a signatory list that doubled to 50 within a day and drew Amazon’s hedged signature recently, leaving Anthropic as the last major holdout.
IBM has publicly aligned itself with that open-weight movement alongside industry heavyweights including Nvidia, Microsoft, Meta, Dell, Palantir, and ServiceNow. But Gunnar resists framing the discussion as a winner-takes-all contest. She says most enterprise customers are asking a more practical question: which model works best for a specific workload—and whether they’ll still have the flexibility to change that answer six months from now. As customers move from proof-of-concept into enterprise production, “they’re asking for continuous evaluation and optimization,” she says, and for transparency into how AI operates inside their environment, because optimization “has become one of the biggest barriers to scaling enterprise AI.”
“Open-weight models, much like open-source software, underpin an open and modern enterprise infrastructure,” she said. “Our goal isn’t to force customers into one camp. It’s to give them choice while providing the technology layer that lets them understand how decisions are being made.”
The Control Layer Verdict
The excitement surrounding frontier models hasn’t disappeared, but executives increasingly want evidence that AI can improve operating margins, accelerate product development, strengthen customer relationships, and survive financial scrutiny. IBM isn’t alone in making that argument. Nearly every major enterprise software vendor now promises governance, observability, hybrid deployment, and agent orchestration.Â
Enterprise AI has reached a point where technical capability no longer appears to be the primary constraint. Organizational willingness may be. If IBM’s central thesis proves correct, the winners of the next AI cycle won’t be the companies with the smartest models—they’ll be the businesses that learned how to govern them.
Victor Dey is a tech analyst and writer who covers AI, data science, startups, and cybersecurity. A former AI editor at VentureBeat, his work also appears in New York Observer, Fast Company, Entrepreneur Magazine, HackerNoon, and more. Victor has mentored student founders at accelerator programs at leading universities including the University of Oxford and the University of Southern California, and holds a Master's degree in data science and analytics.
More from Victor Dey
See all
Future of AICysic CEO Says Paying for an AI Model Doesn’t Prove Your Host Ran It
Future of AIInside Indeed’s AI Hunt For Fake Employers Before Anyone Applies
Future of AIDatabricks Co-Founder Explains How Genie Tackles the Blind Spot In Always-On AI
Enterprise AIWhat Happens to Enterprise Software When AI Does the Work?
Recommended for you
More stories
AI & TechnologyFrom self-driving excavators to AI leasing agents, proptech’s AI moment has arrived
AI & TechnologyAI literacy shouldn’t begin in the workplace. It should begin in primary school.
DataThe Data Centre Is Becoming a Product
AI & TechnologyToo often, AI narrows rather than widens access. Here’s how the tech industry can fix it
AI & TechnologyHow to Review an AI Subscription Before It Renews
AI & Technology