The prevailing narrative around artificial intelligence (AI) has been one of relentless scale. Bigger models, bigger clusters, bigger budgets.
The assumption, largely unchallenged until recently, was that raw parameter count translated directly into competitive advantage.
New research from Omdia suggests it's time to retire that assumption.
According to the latest market study by Omdia, parameter growth in frontier AI models has slowed to around 5 percent annually since 2021, a stark contrast to the more than hundredfold expansion seen between 2019 and 2021.
Enterprise AI Market Development
For executives who have been making infrastructure and investment decisions based on the assumption that AI would keep demanding ever-larger, ever-more-expensive hardware, this finding deserves serious attention.
The race to the top of the model size leaderboard has, at least for now, plateaued. Crucially, Omdia's analysts are not reading this as an AI winter.
Alexander Harrowell, senior principal analyst at Omdia, notes that sustained slowdowns in AI model growth were historically associated with systemic challenges in the field, but that is clearly not the case today.
Something more structural is reshaping the landscape, and understanding it is essential for any organization planning its next wave of AI investment.
Why an AI Shift Matters to The C-suite
The story is not that commercial AI development has stalled. It is that the center of gravity has already shifted within the typical large enterprise, and now within mid-size companies.
The definition of small AI models is evolving quickly, with models in the 7 billion to 14 billion parameter range increasingly replacing those in the 100 million category, while a mid-sized open-source category is gaining traction across development communities and enterprise adoption alike.
This matters for executive budget holders and technology strategists.
The explosion of capable, compact models is democratizing AI deployment. Organizations that previously lacked the infrastructure budget to run frontier-class models can now achieve high-quality outputs with models that run on a fraction of the compute.
The barrier to entry has dropped substantially, and the addressable market for Applied-AI across all organizations has widened as a result. Driving much of this shift is the rapid rise of Agentic AI.
Modern AI systems are increasingly deriving performance from tool use, effectively trading relatively inexpensive CPU compute for more costly GPU resources.
As a result, the CPU-to-GPU ratio is likely to move closer to 1:1. This has profound implications for how CIOs and CTOs plan their infrastructure. The long-held assumption that AI workloads are predominantly GPU-bound is giving way to a more balanced, more economical architecture.
Mid-sized models are gaining traction due to their role as agent coordinators in multi-model systems, as well as their increasing multi-modal capabilities. In practice, this means a well-orchestrated ensemble of smaller, specialized AI models can now rival the performance of a single monolithic frontier system, at a fraction of the cost.
The AI Inference Cost Imperative
Here is where the conversation becomes urgent for C-suite leaders.
Training a large language model is a one-time capital expenditure. Inference, by contrast, is a recurring operational cost that scales directly with usage.
As AI moves from pilot projects to production systems serving thousands of users and automated workflows, inference economics becomes the dominant variable in total cost of ownership (TCO).
The GPU supply chain remains constrained and expensive.
Older GPUs are retaining value and remaining in service, as they continue to offer a cost-effective option for small and mid-sized model inference and disaggregation. This is a signal the market should not ignore.
The industry is beginning to price inference differently, and organizations that tie their AI strategy exclusively to the latest high-end GPU generation are locking themselves into a cost structure that is difficult to sustain at scale.
Alternatives are gaining credibility rapidly. Custom silicon, including Google's TPUs and a growing array of inference-optimized accelerators from newer entrants, is starting to capture market share from GPUs, with hyperscalers' custom chips expected to become increasingly important over the next several years.
For enterprises, the implication is clear: savvy procurement strategies should account for a more diverse hardware and GPGPU software stack, not to assume NVIDIA dominance as a permanent condition.
AI Infrastructure Gets Smarter, Not Just Larger
Three trends are converging to reshape the enterprise and mid-size AI infrastructure opportunities.
First, agents are driving demand for increasingly long context windows, and managing context offload is becoming critically important, with a new cache hierarchy spanning memory and fast storage emerging to support these workloads.
CIOs and CTOs that invest in network and storage architecture alongside compute will be better positioned than those focused on expensive raw GPU capacity alone.
Second, the broader cloud infrastructure market continues to scale at pace.
Global spending on cloud infrastructure services reached $110.9 billion in Q4 2025, reflecting year-on-year growth of 29 percent, marking the sixth consecutive quarter in which the market expanded by more than 20 percent.
Moreover, AI is no longer an experimental overlay on existing cloud budgets. It is the primary driver of infrastructure investment, and that creates real opportunity for organizations with the clarity to align their AI road map with their IT infrastructure strategy.
Third, and perhaps most importantly for operational leaders, AI demand is no longer confined to the most expensive specialized GPU clusters. It is now pulling through substantial requirements for high-performance CPUs, storage, and networking.
Outlook for Refocused AI Infrastructure Investment
The AI infrastructure conversation has broadened. Winners in the next phase of Applied-AI Initiatives will not be those who simply acquired the most GPU capacity, but those who assembled the most cost-efficient, purpose-fit compute architecture for their specific workloads.
The Omdia findings offer a genuinely useful corrective to years of hype around ever-larger models. Smaller, more efficient purpose-built AI is not a consolation prize. It is the direction the global market has chosen.
I believe for CIOs and CTOs advising corporate boards and shaping investment decisions, the message is straightforward: optimize for inference economics, embrace architectural diversity, and recognize that in AI, as in most things, intelligent design consistently outperforms brute force. Therefore, invest wisely.