The AI industry spent the last few years competing on model size, computing power and benchmark scores. The next phase is becoming a much broader contest — around data, AI agents, inference, infrastructure, cost and the ability to turn AI into measurable business value.
For the past few years, the AI race seemed remarkably easy to understand. Build a bigger model, give it more computing power, train it on more data, and then demonstrate that it performs better than the model that came before it.
The scoreboard was dominated by numbers: parameters, GPUs, tokens, context windows and benchmark scores.
And for a while, bigger really did seem to mean better.
The world’s leading AI companies continue to invest enormous amounts in increasingly capable models. Every new generation promises stronger reasoning, better coding, multimodal capabilities and the ability to handle increasingly complicated tasks.
But the industry is beginning to confront a more important question: What happens after you build the model?
An AI model sitting inside a data centre isn’t, by itself, a business solution. It becomes valuable when it can work with an organisation’s data, connect to existing systems, perform useful tasks, operate at an acceptable cost and deliver an outcome that matters.
That is where the AI race is changing.
Bigger Models Still Matter
It would be wrong to suggest that model capability no longer matters. It does. Frontier models remain important for complex reasoning, coding, research and increasingly sophisticated AI agents. The leading AI companies are still competing aggressively to push the boundaries of what their systems can accomplish.
But model size is no longer telling the whole story.
Omdia’s 2026 research found that growth in the size of frontier models has slowed considerably compared with the explosive parameter growth seen during the earlier years of the AI boom. At the same time, smaller and midsized models are becoming increasingly capable and useful for specific workloads.
That doesn’t mean the industry has stopped building powerful models. It means companies are getting better at using different models for different purposes.
A business doesn’t necessarily need the most powerful model available to classify customer messages, extract information from invoices, summarise internal documents or answer routine employee questions. A smaller specialised model may be perfectly adequate for those jobs, while a larger reasoning model can be reserved for situations that genuinely require greater capability.
The result is a move away from a world where one giant model is expected to do everything and toward a more complex ecosystem in which the right model is selected for the right task.
The Real Race Is Moving Beyond the Model
Consider what happens when a company decides to deploy AI across its organisation.
Choosing a model is only the beginning.
The company also needs access to its own data. It needs security and permissions. It needs connections to existing software. It needs infrastructure capable of running the system. It needs ways to evaluate the AI’s output and monitor its behaviour. It needs employees who understand how to work with it. And increasingly, it needs AI agents capable of actually performing tasks.
The model is becoming one layer within a much larger technology stack.
That is why some of the most interesting developments in AI are happening outside the headline model launches.
Snorkel AI, for example, raised $350 million in September at a valuation of $3.5 billion. The company, which began by helping organisations create and manage training data, has expanded into providing completed datasets and reinforcement-learning environments for AI labs and enterprises. Reuters reported that its annualised revenue had reached about $375 million.
The significance isn’t simply the size of the funding round. It is what investors are choosing to fund.
The market is putting substantial capital into the infrastructure required to make AI systems better, rather than only into companies attempting to build the next giant model.
The AI story is therefore becoming less about who builds the brain and more about everything required to make that brain useful.
Data Is Becoming Part of the AI Arms Race
For much of the AI boom, computing power dominated the conversation.
Who has the GPUs? Who can build the biggest data centre? Who can train the largest model?
But as AI systems become more capable, another question is becoming increasingly important:
What are you actually training them to do?
The data requirements of advanced AI systems are becoming more complicated. It is no longer simply about collecting enormous quantities of text. Systems designed for specialised tasks need high-quality examples, expert knowledge, feedback, evaluation criteria and environments in which they can learn how to perform specific activities.
This becomes particularly important as AI moves from answering questions to performing work.
An AI agent that needs to investigate a customer complaint, analyse a financial document or complete a software task requires more than general language ability. It needs relevant context, access to the right tools, appropriate permissions and examples of what successful execution looks like.
Snorkel’s recent expansion into specialised datasets and reinforcement-learning environments reflects this changing demand for AI data.
The implication is significant: data quality and domain knowledge can become competitive advantages even when companies are using the same underlying foundation models.
A business may not need to build its own frontier model. It may instead create a highly effective AI system by combining a capable model with proprietary information, specialised data and a deep understanding of its industry.
The Shift From Chatbots to Agents
Perhaps the biggest reason the bigger-model narrative is becoming insufficient is the rise of AI agents.
The first phase of generative AI was largely about interaction. You asked a question, and the AI generated an answer.
The ambition is now changing.
Instead of simply responding to a request, an AI system can potentially break a goal into multiple steps, access tools, retrieve information, perform actions and return with a completed result.
That might involve searching a database, analysing documents, writing code, checking information, interacting with another application and then reporting the outcome.
The model is no longer operating alone. It is operating as part of a system.
Omdia’s research points to this shift as one reason AI infrastructure requirements are changing. Agentic systems can involve longer context windows, repeated tool calls and a combination of different computing workloads, creating demand across CPUs, GPUs, memory, storage and networking.
This changes what companies need to optimise.
The question isn’t simply whether a model can produce an impressive answer.
It is whether an AI system can complete a useful piece of work reliably.
Inference Is Becoming a Bigger Part of the Story
There is another part of the AI race that receives considerably less attention than model training: inference.
Training is how a model is built. Inference is what happens every time somebody actually uses it.
A company might train a model once, but that model could subsequently handle millions or billions of requests. Every customer-service interaction, coding request, document analysis, voice conversation and AI-agent action consumes computing resources.
As AI moves from experimentation into everyday business processes, the economics of running those systems become increasingly important.
Gartner forecasts that worldwide spending on AI inference workloads will reach $23.3 billion in 2026, compared with $19 billion for AI training. It also forecasts $42.3 billion in spending on AI-optimised infrastructure during the year.
That is an important shift in emphasis.
For years, the infrastructure conversation was largely about building machines powerful enough to train enormous models. Increasingly, the industry also needs infrastructure capable of running those models efficiently at enormous scale.
Training creates the intelligence. Inference turns that intelligence into a service.
The Most Powerful Model May Not Be the Most Useful
This distinction becomes particularly important for enterprise AI.
Imagine a company processing millions of customer interactions every month. Does every interaction need to be handled by the most powerful model available?
Probably not.
A smaller model might handle routine questions. A specialised model could extract information from documents. A larger reasoning model could be called when a case becomes complicated. A human employee could handle situations requiring judgement, empathy or accountability.
The result is an architecture in which several models work together rather than one model doing everything.
That is consistent with the broader shift toward smaller and midsized models identified by Omdia. As these models become more capable, they can take on specialised roles while larger models are reserved for more demanding workloads.
This creates an important business principle:
The smartest AI system isn’t necessarily the one using the smartest model. It may be the one that knows which model to use, when to use it and how much it should cost.
Cost Is Becoming a Competitive Advantage
Eventually, the AI conversation reaches the CFO’s office.
A model can be extraordinarily capable. It can perform well on benchmarks, generate impressive code and reason through complicated problems. But if using it at scale costs too much, its business value becomes difficult to justify.
The next phase of AI therefore has an uncomfortable requirement:
AI has to make economic sense.
Companies will increasingly compare models not only on capability, but also on latency, reliability, token costs, infrastructure requirements and the amount of human intervention still required.
That could make the phrase “best AI model” increasingly meaningless without context.
The best model for software development may not be the best model for customer service. The best model for real-time voice may not be the best model for analysing financial documents. The best model for an AI agent may not be the best model for generating a marketing campaign.
The question is increasingly becoming:
Best for what?
Infrastructure Is Becoming Part of the Product
The AI race is also spreading into the infrastructure layer.
If AI agents are going to perform increasingly complex tasks, the underlying infrastructure has to keep up. Cloud providers and chip companies are developing systems specifically designed around AI inference, agentic workloads and the growing demand for efficient AI computing.
Google’s 2026 infrastructure announcements, for example, have emphasised systems designed for inference and the requirements of an increasingly agentic AI environment.
Gartner’s forecast of more than $42 billion in AI-optimised infrastructure spending during 2026 provides another indication of the scale of this shift.
The result is that the AI race is no longer confined to model laboratories.
It is happening across chip companies, cloud providers, data-centre operators, networking companies, data businesses and enterprise software platforms.
The model may be the part consumers see.
Underneath it sits an enormous industrial ecosystem.
Reliability May Matter More Than Intelligence
There is one more race that doesn’t show up neatly on a benchmark chart:
Can businesses trust the system?
When AI is simply answering a low-risk question, a mistake may be frustrating.
When an AI agent has access to company systems, customer information, financial records or business workflows, the consequences can be much greater.
Companies need to know what the system can access, what it is permitted to change, how its actions are monitored and when a human should intervene.
That makes evaluation and governance increasingly important.
As AI systems become more autonomous, businesses need ways to test them in realistic environments rather than simply checking how well they perform on standard benchmarks.
The question is therefore no longer only:
“How intelligent is this model?”
It is also:
“What happens when we give this model access to our business?”
That may eventually be the more important question.
So, What Is the Real AI Race?
If you look closely at the industry today, it is difficult to describe the competition as a single race.
There is still a model race, with companies competing to build increasingly capable AI systems. But alongside it is a data race, as high-quality, domain-specific and reinforcement-learning data becomes increasingly valuable.
There is an inference race, where companies are trying to deliver AI intelligence faster and more cheaply. There is an agent race, where the objective is shifting from generating answers to completing tasks. There is an infrastructure race, involving chips, cloud platforms, memory, networking and data centres.
There is also a reliability race, where companies need increasingly autonomous systems to behave predictably enough for real-world deployment.
And then there is perhaps the most important race of all:
The race to turn AI capability into measurable business value.
Because customers don’t buy benchmark scores.
They buy outcomes.
The Biggest Model Doesn’t Automatically Win
None of this means bigger models are becoming irrelevant. They aren’t.
Frontier models will continue to push the boundaries of what AI can reason about and accomplish. They will remain an important part of the technology landscape, and the companies developing them will continue to compete intensely.
But intelligence by itself isn’t enough.
A company could have access to an extraordinarily powerful model and still fail to create meaningful business value if its data is poor, its systems aren’t integrated, its infrastructure is too expensive or its employees don’t know how to use the technology effectively.
Another company could use a collection of specialised models, connect them to excellent internal data and workflows, and build something far more useful.
That is why the next phase of AI feels different.
The first question was:
“How intelligent can we make AI?”
The next question is:
“What can we actually build with that intelligence?”
And that is why the AI race is getting much bigger than the models.
Everyone may still be trying to build the biggest brain in the room.
But the real competition is increasingly about what you connect to that brain, what you allow it to do, how efficiently it can do it, and whether it creates something people are actually willing to pay for.
