For the last few years, AI has been obsessed with models. Who's smartest? Who's fastest? Who has the biggest context window? Who wins the benchmark this week? Those questions still matter if you're building frontier models.
For everyone else, there's a more important question now: Can you actually get AI into production?
Because while the model race gets most of the headlines, another race is happening inside the enterprise. It's the race to turn increasingly abundant intelligence into actual work. And right now, most companies aren't very far along.
The Models Got Good. Really Good.
The progress is hard to overstate. Models that would have looked extraordinary a few years ago are now available from multiple providers. Open and proprietary models compete across reasoning, coding, multimodal tasks, and specialized workloads. And the cost of accessing that intelligence has collapsed.
Stanford's 2025 AI Index found that the inference cost of a model performing at roughly GPT-3.5 level on a common benchmark fell from $20 per million tokens in November 2022 to $0.07 by October 2024. That's more than a 280x decrease in about 18 months.
Depending on the task, Stanford found inference prices have been falling between 9x and 900x per year.
The intelligence isn't just getting better. It's getting cheaper. That changes the enterprise AI equation.
The question is becoming less: Can a model do this? And increasingly: Can we build a system around it that does this reliably?
Those are not the same problem.
Enterprises Aren't Waiting to Spend
Companies clearly believe there's something worth chasing.
Menlo Ventures estimates enterprise generative AI spending reached $37 billion in 2025, up from $11.5 billion the year before. More than half of that spending ($19 billion) went to the application layer: the products actually putting AI into people's hands. Infrastructure captured another $18 billion. In other words, the market isn't waiting for some hypothetical future generation of AI before investing.
The money is already moving. And executives expect it to keep moving.
IBM's 2025 survey of 2,000 CEOs found that 61% said their organizations were already adopting AI agents and preparing to implement them at scale. Those CEOs expected the growth rate of their AI investments to more than double over the following two years.
There's just one problem. Spending money on AI and getting value from AI aren't the same thing.
Welcome to Pilot Purgatory
Enterprise AI has no shortage of demos. Ask a model to analyze a contract. Amazing. Connect it to your company's documents. Even better. Build an agent that completes a five-step process. Now we're talking. Then try to deploy it across a company. That's where things get interesting.
IBM found that only 25% of AI initiatives had delivered the expected ROI among the CEOs it surveyed. Just 16% had scaled enterprise-wide.
BCG found a similar divide. In a study of 1,000 senior executives across 59 countries, only 26% of companies had developed the capabilities necessary to move beyond proofs of concept and generate tangible AI value. Just 4% had reached what BCG classified as cutting-edge AI capabilities across functions.
Deloitte's research found that 68% of organizations had moved 30% or fewer of their generative AI experiments fully into production.
Different studies. Different methodologies. Same pattern.
The bottleneck is moving from intelligence to implementation.
A Demo Has One Job. Production Has Hundreds.
This shouldn't be surprising. A demo has to work once. Production has to work tomorrow. And the next day. And when 10,000 people use it. And when an API changes. And when the model fails. And when someone uploads something unexpected. And when Finance asks what it costs. And when Legal asks where the data went. And when Security asks what the agent can access. And when the CEO asks whether any of this is actually making money.
That's a completely different engineering and operating problem. The model is one piece of the system.
Production requires everything around it. Data. Tools. Integrations. Permissions. Memory. Compute. Observability. Human approvals. Failure handling. Security. Governance. Cost controls. And increasingly, orchestration across multiple models and agents.
The impressive part of an enterprise AI system isn't that it can answer a question. It's that it can do useful work reliably when nobody is standing beside it making sure it behaves.
The Best Model Isn't Always the Point
There's another reason the enterprise conversation needs to move beyond the model race. Companies increasingly have choices.
Menlo Ventures' enterprise research estimated that three major providers accounted for most enterprise LLM API usage in 2025, but their shares continued to shift significantly year over year.
That's a useful reminder of how quickly this market moves. Today's default model may not be tomorrow's. And different models are better at different things. One may be best for deep reasoning. Another for coding. Another may be fast and inexpensive enough to classify millions of records. A smaller model may be all you need for a deterministic task. A private model may make more sense for a sensitive workload.
The enterprise architecture that assumes one model will do everything forever is making a very large bet in a market that changes every few months.
The better question isn't: Which model wins? It's: Which model should do this job?
And more importantly, can you change the answer when something better comes along?
The Real Work Is the Workflow
Here's perhaps the most important data point. McKinsey studied which organizational practices were most associated with companies actually seeing bottom-line impact from generative AI. Of the 25 attributes it tested, redesigning workflows had the strongest relationship with EBIT impact. Yet only 21% of respondents whose organizations used generative AI said they had fundamentally redesigned at least some workflows.
That's the gap.
We've spent enormous energy putting AI into existing work. Copilots inside existing applications.
Chatbots beside existing processes. AI features bolted onto existing software. But the biggest opportunity isn't necessarily making the old workflow 20% faster.
It's asking: If AI can reason, use tools and execute work, why does the workflow look like this at all? That's when AI stops being another productivity feature and starts becoming infrastructure.
Agents Raise the Stakes
Agentic AI makes this even more urgent. A chatbot produces an answer. An agent can take an action. That's a massive difference. If an AI writes a bad draft, a human can delete it. If an agent sends the wrong payment, changes the wrong customer record, or triggers the wrong workflow, you've got a different kind of problem. Which means deployment can't just mean giving agents access to more tools.
Companies need to know: What can this agent access? What is it allowed to do? Which actions require approval? What happens when it fails? What did it actually do? How much did the execution cost? And can a human step in when necessary?
Gartner has forecast that more than 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls.
That's not an argument against agents. It's an argument against confusing an agent demo with an agent deployment.
Success Isn’t Defined by the Best Model
Most enterprises aren't going to train the next frontier model. They don't need to. Their competitive advantage will come from what they build around increasingly commoditized intelligence. Their proprietary knowledge. Their data. Their workflows. Their integrations. Their institutional expertise. Their ability to route work intelligently. Their ability to govern execution. And, ultimately, their ability to turn AI into measurable outcomes.
That means the moat moves. From simply having AI to operationalizing AI.
IBM's CEO research offers an interesting clue here: 72% of surveyed CEOs said their organization's proprietary data was key to unlocking generative AI's value, while 68% said an integrated enterprise-wide data architecture was critical to cross-functional collaboration.
The model may be rented. The advantage doesn't have to be.
This Is the Race Rival Is Built For
Rival isn't trying to win the race to build the world's biggest foundation model. We think enterprises have a different problem.
Make the models work.
That means giving companies the infrastructure to connect models to knowledge, tools, and workflows, and turn intelligence into execution. Use different models for different jobs. Build reusable functions and workflows. Give agents access to the knowledge they need. Put humans in the loop when judgment matters. Control how work executes. And avoid rebuilding the entire stack every time the model landscape changes.
Because the model is increasingly becoming one component in a much larger system. The value is in what the system can accomplish.
Stop Waiting for the Next Model
There will be another model. It will be smarter. Then another will be cheaper. Another will be faster. Someone will win a benchmark on Tuesday and lose it on Friday. That isn't going to stop. Enterprises shouldn't either.
The companies that create durable advantages from AI will be the ones that built the infrastructure to keep choosing and changingmodels while the work keeps running.
The first era of generative AI asked: What can the model do?
The next one asks: What can your company do with it?
The model race isn't the finish line. It's the starting gun.