You know OpenAI and Nvidia. These are the AI companies building everything else

But those companies are not the whole story. A broader ecosystem is building the infrastructure, governance layers, and application platforms that make AI commercially viable. These companies do not dominate the news cycle. But the industry increasingly depends on them.

When DeepSeek rattled markets last year, the lesson went beyond model efficiency. It was a reminder that serious AI work is no longer confined to a handful of famous companies and research labs. What follows is a look at some of the companies building the rest of the AI stack.

NOT MADE IN AMERICA

Silicon Valley has long treated frontier AI as a homegrown invention. Mistral AI, building one of the world’s fastest-growing AI businesses from Paris, is challenging that assumption.

Founded in April 2023 by Arthur Mensch, formerly of Google DeepMind, and Guillaume Lample and Timothée Lacroix, both formerly of Meta, Mistral built its early reputation on a provocative claim: Frontier-quality models did not require the compute budgets of American hyperscalers.

Its open-weight models, released for anyone to build on, helped prove the point. Revenue grew from roughly $10 million in 2023 to more than $400 million in annualized recurring revenue by early 2026. In September 2025, Mistral closed a €1.7 billion Series C led by ASML, pushing its valuation to €11.7 billion.

It has since launched Vibe, an agentic platform for research, drafting, and code deployment, and is exploring its own chip design. “Physical infrastructure control matters just as much as underlying model quality for capturing long-term value,” Mensch said.

Tokyo-based Sakana AI is pulling the frontier further still. Founded in 2023 by David Ha, who led Google Brain’s research team in Japan, and Llion Jones, one of the co-authors of the influential research paper Attention Is All You Need, Sakana was built around the idea that AI systems could evolve the way nature does, through collective intelligence and constant adaptation rather than raw scale.

In March 2025, its AI Scientist system became the first AI to have a paper it independently wrote accepted by peer reviewers at a premier machine learning conference. In November, it raised $135 million at a $2.65 billion valuation, backed by MUFG, Lux Capital, and In-Q-Tel, the CIA’s venture arm. Ha has argued that “innovation, based on constraints”—not unlimited compute—is where the real opportunity lies.

Earlier this week, Sakana moved Fugu and Fugu Ultra out of beta and into general availability. The system routes tasks across a pool of frontier AI models instead of training one from scratch, all through a single interface. Sakana’s case for that approach leans on a recent warning sign: On June 12, the U.S. Commerce Department issued an export control directive that forced Anthropic to shut down Claude Fable 5 and Claude Mythos 5, just three days after launch. As of this writing, neither has been restored.

“Relying on a single company’s APIs for critical infrastructure, finance, or governance is a material vulnerability,” the company said in its launch announcement. “This risk is no longer a hypothetical possibility, but a reality.”

Sakana’s own benchmarks compare Fugu Ultra against Fable 5 and Mythos Preview, using whichever of the two scores higher on a given test. Neither model is part of Fugu’s agent pool, since neither is publicly accessible. The comparison excludes Mythos 5, the more capable model suspended alongside Fable 5, with independent analysis showing it outperforming Fugu Ultra on several of the same benchmarks.

THE UNGLAMOROUS INFRASTRUCTURE LAYER

What powers enterprise AI at scale is not only the frontier model at the top of the stack. It is also the infrastructure beneath it, the layer that determines whether AI applications can access, reason over, and act on data in ways that produce real business outcomes.

For most enterprises, that problem starts before the first line of AI code is written. Organizations adding AI to existing systems are often told to move their data out of relational databases, the traditional table-based systems most business software still runs on, and into new infrastructure. In practice, that can mean multi-year projects. At the pace AI is moving, says RavenDB CEO Oren Eini, that timeline is already a liability. “At the current pace of AI, a two-year project is already too late,” he says. “By the time you ship, the rest of the market has already moved on.”

With RavenDB, Eini says AI projects that previously stretched past six months are now being completed in weeks. The drag, he argues, is usually not the AI model. It is the infrastructure around it. Most business databases were never built to let AI search and reason over their data. Even databases that offer a vector index, which lets AI search by meaning instead of exact keywords, only solve part of the problem. “If your database only provides a vector index, all of that burden falls on you,” he explains.

RavenDB, founded in 2009, has spent the last two years embedding vector search, AI agents, and generative AI capabilities directly into the database layer. The point is to let businesses build on top of what they already run instead of replacing it. “You shouldn’t have to rebuild your systems to use AI,” Eini says.

Alation is solving a related problem one layer up. The data intelligence platform counts Truist Bank, Sallie Mae, Cisco, and Daimler Truck North America among its more than 500 customers. Its work starts from a basic reality: Many enterprise AI projects fail at the data layer. Gartner estimates that 85% of AI projects fail due to poor data quality or lack of relevant data. “An agent without business context is just a clever chatbot,” says CEO Satyen Sangani. “It can answer questions. It can’t drive an outcome that someone in finance, operations, or compliance will sign off on.”

The failure mode Sangani keeps seeing has a name: Pilot purgatory. S&P Global found 42% of companies abandoned most of their AI initiatives in 2025, up from 17% the year before. The culprit, consistently, is that organizations started with the technology and then went looking for problems to solve with it. The ones breaking through did the opposite.

“Every agent action, every human correction, every data product becomes knowledge that makes the next decision better,” Sangani says. “The companies that get this right become indispensable, because their value grows the longer customers use them.”

THE ACCOUNTABILITY LAYER

Enterprise AI adoption is moving faster than the governance structures designed to contain it. Between 2023 and 2024, the amount of corporate data uploaded or pasted into AI tools rose by 485%. Gartner projects that over 40% of AI-related data breaches by 2027 will stem from unapproved or improper generative AI use.

Several companies are working at the edge of that accountability problem. Two in particular, operating in different domains, are converging around the same issue: AI sprawl, or what happens when enterprises deploy AI faster than they can govern it.

Reco, an Israeli agent security and AI governance company, was built to solve that problem. Its platform maps the full AI agent footprint across enterprise SaaS environments, showing what is running, what it can access, and how it is behaving. “Across our customer base, organizations typically discover roughly ten times more apps and a hundred times more agents than they expected when they first connect,” CEO Ofer Klein tells Fast Company.

The security leaders getting this right have stopped asking “what AI tools are my employees using?” and started asking a harder question: When an agent has simultaneous access to Salesforce data, SharePoint files, and Slack channels, how do you stop a breach before it starts?

The answer starts with knowing where AI actually lives inside the enterprise. “Autonomy without governance is exposure at scale,” Klein says. “Every major platform your enterprise already runs on is now also an agent platform. That is exactly where we show value.”

Terzo is making a version of the same case in a different part of the enterprise stack: the contracts, invoices, and purchase orders governing how large organizations spend money. According to World Commerce and Contracting, companies lose an average of 9.2% of annual revenue to poor contract management every year, not through fraud, but through missed rebates, lapsed pricing schedules, and supplier agreements nobody was monitoring.

Terzo’s platform connects contracts, invoices, purchase obligations, and ERP data, validating financial activity against contractual terms in real time. According to CEO Brandon Card, one Fortune 50 company found more than $100 million in savings after running its supplier relationships through Terzo’s platform.

As AI agents increasingly handle procurement and financial workflows, they are only as trustworthy as the data they operate on. An AI agent working from inaccurate or unstructured contract data does not just make mistakes. It makes them at scale. Terzo is building the financial truth layer meant to prevent that.

“The companies that win in the next decade will not just have AI models,” Card says. “They will own the financial intelligence layer that allows autonomous agents to move trillions of dollars safely, accurately, and automatically across the global economy.”

BEYOND TEXT

For all the progress AI has made in the last few years, almost all of it happened in text—reading it, generating it, and searching through it. However, while text is structured and easy to feed into a model, voice and video are neither. They’re messy, happen in real time, and until recently, almost none of it was kept around long enough to be useful.

That gap has been enormous, and mostly invisible. Billions of customer service calls are placed every year, and most are logged and forgotten. Hundreds of millions of surveillance cameras run worldwide, and most of what they capture is never watched. The information was there. The tools to use it were not.

That’s changing now. Two companies are closing that gap—one in conversation, and the other in physical space.

PolyAI, founded in London by Nikola Mrkšić, a Cambridge machine learning PhD and first engineer at VocalIQ, builds enterprise voice AI for organizations managing large volumes of customer interactions. The company counts more than 100 enterprise customers and runs more than 2,000 live deployments across 45 languages in over 25 countries. A Forrester study found its customers achieve 391% ROI with an average $10.3 million in savings. Deloitte recognized it as the U.K.’s fastest-growing AI company in 2025.

In December 2025, it raised $86 million in a Series D with Nvidia’s venture arm among the participants. “PolyAI started with a simple idea: Enterprises should sound human,” Mrkšić said at the announcement. “We turned that idea into reality, and it led to something far greater: The emergence of the agentic enterprise.” He has said publicly that within five years, 90% of contact center work will be automated, a claim that would have sounded reckless three years ago and now sounds like a timeline.

Lumana is doing something similar in the physical world. Its AI platform transforms existing security cameras, including those in retail stores, healthcare facilities, and factories, into real-time intelligence systems without new hardware. “Instead of recording footage and hoping someone reviews it later, the system understands what it sees—the behaviors, context, and anomalies—and does so in real time,” says company CEO Sagi Ben Moshe. “A camera isn’t just capturing video anymore. It’s generating intelligence.”

In under 18 months, the California-based company scaled to more than 50,000 cameras and says its platform can reduce non-actionable physical access control alarms by as much as 95%. More recently, it has expanded beyond detection and into agentic AI, with specialized agents that can monitor environments, investigate incidents, verify events, and initiate responses without requiring a human operator to continuously review video feeds.

Adam Scraba, Nvidia’s head of physical AI ecosystem and vision agents, describes Lumana as turning existing infrastructure into “a perceptive, highly scalable virtual workforce of video analytics AI agents that can understand, verify, and summarize what’s happening in the physical world.”

That means cameras stop being a passive cost center, and become instead an active intelligence asset, informing decisions about customer behavior, staffing, and workplace safety. At scale, that could make the physical world as legible to AI as the digital one.

THE REAL AI ECONOMY

The story of AI value creation that most people tell goes like this: Powerful models emerge, developers build on top of them, enterprises adopt, and economic value follows. What that story leaves out is the middle or context layer: the infrastructure, governance, security, and operational intelligence that makes enterprise adoption viable at scale.

McKinsey estimates AI could add up to $4.4 trillion annually to the global economy through enterprise productivity gains alone. But those gains do not materialize just because a company signed an OpenAI contract. They materialize when the data feeding those models is governed and trusted, when the agents operating inside enterprise systems are visible and accountable, when contracts are operationalized rather than stored, and when voice and physical environments are converted into actionable intelligence.

That is the work the companies in this piece are doing. In previous technology cycles, the companies that won over the long term were not always the ones with the most famous products. They were the ones those products could not run without. If history is any guide, the AI cycle will not be different.