As Anthropic Goes Public, Open Source AI Is Closing In Says Google Pioneer
The lead Anthropic built is evaporating, says NEAR co-founder Illia Polosukhin
By Zack Guzman
November 19, 2026
As Anthropic sets a new IPO record to go public at a $2 trillion valuation, one of the researchers who helped lay the foundation for modern artificial intelligence says the industry's biggest players could be overlooking a growing threat: Cheaper, increasingly capable open-source AI models.
Illia Polosukhin, a former Google researcher who co-authored the landmark 2017 paper Attention Is All You Need, which introduced the Transformer architecture underpinning today's leading AI systems, believes the economics that have propelled companies like Anthropic and OpenAI to extraordinary valuations could eventually work against them.
In an interview with Coinage at Solana's Breakpoint conference, the NEAR Protocol co-founder argued that businesses are beginning to question whether they really need to pay for the most powerful AI models when cheaper alternatives can accomplish much of the same work.
"You don't always need frontier for every single problem," Polosukhin told Coinage. "They started looking at open source models because they're cheaper, they're more accessible."
The argument comes at a consequential moment for the artificial intelligence industry. Anthropic's $81.3 billion public offering underscores the enormous expectations investors have placed on proprietary AI technology. When ANT shares popped in their Thursday debut, maybe investors took a sigh of relief. But Polosukhin sees an emerging challenge to the assumption that the companies building the most expensive and sophisticated models will necessarily capture the most value.
For much of the generative AI boom, companies raced to integrate increasingly powerful models from OpenAI, Anthropic and other leading developers. Access to cutting-edge intelligence was widely seen as a competitive advantage worth paying for.
Now, Polosukhin says, businesses are discovering how quickly those costs can accumulate.
"If you have a 100-person team, everybody's using AI every day. It's like $1,000 per person if they're using AI a lot," he said, describing a hypothetical heavy-usage scenario. "It's like $100,000 a month."
For enterprises deploying AI across hundreds or thousands of employees, those costs can become a significant operating expense. And the more AI becomes embedded in everyday workflows, the greater the incentive to find alternatives.
Read More from Coinage:
That creates an opening for smaller, more specialized open-source models that can handle routine tasks without requiring the computing resources of frontier systems.
Polosukhin argues that businesses are increasingly discovering they don't necessarily need the most advanced model for every application.
But Polosukhin's conviction goes beyond economics. Privacy is also becoming an increasingly important part of the calculation.
"I think open source always wins. I've said it for many years," he said. As businesses incorporate AI into sensitive operations, they must consider not only what a model can do, but what happens to the information they provide it.
It's a concern Polosukhin says he is increasingly hearing from users: "I don't want to give up my alpha. I don't want to give up my strategic data," he said. Those concerns are especially relevant for financial institutions, legal services, healthcare providers and technology companies working with proprietary information.
The combination of lower costs, increasing customization and greater control could become a meaningful competitive advantage for open-source alternatives.
The Problem Open Source Still Needs to Solve
There is, however, a fundamental obstacle to Polosukhin's vision.
Training advanced AI systems requires enormous investments in computing power, research and engineering. And the companies developing proprietary systems have a relatively straightforward way to finance those investments: charging customers for access.
"The reason why OpenAI and Anthropic are such a kind of a giant company is because they can monetize it effectively," Polosukhin explained. Open-source developers face a more complicated challenge.
Making models widely accessible can accelerate innovation and adoption, but it doesn't automatically create a sustainable business capable of financing the next generation of research. Polosukhin believes blockchain technology could help close that gap.
Through the AI infrastructure NEAR has built, he envisions systems where developers can make models available for broader use while still receiving economic rewards when those models generate value.
"I think we have the infrastructure to do this because we have the encryption, we have the verifiability," he said. "We have a way to coordinate decentralized compute and kind of push it toward the common good, right? With economic incentives."
NEAR is already building toward that vision with private AI inference, verifiable computing and infrastructure designed to allow autonomous AI agents to coordinate transactions. The network is also exploring ways to connect activity across its AI and blockchain products to demand for its native NEAR token, including buybacks funded by a portion of revenue generated through NEAR Intents.
Still, whether blockchain-based incentives can support the enormous costs of training competitive frontier models remains an open question.
Polosukhin's perspective carries particular significance because he helped develop the technology powering the AI boom in the first place.
After working on Transformers at Google, he left to pursue an idea that now sounds remarkably familiar: Allowing people to program computers using natural language. It's since now come to be called "vibe coding."
"We called it teaching computers to program," he joked, describing work that anticipated today's enthusiasm for AI-assisted coding.
The technology wasn't ready in 2017. Computing resources were limited, and the models lacked the capabilities needed to make the vision practical.
Now, nearly a decade later, the same researcher is betting that the next major transformation won't simply be about making AI more powerful. It will be about making that intelligence cheaper, more accessible and less dependent on a handful of companies.
For Anthropic, its Wall Street debut represents a powerful endorsement of the proprietary AI business model. To Polosukhin, it also highlights the economic challenge open-source developers must overcome to compete.
The irony is that the next serious threat to the world's most valuable AI companies may not come from a competitor spending even more to build a bigger model. It may come from developers proving that, for much of what businesses actually need, a smaller and cheaper model is already good enough.
Subscribe to the free Coinage newsletter to stay up-to-date on all things crypto and finance.