No Registration: The New AI Race – DeepSeek, Regulation, and America’s Place in the Landscape

For this edition of “No Registration,” we bring you our takeaways from the New England Venture Capital Association (NEVCA) panel discussion “The New AI Race: DeepSeek, Regulation, and America’s Place in the Landscape.” With changes in U.S. administrations, the emergence of new Chinese AI models, and ongoing debates about regulation, the panel delivered timely insights on where the AI industry stands and where it’s headed.

Panelists included Kimo Gandall, CEO of Arbitrus.ai, a private court system using AI judges to make dispute settlement more efficient, and Kleida Martiro, Partner at Glasswing Ventures who specializes in early-stage Enterprise AI investments.

U.S. AI Landscape vs. Global Competition

The panel kicked off with a discussion around how the U.S. AI ecosystem compares to global competitors—particularly the E.U. and China.

Martiro, who brings an international perspective to the conversation, noted that the U.S. has a significant advantage when it comes to AI talent concentration. This talent pool remains one of America’s strongest competitive advantages.

However, the regulatory landscape differs significantly across regions. The E.U. has taken a more forward-looking stance with stricter requirements that often impact AI decision-making systems, such as requirements for things like Software Bills of Materials (SBOM) and security standards that are already planned through 2027.

Gandall highlighted how these regulatory differences directly impact startups like his: “A startup like mine would not be able to operate within the E.U. without changing its back end” due to the E.U.’s blanket prohibition on decision-making AI in certain fields, including legal applications.

The China Factor: DeepSeek and Data Security

DeepSeek’s release has sparked important conversations about data privacy and national security. While some view it as a breakthrough, Gandall provided a more measured perspective: “I think DeepSeek is a little overblown, especially with most companies that are deploying LLMs. The cost for tokens has gotten so low that we don’t actually see a large difference.”

The panel highlighted the critical distinction between running open-source models locally versus using hosted API services, with the latter raising significant privacy concerns. This has led several government bodies, including New York City and South Korea, to ban DeepSeek on their devices. 

The panel speculated on China’s strategic reasoning for open-sourcing DeepSeek: “It was released days after Trump was inaugurated… it was intentional, it was strategic for sure,” noted Martiro. The open-source nature could be seen as “a strategic show” to drive adoption in the U.S.

Shifting Regulatory Sands

The panel outlined how U.S. AI policy has evolved with the administration change. Under President Biden, Executive Order 14110 focused on “civil liberties, consumer protection, prevention of bias, and AI safety and trust.” This approach created dedicated AI oversight roles within government agencies and required developing detailed rules around testing and data protection.

In contrast, Trump’s approach prioritizes deregulation to enhance American competitiveness, with less focus on consumer protections and more concern about maintaining technological advantage over China.

Despite these significant policy shifts, most AI startups aren’t dedicating substantial resources to regulatory concerns. They remain focused on fundamental business objectives—securing funding, gaining customers, and demonstrating traction.

“I talked to basically every C10 portfolio company,” Gandall shared. “Nobody’s really thinking about it, to be honest. They’re more thinking, ‘How do I get my next VC check, and how do I make more money?’”

Martiro confirmed this from an investor’s perspective: “We haven’t seen anything change so far from just a conducting business point of view.” She noted that Glasswing’s investment criteria remain focused on core questions like “Where is the AI? How are you building the technology? How are you getting unique access to data?”

For early-stage founders, the day-to-day priorities are securing clients, handling basic business operations, and gaining a foothold in the market—not navigating complex regulatory frameworks. As Gandall put it, “When you’re an early-stage startup founder, you’re doing cold calls, you’re getting clients on board… you just don’t have time” to focus deeply on compliance unless directly compelled by law.

The Rise of Vertical AI and Model Specialization

AI is increasingly moving toward industry-specific applications rather than one-size-fits-all approaches. “Vertical polarization is where it’s at,” Martiro emphasized, noting that companies that customize general AI models for specific industries and use their own unique data will stand out from competitors.

However, the verticalization trend extends beyond just customizing existing models—it’s about developing deep expertise in specific domains like legal, healthcare, or financial services. Companies with access to proprietary data in these verticals have significant competitive advantages, as they can train models that outperform generic alternatives in specialized tasks.

AI Decision-Making and Bias Considerations

The panel tackled the complex issue of bias in AI systems. Gandall made an important distinction between technical bias and ethical bias: “In law, bias is not clear… bias could be theoretically good from a technical angle. It’s really certain types of bias, like racial bias or ideological bias, [that are problematic].”

Regulatory approaches to bias vary significantly. While Biden’s executive order focused on preventing racial and gender biases, Trump’s approach emphasizes addressing ideological biases. For startups developing decision-making AI systems, this shifting landscape requires careful navigation, especially when operating across different jurisdictions with varying requirements for human oversight.

The Future of AI Infrastructure and Compute Limitations

A significant challenge looming over the AI industry is the growing gap between computing demands and available infrastructure. During the Q&A, one audience member pointed out, “The models required 4x processing power, where the … infrastructure is only growing 30 percent.”

This disparity is driving major investments in energy solutions, with Gandall noting that “Microsoft is literally trying to build a nuclear power plant to power their data centers.” Simultaneously, companies are working on more efficient approaches through model distillation and specialized architectures. As Gandall mentioned, “We’re almost seeing the opposite now… things are getting cheaper” as models become more efficient through distillation techniques.

Martiro offered an optimistic view: “I’m hopeful that there is enough technological advance… that we won’t have to use that much infrastructure.” She also pointed out that while foundation model providers need massive computing capacity, most startups are focused on efficiency: “They’re not using as much or at all… because they don’t want extra costs.”