In US or China, firms need AI they can independently verify and trust

When the Magnificent Seven, America's top tech companies, experienced a sudden drop of over US$1 trillion in stock value, the question resurfaced: is artificial intelligence (AI) a bubble? However, the focus on AI has been overshadowed by personal and governance issues at the leading tech giants. The underlying reality is that these multitrillion-dollar companies cannot solely rely on founder mythology and self-regulation.

The public is questioning who should bear the risk and have a say in AI development. This shift reflects the commercial hesitance of the industry, as banks, hospitals, and factories are unwilling to use AI they cannot independently verify and trust. Adoption rates are not merely slow; they are blocked. Companies today are only buying AI that can be easily checked and trusted.

For instance, Anthropic, a leading AI company, is raising funds but is only seeing commercial success through coding, which is easily verifiable. In contrast, legacy industries such as finance, healthcare, manufacturing, retail, and traditional businesses struggle to adopt AI. A June Veeam survey found that while 88% of executives are piloting AI agents, only 7% are truly AI-ready.

This discrepancy highlights the risk associated with AI tools. A majority of workplace AI tools, 71.7%, carry high or critical risk, and 83.8% of enterprise data sent to AI goes to risky tools. A bank that entrusts its underwriting history to an unauditable AI tool is being negligent. Anthropic's November 2 amended S-1 filing reveals the revenue story and trust issues.

While the company generated US$16.96 billion in revenue for the three months ended September 30, with a quarter-on-quarter growth slowing from 144% to 47%, it committed over US$895 billion to compute, against US$132 billion in cash. Internal estimates suggest Anthropic's models might soon exceed a "dramatically accelerated progress" threshold, requiring additional controls that investors are hesitant to embrace.

Furthermore, Anthropic built its "central library" by purchasing print books, scanning, and discarding the paper. Although this is considered fair use in US courts, Anthropic faced accusations of using pirated material, which were settled after it offered US$1.5 billion. The issue of trust in AI development stems from the public's perception of AI as either godlike power or a potential catastrophe.

While Silicon Valley labs have inherited this narrative, they must move beyond it. Wall Street's fear of AI is driving revenue momentum, but when revenue slows and trials stall, legacy industries remain skeptical about incorporating AI into their operations. The same situation exists in China, where AI rules have been established early, but banks and factories are still grappling with trust issues.

Beijing's regulatory framework focuses on content safety and social stability rather than independent technical audits and commercial liability. When Zhipu AI faces scrutiny over unauthorized user code uploads and privacy concerns, the lenient regulatory environment backfires, leading legacy industries to become more cautious. Frontier labs often confuse damage control with accountability.

OpenAI's November 5 National Community Compact, which involves paying full electricity costs, dropping non-disclosure agreements, reporting water use and noise levels, addresses community concerns about data center build-outs. However, a hospital risk committee is more concerned about liability when a model fails than about a lab's water bill.

To unlock the trillion-dollar AI market, US and Chinese AI companies must embrace regulation and create independent third-party auditing institutions. External audits, evaluations, and accepting binding commercial liability from independent bodies are crucial for establishing a checkable standard. By doing so, they can foster trust and unlock the potential of AI in legacy verticals.