China is currently leading the world in the development of the best open-source large language models, surpassing the US in this domain.
The United States suffers from a 'structural disadvantage' in AI development due to siloed research at major labs and the exclusion of academia from foundational model creation, which stifles innovation.
NVIDIA's dominance in AI hardware is primarily defended by its robust software stack, a moat that competitors like Google, AWS, and AMD have found exceedingly difficult to overcome despite producing good hardware.
Current popular methods for evaluating LLMs, such as 'LLM as a judge' and crowdsourced leaderboards, are inherently flawed and suffer from biases like positional, verbosity, and self-preference.
The massive capital investment into AI data centers by hyperscalers is 'very likely' to lead to an overbuilding of infrastructure, echoing the boom-and-bust cycle of the dot-com era.
Over one year ago
Stoica held expectations that the US would lead in open-source model progress and that more viable hardware alternatives to NVIDIA would emerge.
Past year
Stoica observed a reality that contradicted his expectations, noting he was 'pleasantly surprised' by the significant progress in open-source models coming from China and disappointed by the lack of new, viable competitors to NVIDIA's hardware dominance.
Present
Stoica assesses that the US has a 'structural disadvantage' in AI due to its siloed research ecosystem and has lost its leadership position to China in multiple high-tech industries. He also observes a massive build-out of AI data center infrastructure by hyperscalers.
Next few years
Stoica predicts that China will successfully build out its own domestic AI compute infrastructure to overcome US export controls and that the current infrastructure boom will likely result in an 'overbuilding' scenario similar to the dot-com bubble.
▶China's Ascendancy in the AI RaceApr 2026
Stoica argues that China is not merely a competitor but is actively leading in key areas of AI, particularly in the development of the best open-source large language models. He attributes this to structural advantages like sustained state funding for strategic goals and deep, effective collaboration between its tech industry and academic institutions.
This perspective challenges the Western-centric view of AI innovation, suggesting that investors and analysts should re-evaluate geopolitical risk and recognize China's coordinated national strategy as a primary driver of long-term competitive advantage in the sector.
▶The US's Structural Disadvantage in AIApr 2026
Stoica posits that the US AI ecosystem is fundamentally hampered by its siloed nature, where frontier labs duplicate efforts and academia is marginalized from pre-training foundational models due to a lack of compute resources. He warns this structure creates a 'limited diffusion of innovation' compared to China's more integrated approach.
This analysis implies that private-sector dominance alone may be insufficient for long-term US leadership; policy and investment may need to shift towards fostering a more open, collaborative ecosystem that reintegrates academia to maintain a competitive edge.
▶NVIDIA's Enduring Software MoatApr 2026
While acknowledging the high quality of competing hardware from Google (TPUs), AWS (Trainium), and AMD, Stoica identifies NVIDIA's software stack as its true, defensible moat. He notes that viable alternatives have failed to emerge as quickly as he expected, underscoring the immense difficulty and importance of the software ecosystem in the AI hardware market.
For analysts, this reinforces that the AI chip war is not just about performance benchmarks but about the entire developer ecosystem, making NVIDIA's market position far more resilient than a pure hardware comparison would suggest.
▶The Fragility of LLM EvaluationApr 2026
Stoica critiques the current state of LLM evaluation, citing UC Berkeley research on the 'LLM as a judge' technique pioneered with GPT-4. He highlights that these evaluation methods are subject to significant, measurable flaws, including positional bias, verbosity bias, and self-preference bias, making leaderboards an imperfect measure of model quality.
Stoica's skepticism towards existing benchmarks signals a significant market opportunity for more robust, objective, and unbiased evaluation platforms, which is the problem his new company, Elamarena, aims to solve.