a16z Partner Martin Casado: AI Extinction Risk Theory Becoming Marginal
Venture capital firm a16z general partner Martin Casado stated on social platform X that discussions around AI "existential risk" (X risk) are noticeably cooling and becoming more rational based on his frontline observations. He believes that while AI does pose "new cybersecurity risks," the industry as a whole has the capability to respond, and these risks are "likely much lighter than during the early personal computer and internet era." He also noted that such systems can now be "pragmatically and safely" deployed at scale, and recent concerns about extinction-level risks are increasingly viewed as "marginal and overblown". Moderate voices are not as widely disseminated as extreme ones, making it difficult to perceive this trend unless one continues to engage deeply in industry discussions.
Casado's remarks continue a recent public debate with Databricks CEO Ali Ghodsi regarding the framing of AI risks. In an interview released on September 18 titled "Databricks CEO: Stop Scaring People About AI," Ghodsi acknowledged the risks of AI but argued that "extinction-level risk is close to zero" and urged industry leaders not to create public panic with exaggerated language, as such statements could lead to mental health issues. Casado challenged this in the program, questioning how some industry insiders can claim AI poses existential threats while simultaneously seeking IPO allocations, describing this contradiction as a "four-dimensional game that needs to be unpacked."
Casado's own judgment on AI risks has also shifted recently. In late August, he maintained that AI scaling laws are still valid, linking AI capability improvements directly to capital investment, framing it as a "capital scale game." However, he also pointed out that the real risk to watch is not the models going out of control, but the concentration of computing power, talent, and funding in a few leading AI labs. A single decision by these labs regarding access, pricing, or safety protocols could impact the entire downstream dependent industries. Since the end of 2024, he has consistently advocated for "empirical regulation," meaning regulation should target proven actual harms rather than speculative risks that have yet to materialize.
Shortly before Casado's post, AI company Anthropic released a threat intelligence report for September 2026, revealing several real cases of AI being misused for cyberattacks: a hacker group identified as linked to Russia attacked 24 out of 27 targeted organizations within 130 days, involving multiple Ukrainian government departments and defense agencies. The attacking AI agent could autonomously monitor malware detection and automatically recompile code upon detection, leading to the theft of over 300,000 North African national identity records. An opportunistic ransomware gang completed a cloud environment breach in 2 to 3 hours using just 10 AWS nodes, processing 1.8 million Android application packages during the attack. Another hacker group identified as linked to China identified over ten potential zero-day vulnerabilities targeting about 50 global targets within a month.
These cases are seen as direct evidence of the existence of "new cybersecurity risks" posed by AI, but they also confirm Casado's assertion that "the industry is already responding." Major AI companies now generally monitor, disclose, and ban accounts and techniques used for malicious attacks, forming a targeted detection and response mechanism. Regarding the discussion of existential risks, the opposing viewpoint to extinction-level risk argues that similar concerns were also present during the early stages of personal computer proliferation and internet development, where there were large-scale infrastructure failures and economic losses in the billions. However, the industry ultimately managed these risks through continuous security practices rather than pausing technological development itself.
From a capital and discourse perspective, Casado, as a partner at a16z, which holds significant equity exposure in AI startups, downplaying extinction-level risks and emphasizing "safe scalable deployment" objectively helps reduce external pressure for strict preventive regulation on the AI industry, thereby protecting the valuations and financing environment of AI companies in its portfolio. In contrast, those who insist on amplifying the narrative of existential risks (including some AI safety researchers and policy advocacy groups) typically have a louder voice in public discourse. However, according to Casado, this voice advantage is gradually diverging from the actual distribution of frontline discussions within the industry, indicating a degree of differentiation between capital and public discourse agendas.
ABAB AI Insight
Casado's historical behavior shows he has long stood on the side of "technological optimism and regulatory caution"—since the end of 2024, he has consistently advocated for "empirical regulation," opposing one-size-fits-all restrictions based on speculative risks. His recent downplaying of AI existential risks is not an isolated viewpoint but a continuation of this stance on AI issues. Notably, Casado is a serial entrepreneur who founded the network virtualization company Nicira and sold it to VMware. This experience of witnessing emerging technologies transition from controversy to mainstream forms the cognitive basis for his comparison of AI risks to early personal computer and internet security issues.
The pathways for mobilizing capital and discourse are clear— as a partner in a16z, which holds stakes in several leading AI labs and related infrastructure companies, his public statements downplaying extreme risk narratives and emphasizing "safe scalable deployment" objectively help maintain public and regulatory expectations of leniency towards the AI industry, thereby protecting the valuations and future financing environment of its portfolio. Meanwhile, Casado has also publicly pointed out that the real risk has shifted from "whether models will go out of control" to "the concentration of computing power, talent, and funding in a few leading labs." This indicates that capital is also shifting from single bets on leading labs to betting on more dispersed and certain segments like computing infrastructure, adjusting the logic of capital allocation in line with changing risk perceptions.
The historical analogy Casado actively cites is the early stages of personal computers and the internet—during which there were also real security incidents involving hospital system failures, critical infrastructure disruptions, and economic losses in the billions. However, the industry ultimately did not pause technological proliferation but gradually absorbed risks through continuous security practices, detection, and response mechanisms. The latest threat intelligence report from Anthropic detailing real cases of Russian-backed hacker groups, ransomware gangs, and Chinese-backed hacker organizations using AI for autonomous infiltration and identifying zero-day vulnerabilities provides contemporary examples that support this analogy. In terms of industry positioning, the AI sector is currently at a stage where "risks have been proven to exist, but response mechanisms are also maturing simultaneously," which closely resembles the historical process of the internet transitioning from chaos to control.
Essentially, this represents a struggle for regulatory discourse power—the discussion of AI existential risks is diverging from a single dimension of "whether technology is safe" into two parallel narrative lines of "frontline consensus within AI companies" and "public discourse volume." The former tends to be moderate and focuses on manageable specific risks (such as cyberattacks and computing power concentration), while the latter is still dominated by more extreme extinction narratives. This divergence occurs because moderate, detailed risk assessments naturally lack dissemination power, while extreme statements like "AI could lead to human extinction" are more likely to gain attention and spread on social media. The mechanisms of public discourse systematically amplify extreme voices and suppress moderate ones, leading to a structural deviation between the risk consensus perceived by the public and the actual risk consensus formed within the industry.
ABAB News · Cognitive Law
- Moderate judgments lack dissemination power; extreme judgments attract traffic.
- Every technological panic ultimately relies on engineering patience rather than pausing solutions.
- Whoever controls the definition of risk controls the steering wheel of regulation.