Articles of Interest - Week 12/16 - 12/22
- Walker Robinson
- Dec 22, 2024
- 4 min read
Updated: Jan 27
Barath Harithas, CSIS. December 20th, 2024.
A new CSIS report by Barath Harithas argues that the U.S.-China semiconductor battle isn't about chip manufacturing but rather computing power for AI development. Rather than focusing on chip manufacturing capabilities, Harithas suggests the key metric in this battle is concentrated computing power. The U.S. currently leads in (projected by 2025) overall compute power with 14.3 million AI accelerators compared to China's 4.6 million. That being said, the report warns that this apparent advantage could be less decisive than it appears. China has already demonstrated it can centralize the GPUs (around 100,000) needed to match individual U.S. training runs. Furthermore, China is rapidly developing its computing capabilities through initiatives like the National Unified Computing Power Network. The report ultimately emphasizes that national compute totals are less important than each country's concentrated capabilities. Harithas suggests that China could achieve critical mass in AI development even while operating at a significant overall disadvantage.
In the U.S., we commonly use compute power almost as a ranking, with large tech companies regularly bragging about increasing the number of AI chips they've got or their upcoming AI models trained with massive amounts of data. While U.S. export controls aim to maintain U.S. technological supremacy and undermine China's ability to obtain critical tech components, this strategy may prove insufficient should Beijing's strategy prove effective. That strategy, which includes bureaucratic streamlining, computer centralization, strategic opacity, and technological leapfrogging, could enable China to use its computing power more effectively than the U.S. despite the numerical disadvantage. Furthermore, China's focus on alternative technological advancements like neuromorphic computing and compute-efficient AI algorithms could undermine the current U.S. advantages in semiconductor manufacturing. The risk for the U.S. isn't China slowly catching up through "traditional" means but rather redefining the technological battlefield in unexpected ways that circumvent the current U.S. strengths.
Lauren C. Williams, Defense One. December 19th, 2024.
According to senior Pentagon official Ely Ratner, China continues to give the U.S. the cold-shoulder on attempts to engage in high-level military discussions about critical domains, including space, cyber operations, and nuclear defense. While speaking at a CSIS event, Ratner highlighted China's refusal to discuss its military modernization and expressed concerns about its growing nuclear program (which they refuse to explain). The DoD's annual report also emphasized worries over Beijing's focus on "intelligentized warfare" and "cyber-enabled" espionage capabilities.
The lack of strategic dialogue holds some significant implications for global security. Beijing's expanding nuclear program, which deviates from its historically limited nuclear doctrine, is highly concerning. This communication void becomes even more concerning in light of recent tensions in the South China Sea and Taiwan Strait and increasing cyber-attacks by Chinese-backed hackers. We often discuss the need to balance national security and international transparency regarding emerging technology. That balancing act becomes incredibly difficult if countries refuse to play along. The U.S. is not innocent on this front either. I fear that China's refusal to communicate on these fronts may represent their approach to emerging technology. It is clear that China views tech like AI as a method to achieving power on the global stage, and while that is likely a true belief, I hope that Beijing is at least somewhat conscious of the threats that these technologies pose should their development outpace our capability to deploy safely. Ultimately, the U.S. should continue pushing for high-level dialogues and creating crisis communication channels. While they may not be received well today, they could save countless lives in the future.
Benjamin Jensen, Yasir Atalan, Can Mutlu, Jose M. Macias III, CSIS. December 16th, 2024.
A new CSIS analysis explores how countries think about revealing military capabilities during international crises. The report argues that countries make these high-stakes decisions based on three distinct patterns: "reveal to substitute" (showing off new tech instead of direct military action), "reveal to reduce risk" (demonstrating capabilities to discourage aggression), and "reveal to compensate" (unveiling advances to offset perceived weaknesses). Using mathematical models, the researchers found that successful "reveals" depend heavily on specific conditions: relative military balance between the rivals, how well adversaries can detect and assess the new capability, and whether the "reveal" is part of a larger diplomatic strategy. The timing and context of these "reveals" can significantly affect whether they strengthen or potentially undermine deterrence.
The topic of this paper is highly relevant to the subject of emerging technology in modern militaries. As discussed above, China refuses to engage in high-level military discussions about critical domains. This refusal to communicate creates the perfect storm for miscalculation and miscommunication. The CSIS model suggests that successful deterrence through revealing technology requires an understanding of how your adversary will interpret and respond to the reveal. But how can either side accurately gauge this when we can't even get the two countries to communicate? Both powers will need to make decisions about revealing capabilities in an environment of increasing tension and limited dialogue, which makes the decisions even more difficult. The risk of miscalculation grows when reveals are made without clear channels for strategic communication. I worry that the lack of dialogue will undermine effective deterring "reveals" and lead to more significant risks of instigation and misunderstanding.
