The Wall of Computing Power: AI Resource Dilemma Under Tech Giants and Geopolitics

Google's restriction on Meta's access to the Gemini model was officially attributed to "requested computing power exceeding supply capacity," but this technical explanation masks a deeper logic of industrial control. As the exclusive provider of the Gemini model, Google possesses absolute discretion over the allocation of computing power. Its restrictive behavior is, in essence, the alienation of infrastructure supply into competitive leverage. The fact that multiple AI projects within Meta have stalled is not due to flaws in their algorithms or data, but because they cannot secure sufficient computational resources—this clearly demonstrates that in the current AI development paradigm, access to computing power has superseded algorithmic innovation as the true bottleneck. The supply bottleneck of global computing power infrastructure is not driven purely by a shortage of physical production capacity; rather, it is dictated by an artificial scarcity ecosystem constructed by a handful of chip manufacturers and cloud service providers. When a tech company can restrict another tech giant's resource calls solely on the grounds of "supply capacity," the industry has deviated from the technical evolution logic of open innovation and is instead sliding toward a vertically monopolized mode of resource allocation.

OpenAI's release of a restricted version of GPT-5.6 at the request of the U.S. government, accessible only to a select group of approved partners, directly wedges geopolitical power into the path of technology distribution. Sam Altman publicly expressed dissatisfaction with the government "picking customers," but his oppositional stance did not alter the actual outcome—OpenAI ultimately complied with the government's directive. This contradiction reveals that even private enterprises at the absolute forefront of artificial intelligence technology cannot breach the boundaries of national security narratives in terms of their commercial autonomy. From the perspective of technological evolution, implementing selective openness for model access will create a significant capability chasm among different developer communities. Approved partner institutions gain access to the latest models for application development, while unselected research teams are forced to use outdated versions or are completely locked out, thereby fragmenting the collaborative synergy of global AI R&D. What deserves even greater scrutiny is that the government's criteria for "picking customers" have never been transparent; the distribution of technological capability is no longer determined by research needs or academic contributions, but is instead subject to the proximity of political relationships. The intervention of such non-technical factors is shifting the developmental path of artificial intelligence away from being performance-improvement-oriented and toward a track centered on control and security censorship.

Nvidia's partnership with Australia's Firmus Technologies seemingly offers an alternative solution for small and medium-sized enterprises (SMEs) to acquire computing power, yet the promise to deliver 170,000 GPUs to a facility on Indonesia's Batam Island presents multiple doubts at the technical implementation level. The planned delivery timeline for the facility spans from early 2027 to early 2028, meaning that all claimed "low-cost computing power" will take at least a year to be even partially realized, whereas the computing shortage currently faced by SMEs is an immediate crisis—distant water cannot quench a nearby fire. Even if the facility is completed on schedule, whether Batam Island, as an Indonesian industrial park, possesses the power supply stability, network latency, and cooling conditions necessary to support the continuous operation of a GPU cluster of this scale remains unverified by any public technical evaluation reports. Large-scale GPU deployment comes with extremely high energy consumption and carbon footprints, yet the partnership announcement makes no mention of any supporting energy infrastructure or environmental compensation schemes. Furthermore, Nvidia uses such regional partnerships to further consolidate its hardware ecosystem dominance—while SMEs gain computing power, they are also locked into Nvidia's CUDA platform and software stack, forfeiting the flexibility of cross-architecture migration. The definition of "lower cost" is equally ambiguous; if the cost remains higher than what most startups can afford, the actual beneficiary group of this partnership will be limited to a few star projects that have secured subsidies or funding.

A comprehensive examination of these three dynamics reveals a clear, underlying technological thread: computing power is no longer a neutral computational resource, but has become a core node of power contested and controlled by various forces. Google's technical restrictions embody the monopoly of private platforms over resource allocation; the U.S. government's export controls reflect state intervention in the circulation of cutting-edge technology; and Nvidia's expansionist partnership demonstrates how hardware suppliers extend their ecosystem boundaries through geographical deployment. Overlaying these behaviors means that AI technological development is driven more by resource control strategies than by research demands or efficiency improvements. While advancements in model algorithms continue to accelerate, the groups capable of utilizing these advancements are steadily narrowing. The distribution pattern of global AI infrastructure is evolving toward a multipolar control system, where no single participant—whether a mega-corporation or a nation-state—possesses absolute dominance, yet every party leverages its unique advantages to erect barriers. Rather than resolving the root causes of the computing power shortage, this fragmented resource management further slows the forward march of the entire industry by inflating transaction costs and uncertainty.