AI deployment: the macroeconomic divergence, circular capital loops, and geopolitical arbitrage

This is article 2 of 3 in The Silicon Reckoning. Here is article 1.

While enterprise FinOps teams work to reclaim fiscal control at the microscopic level of individual API queries, the massive macroeconomic engines funding the AI infrastructure boom are operating at a scale that dwarfs corporate optimization efforts. The localized microeconomic struggle to control corporate AI budgets, defined by granular FinOps and dynamic routing, is just one reflection of a much larger, increasingly fragile global economic fragmentation. While teams implement prompt caching to slash monthly bills, they are ultimately operating within a macro-level structure defining where intelligence is built and how much it should cost.

This corporate spending retreat is occurring just as the macro-level AI economy faces a systemic, historical tension: a profound divergence where trillions in infrastructure capital expenditure are chasing highly concentrated, circular revenues. This macroeconomic pressure cooker has catalyzed a new form of high-stakes geopolitical competition, driving enterprise buyers toward a radical alternative: open-weight model arbitrage.

The macroeconomic divergence and circular investment loops

The macro-level AI economy is facing a significant discrepancy between infrastructure capital expenditure and real commercial revenue, while business IT teams battle with microeconomic token optimisation. Sequoia Capital has dubbed this structural tension the "600 Billion Question" based on an analysis that indicates the massive capital investment in data centers, chips, and power grid expansions required for the global AI infrastructure buildout will cost about $600 billion in revenue annually.

As a fraction of the world's GDP, the scope of this infrastructure deployment is monumental, matching or surpassing both the Apollo space program and the Manhattan Project. The largest hyperscalers in the tech industry, such as Microsoft, Alphabet, Amazon, Meta, and Oracle, anticipate a total capital expenditure of almost $650 billion in 2026 alone. However, the total revenue from the AI services layer is still quite concentrated and much out of proportion to this expenditure. AI services are expected to bring approximately $25 billion worldwide in 2025, which is less than 10% of the capital used. Additionally, data centers incur annual depreciation costs of almost $40 billion, which greatly exceeds their direct revenue generating, according to financial organisations like Goldman Sachs.

A very delicate, cyclic investment cycle is responsible for a large amount of the stated expansion in the AI industry. Under these arrangements, cash flows among a few well capitalised players rather than from end-user market demand. For instance, the specialised cloud service CoreWeave sourced its actual GPUs through significant capital expenditures back to Nvidia, a fundamental investor in CoreWeave, while concurrently securing massive multi-billion-dollar infrastructure partnerships with OpenAI and Meta. Similar to this, OpenAI has made large, long-term computing commitments, such as $300 billion with Oracle and $38 billion with Amazon, even though its annualised revenue run-rate is much lower than these long-term obligations. Essentially, the startups are funded by the hyperscalers, who then promptly return the money to the hyperscalers to lease cloud infrastructure, giving the appearance of rapidly increasing software income.

The global technology industry is exposed to systemic hazards as a result of this circularity. If anticipated end-user demand does not materialise or if enterprise purchasers continue to aggressively reduce their expenditure commitments in order to prevent variable cost overruns, the entire AI ecosystem is extremely vulnerable to a quick correction. At the macroeconomic level, the story is already beginning to fall apart. According to Jan Hatzius, chief economist of Goldman Sachs, investment spending in AI contributed "basically zero" to the growth of the US GDP in 2025.

This GDP disparity has a simple structural explanation: highly specialised, imported components account for about 75% of data center capital expenditures. The actual silicon is produced and packaged in Taiwan and South Korea when US IT behemoths spend billions of dollars on cutting-edge GPUs. These transactions are recorded as imports in national accounting, which immediately reduces domestic GDP figures. The actual silicon enters Western balance sheets as capital leaves, leaving the domestic economy vulnerable to the hardware's declining value without a commensurate increase in domestic productivity.

Sovereignty, geopolitical friction, and supply chain vulnerabilities

The geopolitical environment in which the global AI cost reckoning is taking place is very unstable, marked by fierce rivalry between superpowers, strict export regulations, and a systemic drive toward technological sovereignty. In 2026, the technology industry functions under a "New Economic Nationalism," in which governments actively intervene to safeguard domestic supply chains, safeguard vital intellectual property, and create politically aligned technology blocs.

The supply chain for semiconductors is a clear example of this geopolitical tension. Using bipartisan legislative measures like the STRIDE Act and the MATCH Act to unite foreign partners under a single restriction regime, the United States has consistently strengthened its export control system to limit China's access to cutting-edge semiconductors and manufacturing machinery. Additionally, it is anticipated that US Section 232 investigations into semiconductors and critical minerals will reach critical junctures in July 2026. This could lead to a new wave of tariffs, domestic content requirements, and increased customer due diligence obligations throughout the manufacturing hubs of Northeast and Southeast Asia.

China has doubled down on an open-source and open-weight model plan as a highly effective geopolitical "jiu-jitsu" in response to these increasing constraints. While Chinese technology giants and research labs, such as Alibaba, Zhipu AI, Moonshot AI, and DeepSeek, have released highly competitive open-weight models (like Qwen, GLM, Kimi, and DeepSeek) into the global developer ecosystem, Western frontier developers, such as OpenAI and Anthropic, operate behind proprietary APIs and actively advocate for domestic regulatory protections.

These models are not obscure research projects; they are highly sophisticated architectures that have rapidly compressed the cognitive gap with Western frontier systems. According to comprehensive Artificial Analysis data, Chinese frontier models have progressed from representing roughly 60% of Western model capability in 2023 to over 90% in 2026.

Disruptive pricing and where enterprises are placing their bets

For global enterprise buyers navigating severe budget constraints, the defining differentiator of these Chinese models is their disruptive pricing. Select Chinese models are up to 50 times cheaper per token than their Western counterparts, with inference costs running 10 to 20 times lower while maintaining comparable performance on software engineering and reasoning benchmarks. Furthermore, because these models are released under open-weight licenses, enterprises can deploy them locally on private infrastructure or via hybrid cloud catalogs. This open architecture allows companies to run, fine-tune, and quantize the models without paying ongoing API tolls to Western providers or exposing sensitive corporate data to cross-border transfer risks.

For neutral global digital hubs like Singapore, this model mobility offers a significant structural benefit. Multinational firms operating throughout APAC are increasingly using Singapore's highly developed, secure data center infrastructure to host localised versions of open-weight models as the US-China technology race heats up. By doing this, they build extremely robust, hybrid AI stacks that are protected from unforeseen geopolitical supply outages, export limitations, or unilateral cancellations of cloud access, guaranteeing ongoing corporate operations in an increasingly fragmented global order.

Conclusion

The AI infrastructure boom's macroeconomics creates structural imbalances that divide the technological landscape into open-weight Eastern architectures and protectionist Western API networks. In order to implement these hybrid technologies without being caught in the crossfire of superpowers, global enterprise customers are increasingly looking for safe, neutral locations. Due to this fact, Southeast Asia is at the epicentre of the upcoming digital adoption phase. Navigating this multi-speed region, however, necessitates looking beyond capital availability to comprehend the sharp contrast between corporate reality and state ambition, the drastic changes undermining the local labour market, and the crucial environmental limitations endangering the development of local infrastructure.

illuminem Voices is a democratic space presenting the opinions of leading Sustainability Thought Leaders, their views do not necessarily represent those of illuminem.

The world needs sustainability knowledge. At illuminem, no interest group or shareholder can influence our work. Thank you for supporting our mission to make high-quality and independent sustainability information free for all. Every contribution helps. Thank you for donating today.