Washington and Beijing want you to believe the world is splitting into two rigid technological empires. They want you to think every nation on earth must eventually pick a side, sign a loyalty oath, and lock itself into either an American silicon stack or a Chinese infrastructure grid.
It makes for dramatic headlines. It sells defense contracts. It flatters the egos of policymakers in two capitals who still view global politics through a 1950s Cold War lens. Don't miss our previous article on this related article.
There is just one problem. The countries supposedly "torn" between these two superpowers are not agonizing over a binary choice. They are running an auction.
The lazy consensus in tech journalism right now treats middle-power nations—from the Gulf states to Southeast Asia, Latin America, and parts of Europe—like helpless bystanders caught in a superpower tug-of-war. That framing completely misunderstands how sovereign capital, compute infrastructure, and modern industrial policy actually work. Nations are not being forced into AI blocs. They are playing Washington and Beijing against each other to extract subsidized data centers, sovereign cloud architectures, and preferential chip allocations that neither superpower would ever offer to an obedient ally. To read more about the context of this, TechCrunch provides an excellent breakdown.
The Fallacy of the Binary Stack
To understand why the "two blocs" narrative collapses under scrutiny, look at the physical reality of building artificial intelligence. You do not buy "an AI bloc" off a shelf. You build a stack composed of distinct layers: raw energy generation, physical real estate, cooling systems, networking hardware, graphics processing units (GPUs), foundation models, and application-layer software.
China and the United States do not hold symmetrical monopolies across these layers.
The United States, through Nvidia, AMD, and the broader design ecosystem backed by Dutch lithography giant ASML and Taiwanese foundry TSMC, holds a near-total chokehold on top-tier training compute. Meanwhile, China dominates the physical supply chains that make data centers possible at scale—high-voltage transformers, battery storage, solar grid integration, fiber-optic cabling, and cost-effective inference hardware for edge devices.
If you are a sovereign wealth fund in Riyadh, a telecommunications giant in Jakarta, or a manufacturing conglomerate in São Paulo, picking one side exclusively is economic suicide. If you align 100% with Beijing, you are starved of H200 and Blackwell-class compute clusters due to U.S. export controls enforced by the Bureau of Industry and Security (BIS). If you align 100% with Washington, you face skyrocketing capital expenditures, multi-year backlogs for electrical transformers, and strict compliance audits that treat your sovereign data like a national security threat.
I have sat in boardrooms across Singapore and Abu Dhabi watching chief technology officers dissect vendor contracts line by line. They laugh at the Western punditry claiming they are "torn." They are doing what smart procurement officers have always done: multi-homing. They buy American silicon for frontier model training, contract Chinese engineering firms to build the solar-powered grid infrastructure cooling those chips, and deploy open-weights models that neither Washington nor Beijing can remotely switch off.
Why Export Controls Created an Arbitrage Market
When the U.S. Department of Commerce tightened semiconductor export restrictions in October 2022 and expanded them repeatedly through 2023 and 2024, the stated goal was to freeze China’s frontier AI development. The unintended consequence was the creation of the most lucrative compute arbitrage market in history.
Every time Washington restricts a chip or blacklists a Chinese cloud provider, it increases the bargaining power of neutral jurisdictions.
Consider what happened in the Middle East. When the U.S. government raised alarms over Emirati AI giant G42’s ties to Chinese hardware vendors like Huawei, pundits declared that the UAE had been forced to bend the knee to Washington. Look closer at the ledger. G42 divested its Chinese hardware holdings, yes—and in exchange, secured a $1.5 billion strategic investment from Microsoft, direct access to advanced Nvidia chips, and a fast-track partnership to build massive data center capacity in the region. Meanwhile, the broader Gulf region continues to welcome Chinese electric vehicle manufacturing, telecommunications infrastructure, and renewable energy joint ventures.
That is not capitulation. That is textbook geopolitical arbitrage.
Middle powers realized early on that Washington’s greatest fear is losing market share to Beijing, while Beijing’s greatest fear is total technological encirclement. By maintaining credible relationships with both sides, a nation turns itself into a swing voter. Swing voters get paid. Loyalists get taken for granted.
The Open-Weights Wrecking Ball
The "AI bloc" theory also relies on a deeply outdated assumption: that whoever controls the proprietary frontier models controls the intelligence layer of the global economy.
That assumption died the moment Meta released Llama and French startup Mistral proved that small, hyper-efficient models could rival closed American systems. By the time Chinese lab DeepSeek demonstrated that frontier-grade reasoning models could be trained at a fraction of traditional Western compute budgets—and released those weights openly—the entire premise of a closed superpower monopoly evaporated.
Let us define our terms precisely, because tech commentators routinely butcher them. Open-source AI requires full transparency of training data, code, and weights under an OSI-approved license. Open-weights AI means the trained parameters of the neural network are publicly downloadable, allowing any organization to run, fine-tune, and host the model on its own servers without sending a single byte of data back to California or Beijing.
You cannot lock a country into an ideological bloc using software that anyone can download onto a local cluster and modify at will.
When a national bank in Brazil or a logistics firm in Malaysia downloads an open-weights model, fine-tunes it on proprietary local data, and runs it on sovereign servers, which "bloc" are they in? Neither. They have achieved sovereign AI—the ability to generate intelligence using domestic infrastructure, domestic data, and domestic talent, completely insulated from foreign API shutdowns or sudden terms-of-service changes.
This is why the hand-wringing over whether countries will choose OpenAI and Anthropic or Baidu and Alibaba misses the point entirely. Enterprise buyers outside the U.S. and China view closed, API-gated models as temporary prototyping tools, not long-term infrastructure. Nobody builds critical national infrastructure on rented APIs that can be throttled by a foreign regulator's pen stroke.
Dismantling the Real Questions Leaders Are Asking
If you strip away the geopolitical theater, decision-makers in emerging tech hubs are asking very different questions than the ones dominating Western op-eds.
Should countries pick the US or China for AI infrastructure?
They should reject the premise of the question. Choosing a single national vendor stack creates a catastrophic single point of failure. If a country relies solely on U.S. hyperscalers, it exposes its domestic economy to extraterritorial sanctions, Cloud Act data subpoenas, and sudden export-licensing freezes. If it relies solely on Chinese vendors, it invites secondary U.S. sanctions that can sever its banking sector from dollar clearing and cut off access to advanced semiconductor maintenance. The winning strategy is architectural decoupling: keep the physical energy and telecom layer diversified, build sovereign data centers owned by domestic entities, and run open-weights software that can migrate across different chip architectures.
Are smaller nations falling permanently behind in the AI race?
Only if they are foolish enough to compete in the wrong race. Training a frontier foundation model from scratch requires tens of billions of dollars in capital expenditure, gigawatts of dedicated baseload power, and thousands of specialized researchers. Roughly four entities on earth can afford to burn that kind of cash without immediate return on investment. For every other nation, trying to build a sovereign rival to GPT-class models is a vanity project that drains treasury funds for zero economic return. The real wealth in AI is not generated by the companies building the engines; it is generated by the industries deploying those engines to compress costs in mining, agriculture, shipping, healthcare, and financial services. A country that skips the foundation-model arms race and focuses ruthlessly on applied industrial deployment will outperform nations ten times its size.
Will U.S. chip controls eventually force global alignment?
Controls buy time; they do not buy loyalty. Every restriction Washington places on global compute exports accelerates two trends: domestic chip self-reliance in China and silicon diversification everywhere else. Inference—the process of actually running trained AI models to answer queries—does not require the same cutting-edge interconnects and high-bandwidth memory clusters as frontier training. As algorithmic efficiency improves, inference is steadily migrating toward cheaper, older-node chips and custom ASICs that fall completely outside U.S. export thresholds. You cannot sanction basic math, and you cannot export-control efficiency gains.
The Hidden Costs of Sovereign Hedging
I will not pretend this multi-aligned strategy is free or effortless. Playing both superpowers against each other comes with severe operational penalties that naive boosters of "AI sovereignty" routinely ignore.
First, architectural fragmentation is brutally expensive. Maintaining separate compliance environments—one audited for U.S. export compliance and another insulated from Western telemetry—requires duplicate engineering teams, redundant security audits, and custom middleware. I have seen telecommunications operators in Southeast Asia burn through 30% of their annual IT budgets simply maintaining air-gapped networks so their American cloud contracts do not cross-contaminate their Chinese fiber infrastructure.
Second, talent drain remains relentless. You can buy H100 clusters with oil money or sovereign debt, and you can license open-weights models for free. You cannot instantly manufacture the senior machine learning engineers who know how to keep a 10,000-GPU cluster from crashing every six hours due to InfiniBand packet loss or thermal throttling. When middle powers build massive data centers without cultivating local systems engineers, they end up flying in expensive Western or Chinese contractors, recreating the exact dependency they were trying to escape.
Third, regulatory whiplash can strand billions in capital. A data center takes three to five years to permit, connect to the electrical grid, and commission. Washington’s export control rules change every six months. A sovereign fund can break ground on a facility assuming it will receive a specific tier of American accelerators, only to have the Bureau of Industry and Security redraw the licensing map halfway through construction. Hedging requires massive financial buffers to absorb stranded assets.
How Smart Operators Actually Build Sovereign Stacks
If you are an enterprise leader, a policymaker, or an investor operating outside the Washington-Beijing axis, stop reading geopolitical threat assessments and start executing a ruthless procurement playbook.
- Separate your training layer from your inference layer. Do not pay a premium for sanctioned, top-tier American silicon to run routine inference workloads. Rent frontier compute in compliant jurisdictions only when you need to fine-tune weights, then export those quantized models onto cost-effective, locally hosted hardware for daily operations.
- Mandate model portability from day one. Never build your core business logic around proprietary API features—such as vendor-specific tool-calling formats or closed memory layers—that lock you into a single provider. Containerize your inference pipelines using open standards like vLLM or Triton so you can swap underlying models and hardware architectures in forty-eight hours if a trade war flares up.
- Invest in power, not just silicon. Chips depreciate to near-zero value within three to four years as newer architectures drop. High-voltage grid interconnects, substations, and cooling infrastructure appreciate for decades. Countries and companies that control cheap, reliable megawatts hold the ultimate trump card; chip vendors will always lobby their governments for export licenses to reach wherever the electricity is cheapest.
- Exploit the talent spillover. Tightening visa restrictions in the United States and brutal work cultures in China are pushing thousands of top-tier AI researchers to look elsewhere. Jurisdictions like Toronto, London, Zurich, Dubai, and Singapore are aggressively absorbing this displaced human capital through fast-track residency and research grants. Attract the engineers, and the compute will follow.
The narrative of a world helplessly fractured into two rigid AI blocs serves only the people selling weapons, sanctions compliance software, and cable news airtime. Out in the real economy, middle powers have looked at the superpower rivalry and realized a fundamental truth: when two giants fight over who gets to rule the future, the smartest move is to sell them both the bricks.