For most of the past decade, the global AI infrastructure market operated on a unified foundation. A data center in Singapore, a cloud region in Frankfurt, and a hyperscale campus in Northern Virginia all ran on the same hardware — Nvidia GPUs. They used the same software frameworks — CUDA, PyTorch, TensorFlow. They connected to the same cloud providers — AWS, Google Cloud, Microsoft Azure. The physical and logical architecture of global AI compute was, for practical purposes, a single ecosystem with regional variations in cost and latency but no fundamental incompatibilities. That unified architecture is ending.
The combination of US export controls on advanced AI chips, China’s retaliatory restrictions on Nvidia hardware imports, and the accelerating development of domestic Chinese AI infrastructure is producing two parallel AI ecosystems that are increasingly incompatible at the hardware, software, and cloud layers. The split is not complete. It is not reversible. And its consequences extend well beyond the bilateral relationship between the US and China.
The Export Control Regime That Started the Split
The mechanism of bifurcation is the export control regime. The US Bureau of Industry and Security issued its global AI Diffusion Rule on January 15, 2025, establishing a three-tier global licensing framework for advanced AI chips, computers, and model weights. The rule groups countries into categories that determine what AI hardware they can access without licence, what requires a licence application, and what is effectively prohibited. The practical effect of the rule, combined with the specific restrictions on Nvidia’s H100, H200, and A100 chips in China, is to make the world’s most advanced AI accelerators unavailable to Chinese AI developers and infrastructure operators through normal commercial channels.
China’s government has responded by blocking hardware imports that US authorities now require companies to license, creating what financial analysts describe as a regulatory sandwich: US restrictions limit what companies can sell, while Chinese restrictions limit what companies can import. Together, those policies have effectively severed the commercial relationship between Nvidia and the Chinese AI infrastructure market.
The Chinese AI Infrastructure Stack That Is Emerging
The infrastructure response within China to the export control regime has been faster and more comprehensive than most Western observers anticipated. Baidu and Huawei together now control nearly 70% of China’s GPU cloud market, defined specifically as cloud services built on domestically designed AI chips rather than imported GPUs, according to a Frost and Sullivan report. That market definition is itself significant — it reflects a deliberate reframing of what constitutes AI compute infrastructure in China, one that treats domestic silicon as the baseline rather than the alternative. The Huawei Ascend 910C and the newly launched Ascend 950PR have become the primary hardware platform for China’s largest AI developers. ByteDance, the parent company of TikTok, has committed $5.6 billion to purchase Huawei Ascend chips, signalling a dramatic and financially significant shift away from Western AI hardware that would have been commercially unthinkable two years ago.
The software ecosystem that runs on Huawei’s hardware is equally distinct from the Western stack. Baidu’s PaddlePaddle deep learning framework, integrated with Huawei’s CANN neural network computing architecture, forms a vertically integrated stack that is the Chinese equivalent of the Nvidia-PyTorch-CUDA ecosystem. Alibaba and Tencent have deployed massive CloudMatrix 384 clusters — Huawei’s domestic equivalent to Nvidia’s GB200 NVL72 rack systems — to power their generative AI services. The CloudMatrix 384 represents not just a hardware alternative but an architectural departure: it uses a different approach to chip-to-chip interconnect, different memory hierarchy, and different power delivery than the Nvidia NVLink-based architecture that underpins Western hyperscale AI infrastructure. Two AI infrastructure stacks are now developing in parallel, and they are optimising in different directions.
The Performance Gap and Its Strategic Implications
The central question for anyone evaluating the strategic consequences of AI infrastructure bifurcation is whether the Chinese domestic stack can close the performance gap with Western hardware on a timeline that matters strategically. As of early 2026, technical benchmarks suggest the Huawei Ascend 910C has reached competitive performance with Nvidia’s restricted H20 GPU for certain inference workloads, while trailing the unrestricted Blackwell architecture by a margin that remains meaningful for frontier model training. The BIS assessed in May 2025 that Huawei developed its Ascend chips in violation of US controls by using controlled US technology in their development, and warned that using such chips could itself violate US export controls — a legal claim that Huawei disputes but that adds regulatory complexity to any third-country operator considering Ascend adoption.
The performance gap is not static. The export control regime has created a captive market for Huawei — the world’s largest pool of AI developers and infrastructure operators, with no commercially viable alternative to Ascend hardware for their highest-volume workloads. That captive market is providing Huawei with the deployment scale, the operational feedback, and the revenue to accelerate its development roadmap at a pace that would not have been commercially achievable if Nvidia had retained unrestricted access to the Chinese market. Western labs including OpenAI and Anthropic continue to scale using unrestricted H200 and Blackwell clusters, while Chinese labs at Tencent and ByteDance are becoming the world’s largest testbeds for non-Nvidia hardware. This bifurcation is producing a permanent divergence in AI model optimisation — Western models optimised for raw memory bandwidth and CUDA parallelism, Chinese models engineered for the specific throughput characteristics of the Ascend architecture.
The Software Moat and Its Erosion
The software dimension of the bifurcation is as consequential as the hardware dimension and has received less analytical attention. Nvidia’s CUDA ecosystem has been the most durable competitive advantage in AI infrastructure, because the accumulated investment of AI developers in CUDA-optimised code creates switching costs that hardware performance advantages alone cannot overcome. That moat is being eroded from two directions simultaneously. Within China, the forced migration to Ascend hardware is producing a generation of AI developers whose primary expertise is in non-CUDA frameworks, who are building optimisation libraries, inference engines, and training pipelines for the Ascend architecture that have no CUDA dependency. As the Chinese AI developer community grows in size and sophistication, the institutional knowledge embedded in its codebase will become a Chinese-ecosystem asset rather than a CUDA asset.
Outside China, the maturation of hardware-agnostic compiler frameworks is reducing CUDA switching costs for Western developers as well. OpenAI’s Triton compiler, which supports Nvidia, AMD, Google TPU, and Amazon Trainium backends, is the most prominent example of a trend toward software portability that weakens the CUDA lock-in argument. The combination of forced non-CUDA development in China and voluntary framework portability in the West is producing a global AI software landscape that is less CUDA-centric than it was three years ago. The strategic implication is that the performance advantage Nvidia maintains in the hardware layer is less durable than it appears, because the software ecosystem that amplifies that advantage is becoming progressively less exclusive.
The Third-Country Dimension
The bifurcation of global AI infrastructure into US-aligned and China-aligned ecosystems creates a strategic choice that every country building significant AI infrastructure must now navigate. The choice is not simply which chips to buy. It is which ecosystem to join, because the hardware choice determines the software stack, the cloud provider relationships, the development community, and the geopolitical alignment that comes with deep infrastructure dependency on either the US or Chinese technology supply chains.
For most major AI infrastructure markets — the Gulf states, Southeast Asia, India, Europe, and Latin America — this choice is deeply uncomfortable. None wants to choose between the US and Chinese technology ecosystems because each maintains important economic and political relationships with both. Saudi Arabia’s Humain program, for example, is building AI infrastructure around Nvidia hardware while simultaneously maintaining diplomatic ties with China that discourage overt alignment with US export control objectives. Singapore’s data center market relies overwhelmingly on Western hardware, yet the country’s broader economic strategy depends on deep engagement with Chinese technology companies and supply chains.
India is actively developing domestic AI semiconductor capabilities to avoid dependence on either ecosystem. But these third-country positions remain inherently unstable. The export control regime aims to eliminate much of the remaining grey area because the AI Diffusion Rule specifically targets third-country channels used to circumvent chip export restrictions on China.
The ASEAN Dimension
Southeast Asia presents the most immediate and commercially significant third-country test case for the bifurcation dynamic. Malaysia became a flashpoint in 2024 when reports emerged that Nvidia chips were reaching China through Malaysian intermediaries, prompting US pressure on the Malaysian government to tighten enforcement of export control provisions. The Malaysian government’s retraction of a reported deal to deploy Huawei Ascend GPUs reflected the direct diplomatic pressure that the US is applying to prevent third-country adoption of Chinese AI hardware as a substitute for Western chips. Indonesia, Vietnam, and Thailand are all building data center infrastructure at significant scale, and all are navigating hardware procurement decisions that carry geopolitical implications that their governments would prefer not to make explicit.
The AI Diffusion Rule’s three-tier country framework formalises the geopolitical dimension of hardware procurement in ways that make it impossible for third-country governments to treat AI infrastructure investment as a purely commercial decision. A country placed in the most restricted tier under the rule faces significant limitations on its ability to access advanced Western AI hardware regardless of its commercial relationships. A country in the intermediate tier can access advanced hardware subject to licence requirements that create regulatory overhead and timeline uncertainty. Only countries in the most permissive tier can access advanced AI hardware through normal commercial channels without additional regulatory friction. The tier assignment is a geopolitical signal as much as a technical determination, and governments are reading it as such.
The Competitive Markets Europe Is Losing Ground To
Europe occupies a distinctive and uncomfortable position in the bifurcation dynamic. As a market with significant AI infrastructure investment, deep data sovereignty requirements, and strong political alignment with the United States on technology policy, Europe might appear to be a natural member of the US-aligned ecosystem. The practical picture is more complicated. European data sovereignty law, particularly the GDPR and its developing AI-specific interpretations, creates requirements for data processing within European jurisdiction that limit the degree to which European enterprises can depend on US hyperscaler infrastructure for sensitive workloads. At the same time, the EU AI Act’s risk-based regulatory framework applies to AI systems regardless of where their underlying infrastructure is located, creating compliance requirements that affect operators using both US and Chinese technology stacks.
The bifurcation creates a specific problem for European operators whose supply chains or customer relationships involve Chinese counterparties. A European manufacturer with significant Chinese operations may need AI infrastructure that can serve both European and Chinese workloads — but the hardware, software, and cloud platforms appropriate for each context are increasingly distinct. Running inference workloads in Europe on Nvidia hardware and in China on Huawei Ascend hardware requires maintaining two separate technology stacks, two separate developer skill sets, and two separate vendor relationships. The operational complexity of this dual-stack approach is substantial, and it is a complexity that did not exist three years ago when a single Nvidia-based infrastructure could serve both markets without fundamental architectural incompatibility.
The European Semiconductor Sovereignty Response
Europe’s response to the AI infrastructure bifurcation has centred on the push for European semiconductor sovereignty — reducing dependency on both US and Chinese chip supply chains by developing a domestic European capability. The EU Chips Act, which committed 43 billion euros to European semiconductor development through 2030, reflects a recognition that depending entirely on US-controlled chip supply chains creates strategic vulnerability that EU member states are not willing to accept indefinitely. The ambition is to build European capacity at advanced process nodes that can serve both commercial AI infrastructure needs and defence applications that cannot, by their nature, depend on foreign-controlled hardware.
The gap between that ambition and current European capability is substantial. TSMC’s Dresden fab, which is the most significant near-term addition to European advanced semiconductor manufacturing, will produce chips at 28nm and 16nm process nodes — multiple generations behind the 3nm and 2nm processes that define the current AI accelerator competitive frontier. A European AI chip designed and manufactured at European fabs using currently available European process technology would trail Nvidia’s Blackwell architecture by a performance margin that makes it commercially uncompetitive for the most demanding AI training workloads. The European semiconductor sovereignty argument is strategically coherent as a long-term objective. It does not resolve the near-term infrastructure dependency that European AI operators face in a world where advanced AI hardware is increasingly controlled by either the US or China.
The Nvidia Revenue and Strategy Implications
The bifurcation of global AI infrastructure creates a strategic challenge for Nvidia that its current market dominance obscures but does not eliminate. The Chinese market represented a significant revenue stream for Nvidia before the export control regime effectively closed it. The successive rounds of export controls, from the October 2022 restrictions on the A100 and H100 through the restrictions on the modified H20 in 2025, have progressively reduced Nvidia’s accessible Chinese market from one of its largest growth opportunities to effectively zero for its most advanced products.
The revenue loss from the Chinese market is partially offset by surging demand in the US, European, and rest-of-world markets that the export control regime is not restricting. However, the strategic cost of losing China as a development partner is less easily offset. China’s AI developer community was one of the most sophisticated and productive contributors to the CUDA-based AI software ecosystem. Chinese researchers at companies like Alibaba, Baidu, ByteDance, and Tencent, as well as at leading Chinese universities, produced a significant share of the optimisation work, open-source libraries, and model architectures that benefited the entire CUDA ecosystem. As that community migrates to Ascend hardware, its future contributions will flow into the Ascend ecosystem rather than the CUDA ecosystem. The long-term effect is a CUDA software ecosystem that is smaller, less diverse, and less innovative than it would have been without the bifurcation.
Nvidia’s Response Strategy
Nvidia’s strategic response to the bifurcation has three components. The first is accelerating its product roadmap in markets it can still access, maintaining the performance leadership that makes switching costs prohibitive for Western AI developers regardless of what alternatives exist. The second is deepening its software ecosystem investments, expanding the CUDA library ecosystem and developer tooling to increase the switching costs that protect its market position. The third is lobbying the US government to shape export control policy in ways that preserve at least some commercial access to restricted markets without undermining the security objectives behind the controls.
None of these responses can fully compensate for the strategic loss of the Chinese market. The bifurcation has permanently reduced the total addressable market for Nvidia’s highest-performance products, created a well-funded and motivated competitor in Huawei with a captive market of the world’s largest AI companies, and initiated a process of CUDA ecosystem erosion that will compound over time as Chinese developers invest in Ascend-native tooling. Nvidia’s current dominance in the markets it can serve is genuine and durable on a multi-year horizon. The bifurcation dynamic that is reshaping the global AI infrastructure market is, however, structurally disadvantageous for any company whose competitive position depends on global ecosystem dominance.
As covered in our analysis of the time-to-power crisis as AI’s hidden scaling ceiling, the physical constraints on AI infrastructure development are compounding in ways that interact with the geopolitical constraints the bifurcation is creating. An industry navigating both simultaneously faces a more complex operating environment than either constraint would produce alone.
The Financing and Investment Implications
The bifurcation of global AI infrastructure creates financing challenges that have not yet been fully incorporated into how infrastructure investment is evaluated and structured. A data center fund that invested in facilities across multiple geographies before the export control era assumed that the hardware, software, and cloud relationships underpinning those facilities were globally compatible. A facility in Singapore built around Nvidia hardware and AWS cloud connectivity could serve customers in the US, Europe, and China without fundamental architectural changes. That assumption is no longer valid, and the portfolio risk created by its failure has not yet been systematically priced into data center asset valuations.
The most exposed assets are those in third countries with significant customer bases on both sides of the bifurcation — facilities in Singapore, Hong Kong, Japan, and the UAE that built their business models around serving both Western and Chinese enterprise customers from a single infrastructure platform. Those facilities are now navigating a choice between maintaining a single hardware and software stack that serves one ecosystem fully and the other inadequately, or investing in a dual-stack architecture that serves both ecosystems adequately but at substantially higher capital and operational cost. Neither option was anticipated in the original investment thesis, and neither is reflected in current asset valuations. The infrastructure investment community is repricing geopolitical risk across many asset classes simultaneously. AI data center assets in third countries with bifurcation exposure will be among the most significantly affected.
The Debt Market Response
Debt markets that finance AI data center construction are incorporating bifurcation risk into their underwriting frameworks more slowly than equity markets, in part because the impact of bifurcation on facility revenues and costs is difficult to model on the timelines that debt underwriting requires. A lender underwriting a 15-year construction loan for a Singapore data center campus needs to assess the probability that the facility’s customer base remains commercially viable across a hardware and geopolitical landscape that neither the lender nor the borrower can forecast with confidence over that horizon. The standard approach to this kind of uncertainty in project finance — requiring conservative revenue assumptions, maintaining covenant structures that trigger review if key tenants exit, and sizing debt at levels that can be serviced even under stress scenarios — is being applied to bifurcation risk as to other risk categories.
What makes bifurcation risk distinctive is that it is not symmetrically distributed across markets. A data center in Northern Virginia faces minimal bifurcation risk because its customer base is almost entirely US-aligned and its hardware infrastructure is built on technology that the US export control regime is designed to protect rather than restrict. A data center in Singapore or the UAE faces bifurcation risk that is material and growing because its business model depends on serving a customer base that spans both ecosystems. Lenders are beginning to differentiate their pricing and covenant structures accordingly, charging risk premiums for third-country exposure that were not part of the pricing framework two years ago. That differentiation will become more pronounced as the bifurcation matures and its revenue implications become more visible in the operating performance of exposed assets.
What Bifurcation Means for Global AI Development
The most consequential long-term implication of AI infrastructure bifurcation is not the competitive relationship between US and Chinese AI capabilities. It is the structural effect on the pace and direction of global AI development when the world’s two largest AI research communities are working on fundamentally different hardware stacks with decreasing cross-compatibility.
AI research has historically benefited from a global commons of shared infrastructure, open-source frameworks, and freely circulating research that allowed discoveries made anywhere in the ecosystem to propagate rapidly to practitioners everywhere. That model of open, globally integrated AI research is being constrained by the bifurcation dynamic. Chinese researchers working on Ascend-optimised models are producing results that cannot be directly reproduced on Western hardware without significant porting effort. Western researchers publishing results achieved on Blackwell clusters are producing work that Chinese developers cannot replicate at the same scale due to hardware restrictions. The cumulative effect of this progressive decoupling is a global AI research community that is less integrated, less mutually reinforcing, and likely less collectively productive than the unified ecosystem that preceded the export control era.
The Infrastructure Investment Implications
For infrastructure investors, operators, and enterprise AI buyers making long-term commitments, the bifurcation creates a new category of risk that existing frameworks do not adequately capture. A data center operator in a third country that builds its infrastructure around Nvidia hardware and Western cloud platforms is making an implicit bet that the US-aligned ecosystem remains the dominant global platform for the duration of the facility’s operating life. A data center operator that builds around Huawei Ascend and Chinese cloud platforms is making the same bet in the other direction. Both bets carry geopolitical risk that no amount of technical due diligence can fully hedge.
The operators who are most exposed are those whose customer base spans both ecosystems — enterprises with significant operations in both the US and China, multinational corporations navigating data sovereignty requirements in multiple jurisdictions, and cloud providers attempting to serve globally distributed enterprise customers from infrastructure that is increasingly subject to jurisdiction-specific hardware restrictions. As covered in our analysis of the hyperscaler consolidation of AI infrastructure, the competitive dynamics of global AI infrastructure investment are concentrating capacity in locations that combine power access, regulatory stability, and operational security. The bifurcation adds a fourth determinant — ecosystem alignment — that will be visible in investment patterns for years.
The unified global AI infrastructure market that existed three years ago is not coming back. The question now is how wide the gap between the two ecosystems grows, how many countries choose a side, and whether governments or companies develop technical and diplomatic mechanisms to manage the consequences of a split that nobody deliberately designed but nobody appears capable of stopping.
