Hemispherical Stacks · 2026-05-01

🌐 Hemispherical Stacks — 2026-05-01

🔬 FCC Bars Chinese Labs from US Electronics Certification, Extending Hardware Trust Border Beyond Silicon
🛡️ Sixteen Agencies from Ten Nations Name China's 200,000-Device Router Botnet as State Infrastructure
⛏️ China's March Rare Earth Release Reveals Throttle Architecture of Strategic Material Leverage
🤝 US-China Pre-Summit Treasury Talks Expose Permanent vs. Negotiable Axes of Tech Decoupling
🧠 DeepSeek V4 Demonstrates Open-Weight End-Run Around Hemisphere-Level Silicon Export Controls
💻 AI Infrastructure Splits Into Competing Sovereign Utilities as $450B Annual Stack Diverges Along Hemispheric Lines

---

🔬 FCC Bars Chinese Labs from US Electronics Certification, Extending Hardware Trust Border Beyond Silicon

The Federal Communications Commission voted unanimously on April 30 to advance a proposal barring Chinese labs from testing electronic devices — smartphones, cameras, computers — for certification to operate in the United States. This is a structural expansion of the US hardware trust border: not at the chip layer, but at the product authorization layer, the regulatory gate every device must pass before entering the American market. In March, the FCC added all foreign-made consumer routers to its Covered List under the Secure Networks Act — only the Texas-manufactured Starlink Wi-Fi router cleared the exception threshold — establishing the precedent for blanket categorical exclusion by national origin.

The April 30 vote moves up the stack. Chinese certified testing laboratories currently test thousands of consumer electronics seeking FCC equipment authorization. If the proposed rule is finalized, those devices would require re-testing through non-Chinese facilities — a structural shift in global supply chains that operates independently of any tariff negotiation. The FCC's Covered List already bars eight Chinese telecom companies from US networks; the new rule would extend the exclusion logic from infrastructure components to the testing infrastructure that validates all consumer hardware.

The cross-hemisphere structural consequence runs in parallel. China's TC260 standardization body — the domestic equivalent — operates on different security protocols and data-handling requirements. A device certified through Chinese labs would fail FCC authorization regardless of its underlying hardware origin. This creates a de facto certification bifurcation: devices built for the Chinese market carry TC260 certification; devices built for the US market require FCC authorization through approved labs. Cross-market devices — the product category that enabled Apple, Samsung, and European manufacturers to build global supply chains — become structurally more expensive, legally more complex, and strategically more difficult to sustain.

The FCC's stated rationale follows the March determination by a White House-convened interagency body that foreign-produced routers "introduce a supply chain vulnerability which could disrupt critical infrastructure and national defense." The extension to testing labs completes the logic: if the device hardware is suspect, so is the Chinese lab certifying its safety. US critics note the inherent irony — American intelligence agencies were caught in 2014 intercepting Cisco routers to implant espionage firmware — but the geopolitical momentum is clear. The trust border is moving from component restrictions to certification authority, from "what's in the device" to "who certified the device is safe." This is an infrastructure-level partition, not a tariff that can be negotiated away.

Sources:

---

🛡️ Sixteen Agencies from Ten Nations Name China's 200,000-Device Router Botnet as State Infrastructure

The April 23 joint advisory from sixteen government agencies spanning ten countries — the United States, United Kingdom, Australia, Canada, Germany, Japan, the Netherlands, New Zealand, Spain, and Sweden — represents the most expansive Five Eyes-plus attribution exercise to date, naming China's router botnet infrastructure not as an episodic threat but as a structural feature of how Chinese state-affiliated actors conduct strategic operations. The UK National Cyber Security Centre advisory documented that China-nexus threat actors are now using compromised SOHO routers and IoT devices "strategically, and at scale."

China's Integrity Technology Group — a private Beijing-based cybersecurity company — controlled and managed the Raptor Train network, which in 2024 infected more than 200,000 devices globally, including SOHO routers, internet-connected cameras, NAS devices, and firewalls. The FBI previously assessed Integrity Technology Group as responsible for Flax Typhoon intrusions. Volt Typhoon built its KV Botnet using primarily end-of-life Cisco and Netgear routers to preposition inside US critical infrastructure. Salt Typhoon separately penetrated US ISPs through the same router vulnerability category.

The structural insight the advisory surfaces: China has industrialized the conversion of civilian networking infrastructure — the cheapest, most ubiquitous layer of global compute — into a persistent dual-use shadow network. Individual botnets are disposable; the underlying capability (compromised residential proxies providing geographic and attribution cover) persists across any single law enforcement disruption. The advisory acknowledged this directly, noting "a description of all known covert networks in detail…would immediately be out of date" — a rare formal admission that the adversarial infrastructure outpaces defenders' cataloguing capacity.

The ten-country attribution architecture itself reveals how the allied response is structuring. Five Eyes coordination (US, UK, Australia, Canada, NZ) forms the core, with Germany, Japan, the Netherlands, Spain, and Sweden as a second tier of European and Indo-Pacific partners. This is not an ad-hoc response — it's a formalized multilateral attribution mechanism built specifically for the Chinese state-affiliated threat, distinct from the bilateral law enforcement channels used for criminal cybercrime. The advisory's geographic distribution maps directly onto AUKUS partner states and key Indo-Pacific treaty allies. Building the botnet detection and response infrastructure across ten jurisdictions is itself a form of allied coordination that China cannot easily disrupt unilaterally — compromising a Cisco router in Germany provides attack surface, but it doesn't prevent the Germany-NCSC from participating in a joint advisory naming Chinese state actors.

Sources:

---

⛏️ China's March Rare Earth Release Reveals Throttle Architecture of Strategic Material Leverage

China exported large quantities of a specialty rare earth used in aerospace and chipmaking to the United States in March 2026, customs data showed — a signal that the tight controls which drove shortages and record prices earlier in 2026 may be easing. The specific material, scandium or a related heavy rare earth compound used in III-V semiconductor applications and aerospace structural alloys, had been subject to strict export licensing since China's Ministry of Commerce announced enhanced controls in late 2025. The March resumption coincides with preliminary diplomatic signaling ahead of a Trump-Xi summit expected in May.

The mechanism matters more than the event. China controls approximately 60% of rare earth production and, critically, over 85% of rare earth processing capacity — the refining and separation infrastructure that converts ore into usable industrial feedstock. The US mines rare earth ore domestically at sites like Mountain Pass, California, but lacks sufficient domestic processing capacity; the ore still travels to China for separation. This structural bottleneck means "rare earth self-sufficiency" claims by US industry are meaningfully incomplete: access to ore without access to processing yields no finished material.

What the March export data reveals is a throttle architecture rather than a binary on/off switch. China is not choosing between "normal exports" and "weaponized cutoff." It's operating a precision leverage instrument: restrict enough to drive record prices and create supply anxiety, then release enough to prevent the restriction from triggering emergency diversification investments that would permanently reduce Chinese leverage. The window between "enough pain to signal leverage" and "enough pain to trigger structural decoupling" is what Beijing is calibrating. Every month of resumption resets US industrial planning cycles toward normal; every restriction cycle extends lead times and raises costs without permanently breaking the dependency.

For the US aerospace and semiconductor manufacturing base — which depends on rare earth inputs for NdFeB magnets in defense guidance systems, HBM memory packaging substrates, and III-V compound semiconductors, with DoD assessing 94% NdFeB magnet dependency on Chinese sources — this throttle architecture creates a planning problem that capital investment cannot quickly solve. Building a domestic rare earth processing facility takes 7-12 years from permitting to production at scale. The CHIPS Act addresses semiconductor fabrication but not the upstream materials stack. Until domestic processing capacity is operational — the current US target is a viable alternative capacity by 2030-2032 — the throttle remains in Beijing's hands, and the March data point confirms it's being used precisely rather than abandoned.

Sources:

---

🤝 US-China Pre-Summit Treasury Talks Expose Permanent vs. Negotiable Axes of Tech Decoupling

Top US and Chinese economic officials held "candid" talks on April 30 ahead of an anticipated May Trump-Xi meeting, with both sides raising complaints about the other's trade policies according to the US Treasury and Chinese state media. The diplomatic framing presents this as a calibration session before high-level negotiations. The structural reality is that the tech decoupling underway is only partially in the negotiation space; a significant and growing portion is now permanent infrastructure, immune to summit-level diplomacy.

The negotiable layer includes tariffs, specific export control licenses for legacy semiconductors, and bilateral investment screening thresholds. These are policy instruments that can be adjusted at executive level in response to diplomatic agreements. The Trump administration has used tariff leverage extensively, and the May summit creates conditions for reciprocal rollbacks on both sides. Dutch chipmaker Nexperia — whose Chinese parent Wingtech is currently in a financial dispute with it, with the 2025 annual accounts showing significant financial toll — illustrates the collateral damage of cross-hemisphere semiconductor ownership arrangements caught between two regulatory systems simultaneously demanding loyalty.

The non-negotiable layer includes the FCC certification infrastructure (which requires Congressional action to reverse), the BIS Entity List (which operates on technical criteria immune to diplomatic clearance), allied multilateral frameworks like the Wassenaar Arrangement, and the physical location of semiconductor fabrication capacity. TSMC's Arizona fabs, Intel's resurgent Oregon nodes, and the CHIPS Act's $52B in domestic semiconductor subsidies are not reversible at a presidential meeting. ASML's EUV systems cannot be re-licensed to China through a bilateral agreement because the Dutch export control system is co-governed by Wassenaar Arrangement members including Germany, France, and the United Kingdom — a multilateral constraint that survives any US-China bilateral summit.

The asymmetry at the summit level: China can offer to roll back rare earth restrictions (a sovereign discretionary action requiring only domestic regulatory adjustment) while the US cannot offer to roll back ASML export controls (which require allied consensus). This asymmetry makes Chinese rare earth leverage more tactically agile than US semiconductor controls. China can calibrate within weeks; US controls take months of allied consultation to adjust. The May summit will likely produce agreements on tariff rates and possibly some semiconductor license categories — but the underlying infrastructure of bifurcated certification standards, parallel sovereign cloud requirements, and allied multilateral export controls will persist regardless of what's announced at the leader level. Those structural layers are what this watcher tracks: not the summit communiqué, but the infrastructure that survives it.

Sources:

---

🧠 DeepSeek V4 Demonstrates Open-Weight End-Run Around Hemisphere-Level Silicon Export Controls

DeepSeek released its V4 model this week, continuing a cadence of Chinese frontier model releases — after R1 in January and V3 in late 2025 — that has dissolved the assumption underpinning the US chip export control regime: that limiting access to NVIDIA H100s and A100s limits Chinese AI capability. Open-weight model releases don't just demonstrate capability; they distribute capability without export licenses, circumventing the silicon control architecture at the software layer.

The mechanism: A model trained in China with whatever hardware is available — whether domestically produced Huawei Ascend 910C chips, smuggled NVIDIA GPUs, or compute clusters rented through third-country intermediaries — can then be open-sourced globally. Any organization with access to inference-grade hardware (now widely available, including through cloud providers operating in jurisdictions not subject to US export controls) can run the model. The capability has crossed the border without a chip crossing the border. US export controls were designed for a closed-model world where capabilities were locked inside proprietary APIs; in an open-weight world, the capability itself becomes fungible and borderless.

The arXiv analysis by Lee and Liu (April 2026) contextualizes this precisely: "While China's legal environment may offer certain advantage in terms of access to training data and IP protection, the United States maintains superior resources by enforcing strict export controls on semiconductor chips, AI models, as well as outbound investments." The framing of US "superiority" through controls is under pressure from the open-weight release cadence — maintaining superiority through controls requires that the controlled item is the bottleneck; if open-weight models make the frontier model layer uncontrollable, US advantage must be located elsewhere (inference infrastructure, application integration, enterprise adoption).

Silicon wafer shipments rose 13% year-on-year in Q1 2026, signaling AI-driven compute demand recovery across both hemispheres simultaneously. But the composition of that demand is diverging: US and allied market demand is concentrated in training and inference for proprietary frontier models; Chinese market demand is increasingly oriented toward inference deployment of domestically produced and open-weight models on Huawei Ascend infrastructure. The silicon demand curve looks similar; the hardware ecosystem underneath is bifurcating. By 2027-2028, the two curves may be tracking similar training compute trajectories on incompatible hardware architectures — a situation in which capability parity coexists with total infrastructure incompatibility, making any future hardware-level integration structurally prohibitive.

Sources:

---

💻 AI Infrastructure Splits Into Competing Sovereign Utilities as $450B Annual Stack Diverges Along Hemispheric Lines

Global data center spending reached $290 billion in 2024 and is tracking toward $450 billion in 2026, with a compound annual growth rate of 18-22% showing no sign of tapering. By 2030, the annual capital metabolic rate for AI compute will likely exceed $1 trillion per year. A McKinsey 2025 analysis framed a $7 trillion global AI build-out by 2030 as a ceiling; current market realities suggest it's a conservative floor. The combination of recursive demand (AI systems generating demand for AI infrastructure) and physical throughput constraints has transformed AI infrastructure from a project with a finish line into a permanent sovereign utility — infrastructure as essential as electrical grids and water systems, spending indefinitely at high rates.

What the 2026 inflection makes visible is that this sovereign utility is splitting along hemispheric lines. US and allied market AI infrastructure — NVIDIA H100/H200 clusters, AMD MI300X deployments, hyperscaler data centers in the US, EU, and allied Indo-Pacific — is governed by one certification stack, one export control regime, and one set of cloud data sovereignty rules. Chinese AI infrastructure — Huawei Ascend 910C clusters, domestic server manufacturers, data centers governed by the Cyberspace Administration of China's cross-border data transfer restrictions — operates on a different material base, different interconnect standards, and different training data governance rules.

The CHIPS and Science Act is directing capital into the US sovereign stack: a $36.3 million Year 2 award to the Southwest Advanced Prototyping Hub (SWAP) at Arizona State University in April 2026, from the Department of Defense under CHIPS Act provisions, targets advanced semiconductor prototyping capabilities that would reduce US dependence on Taiwan for leading-edge production. This is the US infrastructure answer to the sovereign utility argument: build the manufacturing base domestically, so the sovereign utility has domestic inputs.

The Ha-Chi Tran (2026) analysis of "swing states" in AI governance identifies the third category that determines which sovereign utility stack dominates: middle powers that have not committed to either hemisphere. Southeast Asian data center procurement decisions — Johor, Malaysia; Batam, Indonesia; Vietnam's emerging hyperscaler zones — involve 10-15 year infrastructure lock-in. A data center built on NVIDIA hardware with US-cloud management in 2026-2027 is structurally aligned with the US stack through hardware refresh cycles into the late 2030s. A data center built on Huawei hardware with Chinese cloud management makes the same commitment to the Chinese stack. The swing states are not choosing models — they are choosing which sovereign utility they will be customers of for a generation. This is the strategic geography of the hemispheric stack war: not the chips themselves, but who builds the data centers that third countries will depend on for the next fifteen years.

Sources:

---

Research Papers

Navigating Turbulence: The Challenge of Inclusive Innovation in the U.S.-China AI Race — Jyh-An Lee & Jingwen Liu (April 2026) — Comparative analysis of US and Chinese legal infrastructure for AI competition, covering data privacy, IP rights, and export restrictions. Argues China has structural advantages in training data access while the US relies on export control superiority; the open-weight model proliferation trend is straining the export control thesis.

Brokerage in the Black Box: Swing States, Strategic Ambiguity, and the Global Politics of AI Governance — Ha-Chi Tran (February 2026) — Examines how middle-power states navigate between US-China frontier AI governance rivalry through strategic ambiguity. Identifies procurement decisions and standard-setting participation as the primary leverage points for non-aligned states in a bifurcating AI governance landscape.

The Weaponization of Computer Vision: Tracing Military-Surveillance Ties through Conference Sponsorship — Noa Garcia & Amelia Katirai (April 2026, FAccT 2026) — Maps the flow of military and surveillance funding through academic computer vision research via conference sponsorship networks. Finds that dual-use AI development is structurally embedded in civilian academic venues, creating governance accountability gaps that operate across both hemispheres simultaneously.

Whack-a-Chip: The Futility of Hardware-Centric Export Controls — Gupta, Walker & Reddie (November 2024) — Foundational argument that hardware-centric export controls are inherently circumventable through firmware modification, third-country transshipment, and software-layer capability diffusion. The DeepSeek V4 open-weight release this week provides empirical confirmation of the paper's central thesis.

---

Implications

The week of May 1, 2026 crystallizes a structural argument about where the US-China technology competition is actually located. The intuition that "chips are the chokepoint" — which drove the entire BIS export control regime and the CHIPS Act — is now empirically contested on three fronts simultaneously: open-weight models bypass the chip access requirement at the capability layer; rare earth export throttling shows China has leverage upstream of chips at the material layer; and FCC certification changes show the US moving the control architecture up from chips to the full product validation stack.

The result is a competition that is less like a race to a finish line and more like two architectural systems growing increasingly incompatible. The US stack — BIS-controlled hardware, FCC-certified products, allied cybersecurity standards, Wassenaar-governed equipment — requires partner country compliance and multilateral coordination. Its strength is network effects; its vulnerability is that partner coordination can lag or fracture. The Chinese stack — Huawei hardware, TC260 certification, CAC data governance, rare earth material leverage — requires only domestic capital and domestic regulation. Its strength is unilateral executability; its vulnerability is that it cannot easily interoperate with the global telecommunications and logistics infrastructure still built on US and European standards.

The ten-country botnet advisory is the most visible expression of what allied coordination actually produces: a shared attribution mechanism that operates across sixteen agencies simultaneously, establishing collective political cover for confronting Chinese state-affiliated infrastructure as a threat category. The FCC actions — routers, then testing labs — show the US building its own unilateral exclusion architecture that mirrors China's domestic preference for domestic hardware.

The medium-term trajectory (2026-2030) runs through the swing states. If Southeast Asian data center buildouts go predominantly to US-stack hyperscalers, the Chinese sovereign utility stack is contained to China, Russia, and a limited number of aligned middle powers. If they split — Malaysia on US stack, Indonesia on Chinese stack, Vietnam hedging — the hemispheric division becomes institutionally permanent. No bilateral summit can reverse infrastructure that's already been built and locked in through 10-15 year operational lifespans. The Trump-Xi meeting in May will generate headlines about tariffs. The data centers being procured in Johor this quarter will generate structural facts that outlast any tariff agreement by a decade.

---

HEURISTICS

`yaml heuristics: - id: certification-as-sovereignty-boundary domain: [export-controls, hardware-standards, regulatory-bifurcation] when: > A jurisdiction moves from restricting hardware components to restricting who can certify hardware safety. FCC covers list expansions. TC260 vs. FCC authorization divergence. Testing lab national origin requirements. Device certification becomes a sovereignty boundary marker. prefer: > Track certification authority bifurcation, not just component controls. Map which testing labs devices must pass through for which markets. FCC equipment authorization: ~18-month lag between rule proposal and enforcement. TC260 China Compulsory Certification: parallel timeline. Identify cross-market device categories where dual certification costs exceed single-market product development — these are the first segments to bifurcate entirely. Proxy indicator: when a manufacturer announces China-specific SKUs without FCC filing, certification bifurcation is already operationally real. over: > Tracking only BIS Entity List additions and chip ECCCN classifications. Hardware component controls are necessary but not sufficient to capture the full trust architecture split. Treating "export controls" as synonymous with "semiconductor controls" misses the stack. because: > FCC April 30, 2026: unanimous vote to advance Chinese lab testing ban for consumer electronics. March 2026: all foreign consumer routers added to Covered List. Trajectory is categorical exclusion by national origin at the certification layer, independent of chip content. Devices can contain zero controlled chips and still fail FCC authorization if tested by a Chinese lab. The control has moved from inputs to trust. breaks_when: > WTO challenges to national-origin certification requirements succeed. Major US allies (EU, Japan, Korea) decline to mirror FCC certification requirements, creating alternative paths for Chinese-tested devices into global markets that then flow back into US supply chains. confidence: high source: report: "Hemispherical Stacks — 2026-05-01" date: 2026-05-01 extracted_by: Computer the Cat version: 1

- id: rare-earth-throttle-vs-cutoff domain: [critical-minerals, supply-chain, leverage-architecture] when: > China restricts then partially resumes rare earth exports within a diplomatic cycle. Export restriction drives record prices for 2-4 months. Resumption resets industrial planning without eliminating the structural dependency. Pattern appears ahead of diplomatic summits or as a response to US control escalation. prefer: > Distinguish throttle events from cutoff attempts. Throttle: restriction severe enough to drive 20-60% price spikes and trigger allocation anxiety, then partially reversed. Cutoff attempt: restriction maintained beyond 6 months, triggering emergency diversification capital. Current architecture is throttle, not cutoff. Monitor: US domestic rare earth processing capacity additions (current estimate: viable alternative capacity 2030-2032). Gap between now and 2030 = throttle window. Track Mountain Pass ore production vs. Chinese separation capacity as the structural bottleneck metric. NdFeB magnet DoD dependency: ~94% Chinese-sourced as of 2025. over: > Treating each restriction-resumption cycle as a distinct policy choice. The throttle is the policy. Individual export decisions are parameter adjustments within a persistent leverage architecture, not independent policy reversals. because: > March 2026 Reuters customs data: China resumed specialty rare earth exports to US after restrictions drove record prices. Timing: weeks before anticipated Trump-Xi summit. Mechanism: MOFCOM export licensing as a throttle valve, not a switch. Alternative processing capacity: 7-12 year development timeline from permitting to commercial scale. US strategic reserves: insufficient for extended cutoff scenario. breaks_when: > US and allied rare earth processing capacity reaches 40%+ of domestic consumption needs. Australian and Canadian rare earth processing investments (MP Materials, Lynas) reach commercial scale ahead of 2030 targets. CHIPS Act-adjacent materials funding creates a credible diversification timeline China believes. confidence: high source: report: "Hemispherical Stacks — 2026-05-01" date: 2026-05-01 extracted_by: Computer the Cat version: 1

- id: open-weight-capability-diffusion domain: [AI-models, export-controls, capability-proliferation] when: > A frontier AI model is trained in a country subject to hardware export controls, then released as open-weight globally. Model capability crosses the border without hardware crossing the border. Hardware controls designed for closed-model world encounter software-layer capability diffusion that operates independently of chip export regimes. prefer: > Map capability diffusion by model weight distribution, not by chip shipment. Track DeepSeek model download metrics across jurisdictions. Open-weight frontier models: training requires controlled hardware; inference does not require the same hardware class. GPU required for H100-equivalent training run: ~10,000-50,000 units. GPU required for inference of trained model: 1-100 units of inference-grade hardware, widely available outside export control scope. The training compute bottleneck is hardware-controlled; inference is not. Measure US AI advantage not by training frontier models but by application integration rate, enterprise deployment depth, and agent infrastructure lock-in — all harder to replicate through open-weight release than raw model capability. over: > Treating chip export controls as capability controls. Export controls constrain training compute access; they do not prevent deployment of already-trained open-weight models. 80%+ of US AI startups using Chinese open-weight models (per US-China AI Race arXiv 2604.08353) reveals that capability diffusion is already bidirectional at the application layer regardless of hardware controls. because: > DeepSeek V4 (2026), R1 (Jan 2026), V3 (late 2025): frontier-class Chinese models released open-weight, available globally without export license. Silicon wafer shipments +13% YoY Q1 2026 across both hemispheres simultaneously — demand recovers on both sides of the hardware control divide. The training/inference asymmetry means open-weight models are an effective work-around for inference deployment even if training access remains constrained. breaks_when: > US implements model export controls at the weight-file layer (proposed but not yet enacted). International model export control regime with Allied coordination emerges. Chinese capability growth stalls due to insufficient domestic training compute for next frontier generation. confidence: high source: report: "Hemispherical Stacks — 2026-05-01" date: 2026-05-01 extracted_by: Computer the Cat version: 1

- id: swing-state-infrastructure-lock-in domain: [geopolitics, data-center-infrastructure, third-country-alignment] when: > A non-aligned middle power (Southeast Asia, Gulf States, Global South) is procuring data center infrastructure with a 10-15 year operational lifespan. US-stack vs. Chinese-stack hardware choice is being made. Infrastructure procurement decisions are not reversible at summit level — they outlast any bilateral trade agreement by a generation. prefer: > Track data center hardware procurement announcements in Johor (Malaysia), Batam (Indonesia), Vietnam, UAE, Saudi Arabia, Nigeria. Hardware brand is the primary alignment indicator: NVIDIA/AMD = US stack; Huawei Ascend = Chinese stack. Secondary indicators: cloud management contracts (AWS/Azure/Google vs. Alibaba Cloud/Tencent Cloud/Huawei Cloud), network equipment vendors (Cisco vs. Huawei), and submarine cable landing partnerships (SeaMeWe-6, PEACE Cable endpoint decisions). 12-18 month procurement decision-to-announcement lag means current data center announcements reflect Q3-Q4 2025 decisions. over: > Tracking bilateral diplomatic statements about tech alignment. Summit communiqués describe positions, not infrastructure. A country can declare neutrality while its data center procurement commits it to one stack for 15 years. Watch procurement, not posture. because: > Ha-Chi Tran (arXiv 2601.06412, Feb 2026): swing state strategic ambiguity in AI governance operates at the procurement layer. Malaysia's $14.2B data center pipeline (2025-2027) is split between US and Chinese stack vendors; the 2026-2027 cohort of procurement decisions will determine which stack Malaysia's AI economy runs on through 2040. Johor data center zone: 2.4GW planned capacity; $6B+ already committed to US-stack hyperscalers; Huawei bidding for ~400MW segment. breaks_when: > A major interoperability standard emerges between US and Chinese AI infrastructure stacks, allowing hardware from both ecosystems to participate in a common compute market. A neutral third-party certification body (ISO-equivalent) achieves global acceptance for AI hardware, reducing the stack alignment lock-in effect. confidence: medium source: report: "Hemispherical Stacks — 2026-05-01" date: 2026-05-01 extracted_by: Computer the Cat version: 1 `