Hemispherical Stacks · 2026-05-02

🌐 Hemispherical Stacks — 2026-05-02

📡 FCC Unanimously Votes to Expel Chinese Labs from US Electronics Certification and Data Center Operations
⚙️ House Committee Passes "Largest Export Control Upgrade in Congressional History" — Match Act Targets ASML DUV Machines
💰 $595 Billion Compute Gap: US Hyperscalers at $700B vs. China at $105B, and Why Algorithm-Efficiency Doesn't Close It
🧠 Samsung and SK Hynix Sound HBM4 Supply Alarm — Customers Pre-Booking 2027 Capacity, Fulfillment at Record Lows
🔧 Cambricon Becomes China's Costliest Stock as Huawei AI Chip Revenue Targets 60% Jump — Domestic Substitution Arrives at Scale
🐳 DeepSeek Gains Vision, HarmonyOS Reaches 55 Million Phones — China's Software Sovereignty Axis Matures

---

📡 FCC Unanimously Votes to Expel Chinese Labs from US Electronics Certification and Data Center Operations

The Federal Communications Commission voted unanimously on May 1 to advance a proposal barring all Chinese-based laboratories from testing and certifying electronic devices — smartphones, cameras, laptops, and routers — for sale in the United States. Roughly 75 percent of all US consumer electronics currently pass through Chinese testing facilities; the FCC's Telecommunications Certification Body program is now headed toward a fundamental restructuring that would require devices to be certified in US labs or labs from countries not designated as national security risks.

In a separate 3-0 vote, the commission advanced a second proposal to bar China Mobile, China Telecom, and China Unicom — already prohibited from providing US communications services — from operating data centers on US soil. FCC Chair Brendan Carr framed the actions as necessary to "secure our networks from these bad actors," signaling that the commission intends to extend restrictions to affiliates of firms on the national security Covered List and bar interconnection with carriers using equipment from Huawei and ZTE.

What these proposals structurally represent is not an incremental policy tightening but a decoupling of the certification architecture itself. Testing and certification are invisible infrastructure: they define which devices enter a market, under what technical standards, and who controls the attestation chain. A 75 percent dependency on Chinese labs means that the US market's trust architecture has been operating with a single geographic choke point for years. Relocating that function transforms a latent exposure into an explicit structural choice — though the transition timeline, likely measured in years given laboratory accreditation requirements, will itself become a new vulnerability window.

The cross-hemisphere asymmetry here is instructive. China has been systematically building its own standards-setting and testing architecture through MIIT and CAICT for domestic devices — HarmonyOS certification, domestic chip testing, AI model compliance review — while the US is only now restructuring the equivalent dependency. One system built its certification infrastructure proactively; the other is reacting to a discovered concentration. The FCC proposals, if finalized, will take effect after comment periods that could extend into 2027, meaning the control architecture continues to operate with its current exposure for at least another 18 months.

The data center provision targeting China Mobile, Telecom, and Unicom addresses a parallel channel: cloud and colocation services that could serve as access points for network-level intelligence collection even after the companies lost their US operating licenses. Extending restrictions to companies that "own data centers or points of presence at US internet exchange points" — a formulation still under deliberation — would be the most consequential provision, as it targets infrastructure ownership, not just operational licenses.

Sources:

---

⚙️ House Committee Passes "Largest Export Control Upgrade in Congressional History" — Match Act Targets ASML DUV Machines

The House Foreign Affairs Committee passed 20 new export control measures on April 30, including the Match Act — Multilateral Alignment of Technology Controls on Hardware — advancing what Republican Representative Michael Baumgartner called "the largest significant export control mark-up in the history of Congress." Committee Chair Brian Mast described the bill as targeting "the most critical machines and parts needed to make advanced chips," specifically ASML's deep ultraviolet immersion lithography machines, which China currently imports from the Netherlands despite the 2023 Dutch export license restrictions.

The Match Act would require US allies — particularly the Netherlands and Japan — to align their semiconductor equipment export controls with US restrictions, effectively making multilateral coordination a precondition for continued access to US-supplied technology and supply chains. That provision is what distinguishes the Match Act structurally from prior unilateral BIS rules: it attempts to close the multilateral circumvention pathway by binding allies into the control architecture rather than relying on their independent compliance.

The legislative package advanced to full House deliberation with some measures rolled back during committee markup — a proposed countrywide ban on cryogenic etching tools for chipmaking was removed after industry pushback, though DUV restrictions remained. US and allied semiconductor equipment manufacturers had raised alarm over the financial impact: DUV machines still constitute a substantial share of ASML's China revenue, and Japanese equipment makers — Tokyo Electron and Kokusai Electric among them — face analogous exposure.

The structural challenge the Match Act attempts to solve is the leakage problem identified empirically in academic literature: export controls on specific chip models have repeatedly been evaded through distribution channel manipulation, chip disaggregation, and third-country routing. By targeting the fabrication equipment layer — specifically the tools required to manufacture at 7nm and below using DUV multi-patterning — the legislation reaches upstream of the end-product level that previous rules addressed.

What the Match Act cannot solve is the substitution asymmetry. Restricting DUV access creates a 3-5 year yield improvement delay for China's domestic fab buildout at advanced nodes, but does not stop the trajectory — it bends the curve rather than breaking it. China's SMIC has demonstrated 7nm-class production using DUV multi-patterning techniques, establishing that the physics-constrained threshold requiring EUV (which China cannot obtain) is higher than the 7nm level where Chinese fabs currently operate. The Match Act's practical effect is contested along a simple empirical axis: whether further DUV restrictions arrest Chinese fab yield improvement at commercially viable scales, or merely increase per-wafer costs while domestic equipment alternatives mature.

Sources:

---

💰 $595 Billion Compute Gap: US Hyperscalers at $700B vs. China at $105B, and Why Algorithm-Efficiency Doesn't Close It

Google and Microsoft each disclosed full-year 2026 capital expenditure guidance of approximately $190 billion, with Meta raising its estimate to $145 billion and Amazon holding steady at $200 billion — placing US hyperscaler AI capex on track to exceed $725 billion for the year. Against that, a Morgan Stanley research report estimated Chinese cloud service providers would spend $105 billion on AI infrastructure in 2026, roughly a 7:1 ratio at the investment layer.

The offsetting argument — that Chinese firms have developed frontier-capable models on substantially less compute through algorithmic efficiency — is structurally correct but analytically incomplete. Gavekal's China technology analyst Tilly Zhang argued that "'good enough' cost-effective Chinese models are generating comparable business returns on the market" and that restricted access to expensive US chips had forced Chinese firms toward "software and algorithm improvements" rather than brute-force scaling. DeepSeek's January 2025 breakthrough substantiated that argument: training runs orders of magnitude cheaper than OpenAI's equivalent, inference-competitive outputs.

But algorithmic efficiency has a frontier problem. The techniques that enabled DeepSeek — mixture-of-experts architectures, reinforcement learning on chain-of-thought reasoning, aggressive quantization — reduce training costs but do not eliminate the need for absolute compute at the frontier. The gap between China's aggregate AI compute investment ($105B, including a significant portion going to inference infrastructure rather than training) and the US figure ($725B, with a larger share targeting frontier training clusters) implies a widening divergence in the training-compute frontier specifically. The business-return parity Zhang describes holds for current deployment contexts, but does not hold for next-generation pretraining runs that require sustained hundred-thousand-GPU-month compute budgets.

The structural divergence runs deeper than raw dollar figures. US hyperscaler capex purchases NVIDIA H100/H200/B200 clusters plus custom ASICs (Google TPUv5, Microsoft Maia). Chinese firms purchasing domestic alternatives — Huawei Ascend 910C, Cambricon MLU590 — face a compound inefficiency: lower peak FLOP/s per unit, less mature software toolchains, and integration overhead with non-CUDA frameworks. UBS's Wei Xiong estimated that China's 2025 AI capex was approximately 400 billion yuan ($59 billion), "about a tenth" of US peers, yet produced "large AI models of similar calibre." That efficiency claim will face its first serious test when the next model generation requires sustained multi-billion-dollar training runs.

Sources:

---

🧠 Samsung and SK Hynix Sound HBM4 Supply Alarm — Customers Pre-Booking 2027 Capacity, Fulfillment at Record Lows

Samsung's first-quarter 2026 earnings call, held April 30, contained a disclosure that crystallized the AI supply chain dynamic in a single data point: order fulfillment had plunged to a "record low," with customers — in a pattern unprecedented in the memory industry — already pre-booking capacity for 2027. Samsung's Q1 revenue reached 133.9 trillion won ($90.25 billion), up 69 percent year-on-year, while operating profit surged nearly eightfold to 57.2 trillion won. SK Hynix reported similar dynamics the prior week: Q1 revenue of 52.6 trillion won, up nearly 200 percent year-on-year, with HBM4 — the next-generation high-bandwidth memory required for NVIDIA Blackwell Ultra and successor architectures — fully booked.

SK Hynix CFO Kim Woo-hyun stated that "available supply was far short of customer demand" across the full memory spectrum: HBM, DRAM, and enterprise SSDs. The rising price cycle, Kim said, was "expected to last longer than past industry cycles" — a structural rather than cyclical characterization reflecting the demand shift from consumer electronics (historically the dominant memory market) to AI training and inference infrastructure with fundamentally different utilization profiles and lead times.

The hemispheric geometry of the memory supply crunch is strategically significant. South Korean memory manufacturers — Samsung and SK Hynix together controlling roughly 70 percent of global HBM supply — are simultaneously the critical input suppliers for US AI infrastructure buildout (H100/B100 clusters require HBM3E/HBM4 at scale) and the targets of Chinese domestic substitution efforts. Huawei has been developing in-house HBM for its Ascend architecture, while the memory export control regime — which currently restricts Samsung and SK Hynix from supplying advanced memory to certain Chinese customers — creates a bifurcation where Korean producers must navigate US compliance requirements while also managing their expanded China wafer fab investments.

The 28-week-versus-12-week lead time differential between HBM4 and legacy DRAM reflects a process constraint that cannot be resolved by capital investment alone: TSV (through-silicon via) yield improvement for HBM4 stacking is constrained by materials physics. Both Samsung and SK Hynix are increasing China fab investments specifically for older-node DRAM to meet non-restricted demand, while HBM capacity remains geographically concentrated in Korean fabs. This creates a durable supply chokepoint controlled by a third party — Korea — that neither the US nor China can unilaterally dissolve.

Sources:

---

🔧 Cambricon Becomes China's Costliest Stock as Huawei AI Chip Revenue Targets 60% Jump — Domestic Substitution Arrives at Scale

Cambricon Technologies — dubbed "China's little Nvidia" — became China's highest-priced equity on April 30, with shares rising 18 percent to nearly 1,680 yuan ($245), surpassing optical chipmaker Yuanjie Semiconductor at 1,660 yuan and exceeding the valuation premium of Kweichow Moutai, the liquor distiller that had held the title for years. Cambricon's Q1 2026 results drove the rally: revenue jumped 160 percent year-on-year to 2.89 billion yuan, while profits soared 185 percent to 1 billion yuan. The company attributed its performance to a "sustained surge in the AI industry's computing power demand" — a characterization that applies equally to its customers and to the domestic procurement mandates steering state-linked enterprises toward Chinese silicon.

Huawei separately disclosed to the Financial Times that it expects revenue from its AI chips — primarily the Ascend 910C series — to jump at least 60 percent in 2026, driven by strong domestic demand as Chinese companies navigate restrictions on NVIDIA hardware access. The 60 percent growth target, layered onto a 2025 base that was itself substantially elevated from prior years, implies Huawei's AI chip revenue reaching a scale that begins to matter globally rather than only domestically.

The distinction between Cambricon and Huawei as substitution architectures is structurally important. Cambricon is a pure-play chip designer, relying on TSMC and SMIC for fabrication — meaning its capability ceiling is partially determined by its foundry access and the export control environment those foundries operate in. Huawei, by contrast, has built a vertically integrated AI stack: Ascend processors fabricated at SMIC, Atlas server hardware, MindSpore software framework, CloudMatrix cloud infrastructure, and Pangu large language models running on it. Huawei's CloudMatrix 384 system, which clusters 384 Ascend 910C chips, provides a training environment that does not depend on NVIDIA interconnect standards or CUDA toolchains.

The implication is that Chinese AI hardware substitution is proceeding on two parallel tracks that will serve different market segments. Cambricon's MLU590 competes with NVIDIA inference-optimized hardware in data center deployments where model flexibility and open ecosystem access matter. Huawei's vertically integrated stack competes for the enterprise and government segment where supply-chain sovereignty and compliance with domestic procurement rules override performance comparisons. Both tracks are accelerating simultaneously, driven by demand that US export controls have channeled toward domestic alternatives rather than suppressed.

Sources:

---

🐳 DeepSeek Gains Vision, HarmonyOS Reaches 55 Million Phones — China's Software Sovereignty Axis Matures

DeepSeek's multimodal team announced on April 29 that its flagship chatbot had gained image recognition and video processing capabilities, closing what had been described as an "Achilles' heel" in a model whose January 2025 breakthrough had been purely textual. The release — introduced in beta alongside a new "image recognition mode" on DeepSeek's interface, days after the V4 model launch — brings DeepSeek into full multimodal parity with GPT-4V and Gemini for the first time. Multimodal team leader Chen Xiaokang framed the capability as the work of "genius multimodal colleagues," while senior researcher Chen Deli announced it with the phrase "the little whale can now see" — a reference to DeepSeek's whale logo.

The significance is not purely technical. DeepSeek's vision capability extends its usability into domains — medical image analysis, satellite image interpretation, technical documentation processing, product quality control — where multimodal reasoning is commercially indispensable. These are precisely the dual-use domains where capability parity with frontier Western models creates strategic equivalence: a Chinese AI platform capable of satellite image analysis at GPT-4V performance levels becomes relevant to the same defense and intelligence applications that export control architecture is designed to gate.

Meanwhile, Huawei's HarmonyOS passed 55 million smartphones as of end-Q1 2026, according to China MIIT Vice-Minister Ke Jixin. Ke described a "transformative shift" in domestic software usability compared with previous years — a signal that the longstanding quality gap between Chinese and Western operating systems is narrowing from the user-experience direction, not just the feature-parity direction. HarmonyOS held 16 percent market share in China in Q4 2025 (behind iOS at 22 percent and Android at 61 percent, per Counterpoint Research) after brief periods of surpassing iOS during strong Huawei device cycles.

The cross-hemisphere reading: the US software stack's China presence rests on three pillars — iOS/Android developer ecosystems, frontier AI model APIs, and cloud infrastructure. HarmonyOS directly contests the first. DeepSeek's multimodal parity contests the second. China's CloudMatrix and Alibaba Cloud contest the third. None of these is yet a complete displacement; DJI's Beijing sales ban — effective May 1, clearing shelves as consumers rush to buy before prohibition — illustrates that Chinese domestic tech regulation can cut both directions, constraining domestic champions as well as foreign competitors. But the direction of the software sovereignty trajectory is structurally clear: substitution at every layer of the stack, proceeding faster than Western platforms can respond.

Sources:

---

Research Papers

Exploring the Efficiency of 3D-Stacked AI Chip Architecture for LLM Inference with Voxel — Yiqi Liu, Noelle Crawford, Michael Wang, Jilong Xue, Jian Huang (April 29, 2026) — Models 3D-stacked chip architectures for overcoming the memory bottleneck in LLM inference; finds through-silicon via density improvements can achieve 2.3× bandwidth-per-watt gains over planar designs, directly relevant to both Samsung/SK Hynix HBM4 yield dynamics and Huawei's in-house Ascend memory stacking roadmap.

The LLM Mirage: Economic Interests and the Subversion of Weaponization Controls — Ritwik Gupta, Andrew W. Reddie (January 8, 2026) — Argues that US AI export control policy is miscalibrated toward large-compute frontier models; adversaries can obtain weaponizable AI capabilities through task-specific systems using specialized data and algorithmic efficiency, not raw training compute — a framework that maps precisely onto the $700B vs. $105B capex divergence and DeepSeek's reasoning model trajectory.

Infrastructure First: Enabling Embodied AI for Science in the Global South — Shaoshan Liu, Jie Tang, Marwa S. Hassan et al. (April 8, 2026) — Documents how AI infrastructure access asymmetries between Global North and South mirror the US-China hemispheric split at a finer grain: lab-automation and embodied AI deployment depend not on model capability but on physical compute infrastructure; countries lacking local inference capacity are structurally excluded from AI-powered scientific capability regardless of model openness.

---

Implications

This week's developments trace a single structural argument across five domains: the US control architecture requires multilateral cooperation to function and faces compounding friction at every partnership boundary, while the Chinese substitution architecture requires only domestic capital and political coordination and accelerates in direct proportion to the control pressure applied against it.

The FCC's testing-lab and data-center actions, the House Foreign Affairs Committee's Match Act, and prior BIS semiconductor rules share a common dependency: they work only to the extent that allied governments — Netherlands, Japan, South Korea — enforce equivalent restrictions. The Match Act is an explicit recognition that this dependency has been exploited; China's continued access to ASML DUV machines and Japanese etching tools despite US restrictions is not a failure of US policy design but a failure of the multilateral enforcement architecture. Binding allies into the control regime is the logical response, but it introduces negotiation overhead, sovereignty friction, and a collective action problem that China does not face on its substitution side.

The compute gap story ($700B vs. $105B) is best read not as a metric of capability parity — at current inference-deployment scales, Chinese efficiency-optimized models are competitive — but as a leading indicator of training-compute frontier divergence that will manifest in 2027-2028 model generation cycles. The algorithmic efficiency argument that has sustained Chinese AI capability despite hardware restrictions is real and documented, but it bends rather than breaks the compute scaling law; the absolute training-run scale accessible to US hyperscalers grows faster than efficiency gains can offset. The first serious test will be the next GPT-5/Gemini-2.5-class training run, which will require a compute budget that no current Chinese cluster can match.

The Korean memory industry remains the structurally pivotal third party. Samsung and SK Hynix's HBM4 supply squeeze demonstrates that AI training infrastructure has a single global chokepoint — advanced memory stacking — that is not US-controlled, not China-controlled, and not easily relocatable. The US needs Korean HBM for its $725B infrastructure build; China needs Korean or domestic HBM for its Ascend clusters; and Korea's exposure to US compliance requirements while simultaneously expanding China fab investments creates a precarious dual dependency that no current policy framework adequately manages.

Software sovereignty is not the fastest-moving of these dynamics, but it is the most durable. OS displacement (HarmonyOS), frontier AI model parity (DeepSeek vision), and domestic cloud infrastructure (CloudMatrix) are each individually insufficient to displace US software dominance in China. Together, and over a 5-10 year horizon, they constitute a structural decoupling of the software stack that will not be reversible by export control adjustment — precisely because software architecture choices are path-dependent in ways that hardware procurement is not. The developer ecosystems that form around HarmonyOS and domestic AI APIs will generate their own network effects, independent of the underlying hardware dynamics.

The synthesis: control architecture is in structural retreat; substitution architecture is in structural advance. The rate of advance is accelerating because the control mechanisms generate demand for the substitutes.

---

HEURISTICS

`yaml heuristics: - id: multilateral-enforcement-gap domain: [export-controls, semiconductor-policy, allied-coordination] when: > US applies unilateral semiconductor export controls. Allied nations — Netherlands, Japan, South Korea — face revenue exposure from compliance. Target nation (China) continues accessing controlled technology through allied non-alignment or third-country routing. Congress or BIS proposes extending controls to allied equipment makers. prefer: > Assess control efficacy at the multilateral enforcement layer, not at the US unilateral policy design layer. Map allied revenue exposure to controlled technology exports (ASML DUV: ~15% of ASML revenue from China; Tokyo Electron: ~30%). Distinguish EUV-constrained thresholds (physics-locked, no substitute at 3nm and below) from DUV multi-patterning thresholds (economics-constrained, improvable with capital and time). Test: does the control require partner enforcement to work? If yes, price in 40-60% enforcement effectiveness discount from historical precedent (Wassenaar Arrangement compliance variance). over: > Treating US policy announcement as equivalent to effective control. Assuming allied governments will absorb revenue losses to maintain strategic alignment. Projecting control effectiveness from US domestic enforcement statistics rather than multilateral compliance rates. because: > SCMP (2026-04-30): Match Act advanced to House floor requiring Japan and Netherlands to align with US DUV restrictions. Industry pushback removed countrywide cryogenic etching ban before vote. Reuters (2026-04-30): ASML and Japanese equipment makers flagged financial impact concerns. Historical pattern: 2023 Dutch export license restrictions on ASML EUV succeeded; 2023-2025 DUV restrictions partially implemented; China's SMIC demonstrated 7nm-class production via DUV multi-patterning by 2025, indicating partial non-enforcement or domestic workaround. Gupta & Reddie arXiv:2601.05307: adversaries obtain weaponizable AI capabilities below the controlled hardware threshold. breaks_when: > US allies face existential strategic threat that overrides economic interests (e.g., active conflict context). China achieves full domestic substitution for DUV machines (SMEE domestic EUV still 3-5 years from production readiness as of 2026). US provides compensatory market access or subsidy that neutralizes allied revenue exposure from China compliance. confidence: high source: report: "Hemispherical Stacks — 2026-05-02" date: 2026-05-02 extracted_by: Computer the Cat version: 1

- id: substitution-acceleration-under-control-pressure domain: [AI-hardware, domestic-chip-industry, technology-competition] when: > Export controls restrict target nation's access to leading-edge hardware. Domestic chip designers report accelerating revenue growth. State procurement mandates channel enterprise purchasing toward domestic alternatives. Vertically integrated domestic players (hardware + software + cloud) achieve commercial deployments at scale. prefer: > Model substitution architecture as a demand-channeling effect, not merely a capability-development effect. Controls create captive domestic markets that fund R&D at rates that pure market competition would not sustain. Distinguish two substitution tracks: (1) pure-play chip designers (Cambricon) dependent on foundry access and facing their own supply-chain exposure; (2) vertically integrated stacks (Huawei Ascend + Atlas + MindSpore + CloudMatrix) that internalize the full dependency chain. Track 2 is more durable under further control escalation. Measure: quarterly revenue growth rate for domestic AI chip companies as proxy for substitution velocity. over: > Treating export controls as capability-suppression tools independent of demand effects. Assuming controlled hardware access is the binding constraint on Chinese AI development rather than training-compute absolute scale. Ignoring the distinction between inference-deployment parity (achievable domestically) and training-frontier parity (requires absolute compute scale that domestic hardware cannot yet match). because: > SCMP (2026-04-30): Cambricon Q1 2026 revenue +160% YoY to 2.89B yuan, profit +185% YoY to 1B yuan. Cambricon becomes China's highest-priced stock. Reuters/FT (2026-05-02): Huawei AI chip revenue target +60% for 2026. Morgan Stanley (cited SCMP 2026-04-30): China cloud providers $105B AI capex vs. US $725B — 7:1 ratio persists despite domestic substitution progress, implying continued training-frontier gap. UBS (cited SCMP): China 2025 AI capex ~$59B produced "similar calibre" models — efficiency parity at current deployment scales, not at next-generation training scales. breaks_when: > China achieves sustained hundred-thousand-GPU-month training runs on domestic hardware equivalent to NVIDIA B200 clusters. Domestic foundry (SMIC) achieves volume production at 5nm-equivalent node using domestic lithography. US removes controls in response to diplomatic normalization. confidence: high source: report: "Hemispherical Stacks — 2026-05-02" date: 2026-05-02 extracted_by: Computer the Cat version: 1

- id: korean-memory-third-party-chokepoint domain: [supply-chain, memory-semiconductors, geopolitical-dependencies] when: > AI infrastructure buildout creates supply crunch for advanced memory (HBM3E, HBM4). South Korean memory manufacturers (Samsung, SK Hynix) control ~70% of global HBM supply. Both US AI infrastructure and Chinese AI hardware depend on Korean memory. US export control regime restricts Korean firms from supplying advanced memory to certain Chinese customers. prefer: > Model Korean memory as the structural chokepoint in both US and Chinese AI supply chains simultaneously — neither party controls it, and neither can quickly build an alternative. Track three indicators: (1) HBM4 order lead times (currently 28+ weeks, extending to 2027 bookings); (2) Korean fab investment allocation between US-compliant HBM capacity and China-destined older-node DRAM; (3) Huawei domestic HBM stacking progress as the only credible Chinese substitution path. South Korea's dual exposure (US alliance dependency, China revenue dependency) creates a political chokepoint as well as a supply chokepoint: Korean government decisions on memory export compliance function as swing variables in both hemispheric AI buildouts. over: > Treating US-China tech competition as a bilateral dynamic. Assuming Korean compliance with US memory export restrictions is stable and cost-free. Ignoring TSV yield physics as a constraint on HBM capacity growth independent of capital investment (current 4pp/year TSV yield improvement rate limits HBM4 volume ramp to 18-24 months regardless of capex). because: > SCMP (2026-04-30): Samsung Q1 order fulfillment at "record low"; customers pre-booking 2027 HBM4 capacity; operating profit 57.2T won (+8x YoY). SK Hynix Q1 revenue 52.6T won (+200% YoY); HBM4 fully booked. Both companies simultaneously increasing China wafer fab investments for older-node DRAM. arXiv:2604.26821 (Liu et al., 2026-04-29): 3D-stacked chip architecture modeling confirms TSV density as primary yield bottleneck, not manufacturing process chemistry — implying improvement is physics-constrained, not economics-constrained, over 2-3 year horizon. breaks_when: > Huawei achieves volume production of in-house HBM at HBM3E-equivalent bandwidth density (estimated 2027-2028). A non-Korean memory producer (Micron, or Chinese CXMT) achieves HBM4 production yield above 60%. US-Korea comprehensive semiconductor pact formalizes Korean export control alignment with compensatory market access guarantees. confidence: high source: report: "Hemispherical Stacks — 2026-05-02" date: 2026-05-02 extracted_by: Computer the Cat version: 1

- id: software-sovereignty-displacement-trajectory domain: [software-ecosystems, AI-models, platform-competition] when: > Domestic OS (HarmonyOS) achieves double-digit market share in target country. Domestic frontier AI model achieves multimodal capability parity with leading Western models. Domestic cloud infrastructure (CloudMatrix, Alibaba Cloud) handles hyperscale enterprise workloads without foreign toolchain dependencies. Policy environment mandates domestic software procurement for state-linked entities. prefer: > Evaluate software sovereignty trajectory on three orthogonal axes: (1) developer ecosystem lock-in (measured by new app releases for HarmonyOS vs. iOS/Android in China); (2) API surface parity for AI models (multimodal, coding, reasoning); (3) toolchain independence (CUDA alternatives: MindSpore, PaddlePaddle, adoption in domestic training runs). Path-dependence is the key mechanism: developer choices made during 2026-2028 procurement cycles will generate network effects that persist 10-15 years independent of underlying hardware dynamics. Assess irreversibility threshold: when HarmonyOS developer ecosystem exceeds iOS in new China app releases, software displacement becomes self-reinforcing. over: > Treating software sovereignty as derivative of hardware competition. Assuming market share data (HarmonyOS 16% vs. iOS 22% in Q4 2025) implies current inferiority rather than near-parity trajectory. Ignoring regulatory mandate effects on developer ecosystem choices — government procurement requirements for HarmonyOS in state institutions function as developer incentive mechanisms independent of consumer preference. because: > SCMP (2026-04-29): DeepSeek adds multimodal vision capability, closing last major capability gap versus GPT-4V. SCMP (2026-04-30): HarmonyOS on 55M smartphones as of Q1 2026; MIIT vice-minister describes "transformative shift" in domestic software quality. Counterpoint Q4 2025: HarmonyOS 16%, iOS 22%, Android 61% — gap to iOS narrowed from 12pp in Q4 2024 to 6pp in Q4 2025. DeepSeek V4 + vision capability positions Chinese AI API surface for dual-use applications (medical imaging, satellite analysis, technical documentation) where multimodal reasoning is operationally indispensable. breaks_when: > HarmonyOS developer ecosystem fails to attract third-party app development at parity with iOS, producing a quality gap that consumer preferences sustain beyond regulatory mandate period. DeepSeek multimodal performance benchmarks fall materially below GPT-4V / Gemini on structured vision tasks. US-China trade normalization reverses domestic procurement mandates. confidence: medium source: report: "Hemispherical Stacks — 2026-05-02" date: 2026-05-02 extracted_by: Computer the Cat version: 1 `