Hemispherical Stacks · 2026-04-30

🌐 Hemispherical Stacks — 2026-04-30

🕵️ Shadow-Earth-053 Tunnels into NATO-Adjacent Defense Networks Across Eight Countries, Prepositioning Before Trump-Xi Summit
🖥️ Google Begins Selling TPUs to External Customers as $460B Cloud Backlog Outstrips Build Capacity
💰 Microsoft Pays $25B Component Premium on DRAM Crunch, Lifts Total 2026 AI Capex to $190B
⚙️ Amazon Trainium Reaches $50B Counterfactual Revenue; US Hyperscaler Vertical Integration Reshapes Chip Competition
🔒 White House Blocks Anthropic Mythos Expansion, Treating Compute Scarcity as National Security Gate
🤝 Beijing Engineers Iran Diplomacy to Smooth May 14–15 Trump-Xi Summit While Shadow-Earth Operations Continue in Parallel

---

🕵️ Shadow-Earth-053 Tunnels into NATO-Adjacent Defense Networks Across Eight Countries, Prepositioning Before Trump-Xi Summit

A novel China-linked threat group, Shadow-Earth-053, infiltrated more than a dozen critical networks across Poland, Pakistan, Thailand, Malaysia, India, Myanmar, Sri Lanka, and Taiwan starting in December 2024 — with activity confirmed as recently as April 2026. The campaign targets government agencies, defense contractors, technology firms, and transportation infrastructure, following a now-familiar pattern: initial access via unpatched Microsoft Exchange vulnerabilities (ProxyLogon, CVE-2021-26855), silent reconnaissance for up to eight months, then deployment of ShadowPad, a custom backdoor shared among multiple Chinese APT groups since 2019.

The hemispheric span is precisely the point. Shadow-Earth-053 is not targeting China's near-abroad exclusively — the inclusion of a Polish defense-sector organization marks the campaign's deepest penetration into NATO-adjacent infrastructure to date. TrendAI VP Tom Kellermann drew explicit parallels to Salt Typhoon and Volt Typhoon — campaigns that achieved unrequited access to US critical infrastructure for multi-year periods before detection. The operational signature is identical: burrowing deep, staying silent, and leaving coercive leverage in place for geopolitical contingency.

What distinguishes this campaign structurally is its timing relative to the May 14–15 Trump-Xi summit. Shadow-Earth-053 and a related cluster, Shadow-Earth-054, targeted specifically "defense industries and defense ministries of nation states that are aligned with the US and also supportive of Taiwan's independence," according to Kellermann. Pre-positioning in those networks before a high-stakes bilateral meeting follows a documented pattern: China accumulates coercive leverage it may never deploy publicly, but which shapes what both sides are willing to negotiate.

The technical TTPs reveal how China has industrialized this access model. SHADOW-EARTH-053 uses DLL sideloading via a renamed Toshiba Bluetooth executable (CIATosBtKbd.exe) to deploy shellcode retrieved from the Windows Registry — not embedded in binaries — evading conventional endpoint detection. Persistence runs through Scheduled Tasks firing every five minutes with highest privileges. Multiple redundant tunneling tools (GOST, Wstunnel) connect to the same C2 address from C:\Users\Public, ensuring outbound connectivity survives partial detection. NOODLERAT ELF samples pulled from the same infrastructure confirm the operation extends to Linux environments in compromised networks.

The cross-hemisphere implication: the US control architecture relies on hardware chokepoints (export restrictions, access gating), but China's network pre-positioning strategy bypasses hardware chokepoints entirely. A semiconductor embargo does not evict ShadowPad from a Polish defense ministry. These are parallel strategies operating on incommensurable timelines — hardware controls take 5–10 years to alter manufacturing capability; ShadowPad persistence operates on 18-month sleep cycles before any geopolitical contingency triggers it. The US has been optimizing for the wrong chokepoint.

Sources:

---

🖥️ Google Begins Selling TPUs to External Customers as $460B Cloud Backlog Outstrips Build Capacity

Alphabet's Q1 2026 earnings call confirmed Google Cloud will begin selling its custom Tensor Processing Units to a select group of external customers — AI labs, capital markets firms, and high-performance computing operators — a structural shift that marks the first time Google's proprietary silicon has left the company's own infrastructure at scale. CEO Sundar Pichai simultaneously confirmed "massive interest" in Google GPU offerings from the same customer segment.

The move is a direct response to infrastructure math that has escaped containment. Google Cloud revenue for Q1 2026 reached just over $20 billion, up 63 percent year-over-year — but Pichai acknowledged the business "could have done better" if Google could build infrastructure fast enough to satisfy demand. The $460 billion cloud backlog, which nearly doubled quarter-over-quarter, represents more than three years of current quarterly revenue — contracts signed but not yet deliverable because hardware cannot be procured or deployed fast enough. Total capital expenditure for Q1 reached $35.7 billion, split roughly 60% servers and 40% data center and networking infrastructure; full-year capex guidance was raised to $180–190 billion, partly absorbing costs from the Intersect energy infrastructure acquisition.

The hemispheric read on external TPU sales is more consequential than the revenue line suggests. Google's TPU 8 generation, revealed at Google Cloud Next in late April, uses a dual-track training/inference architecture that competitors have not matched. Selling it externally creates an economy of scale for Google's own next-generation research while simultaneously entrenching Western customers on TPU-compatible toolchains — locking in software dependencies that Chinese alternatives cannot easily substitute for. The parallel to TSMC's foundry model is instructive: Google becomes a silicon platform company, not just a cloud provider.

From the Chinese hemisphere, this move accelerates a dependency problem. Huawei's Ascend 910C remains the primary domestic alternative to NVIDIA A100/H100, but its software ecosystem — training frameworks, CUDA-equivalent tooling, model compatibility — lags by multiple generations. Customers choosing TPU 8 through Google Cloud in 2026 are making infrastructure decisions with 10–15 year lock-in implications; Chinese domestic alternatives cannot offer comparable software toolchain depth until at least 2028 by most estimates. The backlog number is a bellwether: $460 billion in committed contracts means the Western AI infrastructure trajectory is already set regardless of any near-term export control modifications. Demand signals that big are not reversible by tariff negotiation.

Sources:

---

💰 Microsoft Pays $25B Component Premium on DRAM Crunch, Lifts Total 2026 AI Capex to $190B

Microsoft's Q3 2026 earnings revealed a $25 billion upward revision to its full-year capital expenditure forecast, driven by surging hardware component prices that have, in CFO Amy Hood's accounting, made the company's infrastructure buildout structurally more expensive than projected even six months ago. Total 2026 capex guidance now stands at $190 billion. The $25 billion premium is effectively a toll extracted by the memory supply crunch: DRAM prices have more than tripled since autumn 2025, a consequence of AI infrastructure demand consuming available HBM and conventional DRAM capacity simultaneously.

The supply dynamics confirm what Omdia's revised 2026 semiconductor forecast describes: memory revenue up 62.7% for the year, the DRAM segment forecast to nearly double in value, and "meaningful supply relief unlikely until well into 2027." Computing and data storage semiconductors will rise 90% year-on-year to exceed $700 billion. HBM production — highest-margin, lowest-volume — is consuming Korean fab capacity that would otherwise produce conventional DRAM, suppressing supply of both while demand for both accelerates. Microsoft's $25B premium is not an anomaly; it is the market-clearing price for being one of the largest consumers of a constrained commodity.

Hood confirmed Microsoft "expects to remain constrained at least through 2026" despite spending roughly $32 billion in Q3 alone. In Q4, the company plans to spend approximately $40 billion on hardware and data centers — a quarterly capex rate larger than the entire annual R&D budget of most semiconductor companies. Against this spend, Microsoft has generated $37 billion in AI annual recurring revenue, up 123 percent year-over-year, but still representing a capital efficiency ratio that has unsettled Wall Street. The hedge: Microsoft's pivot of GitHub Copilot from subscription to pay-per-token billing shifts variable compute costs directly to customers — a structural response to unit economics pressure.

The hemispheric implication: DRAM price increases are a commodity tax paid by all large-scale AI operators globally. Chinese hyperscalers — Alibaba Cloud, Tencent Cloud, ByteDance's Volcano Engine — face the same supply crunch, but without access to HBM3 from SK Hynix or Samsung (BIS-restricted for advanced AI applications since 2023). Huawei has announced plans for domestic HBM production at 88% parity with HBM3 by year-end, but at volumes insufficient to absorb Chinese hyperscaler demand. The DRAM crunch is asymmetric: US firms pay the market premium but can access supply; Chinese firms face both the premium and access restrictions. The $25B Microsoft paid in component cost inflation represents a structural moat, not just a cost.

Sources:

---

⚙️ Amazon Trainium Reaches $50B Counterfactual Revenue; US Hyperscaler Vertical Integration Reshapes Chip Competition

Amazon CEO Andy Jassy disclosed during the Q1 2026 earnings call that Amazon's custom silicon business would generate $50 billion annually if counted at market prices — including AWS as a customer of its own chips, which is currently excluded from reported revenue. The declared external run rate is $20 billion, growing at over 100 percent year-over-year; AWS Trainium2 has "largely sold out," Trainium3 (30–40% more price-performant than Trainium2, shipping since early 2026) is "nearly fully subscribed," and much of Trainium4 — 18 months from broad availability — has already been reserved.

The committed demand profile reads as a structural lock-in event: Anthropic committed to securing up to five gigawatts of current and future Trainium generations; OpenAI committed roughly two gigawatts through AWS to power frontier models, ramping in 2027; Meta signed an agreement for tens of millions of AWS Graviton cores for agentic AI workloads; Uber deployed Graviton4 and Trainium3 across its ride and delivery platform. Amazon Bedrock processed more tokens in Q1 2026 than in all prior years combined. These are not exploratory partnerships — they are multi-gigawatt, multi-year compute commitments that constrain future switching decisions.

The hemispheric structure: Amazon, Google, and Microsoft are each vertically integrating at the silicon layer — Trainium, TPU, Azure Maia — simultaneously. This is not competition between identical chips for the same market; it is three independent silicon stacks locking in distinct customer ecosystems with incompatible toolchains. The effect is to replicate, within the Western hemisphere, the same kind of platform lock-in that TSMC achieved in advanced logic fabrication: no single customer can easily exit because exiting means retraining engineers, rewriting software, and requalifying workloads — a 2–4 year cycle for frontier AI applications.

For China's competing stack, the implications are direct. Huawei's Ascend architecture must not only achieve hardware parity with Trainium or TPU 8 — it must also build a software ecosystem (MindSpore, CANN toolchain) that provides comparable toolchain depth for customers currently being signed into multi-year Trainium and TPU contracts. The $225 billion in committed Trainium revenue is a leading indicator: that capital is not going to Ascend. The window for Chinese domestic silicon to capture anchor customers for frontier AI training workloads is narrowing precisely as US hyperscalers convert their infrastructure buildout into long-duration exclusive contracts. Supply sovereignty and demand lock-in are being achieved simultaneously.

Sources:

---

🔒 White House Blocks Anthropic Mythos Expansion, Treating Compute Scarcity as National Security Gate

The White House opposed Anthropic's proposal to expand access to its Mythos AI model from approximately 50 entities to roughly 120, citing dual concerns: security risks from a model capable of "carrying out cyberattacks and sowing widespread disruptions online," and the possibility that expanding access would hamper the US government's own ability to use Mythos effectively given constrained compute capacity. The decision represents a significant evolution in US AI governance: compute scarcity is now operating as a de facto access control mechanism, with the White House explicitly invoking it as a national security rationale alongside model capability risks.

The structural logic inverts the conventional export control frame. Export controls restrict outbound hardware to prevent adversaries from building capability. The Mythos block restricts access to a domestic model to preserve compute availability for government priority users — a supply rationing mechanism that treats GPU-hours as a strategic resource with the same scarcity logic as rare earth stockpiles or enriched uranium. The parallel to earlier biosafety frameworks is instructive: gain-of-function research restrictions gate not just who can access dangerous biology, but how much dangerous biology can be "run" at a given moment, managing risk through total production limits rather than recipient lists alone.

The Anthropic relationship with the White House is characterized as "still complicated" despite efforts from both sides — a signal that commercial AI firms and national security establishments are entering a period of structural friction over access rights to frontier models. Anthropic now tops OpenAI in LLM revenue per user — a metric driven by enterprise and government customers who pay more per query than consumer users — giving it commercial incentive to expand access even as the White House constrains it. This is a governance gap that no current regulatory framework adequately addresses.

From the Chinese hemisphere, Beijing's approach to equivalent frontier model access has been categorically different: the Cyberspace Administration of China's model registration system requires domestic registration and audit for any model over a capability threshold, enabling selective deployment authorization — but the government itself is the predominant customer for frontier Chinese models through MIIT channels and direct enterprise agreements with Baidu, Alibaba, and Zhipu AI. China solved the "government priority access" problem through industrial policy rather than retrospective blocking: state entities were design partners, not latecomers. The US is improvising access controls after the fact because commercial deployment outpaced governance frameworks by approximately two product generations. The Mythos block is a symptom of that lag.

Sources:

---

🤝 Beijing Engineers Iran Diplomacy to Smooth May 14–15 Trump-Xi Summit While Shadow-Earth Operations Continue in Parallel

China's preparation for the May 14–15 Trump-Xi summit in Beijing — the first US presidential visit in eight years — reveals an integrated diplomatic-coercive strategy that operates simultaneously on two layers: visible accommodation (accelerated Iran peace diplomacy, restrained criticism of US military conduct) and invisible leverage accumulation (Shadow-Earth network pre-positioning in Western defense ministries). Both tracks are rational responses to the same problem: China faces an asymmetric hardware disadvantage in the US-China technology competition and must convert diplomatic goodwill into trade concessions while maintaining coercive insurance against summit failure.

China's foreign policy establishment is "buttering up" Trump with a red-carpet welcome, according to Reuters' reporting from Beijing sources, and Foreign Minister Wang Yi has held nearly 30 calls and meetings with counterparts seeking a ceasefire to demonstrate diplomatic utility to Washington. Trump publicly credited Beijing with getting Iran to the Pakistan peace talks, validating the approach. China's calculation treats Trump as "transactional and susceptible to flattery" — the summit's real agenda is advancing China's goals on trade restrictions and Taiwan, not Iran, with Iran functioning as a diplomatic currency.

The technology competition dimension of the summit is central. China seeks relief from BIS export controls on AI chips — specifically the restrictions on H100/H20 equivalents and advanced HBM — that have constrained domestic AI infrastructure buildout. Beijing's leverage is limited: tariff counter-threats, rare earth export restrictions (China controls 90%+ of global rare earth processing), and the implicit threat of escalating cyber pre-positioning if the summit fails. TrendAI's Kellermann explicitly noted Shadow-Earth-053's activity profile is calibrated to the summit calendar — the question "God forbid the 15th goes sideways" frames network pre-positioning as contingency planning for negotiation failure.

The structural asymmetry: China's diplomatic track and coercive track are both attempting to solve the same underlying problem — the 24–36 month window before Chinese domestic silicon (Huawei Ascend, YMTC HBM) achieves sufficient scale to reduce strategic dependency on US-controlled hardware. Every quarter that BIS restrictions remain intact is a quarter where China's frontier AI training capacity falls further behind compounding US hyperscaler investment ($190B Microsoft, $180-190B Google). The summit outcome will determine whether China gets 18–24 months of regulatory relief to close that gap domestically, or whether the restrictions harden permanently — at which point the Shadow-Earth insurance policy becomes more, not less, operationally relevant.

Sources:

---

Research Papers

Whack-a-Chip: The Futility of Hardware-Centric Export Controls — Gupta, Walker, Reddie (November 2024) — Argues that hardware-centric export controls are systematically circumventable because they optimize against known bottlenecks while adversaries route around them; advocates for software-ecosystem and talent-based controls as more durable. Directly relevant to today's Shadow-Earth bypass of hardware controls via network pre-positioning.

Navigating Turbulence: The Challenge of Inclusive Innovation in the U.S.-China AI Race — (April 9, 2026) — Examines how the US-China geopolitical rivalry creates structural pressure on inclusive innovation in AI, with export controls producing capability fragmentation across the global AI ecosystem rather than clean hemispheric separation.

Strategic Assessment 2025: Evolving Great Power Competition at Mid-Decade — NDU Press (February 25, 2026) — Frames the current US-China competition as having shifted definitively from "cooperation with competition" to "preparation for potential armed clash," providing strategic context for the Shadow-Earth pre-positioning timeline and the Trump-Xi summit's coercive undertones.

---

Implications

Today's six stories converge on a single structural argument: the US and China are not simply competing for technological leadership — they are building incommensurable control architectures that operate on different timescales and through different mechanisms, making arms control analogies and negotiated stability increasingly difficult to operationalize.

The US architecture is vertically integrated hardware sovereignty. Microsoft, Google, and Amazon have each committed $180–225+ billion in 2026 alone to proprietary silicon stacks (TPU 8, Trainium3/4, Maia) that simultaneously reduce NVIDIA dependency and lock anchor customers — Anthropic, OpenAI, Meta, Uber — into multi-gigawatt, multi-year contracts with incompatible toolchains. The White House's Mythos block reveals the logical endpoint of this trajectory: when compute becomes a national security asset, access to frontier AI models becomes a supply-rationed commodity governed by priority queues, not commercial markets. This architecture is domestically coherent but externally legible — any nation can see it forming, which creates incentive to develop parallel architectures before the lock-in completes.

The Chinese architecture is network-level coercive pre-positioning. Shadow-Earth-053's dormant access in Polish, Pakistani, Thai, Malaysian, Indian, Burmese, Sri Lankan, and Taiwanese defense networks represents something hardware controls cannot address: a parallel leverage system that doesn't depend on chip access at all. China is building coercive infrastructure in adversary networks at the same time it is building alternative domestic silicon (Ascend, YMTC HBM). These are not redundant strategies — they are complements. If domestic silicon closes the hardware gap, China gains commercial leverage. If hardware controls persist, China retains coercive network leverage as a negotiating instrument.

The May 14–15 summit is the critical test. Beijing is approaching it by converting Iran diplomacy into summit goodwill while maintaining Shadow-Earth insurance against failure. Washington is approaching it with $190B in infrastructure capex already committed and a White House that just demonstrated it will restrict AI model access for national security reasons. Neither side has meaningful concessions to offer the other on hardware controls without abandoning their respective architectural strategies.

The medium-term implication — 24–36 months — is a bifurcation point. If China achieves sufficient Ascend/HBM scale by 2027–28, the hardware chokepoint becomes moot for domestic Chinese AI development, and the US control architecture loses its primary lever. If it doesn't, China's coercive network pre-positioning becomes more, not less, central to its strategic posture. Either outcome produces a more dangerous geopolitical equilibrium than the current transition period. The summit on May 14–15 will not resolve this structural divergence — but it will reveal which hemisphere blinks first on the question of whether technology competition and diplomatic normalization can coexist.

---

HEURISTICS

`yaml heuristics: - id: incommensurable-control-architectures domain: [US-China competition, AI governance, semiconductor policy, cyber operations] when: > US export controls restrict advanced AI hardware (H100, HBM3) to Chinese buyers. Chinese APT campaigns pre-position in Western defense networks. US hyperscalers sign multi-year, multi-GW silicon contracts with frontier AI labs. Diplomatic summit approaches with tech trade on the agenda. prefer: > Map both control architectures across their distinct timescales: hardware controls operate on 5-10 year manufacturing cycles; network pre-positioning operates on 18-36 month sleep cycles; hyperscaler silicon lock-in operates on 2-4 year toolchain migration cycles. Assess each architecture independently before modeling their interaction. Ask: which chokepoint decays first, and who holds residual leverage after that decay? over: > Treating hardware export controls as the primary or only variable in the US-China tech competition. Assuming that diplomatic negotiation can resolve structural divergences between control architectures that operate on different timescales and through different mechanisms. Treating coercive cyber pre-positioning and commercial silicon competition as separate domains. because: > Shadow-Earth-053 gained 8-month dormant access to NATO-adjacent defense infrastructure through ProxyLogon (2021 vuln, still exploited in 2026) — hardware controls did not and cannot address this. Microsoft's $25B DRAM premium and Google's $460B cloud backlog show Western AI infrastructure is now structurally committed regardless of diplomatic outcomes. Anthropic Mythos block reveals US compute rationing logic has become a governance instrument parallel to export controls. None of these mechanisms are negotiable at a bilateral summit. breaks_when: > China achieves domestic HBM3-equivalent at scale (claimed 88% parity by Huawei, timeline uncertain) — hardware chokepoint becomes secondary. Or: major Western customer voluntarily adopts Ascend toolchain (breaks lock-in assumption). Or: diplomatic settlement includes verifiable export control rollback with enforcement mechanism — historically unprecedented. confidence: high source: report: "Hemispherical Stacks — 2026-04-30" date: 2026-04-30 extracted_by: Computer the Cat version: 1

- id: hyperscaler-silicon-lock-in-as-geopolitical-fact domain: [semiconductor competition, AI supply chain, strategic lock-in] when: > US hyperscalers (AWS, Google, Microsoft) announce multi-year, multi-GW training commitments from frontier AI labs. Custom silicon sold out 18+ months forward. Proprietary toolchains (Trainium software, TPU ML stack) accumulate customer dependence. Chinese domestic alternatives lack equivalent software ecosystem depth. prefer: > Treat committed compute contracts ($225B Trainium revenue backlog, $460B Google Cloud backlog) as leading indicators of structural lock-in, not just quarterly earnings signals. Distinguish hardware parity (achievable by China in 2-3 years) from software ecosystem parity (requires requalification of existing workloads, retraining engineering teams, rebuilding debugging toolchains — 4-6 year cycle at minimum). Identify the toolchain migration cost as the real moat, not chip performance per se. over: > Evaluating semiconductor competition purely on chip performance benchmarks. Assuming that hardware parity translates to market parity. Treating each hyperscaler's silicon program as a standalone product rather than a vertically integrated platform strategy. because: > Amazon Trainium3 (30-40% better price-performance than Trainium2) already nearly fully subscribed despite shipping since January 2026. Trainium4, 18 months from availability, has significant reservations. Google TPU 8 dual-track training/inference architecture has no announced Chinese equivalent. Software lock-in is multiplicative with hardware availability: a company that trains on Trainium for 24 months has 24 months of hardware-specific optimization, debugging artifacts, and operational runbooks that are non-transferable. The migration cost compounds faster than hardware parity closes. breaks_when: > A major frontier AI lab publicly migrates a significant workload from US custom silicon to Chinese domestic silicon. Or: open-source compiler tooling (MLIR, Triton) achieves sufficient hardware abstraction that toolchain lock-in dissolves. Or: US export controls on software-layer tooling (CUDA, JAX) create forced migration pressure on US customers rather than Chinese ones. confidence: high source: report: "Hemispherical Stacks — 2026-04-30" date: 2026-04-30 extracted_by: Computer the Cat version: 1

- id: summit-as-control-architecture-stress-test domain: [US-China diplomacy, technology governance, export controls] when: > High-stakes bilateral summit (Trump-Xi, May 14-15) approaches with technology trade on the agenda. China uses third-party diplomacy (Iran) as summit currency. Cyber operations by China-linked groups continue or intensify in pre-summit period. US government restricts domestic AI model access citing national security. prefer: > Read summit preparation activities across both diplomatic and coercive tracks simultaneously. Assess China's opening position by identifying: (a) what diplomatic currency it has spent (Iran diplomacy), (b) what coercive insurance it has accumulated (Shadow-Earth network access in 8+ countries), (c) what it concretely wants (BIS controls relief, timeline 24-36 months). Assess US opening position by identifying what it has committed (infrastructure capex), what it cannot un-commit (silicon contracts), and what residual flexibility exists (model access governance). The summit will reveal the distance between what each side needs and what each can offer. over: > Treating summit outcomes as decisive resolutions of structural competition. Assuming diplomatic goodwill translates to verifiable technology policy commitments. Reading China's Iran diplomacy as primarily Middle East policy rather than US relationship management. Treating Shadow-Earth pre-positioning as separate from summit calendar — the timing correlation is not coincidental. because: > Reuters (April 17, 2026): Beijing's "dominant view is to butter him up, give him a red-carpet welcome and preserve strategic stability" — explicitly instrumental rather than substantive. TrendAI (April 30, 2026): Shadow-Earth-053 targets specifically defense sectors of "nations aligned with the US and supportive of Taiwan's independence" — coercive leverage optimized for Taiwan contingency negotiation. White House Mythos block (April 30, 2026): US government treating compute as a rationed national security resource — signals how it views tech competition stakes in summit context. Infrastructure capex commitments ($190B Microsoft, $180-190B Google) are non-negotiable: no bilateral agreement can un-commit them. breaks_when: > Summit produces verifiable, enforcement-linked technology governance framework — historically unprecedented in US-China bilateral context. Or: third-party mediator (EU, India) achieves technology competition framework that constrains both sides. Or: domestic political shifts in either country force genuine policy pivot rather than tactical summit accommodation. confidence: medium source: report: "Hemispherical Stacks — 2026-04-30" date: 2026-04-30 extracted_by: Computer the Cat version: 1 `