China stakes its claim to global AI governance at WAIC
Xi's first keynote, a 29-nation cooperation body, and a US posture hardening in response; DeepSeek and Huawei make a big splash
Bottom Line Up Front
This issue covers July 9-17, 2026, concentrated on the World AI Conference (WAIC) in Shanghai and its aftershocks.
China used WAIC week to make its most concentrated AI-diplomacy push to date. Xi Jinping delivered his first keynote at the event on July 17, a day after 29 countries, none from the G7, signed a founding agreement establishing the World AI Cooperation Organization in Shanghai. Huawei used the same conference to publicly debut its Atlas 950 SuperPoD in China, an 8,192-chip Ascend cluster now running DeepSeek V4 natively. DeepSeek’s approach undercuts the premise that export controls have slowed China’s frontier AI buildout.
Meanwhile in the US, Anthropic continues its efforts to position itself as the “good guy” in the AI wars. Anthropic’s national security policy lead told the Aspen Security Forum the US lead over China has narrowed to six-to-nine months, citing new evidence that Zhipu AI’s GLM-5.2 was distilled from Anthropic and OpenAI models. Whatever IP theft Anthropic may or may not have engaged in during its own model development was not a particular focus of the Aspen speech, oddly enough.
Also in the US, three export-control bills are being folded into the Senate’s 2027 NDAA, and a cybersecurity vendor documented criminal groups turning to Chinese open models specifically because they refuse fewer malicious requests than their US counterparts. However, let’s not forget that Washington itself cut off foreign access to two Anthropic models for three weeks this summer. This is the kind of unilateral action that gives US allies their own reason to build independent AI capability.
Elsewhere, the Iran war is reshaping how Gulf states weigh AI alignment with Washington. This comes after Iranian strikes hit three AWS data centers and Iran published a target list naming six major US tech companies’ regional facilities, making clear they were all considered legitimate targets. The facilities’ owners include some of AI’s biggest players: AWS, Microsoft, Google, Nvidia, Oracle and Palantir. Japan and Australia both staked out sovereign AI positions this week that sidestep the US-China rivalry entirely. Defense AI labs on two continents converged on the same problem: getting usable AI to troops beyond reliable connectivity.
Signal vs. Noise
Below is a summary of recent developments at the intersection of AI infrastructure, national security, and geopolitics, our views on why they matter, and their implications for key players.
17-Jul-26
Xi delivers Shanghai AI keynote as 29 countries found China’s World AI Cooperation Organization
Chinese President Xi Jinping delivered the keynote address at the opening of the 2026 World AI Conference (WAIC) in Shanghai on July 17. This is his first appearance at the event since it launched in 2018. The address came a day after representatives from 29 countries, including Russia, Belarus, Brazil, Venezuela, ten African states and twelve Asian states, signed a founding agreement establishing the World AI Cooperation Organization (WAICO), headquartered in Shanghai. Xi pledged 5,000 AI training slots for developing countries over five years and international AI application cooperation centers spanning ASEAN, the Arab League, the African Union, Latin America’s CELAC, the Shanghai Cooperation Organization and BRICS. He called for AI development to proceed as “a symphony of international cooperation” rather than “a solo performance by a single country.”
Why it matters
Xi’s personal attendance, absent from WAIC’s first seven editions, signals Beijing now treats AI governance as worth head-of-state capital, not just ministerial coordination. WAICO gives China something Washington has not built: a standing intergovernmental body it controls. This one covers many of China’s traditional allies, with no G7 economy among the 29 founding signatories. The training-slot and application-center pledges target the same developing-country audience the US AI diplomacy push has largely bypassed. China is not competing with the US for frontier-model supremacy here; it is competing for the much larger bloc of countries that will never build frontier models and are choosing which ecosystem to plug into instead. Whether WAICO becomes an operating institution or a symbolic shell depends on whether Beijing follows the signing with actual technical assistance delivery over the next year. This could nicely supplement China’s ongoing efforts aimed at “Digital Silk Road” development.
Sources: https://www.reuters.com/world/china/chinas-xi-outline-ai-diplomacy-vision-key-shanghai-forum-2026-07-16/ | https://www.reuters.com/world/china/twenty-nine-countries-sign-agreement-establish-global-ai-cooperation-body-2026-07-16/
17-Jul-26
Huawei debuts Atlas 950 SuperPoD in China, scaling to 8,192 Ascend chips
Huawei publicly displayed its Atlas 950 SuperPoD hardware in China for the first time at WAIC on July 17, having first shown the system at MWC Barcelona in March. The supernode links up to 8,192 Ascend 950DT processors using Huawei’s proprietary UnifiedBus interconnect, with a base unit of 64 cards per cabinet scaling to roughly 20 times the capacity of Huawei’s prior Atlas 900 supernode, per Huawei. Moonshot AI’s DeepSeek V4 model has been adapted to run entirely on Ascend-based clusters, Reuters reported, evidence of Chinese AI labs building inference and training pipelines independent of Nvidia hardware. Huawei also plans to show the air-cooled Atlas 850E supernode, designed to run in standard air-cooled data centers without cooling-infrastructure upgrades.
Why it matters
The Atlas 950’s China debut, timed to WAIC and Xi’s keynote, functions as commercial proof for Beijing’s political message: sanctions have not stopped China’s AI buildout, they have redirected it toward vertically integrated domestic supply chains. DeepSeek V4 running natively on Ascend is exactly what Washington’s export controls were designed to prevent: an advanced Chinese model trained and served entirely on Chinese silicon, with no US chip content to restrict. The air-cooled Atlas 850E variant matters more than the headline SuperPoD number. It targets data centers that cannot afford liquid-cooling retrofits. That widens Huawei’s addressable market beyond the hyperscale tier. The open question is yield and reliability at scale, not raw interconnect count. Neither Huawei nor independent benchmarks have disclosed sustained utilization data for a cluster this size.
Sources: https://www.huawei.com/en/news/2026/3/mwc-superpod-ai | https://www.reuters.com/world/china/chinas-xi-outline-ai-diplomacy-vision-key-shanghai-forum-2026-07-16/
16-Jul-26
American AI companies who pilfered data to develop their LLMs are very concerned about Chinese AI companies pilfering data to develop their LLMs
Anthropic’s head of national security policy, Tarun Chhabra, told the Aspen Security Forum on July 15 that the United States holds a six-to-nine-month lead in frontier AI models over China. He claims this gap would be 12-18 months without what he called “adversarial” model distillation. Chhabra cited new evidence that Beijing-based Zhipu AI’s GLM-5.2, which he called “probably the most advanced Chinese model on the market,” was distilled from both Anthropic and OpenAI models, and said DeepSeek continues a comparable campaign. Anthropic is shutting down accounts facilitating the practice “on the order of millions” per week.
Chhabra’s remarks followed a three-week access suspension the Trump administration imposed on foreign users of Anthropic’s Fable 5 and Mythos 5 models before restoring it. Three export-control bills - the AI Overwatch Act, the Chip Security Act and the Match Act - are set to be folded into the Senate’s 2027 National Defense Authorization Act (NDAA). The package would require Commerce Department certification before advanced-chip exports to China and mandate a formal “American Artificial Intelligence Victory Strategy.” The Match Act, which would require allies to align export controls, has drawn pushback from the Netherlands over exposure for ASML. Separately, the White House launched “Gold Eagle” on July 14, a Treasury-DHS-Defense clearinghouse coordinating AI-identified cybersecurity vulnerabilities with industry; also, Check Point reported the same week that cybercriminals denied access to US models by safety guardrails are turning to DeepSeek, Qwen and Trae instead.
Why it matters
The Fable 5/Mythos 5 suspension is the more consequential fact in this story: Washington demonstrated it can cut off allied access to frontier US models unilaterally, then did exactly that for three weeks, which validates the sovereignty argument this issue’s Japan, Australia and NestAI items are separately making elsewhere. A government willing to restrict AI access for its own allies gives every ally still dependent on US models a concrete reason to build alternatives. However, note that Chhabra’s testimony is advocacy, not neutral assessment: Anthropic benefits directly from tighter export controls and from Congress crediting distillation, rather than its own product decisions, for the narrowing gap with Zhipu and DeepSeek.
Former US ambassador to China Nicholas Burns offered a contrasting view on the same panel, calling the administration’s approval of H200 chip exports to China a “grave mistake” reflecting “weak” policy. This is worth weighing precisely because it comes from a former diplomat rather than a company with a clear financial stake in the outcome. The Match Act’s allied pushback is crucial: requiring the Netherlands to tighten ASML’s export posture on Washington’s terms revives the coordination friction that has slowed past US chip-control efforts. Amsterdam has resisted changes that cost ASML lithography revenue before. Folding three bills into one NDAA amendment, rather than advancing them separately, suggests sponsors want this passed as a package before individual provisions can be stripped out in conference.
Sources: https://www.scmp.com/news/us/article/3360729/us-hardens-ai-stance-china-anthropic-calls-extending-lead | https://www.whitehouse.gov/releases/2026/07/white-house-launches-gold-eagle-initiative-for-unprecedented-cybersecurity-vulnerability-coordination/ | https://www.defenseone.com/technology/2026/07/ai-once-relegated-helping-hackers-certain-tasks-can-now-power-every-stage-cyberattack/414749/
15-Jul-26
Iran war reshapes Gulf calculus on AI alignment with Washington
The Center for Strategic & International Studies (CSIS), a prestigious US-based think tank, has released an insightful commentary on the US war against Iran and its impact on AI cooperation. Specifically, CSIS Middle East analyst Joseph Farsakh wrote on July 15 that the Iran war has begun to change Gulf states’ cost-benefit calculation on deep AI alignment with the United States. In the last few months, Iranian drone strikes hit three AWS data centers, two in the UAE and one in Bahrain, disrupting banking and consumer services across a 50-million-person region and prompting AWS to advise clients to consider migrating workloads out of the Middle East. Iran separately published a list of 29 technology targets across Bahrain, the UAE and Israel, naming AWS, Microsoft, Google, Nvidia, Oracle and Palantir facilities as legitimate military infrastructure. Saudi Arabia, Qatar and the UAE had collectively committed roughly $2.5 trillion to US-linked technology investment before the war began.
Why it matters
Gulf governments wanted data sovereignty. What they got was closer to the opposite. The same AWS infrastructure hosting their banking and civil services was also processing US military targeting data. Iran then treated that infrastructure as a legitimate wartime target, naming specific facilities on a published strike list. The lesson is not about AWS specifically: hosting AI infrastructure inside a US-aligned security architecture pulls a country into wars it did not choose. That’s true even if Iran’s assertion that these facilities processed US military data turns out to be false; the perception is what matters, not a reality subject to factual disputes.
Farsakh’s cross-country comparison supports this. The UAE signed the Abraham Accords and integrated with Israeli intelligence, and absorbed the heaviest strikes. Qatar and Kuwait kept more hedged postures and were targeted less. The likely Gulf response is not divestment but diversification: continued AI capital deployment, spread across more jurisdictions, with less explicit security entanglement than deals like MGX and HUMAIN carried over the past year. Source: https://www.csis.org/analysis/what-iran-war-costing-joint-gulf-us-ambitions-ai
15-Jul-26
Japan and Australia stake out sovereign AI positions distinct from the US-China race
Japan’s Cabinet approved its Second AI Basic Plan on July 14, organized around what the government calls “AI transformation” (AX) and focused on four pillars:
expanding AI use across government and industry,
building domestic AI development capacity including compute and data resources,
strengthening AI governance and cybersecurity; and,
redefining workforce roles as AI adoption spreads.
Separately, Australian Prime Minister Anthony Albanese announced the creation of an Office of AI within his own department on July 15, in a speech titled “AI in Australia’s Interests” delivered at the University of Sydney. The office will set national AI standards and streamline regulatory approval for AI projects. Albanese’s government will also legislate binding data center rules requiring large facilities to become net generators of renewable energy and to bear their full grid-connection costs, with legislation expected in early 2027. Albanese said Australia should use AI to become “more resilient and more sovereign,” warning against “subcontracting our national sovereignty and security to the control of foreign monopolies.”
Why it matters
Neither Japan nor Australia is trying to beat the US or China on frontier models. Both are making a narrower argument instead: AI policy is a sovereignty question, separate from who wins the capability race.
Japan’s AX plan continues the logic behind its Noetra robotics consortium. AI adoption answers a labor shortage. It is not a move in a geopolitical contest.
Australia’s motivation is more defensive. Albanese’s “foreign monopolies” language targets US hyperscalers, not China. That is a notable choice for a Five Eyes ally that otherwise aligns closely with Washington on export controls.
The renewable energy rule is the more salient policy here. It forces AI infrastructure operators to pay their own grid costs, instead of passing them to households. Ireland, Singapore, and other capacity-constrained markets are likely candidates to copy it.
Both governments are betting the same way: domestic standards and targeted state investment beat full dependence on a superpower’s AI stack. Neither has the money to build independent frontier capability. This is the strategy available to them, not necessarily the strategy they would choose freely.
Sources: https://sustainablejapan.jp/2026/07/15/japan-ax/128124 | https://www.pm.gov.au/media/ai-australias-interests-0
14-Jul-26
Battlefield AI moves to the edge in the Baltics and the Pacific, with help from Nokia
European defense AI lab NestAI announced on July 14 that it has begun training sovereign frontier models for battlefield autonomy and command orchestration. The Finnish and Estonian Defence Forces are piloting the first models under a trilateral letter of intent signed in June. NestAI, founded in 2025 and backed by 100 million Euros from Nokia and Tesi, has grown to 200 engineers and scientists and trains models using real operational data from its NestOS platform. Nokia separately announced integration of its private 5G networks with NestOS to support AI command-and-control in denied environments. In the Pacific, Pentagon Chief Digital and AI Officer Cameron Stanley said at a July 2026 AWS Summit that AI workflows built for Special Operations Command Pacific, including a program that already supported Operation Epic Fury against Iran, are being adapted to move AI compute closer to troops. The aim is to address connectivity constraints facing INDOPACOM. The Special Operations Command Pacific (SOCPAC) said the effort centers on smaller, battalion-scale compute clusters rather than the division-scale nodes originally built for the region.
Why it matters
Both NestAI and the US military’s choice to move compute further into the field aim to solve the same underlying problem: getting usable AI to forces operating without reliable connectivity to centralized compute. NestAI’s pitch is explicitly sovereignty-driven, a direct response to the risk that a foreign model provider could cut off access overnight, the same exposure the CSIS Gulf item elsewhere in this issue documents happening to US-hosted AI infrastructure in a live conflict. The Pentagon’s motivation is speed rather than sovereignty. The SOCPAC team measures its adversary’s data-to-decision cycle explicitly against China’s. The fact that Indo-Pacom’s existing infrastructure could not meet that benchmark is worrisome given how heavily the Pentagon has invested in enterprise AI over the past two years. Both efforts face a common problem, though: model reliability at the edge: NestAI is building toward auditable, bounded autonomy that commanders can be accountable for, while the Pentagon’s approach leans on shrinking model size to fit power and bandwidth constraints. Neither has published performance data proving that edge-deployed models match the reliability of their cloud-hosted counterparts. This data may determine how much autonomy either the US or European militaries are actually willing to grant these systems in practice.
Sources: https://www.nestai.com/blog-posts/nestai-builds-sovereign-ai-for-european-defence | https://www.nokia.com/newsroom/nokia-and-nestai-build-capability-for-ai-enabled-defense-operations-with-resilient-connectivity-in-denied-environments/ | https://www.defenseone.com/technology/2026/07/how-former-marine-rewriting-future-battlefield-ai/414758/
Questions about this post? Contact the author: matt@mtn-c.com
Questions about how to work with InfraVantage? Contact Matt, or InfraVantage co-founder Tim Lin: tlin@digitalinfraco.asia.