Kazakhstan’s international arrangements and Singapore’s cooperation with Anthropic show how AI partnerships address infrastructure, technology access, testing and domestic skills.
What countries seek from AI cooperation
International AI partnerships offer countries routes to computing, supply-chain cooperation, research, governance and skills. Kazakhstan’s participation in separate US-led and China-initiated arrangements, alongside Singapore’s cooperation with Anthropic, illustrates that range. These announcements describe commitments and objectives; they do not, by themselves, establish delivered investment or technological independence.
This is analysis of ongoing developments as at 25 October 2026, not a new agreement between the United States and China. An Associated Press analysis published that day, also republished by WPXI, examines the wider partnership landscape. Those publications carry the same reporting, rather than two independent accounts.
International AI partnerships: Kazakhstan’s two tracks
Kazakhstan’s ministry records its accession to Pax Silica in Washington on 25 June 2026. Zhaslan Madiyev signed the accession declaration and a joint statement on the AI Opportunity Partnership.
Less than a month later, Kazakhstan appeared among the signatories to a different arrangement. China’s foreign ministry says representatives of 29 countries signed the agreement establishing the World Artificial Intelligence Cooperation Organization, or WAICO, in Shanghai on 16 July 2026. Laos, Pakistan, Russia and Indonesia were also among the signatories.
The Chinese ministry describes WAICO as an independent intergovernmental organisation headquartered in Shanghai. Its stated purposes include international AI cooperation and global governance, with beneficial, safe and fair development grounded in the purposes of the UN Charter.
The dates matter: these were June and July events, not agreements first signed on 25 October. Kazakhstan’s participation documents two partnership tracks. It does not establish a joint US–China initiative, nor demonstrate what either arrangement has delivered to participants.
The arrangements also have different stated emphases. Kazakhstan’s account of Pax Silica highlights supply chains and production. China’s description of WAICO emphasises an intergovernmental organisation concerned with cooperation and governance. Membership alone does not explain how those ambitions will translate into domestic projects.
Supply chains and the ambition to build local capability
Kazakhstan describes Pax Silica cooperation as covering critical minerals, semiconductors, data centres, energy, manufacturing, research and talent. Its stated ambition is to move beyond raw-material exports towards processing and higher-value technology production.
The AI Opportunity Partnership identifies three priorities: innovation-friendly regulation; cooperation across research, energy, computing, minerals and talent; and private-sector mobilisation. The latter includes venture investment, research partnerships, data centres, computing access and trusted models.
Possible educational cooperation includes computational materials science, AI in mining and processing, data security, data-centre engineering and industrial AI. These are prospective programme areas, not evidence of completed training.
These categories distinguish physical equipment from the software it runs. Hardware includes semiconductors and computing equipment; models are software systems used to perform AI tasks. Technology access can involve both, but access to one does not establish access to every other component.
Local capability is therefore a broader objective than obtaining equipment. In Kazakhstan’s account, it also encompasses research, engineering and industrial knowledge. The educational topics connect AI to existing industrial activities, including mining and processing, rather than only to developing general-purpose models.
The announcement sets out intended areas of cooperation, rather than a record of completed production or skills development. It supplies no basis for comparing the performance of particular chips or models. Its significance lies in the breadth of the stated ambition: linking resources, computing and expertise within the same cooperation agenda.
Data centres turn partnership plans into infrastructure choices
Kazakhstan’s infrastructure announcements illustrate why operating equipment must be distinguished from proposed capacity. In its account of an NVIDIA meeting on 1 October 2026, the presidency said two NVIDIA-based supercomputing clusters had launched in the country.
The same account described Data Center Valley as envisaging approximately 300 megawatts of computing infrastructure, with potential expansion to one gigawatt. That is a description of infrastructure ambitions, not confirmation that the proposed campus or its full expansion was operating.
A separate announcement has a different scope. On 28 August 2026, KAZAKH INVEST described discussions with Singapore-based Energy Capital Global about an AI data-centre campus at the Ekibastuz site. A first phase of approximately 100 MW of IT load, potentially expanding to 1 GW, was under consideration. ECG and KT-Telecom signed a memorandum of understanding.
These figures should not be added together as though they describe separate, fully confirmed construction phases. The August proposal and October infrastructure description concern different scopes. Their shared reference to possible 1 GW expansion does not resolve how the plans relate or establish completed construction.
The units also need care. A megawatt measures power, while one gigawatt equals 1,000 megawatts. Neither unit directly measures the speed, accuracy or usefulness of an AI model. A figure for IT load is not a count of installed processors.
Three different stages appear in these accounts: a memorandum recording cooperation, a proposed campus with capacity under consideration, and clusters that the presidency says had launched. Keeping those stages separate gives a clearer picture than treating every announcement as operational computing.
The accounts do not establish a single completed investment programme. Nor do the proposed power figures demonstrate a technological breakthrough. The presidency’s statement about launched clusters is narrower than a claim that all of Data Center Valley is running.
Singapore’s partnership extends to testing and skills
Singapore provides a different example of international AI partnerships. On 14 October 2026, Anthropic announced its Singapore office and a memorandum of understanding with the Ministry of Digital Development and Information. The company described it as its fifth Asia-Pacific office.
According to Anthropic, the memorandum covers AI safety, security and model evaluation; research into adoption and its effects; and responsible adoption in public and private organisations. Planned technical exchanges include testing standards and incident response.
That scope differs from a data-centre capacity announcement. It concerns how systems are assessed and used, as well as how institutions develop the knowledge needed to work with them. The memorandum does not itself report measured adoption benefits.
Anthropic traces its Singapore evaluation work to a sandbox in 2023 and a subsequent red-teaming project conducted in English, Mandarin, Malay and Tamil. Red-teaming means deliberately testing a system for weaknesses. The company’s account identifies the languages involved, without establishing that all possible weaknesses were found or resolved.
The announcement also identifies prospective skills work, support for smaller businesses and capability development among public officers. These planned activities should be distinguished from the earlier evaluation work. An office opening, a previous testing project and future cooperation are different developments, even when described in the same announcement.
Technology access and the limits of AI independence
The strategic question is whether partnerships can provide technology access without leaving countries unable to act independently. Anton Leicht and Sam Winter-Levy address that tension in a Carnegie paper dated 28 September 2026.
The authors argue that middle powers such as Germany and India face difficulties developing frontier AI independently, while also facing risks from dependence on the United States or China. Frontier AI refers here to systems at the leading edge of development.
Their proposed response is to maintain capacity for a concentrated frontier-development effort while hoping not to need it. That is a strategic recommendation, not evidence that the countries discussed have already built an independent alternative.
The paper’s opening identifies security-driven restrictions on access to leading models and computing shortages that constrain customer usage. These are the authors’ analytical findings. They should not be read as proof that every country has lost access to advanced systems.
China’s conference chair’s statement presents WAICO as intended to help Global South countries develop AI innovation, applications and governance capabilities. It says the organisation should be open to all countries and communicate with other international organisations about possible cooperation. Those are stated intentions, not demonstrated outcomes.
AI independence is therefore a separate question from membership or access. Kazakhstan’s announcements combine supply-chain objectives with infrastructure ambitions; Singapore’s example includes testing and skills. Neither establishes that cooperation removes dependence on external equipment or software.
To assess international AI partnerships, distinguish what has been signed, what participants say is operating and what remains proposed. Technology access and local capability are related aims, but neither follows automatically from an agreement. The next meaningful evidence is delivery: infrastructure brought into use, research undertaken and skills programmes completed.
Questions about international AI cooperation
Is this a new AI agreement between the United States and China?
No. This analysis concerns separate international arrangements, including Kazakhstan’s participation in Pax Silica and the agreement establishing the World Artificial Intelligence Cooperation Organization.
When did Kazakhstan join the two arrangements?
Kazakhstan’s ministry records its Pax Silica accession on 25 June 2026. China’s foreign ministry records Kazakhstan among the signatories to the WAICO establishment agreement on 16 July 2026.
What does Kazakhstan say Pax Silica cooperation covers?
Kazakhstan describes cooperation spanning critical minerals, semiconductors, data centres, energy, manufacturing, research and talent.
Do the announcements establish completed local training programmes?
No. Kazakhstan’s ministry identifies potential educational cooperation, including data-centre engineering and industrial AI, but does not establish completed training outcomes.
Is Kazakhstan’s proposed 1 GW data-centre expansion already operating?
The announcements do not establish that. Kazakhstan’s presidency describes potential expansion separately from its statement that two NVIDIA-based supercomputing clusters had launched.
What does Anthropic say its Singapore memorandum covers?
Anthropic says it covers AI safety, security, model evaluation, adoption research and responsible adoption, with planned technical exchanges on testing standards and incident response.
Which languages featured in Anthropic’s earlier Singapore red-teaming work?
Anthropic identifies English, Mandarin, Malay and Tamil. Red-teaming means deliberately testing a system for weaknesses.
Do international partnerships guarantee AI independence?
No such guarantee is established. Carnegie authors Anton Leicht and Sam Winter-Levy describe both the difficulty of independent frontier AI development and the risks of relying on external suppliers.