Sovereign AI gives countries local control over sensitive data, models, applications, and compute instead of relying entirely on foreign providers.
2
The time and cost of building AI infrastructure depend on its scale, chip requirements, power supply, and whether an existing data center can be repurposed.
3
US protectionism is pushing governments and companies to find local and regional AI providers, creating opportunities for smaller teams and software companies outside the US.
Summary
Frank Meehan argues that sovereign AI has become a national priority because governments want sensitive defense, health, finance, and agriculture data to stay inside their borders. He contrasts large, liquid-cooled facilities that can take 18 to 24 months with smaller inference-focused deployments that can be built by repurposing existing data centers. Governments need to coordinate with telecoms, infrastructure providers, local companies, research institutions, and corporate customers before committing to GPUs and power. Meehan expects regional cooperation, especially when a country has excess low-cost power or more compute capacity than its domestic market can use. He also sees US protectionism accelerating demand for local software and infrastructure. That shift could help small teams build applications for local problems, test them quickly, and expand regionally. The conversation also covers the risks of foreign control over critical AI systems and Meehan's earlier experience investing in DeepMind and building Siri before Apple's acquisition.
Sensitive national data needs local AI infrastructure
Frank Meehan says governments want to place defense, finance, agriculture, and health data into AI models without sending it to the US, China, or neighboring countries. He argues that regional cloud hubs do not provide enough localization because a country may still be using infrastructure located elsewhere. Recent political developments have made governments less willing to rely on large foreign providers. Sovereign AI therefore includes local models, applications, security controls, and infrastructure. Meehan describes this as a major market outside the US because governments need AI systems they can control and secure within their own borders.
Infrastructure timelines change sharply with scale and workload
Meehan separates large training facilities from smaller deployments focused mainly on inference. A new liquid-cooled facility using the latest Nvidia Blackwell chips may take 18 to 24 months and requires specialized partners, chip access, power, and cloud management. He contrasts that with Indonesia, where an existing Hutch Indo data center was upgraded for H200 and H100 chips. Nvidia helped with the work, and the facility became operational in four months. Smaller five- to ten-megawatt sites can fit the needs of countries that do not need to train the largest models. The right design depends on local demand and the workload.
Each country needs a local operating and financing model
Frontier One AI works with governments to define the requirements for each market. Meehan says projects need a local company, local talent, infrastructure partners, customers, and a financing plan. A government might fund all the facility, fund part of it, or guarantee an agreed share of the compute. He gives a typical range for markets in Asia, the Middle East, and Latin America: a five- to ten-megawatt facility, 4,000 to 7,000 GPUs, and $100 million to $200 million. The model also has to account for remote locations, limited power, heat, dust, and the applications that customers will actually run.
Secure government AI should still give local companies access
Meehan says governments want to use AI to improve systems such as healthcare, while allowing local startups to build applications on top of protected data. A health system might provide controlled access to a company digitizing health records, but the data must remain secure and local. He distinguishes this from anonymized data used for research, where wider access can help researchers study rare diseases. The sensitive case is an internal system that accesses an identifiable person's records, models them, and returns a result. That requires local controls around the model, vector database, cybersecurity platform, and application.
US protectionism is creating demand for local and regional AI suppliers
Meehan argues that US political decisions have pushed countries toward localization across trade, software, finance, defense, and government systems. He says some German banks do not want US companies in their technology stack, including GitHub. This creates room for local companies to provide parts of the AI stack and for regional firms to solve requirements such as language and regulation. He expects finance, defense, and government to move first because those sectors are especially sensitive to foreign dependence. In his view, the change is opening opportunities for developers and companies outside the US.
Small teams can build and test local products faster
Meehan says AI tools let small teams prototype products with more functionality and much faster feedback than the older venture model, where a company raised millions and spent nine to twelve months building an MVP. Engineers can keep their jobs while testing a product and leave only when the idea shows traction. He cites Lovable and Cursor as examples of product-led companies that can scale without large sales and marketing teams. He also thinks venture capital firms should evaluate a small group that has already shipped and is growing, rather than requiring a large organization before investing.
Regional compute markets will depend heavily on power
Meehan expects countries to share excess AI capacity across regions when their domestic demand is too small for a large facility. France could make a large site available to other European countries while reserving some racks for highly local workloads. He thinks regional cooperation will vary across Europe, Latin America, Africa, and Asia. The deciding factor may be cheap and available power. He points to Bhutan's unused hydroelectric potential and El Salvador's geothermal ambitions as examples of places that could attract data centers. Governments still need to guarantee enough demand to make the investment financeable.
DeepMind and Siri show the value of backing deep technology early
Meehan recalls investing in DeepMind around 2011 while working with Horizon Ventures in Hong Kong. He says the investors focused on long-term deep technology when most funding was going toward social networks, even though the business model was unclear. He also describes Siri as a research team from SRI International that launched on the App Store in 2010 with functions such as reservations and cinema bookings. Steve Jobs saw the product shortly after launch, and Apple acquired it four weeks later for $200 million. Meehan uses these stories to argue that investors should recognize technical opportunities before conventional revenue signals appear.
"If I'm a government and I want to deploy AI into a model, I want to put my data, my defense data, my finance data, my agriculture data, my health data into a model, that data cannot go to the US. It cannot go to China. It cannot go to my neighbor. It has to be in my country."Frank Meehan03:30
Who should watch
You are designing AI infrastructure for a government or regulated organization and need to decide what must remain local.
Your country or company is considering a GPU facility and you need to think through power, customers, financing, workload, and regional demand.
You are building an AI product with a small team and want to understand how local problems, fast prototyping, and regional expansion can replace the older venture path.
