South Korea's $597bn AI budget meets the energy and water wall
South Korea is proposing a record national budget of $597 billion for 2027, with AI infrastructure at its centre. A semiconductor boom driven by global AI demand has reportedly doubled the country's tax revenues, giving Seoul the fiscal headroom to launch a $118 billion Future Response Fund earmarked for GPU procurement, AI data centres, and subsidised AI access for citizens. The plan represents one of the most explicit state-level commitments to AI-led economic transformation seen from any mid-sized economy, and it is already running into the hardest limits in industrial infrastructure.
One planned chip cluster alone is projected to require electricity equivalent to 80% of its host region's entire current annual consumption. Water demand for cooling at scale adds a second physical constraint that no amount of sovereign capital can immediately resolve. The episode is a sharp illustration of a pattern now visible across AI buildouts globally: digital ambition consistently outruns grid and water capacity, forcing planners to treat energy and utility infrastructure as a first-order AI policy problem rather than a downstream one.
Manus AI and the reach of Chinese regulation
Running parallel to Seoul's fiscal manoeuvre is a story with direct implications for any cross-border AI investment thesis involving Chinese-founded companies. Meta's roughly $2 billion acquisition of Manus AI, the agentic AI startup that attracted considerable attention for its autonomous task-completion capabilities, has been unwound after Chinese regulators blocked the transaction months after it closed. Manus was required to sever data ties with Meta and delete shared customer information.
The episode carries a pointed message for capital allocators: a Singapore headquarters does not place a Chinese-founded company beyond Beijing's regulatory reach. For investors underwriting AI acquisitions across the US-China technology boundary, the Manus unwind is a live case study in jurisdiction risk, the kind of due-diligence variable that sovereign wealth funds and cross-border PE houses are now pricing explicitly into deal structures, particularly for agentic AI assets where data sovereignty is intrinsic to the product.
The convergence read-across
Both stories connect to a single macro trend: the AI infrastructure layer is becoming a geopolitical asset class. South Korea's move sits within a broader APAC sovereign-capital race, Japan's fumds for domestic semiconductor capacity, India's GPU subsidy scheme, and Gulf state data-centre buildouts all reflect the same structural shift: governments treating compute as a strategic reserve, not a market outcome.
The energy constraint Seoul faces is not unique to Korea. Every large-scale AI compute cluster under construction globally, from Northern Virginia to Riyadh to Osaka, is generating similar grid stress. The capital implication is significant: investors who have been underwriting AI infrastructure plays purely on compute demand need to model energy and water access as a binding constraint on delivery timelines, not an operating cost line. Utilities, grid-scale storage, and water-recycling infrastructure are quietly becoming beneficiaries of the AI capex cycle, even when they appear in no AI-themed fund mandate.
The Manus episode adds a regulatory-risk dimension that compounds the infrastructure challenge. As agentic AI becomes the dominant paradigm for enterprise automation, a theme visible across the Asia Tech Podcast's broader episode, covering everything from AI-assisted legal work to AI-managed bike-sharing logistics, the question of which regulatory regime controls an agent's data and decision-making will increasingly determine where investment can safely flow. The combination of physical constraints and jurisdictional complexity suggests the next phase of AI infrastructure investment will be far less fungible across geographies than the first wave of cloud buildout was.