Sovereign AI: Nation-States Are Now Playing at the Table

May 20, Singapore. Opening day of ATxSG (Asia Tech x Singapore).

Three things happened at once:

$OpenAI(OpenAI.NA) OpenAI announced its first Applied AI Lab outside the US, right here in Singapore — committing US$234 million.
$NVIDIA(NVDA.US) NVIDIA announced a research center plus a physical AI testbed in Singapore.
$Alphabet(GOOGL.US) Google signed a new round of national-level AI partnership with the Singapore government.

Three US AI giants. One city. One day. Synchronized moves.

This isn't a coincidence — it's an era surfacing into view. The name of this era is Sovereign AI.

You might be asking: haven't governments always been doing AI? What's different this time? The answer — previously, governments were users of AI. Now, they're becoming AI infrastructure builders.

It's not "AI built by the government"

The most common misunderstanding: Sovereign AI is not about "the government developing its own model."

If a country builds its own large model but it runs on US cloud, on US-made GPUs, with data stored on US servers — it's still not "sovereign." It's just a technical achievement.

What Sovereign AI actually defines is three layers:

Data sovereignty — critical national data (government records, medical files, financial transactions, citizen PII) must sit on domestic infrastructure
Compute sovereignty — critical AI workloads run on the country's own soil, its own power grid, its own data centers
Model sovereignty — not entirely dependent on services from a handful of US firms like OpenAI, Anthropic, or Google

An analogy: it's like the shift from "using American telephones" to "building your own telephone network." Long-distance calls used to route through AT&T; now every country has realized it needs its own switches, its own copper lines, its own base stations.

What Singapore announced today is a textbook case. Getting OpenAI to set up a lab locally isn't about helping OpenAI sell more ChatGPT in Singapore — it's about ensuring that "how to use AI" actually happens on Singapore soil.

"Nobody needs an atomic bomb, but everyone needs AI"

That line is from Jensen Huang — said at the 2026 World Government Summit in Dubai, to an audience of leaders from dozens of countries.

For the past 70 years, no single technology has gotten a national leader to hear a line like that without pushing back. But when this one landed, the room went silent — because it captured precisely what's been happening between 2024 and 2026.

Four forces have come together to push this moment forward.

AI shifted from "software product" to "infrastructure." Ten years ago AI was an app — install it, uninstall it, whatever. Now it's embedded in power grid dispatch, medical diagnostics, military command, financial transactions. Once it's at those layers, it becomes national infrastructure — like electricity. You can't outsource your power grid to a foreign company.

Hyperscaler concentration became a risk too. Over 80% of global AI compute runs on three US firms: Microsoft Azure, Google Cloud, AWS. Every country's critical workloads end up on these three. That level of dependency has no precedent in the pre-AI era.

US export controls ironically pushed countries to build their own. Since 2023, restrictions on high-end GPUs have tightened — Saudi Arabia, UAE, India, and parts of Southeast Asia are now on the "license required" list. The irony: the more the US controls, the more these countries want to build domestically. If you can't be sure you'll get the chips, you stock up and DIY.

Data sovereignty laws are tightening too. EU GDPR, China's Data Security Law, India's DPDP Act, Singapore's PDPA — more and more laws explicitly require that critical data not leave the country. Data can't cross borders = models have to be trained locally.

Add these four forces together, and starting in 2024, Sovereign AI moved from a concept to a hard line item in national budgets.

Who's at the table?

The Middle East duo are the loudest players.

Saudi Arabia — In May, during Crown Prince Mohammed bin Salman's (MBS) Washington visit, the US-Saudi AI investment commitment was lifted from US$600 billion to US$1 trillion. Behind it: PIF (Saudi's US$1 trillion sovereign wealth fund) plus the newly established national AI company HUMAIN. NVIDIA struck a deal with Saudi to deploy a 5,000-unit Blackwell GPU "AI factory." HUMAIN's target is wild — to supply 6% of global AI compute by 2034.

UAE — Abu Dhabi has G42 (a state-backed AI company founded in 2018). The Stargate UAE project, joining NVIDIA, Oracle, Cisco, and SoftBank, is building a 1 GW data center cluster. Microsoft took a US$1.5 billion stake in G42.

Singapore is the Southeast Asia hub. Today's announcements — the OpenAI lab, NVIDIA research center, the ASPIRE 2B national supercomputer expansion — layered on top of the Smart Nation 2.0 national strategy, are turning Singapore into the clearing house for AI compute in the region. Every ASEAN country that wants AI but lacks the infrastructure to build its own will route through the Lion City.

The second tier is already in the game: India (the US$1.2 billion India AI Mission), Japan (NEC + Sakana + SoftBank), France (Mistral + government), Canada, the UK — every one of them is putting money down to build their own compute stack.

Honestly, this is no longer a bipolar "US vs China" game. It's becoming a multi-polar geopolitical map.

Three ceilings

The story flows nicely, but the costs are real.

Money — Building a 1 GW data center, deploying tens of thousands of GPUs, training the engineers — you're looking at US$10 billion just to start. That's not a game mid-tier countries can play. Which is why the ones actually at the table are oil producers, city-state economies, or large population-dividend countries.

Power — A 1 GW data center equals the total electricity consumption of a mid-sized city. Singapore is so land-constrained that even the grid is tight, which is why it passed the Digital Infrastructure Act specifically for data centers, requiring all new builds to hit energy efficiency standards.

Chips — This is the most delicate layer. Even the most ambitious Sovereign AI plan still has to buy NVIDIA GPUs, use TSMC for fab, and deploy Broadcom networking. What you can make sovereign is the geographic location of the data center, the grid access, the operating rights — but for chip sovereignty itself, only the US and Taiwan get a say.

Strictly speaking, all "Sovereign AI" today is really "semi-sovereign." One US export license revocation, and the whole plan can stall. Saudi Arabia and UAE have already tasted this in 2024 — the US temporarily restricted high-end GPU exports to the Middle East, and G42 had to publicly commit to "no collaboration with China-linked AI companies" before getting its license.

Back to that Singapore scene

OpenAI setting up a lab in Singapore and committing US$234 million isn't a unilateral expansion decision by OpenAI — it's OpenAI putting real money on the table to acknowledge to the Singapore government: "We accept that you need to own how AI gets deployed on your own soil."

This is what the Sovereign AI era is changing. US AI giants are no longer untouchable suppliers — they're starting to become "contractors" within each country's sovereign AI strategy.

Ten years ago, globalization was about goods. Five years ago, it was about data. Now, globalization is about sovereign infrastructure nested into each other.

The nation-state players are at the table. This game is too big — not even the hyperscalers can eat it all.