The Sovereign AI Race: How Riyadh, Abu Dhabi and Singapore Are Building National Models

On 13 May 2025, at the U.S.-Saudi Investment Forum in Riyadh, HUMAIN, the four-day-old AI subsidiary of Saudi Arabia’s Public Investment Fund, signed a deal with NVIDIA for an initial 18,000 GB300 Grace Blackwell superchips and a forward pipeline of “several hundred thousand” more. Within six months the company had layered on a $10 billion AMD partnership for 500 megawatts of MI-series compute, an AWS cloud build, contracts with Cisco and Qualcomm, and a target of 600,000 NVIDIA units deployed over three years. HUMAIN existed on paper for less than a week before it became one of the three largest GPU buyers on earth. The speed is the story. Three states, Saudi Arabia, the United Arab Emirates and Singapore, have moved decisively from the rhetoric of “sovereign AI” into the harder business of buying it, and the absence of a credible European entrant has become the defining strategic fact of the decade.

Riyadh’s Stack. HUMAIN is the corporate vehicle for what Saudi Arabia now treats as a $70 billion AI capital programme, sequenced alongside the broader Vision 2030 industrial portfolio. The model layer is ALLaM 34B, an Arabic-first foundation model trained on more than 500 billion Arabic tokens with native handling of Saudi, Egyptian, Jordanian and Levantine dialects; it powers HUMAIN Chat, launched in 2025 as a sovereign alternative to ChatGPT for Gulf government and enterprise workloads. Saudi pilots already include Sawtak, an in-court transcription system. The infrastructure layer is more ambitious than the model. The first GB300 supercomputer alone, before any of the AMD megawatts come online, places HUMAIN’s compute footprint above all but a handful of Western hyperscalers. PIF is funding this directly, which means the Saudi state is short-circuiting the usual venture-capital, GPU-allocation and export-licence frictions that have constrained every European attempt to do the same thing. The licences themselves were granted under the post-AI-Diffusion-Rule framework negotiated with Washington in spring 2025, which formally elevated Riyadh and Abu Dhabi to tier-one customer status for advanced US accelerators.

Abu Dhabi’s Two Pincers. The UAE’s bet is structured around two distinct vehicles. G42, chaired by Sheikh Tahnoon bin Zayed Al Nahyan, took a $1.5 billion minority investment from Microsoft in April 2024, divested its declared Chinese-technology holdings as the price of the deal, and now functions as the operating arm. The Technology Innovation Institute supplies the model IP, principally the Falcon series, with Falcon 3 trained on 14 trillion tokens and a Falcon 4 multimodal generation released through 2025 covering vision and OCR. The second pincer is capital. MGX Fund Management, established by Abu Dhabi in 2024, targets $100 billion in assets, plans roughly $10 billion of annual deployment and is a founding partner in the global Stargate consortium alongside OpenAI, Oracle and SoftBank. Stargate UAE, announced May 2025, will sit on a 10-square-mile, 5 gigawatt US-UAE AI campus in Abu Dhabi, with a 1 GW cluster built by G42 and operated by OpenAI and Oracle. The first 200 MW phase comes online in 2026. The numbers attached to it run to roughly $20 billion for the 1 GW build alone. By any measure other than the US itself, this is the largest single AI infrastructure project on earth, and the most expensive non-American AI bet ever placed.

Singapore’s Inversion. Singapore has chosen the opposite posture. Rather than chase capex parity with the Gulf, the city-state is positioning itself as the neutral compliance, inference and Southeast Asian linguistic layer that the duopoly cannot serve. National AI Strategy 2.0, launched in late 2023, anchors a programme worth more than S$1 billion over five years, including a S$500 million high-performance compute allocation and a S$270 million classical-plus-quantum supercomputer due online late 2025. Smart Nation 2.0, announced October 2024, adds a S$120 million “AI for Science” envelope. The model is SEA-LION, now in its Qwen-Sea-Lion-v4 iteration after migrating in late 2025 from a Meta Llama base to Alibaba’s Qwen3-32B, trained on more than 1 trillion tokens across 13 Southeast Asian languages and dialects. It is the leading open-source model under 200 billion parameters on the regional SEA-HELM benchmark. The AI Verify Foundation, the world’s first government-backed open-source AI testing toolkit, is the regulatory complement, designed to be exported to ASEAN partners as a standard. Singapore is not trying to outspend HUMAIN. It is trying to make every export licence and compliance audit east of Suez pass through its jurisdiction.

The Asian Periphery. Around these three poles, three more programmes deserve serious attention. South Korea launched a sovereign foundation-model competition in 2025 with a ₩530 billion envelope, selecting five consortia, LG AI Research, SK Telecom, Naver, NC AI and Upstage. LG’s K-EXAONE, a 236-billion-parameter model, headlines the field; the government intends to narrow the field to two finalists by 2027. Japan’s bet is more diffuse but better funded over time, anchored on the NEDO GENIAC programme that provides government compute to Sakana AI, Preferred Networks and others. Sakana itself raised a $135 million Series B in November 2025 at a $2.65 billion valuation, an explicit pitch on “sovereign AI solutions that reflect national cultures and values” for finance and defence. India is operating at a smaller unit cost but a similar ambition. The IndiaAI Mission allocates ₹10,372 crore, roughly $1.25 billion, over five years; the IndiaAI Compute Portal already reports 38,231 GPUs deployed, against a 100,000-GPU target by end-2026. Sarvam AI alone has been allocated 4,096 H100s with a ₹98.7 crore subsidy to train a 70-billion-parameter Indic model.

The European Silence. Set against this, the European position is unflattering. The Commission has earmarked €20 billion of public funding for up to five AI gigafactories within a broader €200 billion InvestAI ambition. Mistral, the only credible French frontier-model house, has confirmed a €1 billion capex plan for 2026, a €1.2 billion Swedish data centre and a Paris facility built on 13,800 NVIDIA chips. Aleph Alpha, the German equivalent, has effectively conceded the foundation-model layer and is in merger talks with Cohere on terms valuing the combined entity near $20 billion, a transatlantic structure that quietly admits European stand-alone is no longer the plan. The EU AI Act, fully in force in 2026, imposes compliance overhead that Riyadh, Abu Dhabi and Singapore have explicitly chosen not to mirror. The aggregate European sovereign-AI spend, weighted for execution risk and timeline slip, is not credible against a single HUMAIN-NVIDIA contract, let alone Stargate UAE. The German political economy, in particular, has produced no equivalent to PIF, MGX or Temasek; the necessary capital simply does not move at the necessary speed.

The Verdict. Sovereign AI has become the most coherent industrial-policy convergence of the decade, and it has done so without the United States, the European Union or China formally orchestrating it. The Gulf is buying its way into the third pole of compute through a willingness to deploy state capital at hyperscaler velocity, accept American security framings on Chinese hardware, and let Microsoft, OpenAI and NVIDIA operate inside national champions. Singapore is buying its way into the regulatory and linguistic layer that the third pole will need. Korea, Japan and India are running the same play at smaller scale but with greater domestic depth. The strategic objective in every case is identical: compute on national soil, models trained on national data, behaviour aligned to national priorities, and export licences in national hands. Europe’s absence from this list is no longer a gap that can be closed by a gigafactory press release. By the time the first €20 billion in EU public capital lands on the runway, HUMAIN and G42 will already have deployed more accelerators than every European AI lab combined. The third pole is being built. The question for Brussels and Berlin is whether they intend to be customers of it, or simply spectators.

Read our full Report Disclaimer.