A Singapore AI-Sovereignty Roadmap

SG_AI_Sovereignty_Three_Phase_Strategy_Parliamentary.pdf

Download PDF • 296KB

ONE PERSON'S SUGGESTED POLICY PAPER

From Platform Tenant to Productive Owner: A Three-Phase Strategy for Singapore's AI Sovereignty

Foundation, Activation, Sovereignty — and the Energy Base That Must Carry Them

Tat Yuen  ·  Bixbe Consulting  ·  July 2026

Executive Summary

Mr Speaker, Sir — I set out in this paper a case that this House will, I believe, find increasingly difficult to set aside: that Singapore's transition into the AI economy will be determined not by how quickly our businesses adopt artificial intelligence, but by whether they come to own any productive stake in it. This is a distinction with material consequence. Adoption without ownership is a well-run tenancy. It is not sovereignty.

This paper sets out a three-phase national strategy — Foundation, Activation, Sovereignty — running from 2027 to 2040, and argues for two structural commitments that this House should regard as inseparable from that strategy rather than adjacent to it. The first is the AI Workflow Cooperative, a tripartite mechanism through which Singapore's small and medium enterprises move from consuming AI to co-producing it. The second is a deliberate, three-part model sourcing portfolio — combining home-grown distilled models, open-source foundation models, and licensed access to leading foreign frontier models — designed explicitly to prevent the vendor lock-in that would otherwise recreate, under a different name, the very dependency this strategy is meant to escape.

I close, deliberately, on the foundation beneath all of it: energy. No phase of this strategy — not the cooperative, not the sovereign compute build-out, not the model portfolio — survives contact with an unresolved energy position. It is my submission to this House that the energy decision Singapore takes between now and 2030 will determine, more than any single AI policy this Government adopts, whether the sovereignty described in this paper is achievable at all.

The Core Argument

AI adoption is a statistic. AI ownership is a structural transition. Singapore has made excellent progress on the former. This paper is concerned, in its entirety, with the latter — and with the energy and institutional foundations that ownership actually requires.

1. The Strategic Context: Why This Debate Cannot Wait

Sir, this House is well acquainted with the extensive-growth model that has served Singapore for half a century — imported labour, imported capital, and a land-scarce economy in which property rent has structured both our economics and, over time, our psychology. That model has been managed with considerable skill. It is not, however, the model on which the AI era will be won.

Layered onto this legacy structure is a second, newer form of dependency. Singaporean enterprises — and increasingly our public institutions — now run substantial parts of their operations on platforms owned, priced, and governed from outside our borders: cloud infrastructure, enterprise software, and now large language models. Every dollar spent in this manner is a dollar that does not compound domestically. It compounds in the balance sheets of firms headquartered elsewhere. This is not, I stress, a criticism of the firms in question, several of whom are valued partners of Singapore's economy. It is a description of a structural pattern that this House would be remiss to leave unexamined.

Honourable Members will recognise the term technofeudalism, associated with the economist Yanis Varoufakis, used to describe an economic order in which a small number of platform owners extract rent from users who never accumulate ownership of the infrastructure they depend upon. I do not invoke this term for its rhetorical value. I invoke it because it names, with useful precision, a risk that Singapore's own economists — including Professor Linda Lim of the Atlantic Council, writing on the limits of state-led extensive growth — have separately identified from a different analytical tradition. When two independent lines of economic reasoning converge on the same structural concern, this House should take note.

The strategy that follows is organised in three phases. I will take each in turn, in some detail, because the substance of this strategy lies in its sequencing, not merely in its ambition.

2. Phase 1 — Foundation (2027–2030)

The task of Phase 1 is to build the physical and institutional rails on which the later phases depend, and — just as importantly — to establish the political and public narrative permission for the harder decisions that Phase 2 will require. Four workstreams anchor this phase.

2.1  A National AI Utility Layer

Singapore has, through AI Singapore and the National Supercomputing Centre, already built meaningful institutional foundations. Phase 1 consolidates and extends this into a genuine utility model — shared, subsidised compute and middleware access that allows small and medium enterprises to adopt AI tools without each having to individually negotiate terms with foreign platforms. IMDA would administer data trusts; AI Singapore would provide the model layer; SkillsFuture funding would be redirected toward practical AI tool access rather than credential accumulation alone.

2.2  Subsidised SME Access

It would be unreasonable to ask Singapore's 280,000-odd SMEs to make a sovereignty-minded choice if the sovereign option costs materially more than the incumbent foreign platform. Phase 1 therefore includes a time-limited subsidy that closes this cost gap — not indefinitely, but long enough to establish an adoption base that Phase 2 can convert into genuine co-production.

2.3  Honest Public Accounting of Platform Dependency

This House should, in my respectful submission, support the publication of a clear, quantified account of what foreign platform dependency currently costs Singapore's economy in aggregate outflow — cloud spend, software subscriptions, and AI API costs leaving Singapore annually. This is not an exercise in alarm. It is fiscal transparency, and it is also, I would note, the political groundwork without which Phase 2's cooperative model will struggle to gain public traction.

2.4  Beginning the Cultural Work

Members will be familiar with the phenomenon colloquially known as kiasuism — an aversion to risk and loss that has, in a great many respects, served Singapore well. I would ask this House to consider it, in the context of this strategy, as a structural variable rather than an immutable trait of national character. It has been redirected before. It can be redirected again, through sustained Government signalling, curriculum reform in our schools, and a deliberate reframing of the competitive question — from "how do we avoid losing" to "how do we build the right to lead."

Phase 1 Success Test

By the end of 2030: the national AI utility layer is operational with demonstrable SME uptake; the platform-dependency audit is published; and the cooperative pilot groundwork with NTUC is underway in at least two sectors. Phase 2 should not be launched ahead of these conditions being met.

3. Phase 2 — Activation (2031–2035): The AI Workflow Cooperative

Sir, Phase 2 is where this strategy either succeeds or fails, because it requires something considerably harder than infrastructure spending: it requires Singaporean SMEs to change their relationship with AI from one of consumption to one of production. This does not happen through subsidy alone. It requires a structural vehicle, and I submit to this House that the vehicle already exists in Singapore's own institutional history — the cooperative.

3.1  Institutional Rationale for the Cooperative Model

This House need not be persuaded that the cooperative is a foreign or untested model for Singapore. NTUC FairPrice, NTUC Income, and our credit cooperatives have organised collective economic power in this country for decades, well before Singapore could compete individually on capital alone. The AI Workflow Cooperative applies this same institutional logic to the AI economy.

Three properties recommend NTUC specifically as the sponsoring institution, and I would note that no alternative statutory body currently possesses all three in combination. First, NTUC already holds sectoral trust relationships across every target industry — relationships a new agency would require years to replicate. Second, cooperative economics is NTUC's founding institutional identity, not an adjacent function. Third, and perhaps most importantly for this House's purposes, the tripartite framework — Government, business, and labour, working in coordination — gives the cooperative a form of political legitimacy that a purely commercial vehicle could never claim.

3.2  How the Cooperative Would Function

The mechanism, in brief: a group of SMEs within a given sector — food and beverage, logistics, or professional services, for instance — pool their operational data through the cooperative structure. A sector-adapted AI workflow model is trained on this pooled data. Crucially, member businesses do not surrender their raw data to one another or to a third party; the cooperative's governance structure ensures that individual firms contribute to a shared training process whose output — the resulting workflow model — is jointly owned, while the underlying data remains the contributing firm's own. The cooperative, not a foreign platform and not a single dominant Singapore entity, holds the resulting asset.

3.3  The Benefits, Stated Plainly

It is worth setting out, in some detail, precisely what this model delivers, because the benefits accrue differently to different stakeholders and this House should weigh all of them.

For the individual SME

• Cost pooling: compute and model training costs are shared across the cooperative membership rather than borne individually, materially lowering the entry cost to genuine AI capability compared to enterprise-grade solo procurement.

• Data asset accumulation: rather than generating data that enriches a foreign platform's model with no return to the business itself, the SME accumulates a durable, jointly-owned productive asset over time.

• Sector-adapted performance: a workflow model trained on fifty businesses' worth of sector-specific data will, in nearly every practical case, outperform a generic foreign model applied to a single business — a competitive advantage kiasuism-driven caution has, until now, left largely unclaimed.

For the worker

• A defined path to genuine economic participation, not merely employment continuity, through worker-ownership components built into the cooperative's governance structure — consistent with NTUC's founding mission.

• Reskilling pathways tied directly to cooperative operation, including a proposed AI Workflow Cooperative Steward role — a WSQ-aligned competency track for workers who take on cooperative governance and technical stewardship functions.

For the state

• A structural alternative to the two failure modes that a less deliberate policy risks: continued platform-tenancy on one hand, or the emergence of a single dominant domestic AI champion — a homegrown techno-lord in a different flag — on the other.

• A tripartite delivery mechanism that does not require the creation of a new statutory board, and which builds on institutional trust already established over decades.

• A genuine productivity dividend that, unlike platform-mediated productivity gains, is measurable and partially retained within the domestic economy.

3.4  Addressing the Anticipated Objection

Sir, I do not wish to understate the principal objection this model will face, because I expect several Members will raise it directly: SME owners are, understandably, reluctant to share commercially sensitive operational data, even with fellow members of their own sector. This is not an irrational concern, and it should not be dismissed as mere kiasuism.

The answer lies in the cooperative's governance design, not in asking business owners to simply trust the process. Member firms do not share raw data with competitors. They contribute to a governed training pipeline whose output is a shared model; the underlying data remains under the contributing firm's control throughout. Properly explained — and this is a communication task the tripartite partners must invest in directly — the competitive reframe becomes clear on its own terms: the SME that joins trains its workflow on fifty businesses' worth of collective experience; the SME that abstains trains on one. Declining to participate is not, in this light, the cautious choice. It is the choice that compounds disadvantage year over year.

This reframing should, in my respectful view, be carried at the level of ministerial and NTUC leadership communication directly, not left to sector-level marketing materials alone.

4. Phase 3 — Sovereignty (2036–2040)

Where Phase 1 builds infrastructure and Phase 2 builds productive ownership at the enterprise level, Phase 3 is concerned with converting accumulated cooperative and sovereign compute capability into exportable national capacity — the point at which Singapore moves from defending against dependency to generating value regionally on its own terms.

4.1  Singapore-Origin Intellectual Property

By this phase, the workflow models developed within mature cooperatives — sector-adapted, tested, and refined over several years of operation — represent genuine intellectual property. Phase 3 is the point at which this House should expect to see the first wave of Singapore-origin AI workflow products exported regionally, particularly into ASEAN markets undergoing similar digitisation pressure without Singapore's institutional head start.

4.2  A Partially Socialised Productivity Dividend

As cooperative workflows mature and compound in value, this House should consider mechanisms — potentially administered through the Central Provident Fund system, or a dedicated cooperative dividend structure — through which a portion of the productivity gains generated by domestically-owned AI capability is returned to the broader citizenry, rather than accruing solely to cooperative members or, worse, migrating abroad as it does today under the pure-platform model.

4.3  Pluralist Ownership Architecture

I return here to a principle this House should treat as non-negotiable across all three phases: the objective is distributed, plural ownership — many cooperatives, many sectors, open standards governing interoperability between them — not the concentration of AI productive capability within a single Government-linked or statutory entity. Domestic technofeudalism, dressed in the national flag, is not a compromise outcome of this strategy. It is the failure mode this strategy exists to prevent.

4.4  Singapore as ASEAN's Governance Anchor

Sir, Singapore cannot achieve technological sovereignty at the scale of 5.9 million people acting alone — this is a matter of arithmetic, not ambition. The ASEAN region, by contrast, holds close to 680 million people navigating precisely the same structural dependency this paper describes, without Singapore's institutional depth, legal clarity, or financial infrastructure. Phase 3's most strategically leveraged move is for Singapore to position itself not as a hyperscaler competitor — a contest Singapore cannot realistically win — but as the region's trusted interoperability and AI governance hub, anchored in bilateral, mutual-benefit data and compute-sharing arrangements with Indonesia, Malaysia, Vietnam, and Thailand.

5. The Model Sourcing Portfolio: A Deliberate Hedge Against Lock-In

Sir, I turn now to a question that this House has, to date, addressed only implicitly: from where should Singapore's AI capability actually be sourced? I would submit that the answer must be a deliberate portfolio of three sources, held in combination, precisely because reliance on any single source recreates the very dependency this entire strategy is designed to escape.

5.1  Home-Grown, Distilled Models

The first component is a domestically developed layer of smaller, distilled models — models trained or fine-tuned on Singapore- and sector-specific data, derived from larger foundation models but compact enough to be hosted on sovereign or cooperative-owned compute at reasonable cost. This is the layer the AI Workflow Cooperative model, described above, is directly designed to produce. It is also, I would note, the layer least susceptible to foreign platform disruption, because Singapore retains both the training data and the resulting model weights.

This House should not, however, mistake this layer for a complete solution on its own. Distillation produces models that are efficient and well-adapted to narrow, well-defined tasks. It does not, at Singapore's scale of investment, produce genuine frontier-capability foundation models — the very large, general-purpose systems that require capital expenditure well beyond what a nation of our size can prudently commit to reproducing independently.

5.2  Open-Source Foundation Models

The second component is deliberate, systematic adoption of leading open-source and open-weight foundation models — models whose weights can be downloaded, self-hosted on sovereign compute, and modified without ongoing licensing dependency on a single foreign vendor. This layer gives Singapore genuine architectural control: the ability to run, inspect, and adapt capable general-purpose models entirely within our own infrastructure, without a continuing rental relationship. Government agencies and cooperative infrastructure alike should treat open-weight model support as a first-order procurement criterion, not an afterthought to be considered after a proprietary vendor has already been selected.

5.3  Licensed Access to Frontier Foundation Models

The third component is continued, clear-eyed commercial access — via API — to the small number of leading frontier foundation models, predominantly developed by United States firms, that represent the current global capability ceiling. This House should not, in my respectful submission, treat this as a concession or a failure of sovereignty. No nation of Singapore's size will independently train a frontier-capability model competitive with the leading global labs within the timeframe this strategy covers, and pretending otherwise would divert capital away from the layers where Singapore can realistically build genuine ownership. The correct posture is not autarky. It is informed, contractually protected access — negotiated with data portability, usage transparency, and exit provisions built into procurement terms from the outset, so that this layer remains a chosen input rather than an unexamined dependency.

5.4  Why the Combination — and Not Any Single Layer — Is the Point

The strategic logic of holding all three layers simultaneously, rather than committing to any single one, deserves to be stated explicitly to this House, because it is the layer of the strategy most likely to be quietly abandoned under budgetary pressure in favour of the path of least resistance — full reliance on a single foreign API provider.

Layer	What It Provides	What It Cannot Provide Alone
Home-grown distilled models	Sector-specific performance; full data and model ownership; lowest ongoing cost	Frontier-level general capability; large upfront training compute
Open-source foundation models	Architectural control; no licensing dependency; portability across compute providers	Guaranteed parity with the most capable proprietary frontier systems
Licensed frontier model APIs	Access to the current global capability ceiling without capital outlay	Guaranteed long-term pricing, terms, or access continuity set by a foreign provider

History offers this House a directly relevant precedent from an adjacent domain: export control regimes governing advanced semiconductor hardware have, on more than one occasion in recent years, altered global access terms for the compute that underpins AI development, with limited notice and no recourse for the affected purchasing nations. Model access carries an analogous exposure — through pricing changes, licensing revisions, or export policy shifts made unilaterally by the provider's home jurisdiction. A portfolio held across all three layers is not a hedge against unlikely events. It is a hedge against a demonstrated category of risk.

The Non-Negotiable Principle

Vendor lock-in — whether to a single cloud provider, a single foundation model, or a single proprietary API — should be treated by this House as a structural risk equivalent in kind to the platform dependency this entire strategy is designed to correct. Procurement policy across every phase of this strategy should require multi-model, multi-vendor portability as a condition of public funding, not an optional enhancement.

6. The Foundation Beneath All Three Phases: Energy Independence

Sir, I have, until this point, deliberately withheld a constraint that governs every phase and every layer described above, because I wished this House first to appreciate the full ambition of the strategy before confronting the single condition on which its feasibility ultimately rests. That condition is energy.

Every element of this paper — the national AI utility layer of Phase 1, the cooperative compute pools of Phase 2, the sovereign IP and regional anchor role of Phase 3, and every layer of the model sourcing portfolio in Section 5 — depends on sustained, reliable, scaled electrical power. Distilled models require training compute. Open-weight models require hosting infrastructure. Even API-based access to frontier models depends on data-centre capacity for the Singapore-side infrastructure that supports cooperative and public-sector integration. None of this runs on policy documents. It runs on electricity.

Singapore generates in excess of 90% of its electricity from imported liquefied natural gas. This has been, for decades, a stable and economically sound arrangement. It also means that the entire strategy described in this paper sits, at its foundation, on an energy supply chain running through international shipping lanes over which Singapore has no direct control. This House will recall the disruption to Strait of Hormuz shipping earlier this year arising from the conflict between the United States and Iran — a disruption that produced measured shortages in petrochemical feedstock and cooking gas supply across our region, as documented in the International Energy Agency's Southeast Asia Energy Outlook 2026. That the disruption did not more severely affect Singapore directly should not, I would caution this House, be read as evidence of resilience. It should be read as a demonstration, at relatively low cost, of an exposure that a compute-intensive AI economy will make considerably more consequential with each passing year.

Singapore's regional power import initiative — the Laos–Thailand–Malaysia–Singapore Power Integration Project, recently expanded to 200 megawatts of import capacity — represents genuine and well-executed diplomatic progress, and this House should recognise it as such. It should also, however, be sized honestly: 200 megawatts represents approximately 1.5% of Singapore's peak electricity demand, and it is delivered through transmission infrastructure across three neighbouring jurisdictions rather than through capacity Singapore itself controls outright.

This House has before it, in effect, a choice that mirrors decisions this Government has made successfully in the past — NEWater, and the reclamation programme that built Changi into the region's leading air hub both required capital commitments that appeared, at the time of decision, considerably more uncertain than they appear in hindsight. I would submit that the energy decision before this House now belongs in that same category. Singapore has commissioned feasibility studies into small modular reactor technology. The technology landscape, and the case for treating 2028 as a genuine decision point rather than a further extension of study, has been set out in detail in a companion paper submitted to this House. I will not repeat that analysis in full here, save to state its conclusion plainly: a sovereign compute strategy without a resolved energy pathway is not a sovereignty strategy. It is a capital commitment made in the hope that someone else's energy policy does not change.

Closing Submission

Sir, I ask this House to treat the 2028 energy decision gate with the same seriousness it has rightly extended to the AI Workflow Cooperative framework and the sovereign compute programme described in this paper. The Foundation, Activation, and Sovereignty phases of Singapore's AI strategy are, each of them, well-conceived. None of them will bear their intended weight if the energy question beneath them remains, in 2028 as it does today, still under study rather than decided.

7. Summary of Recommendations to This House

• Endorse the three-phase strategy — Foundation (2027–2030), Activation (2031–2035), Sovereignty (2036–2040) — as the organising framework for Singapore's national AI and digital economy transition.

• Direct NTUC and EnterpriseSG to establish a joint working group on the AI Workflow Cooperative framework within six months, with pilot cooperatives launched in food and beverage and logistics sectors ahead of Phase 2's formal commencement.

• Require, as a condition of public AI infrastructure funding at every phase, a demonstrated multi-model, multi-vendor sourcing posture spanning home-grown distilled models, open-source foundation models, and licensed frontier model access — with explicit data portability and exit provisions in all foreign vendor contracts.

• Establish a Cabinet-level energy decision gate for 2029, converting current SMR feasibility work into either a firm construction commitment or a clearly reasoned deferral, in coordination with continued scaling of regional power import capacity.

• Commission a public accounting of platform-rent leakage and energy-import exposure, published together, so that this House and the public may evaluate both dependencies with the same rigour and on the same timeline.

Prepared by Tat Yuen, Bixbe Consulting. This paper draws on the Varoufakis technofeudalism framework, Linda Lim and Pang Eng Fong (Academia.sg) on Singapore's growth model, the Atlantic Council's commentary on state-led expansion, the IEA Southeast Asia Energy Outlook 2026,  public EMA/Keppel disclosures on the LTMS-PIP programme, and independent research.