Temasek Holdings
Temasek Holdings (Private) Ltd, Singapore; CEO Dilhan Pillay Sandrasegara. Singapore's state-owned active investor, run more like a high-conviction fund than an index tracker.
Temasek runs its US 13F as a high-conviction thematic book, and the AI/semi/compute thesis is explicit in the top holdings: Nvidia (~5.4%), Broadcom (~3.6%), Alphabet and Microsoft sit alongside its BlackRock/Visa/Mastercard financial-platform core. Beyond listed equity, Temasek is a co-buyer (with Nvidia, BlackRock, Microsoft) in the ~$40B+ Aligned/AI data-center mega-deal and a backer of PsiQuantum (quantum compute, with Nvidia's NVentures). It is the deep-tech / frontier-compute end of the Singapore pair — more willing than GIC to own the listed AI-semi names directly.
- Not open — Not an open-ended vehicle — analysis only, not an available allocation.
Top holdings
~$5.1B. Largest US holding; financial-infrastructure core (also AI-infra co-investor)
~$2.0B. Payments platform
~$1.6B. Core AI-semi conviction position
~$1.6B. Payments
~$1.4B. TPUs / Gemini AI
~$1.1B. Custom AI ASIC / networking silicon
Recent moves
Q1 2026 13F: US book ~$31B, down ~3.5% QoQ; AI-semi names (NVDA, AVGO) held as core. Off-13F: part of the investor group (with Nvidia, BlackRock, Microsoft) buying a major AI data-center operator in a ~$40B+ record deal; participated in PsiQuantum financing alongside Nvidia NVentures.
Loading disclosed positions…
Our take
Why it matters: Temasek's 13F is one of the cleaner sovereign reads on active AI-semi conviction — Nvidia and Broadcom are deliberately chosen top-6 positions, not index residue, and it pairs them with direct AI data-center and quantum-compute bets. Edge/limits: the US sleeve (~$31B) is a fraction of the total portfolio and skews to financial-platform names (BlackRock/Visa/Mastercard), so AI is a meaningful slice, not the whole thesis. Honest caveat: the 13F is long US-listed equity only — no shorts, foreign listings (its big Asian holdings), or private deep-tech holdings appear; the AI data-center and PsiQuantum exposure is off-13F and sized from press.
The headline is not S$518B — it is two disclosures.
First, the AI allocation: ~6% today, framed as directional, up to 15% by 2031, pace dependent on valuations. We take the caveat at face value and still think it binds weakly: a public weight with a date creates institutional momentum and a reputational cost to under-delivering, and management paired it with the claim that the other 85% must focus on AI adoption. A return hurdle disciplines entry price; a weight-and-date does not. Second, FY2026 completes the extension of mark-to-market to the unlisted book. That is a real governance improvement, and it makes the S$518B-vs-S$469B comparison properly like-for-like, because the prior base was restated (published S$434B, S$469B on an MTM basis). What it does not make comparable is the multi-year series: the 10-year TSR moved from a published ~5% at FY2025 to 7.1%, and much of that step is the restatement, not a year of return. And roughly half the portfolio is unlisted, marked to fair value via comparables and DCF, not to an observable price.
Thesis
Temasek is a state-owned direct investor, not a diversified sovereign fund. Net portfolio value was S$518B at 31 Mar 2026, up S$49B against a restated S$469B base, with 1-yr TSR 10.5% in SGD. Singapore Temasek Portfolio Companies are 43% of the book by segment, Singapore is 27% by underlying geographic exposure, and three names — DBS 9%, Singtel 8%, PSA 5% — are 22% of the whole. The record is a concentrated read on Singapore Inc. plus a growing global-direct sleeve. It is closed to outside capital; the useful question is whether the strategy's logic holds, not whether one can buy it.
Three segments at 31 Mar 2026: Singapore Temasek Portfolio Companies 43%, global direct investments 38%, partnerships/funds/AMCs 19%. Roughly 50% listed, 50% unlisted. Underlying geographic exposure: Singapore 27%, Americas 26%, China 17%, EMEA 12%, India 7%. AI exposure ~6% today, stated as directional toward up to 15% by 2031 across power, semis, cloud and models. Permanent capital, one shareholder, no external client.
Assessment
- Permanent capital with no redemption risk lets it hold unlisted and infrastructure assets through drawdowns — the one edge a client-facing fund cannot replicate.
- Disclosure is strong for a sovereign vehicle: TSR at multiple horizons, segment mix, underlying geographic exposure, named top holdings, and a completed book-to-MTM transition.
- AI exposure is expressed across the stack — power, semis, cloud, foundation models, enterprise software — rather than a single-layer bet on model vendors.
- Restating the prior-year base to S$469B rather than comparing S$518B against the old S$434B is the honest presentation choice, and it is disclosed.
- The AI target is disclosed as directional and valuation-dependent, but a public portfolio weight plus a date still creates buying momentum. We would rather see a return hurdle than a 2031 waypoint.
- Concentration: 43% of the portfolio is Singapore Temasek Portfolio Companies and 22% sits in three names. Underlying Singapore exposure is lower at 27%, but the segment lens is the ownership reality.
- Roughly half the book is unlisted, marked to fair value via comparables and DCF. Better than book value, but model marks are self-reported and this method has no full-cycle test behind it.
- The 10-year TSR moved from a published ~5% to 7.1% substantially via the MTM restatement, not a year of returns. The multi-year series is not comparable across the method change.
- TSR on a vehicle whose sole shareholder can inject or withdraw capital is not cleanly comparable to an index total return; the peer comparison is looser than it looks.
- 2026-07July 2026 tested that breadth and it proved the worse exposure: the SOX fell 20.6%, memory 29-47%, wafer-fab equipment 28-30%, Vertiv 27.9%, neoclouds 28-31%, while Nvidia +0.3% and Broadcom +3.1% held.
Record
FY2026 (year to 31 Mar 2026): NPV S$518B from a restated S$469B base, 1-yr TSR 10.5% SGD — 12.9% on a constant-currency basis, so a ~2pp drag from a strong SGD — with S$51B invested against S$31B divested. Longer horizons as reported this year: 5-yr 4.6%, 10-yr 7.1%, 20-yr 6.8%. The comparison that matters: at FY2025 Temasek published a 10-yr TSR of ~5% and a 20-yr of ~7% on the old basis, so this year's 7.1% is not a decade of compounding improving — it is substantially the S$35B unlisted uplift from the method change flowing through the series. Attribution for the year as disclosed: listed Singapore portfolio companies plus divestment gains, offset by currency. Read plainly: a strong reported year, a decade harder to call index-beating than 7.1% suggests, and a Singapore-cycle contribution the segment mix makes hard to separate from skill.
Risks & fit
- Singapore-cycle correlation: the vehicle that cushions the national budget is itself levered to the national economy, so the hedge weakens exactly when it is most needed.
- Fiscal dependence — investment returns fund a material share of Singapore government spending — creates pressure toward reported-return smoothness.
- AI exposure across semis, cloud, power and foundation models is one factor risk wearing several tickers.
- Unlisted fair-value marks are model-dependent and lag listed repricing; the newly completed MTM method has not been tested through a drawdown.
- An explicitly state-owned buyer faces foreign-investment-review risk on Western strategic and semiconductor assets.
Our read is that the record is substantially Singapore-cycle beta plus a not-yet-proven global build-out, and that the new 10-yr figure flatters the decade. It would be wrong if, over three to five years, the global-direct and partnership sleeves out-earn the Singapore portfolio-company book by a clear margin while the Singapore segment weight falls below ~35% (from 43%) and underlying exposure below ~20% (from 27%). It would also be wrong on marks if the now-fully-MTM unlisted book takes a visible, promptly-reported writedown in the next drawdown instead of drifting down slowly.
Analytically useful to anyone tracking sovereign capital flows into AI infrastructure, to Singapore-listed equity analysts (Temasek is the reference shareholder in DBS, Singtel and PSA, so its stance is a governance input), and to allocators benchmarking sovereign-fund disclosure and valuation practice. Temasek is closed to outside capital — there is no vehicle to access, and nothing here is a view on whether to hold or avoid any security it owns.
No external management or performance fee — Temasek invests its own balance sheet for a single shareholder. The fee load is indirect: 19% of the portfolio sits in partnerships, funds and AMCs carrying conventional external-manager economics.