GIC (Government of Singapore Investment Corporation)

GIC's only published metric, the 20-year annualised real return (3.8% to 31 Mar 2025), is less investor communication than fiscal input: it underpins Singapore's Net Investment Returns Contribution, which lets the government spend up to half of expected long-term real returns in the Budget. Read that way, the opacity and smoothing are policy design, not evasion. Its AI expression, private compute and power infrastructure plus co-leading Anthropic's $30B Series G, is a trade only a balance sheet that never marks itself daily can hold.

Total-portfolio management of Singapore's reserves across equities, fixed income and real assets (51/26/23 percent at 31 Mar 2025), with a large internal private-markets book. AI runs through the picks-and-shovels layer, hyperscale data-centre and telecom platforms (Equinix xScale, EdgeCore, CETIN), plus selective frontier-model equity, while management publicly flags listed-AI overvaluation. Closed to outside capital, internally run, no external fee load.

• Horizon genuinely matches liabilities: no redemption risk, no annual performance fee clock, so it can hold illiquid compute and power infrastructure through a drawdown.
• Picks-and-shovels framing is the right instinct: contracted data-centre and telecom-infrastructure cash flows survive a model-layer shakeout better than application equity.
• Internal management at roughly $936B scale (third-party estimate) avoids the 2-and-20 drag that would consume most of a 3.8% real return.
• Publicly naming valuation overshoot, with explicit reference to Nikkei 1989 and Nasdaq 2000, is unusually candid for a sovereign and creates an accountable record.

• The data-centre book is sold as diversifying but its tenants are the same hyperscalers whose capex is the AI bet. Correlated exposure dressed as picks-and-shovels.
• Co-leading Anthropic's Series G at $380B while warning on AI valuations is a tension the disclosure regime never forces GIC to reconcile publicly.
• 20-year rolling reporting means a bad decade would take years to surface, and by then the responsible decision-makers have rotated out.
• No position-level disclosure and no meaningful 13F. Every holding here is a press sighting or a deal announcement, not an audited line item; unsuccessful deals are simply never reported.
• AUM itself is a third-party estimate. GIC has never confirmed it, reportedly to avoid inviting speculation against the SGD, so even the denominator is unverifiable.

• AI capex cycle turning: hyperscaler lease renewals repricing down would hit the data-centre platforms and the Anthropic mark at the same time.
• Private-mark staleness: a frontier-model down round would force a write-down years after the fact, with no interim signal.
• Political constraint: reserve losses are domestically explosive in Singapore (the 2008 UBS and Citigroup episodes), which can bias toward de-risking at the wrong moment.
• Budget dependency: the NIRC formula ties a real fiscal line to expected long-term returns, so a sustained real-return decline is a policy problem, not just a portfolio one.
• Scale: at roughly $936B, deployment options narrow to large private deals where the competitive set is other sovereigns bidding the same assets.