I
believe a debt-financed capital loop is inflating an American AI bubble to a
scale the underlying cash flow does not support. I believe that bubble sits inside an economy
that no longer carries the fiscal reserves it once had to absorb a
correction. I believe the same debt
trajectory is pushing Treasury yields to levels the market itself is beginning
to call structural rather than cyclical.
I believe China’s chip constraints, intended to slow its AI progress,
instead forced a more efficient architecture that may prove the more durable
advantage. I am making the case that
Singapore’s sovereign funds should reassess their exposure accordingly.
A
circular financing loop, estimated at over US$800 billion, links Nvidia, AI
laboratories, and cloud providers in a structure that books related-party
spending as independent revenue. OpenAI
alone has committed US$1.15 trillion through 2035 against a projected US$14
billion loss in 2026. Nearly 40 per cent
of the S&P 500’s market capitalisation is now under direct AI
influence. Lucent, Nortel, 1929, and
2008 each followed the identical pattern: concentrated leverage mistaken for
collateral, collapsing once independent cash flow failed to materialise. Federal debt has crossed US$40 trillion, and
the Federal Reserve raised rates in September 2026 rather than cutting them,
removing the monetary cushion that absorbed the 2001 dot-com correction. The 10-year Treasury yield crossed 5 per cent
the same month, and every Treasury intervention attempted in 2026 produced only
temporary relief before yields resumed climbing. Chip sanctions forced Chinese laboratories
toward architectures requiring a fraction of the compute of Western models, at
a fraction of the cost. US data centres
already consume as much electricity as Ireland’s entire national grid, with
household electricity bills already rising to fund the buildout. The top 1 per cent of US households now hold
31.7 per cent of national wealth, concealing fragility beneath resilient-looking
aggregate consumer data. Each section
below closes with a specific implication for Singapore’s funds.
Part One: The Capital Loop Feeding an Unsustainable Bubble
Nvidia
invests billions into AI laboratories such as OpenAI and Anthropic. Those laboratories sign compute contracts
with cloud providers, including Microsoft, Oracle, and Amazon Web Services. Those providers spend a large share of that
revenue buying chips back from Nvidia.
Cash leaves Nvidia’s balance sheet as an investment. It returns as revenue, having toured through
two or three other balance sheets along the way. Analysts have identified over US$800 billion
moving through this loop, and deals of this scale fuel circularity concerns.
OpenAI
has committed US$1.15 trillion across seven vendors through 2035, while
projecting a US$14 billion loss for 2026, nearly triple the prior year’s
loss. Big Tech tripled annual AI capital
spending from US$150 billion in 2023 to over US$500 billion in 2026. Nearly 40 per cent of the S&P 500’s
market capitalisation is under direct AI influence, a concentration ratio
without precedent outside the dot-com peak.
Forward price-to-earnings on the index is 22.85 times, within a fraction
of the 2020 peak of 23.11.
Revenue
booked from a related-party compute contract is not the same as revenue earned
from an independent customer with an independent reason to keep paying. When one node in this loop misses a covenant
or writes down an asset, the loss does not stay contained to that balance
sheet. It removes demand from every
counterparty whose revenue projections assumed the spending would continue
uninterrupted.
Temasek
Holdings reported a net portfolio value of S$518 billion, US$401 billion, for
the year ended 31st March 2026.
The United States accounts for 26 per cent of that portfolio. AI-related exposure is 6 per cent and is
committed to rise to 10 to 15 per cent by 2031.
Temasek Holdings holds stakes in both OpenAI and Anthropic, and was, per
Reuters, in active discussions to increase its OpenAI position further. Chief Executive of Temasek Global Investments
Chia Song Hwee has told reporters overvaluation in AI is “unavoidable” and that
nobody, Temasek Holdings included, can predict when a correction arrives. Temasek Holdings has declined to disclose
specific stakes or their effect on overall performance, leaving me unable to
independently assess how concentrated this exposure has become beneath the
reported 6 per cent figure. I believe a
portfolio review distinguishing genuine end-user revenue from circular
vendor-financed revenue should precede any further allocation toward the stated
target.
Part Two: The Historical Parallel, and the Write-Down Cascade
Lucent
Technologies and Nortel Networks ran an almost identical loop during the
dot-com era, lending customers money to buy their equipment and booking the
proceeds as revenue on both sides. When
real demand failed to match financed demand, both collapsed in the same
downturn. Lucent’s market value fell
from over US$258 billion at its 2000 peak to a fraction of that within two
years. Nortel filed for bankruptcy in
2009.
The
1929 crash carried the same architecture in a different sector: margin-financed
speculation, concentrated in a narrow set of favoured stocks, collapsing once
credit tightened. The 2008 crisis ran
the identical script through mortgage-backed securities. Lehman Brothers held US$639 billion in assets
at its collapse, and the resulting deleveraging wiped out an estimated US$19.2
trillion in US household wealth by 2011.
Every one of these episodes shares the same signature: concentrated
leverage, mistaken for collateral, until the underlying cash flow failed to
materialise.
MAS
has built genuinely sophisticated AI governance infrastructure. The MindForge AI Risk Management Toolkit,
released in April 2026, was developed with 24 institutions, including DBS,
OCBC, UOB, GIC, and Temasek, alongside Nvidia, AWS, Google Cloud, and
Microsoft. Director Alan Lim of MAS’s
Financial Infrastructure and AI Office described it as moving institutions “from
theory to practice.” This framework
governs operational AI risk, how banks deploy AI internally, not the separate
question of what happens to Singapore’s banking balance sheet if the AI sector
itself suffers a credit event. GIC led
Anthropic’s US$30 billion Series G round, valuing the company at roughly US$380
billion. GIC took a leading role in
financing a private AI company at a valuation built on projected revenue rather
than demonstrated cash flow. I believe
MAS’s stress-testing framework for DBS, OCBC, and UOB should explicitly model
an AI-sector credit event as a distinct scenario, separate from MindForge’s
operational governance work.
Part Three: Why America Cannot Absorb This the Way It Absorbed the
Dot-Com Crash
The
dot-com correction in 2001 hit an economy with far more fiscal room. Federal debt was below 60 per cent of GDP,
and the government ran a surplus as recently as fiscal year 2000. That room no longer exists. National debt has crossed US$40 trillion. The Congressional Budget Office projects debt
held by the public rising from 101 per cent of GDP this year to 120 per cent by
2036. Net interest reached US$963
billion over ten months of fiscal 2026, US$3.18 billion a day.
The
Committee for a Responsible Federal Budget estimates that if yields stay just
80 basis points above baseline, annual interest payments reach US$2.7 trillion
by decade’s end, exceeding combined Medicare and Social Security spending. The Federal Reserve raised rates on 16th
September 2026, its first hike since 2023, because inflation had run above
target for years and the central bank had no room to accommodate a fresh shock
stacked on its own.
Singapore
held US$280 billion in US Treasury securities as of April 2026, according to US
Treasury International Capital data, separate from GIC’s estimated US$936
billion to US$1.16 trillion and Temasek Holdings’ S$518 billion portfolio, both
carrying further undisclosed US exposure.
Singapore’s Ministry of Finance will not disclose GIC’s exact assets,
because doing so would “amount to publishing the full size of Singapore’s
financial reserves,” what it calls Singapore’s “war chest.” Then MAS Managing Director Ravi Menon
described the reserves’ three functions in 2019: a crisis buffer, an income
stream, and a foundation for exchange-rate confidence. A Treasury market seizing up around an
AI-sector credit event does not merely threaten the market value of these
holdings. It threatens the function they
exist to perform during a crisis. GIC’s 20-year
annualised real return fell to 3.4 per cent as of March 2026, a six-year
low. I believe MAS and MOF should explicitly
model a US Treasury market dysfunction scenario tied to an AI-sector credit
event, before market conditions force the question.
Part Four: The Structural Yield Problem
The
10-year Treasury yield crossed 5 per cent in September 2026. The 30-year yield reached its highest level
since 2001 the month before, with auction demand weaker than dealers expected. The Treasury Borrowing Advisory Committee
flagged a US$1.45 trillion funding shortfall for fiscal 2027 to 2028, a warning
from the dealers who underwrite these auctions themselves.
Treasury
Secretary Scott Kenneth Homer Bessent attempted three separate interventions
across 2026: a joint yen operation with Japan structured to avoid selling
Treasuries outright; a bond buyback funded by selling short-dated debt; and a
described strategy of trading on “asymmetric information,” a hedge fund framing
applied to sovereign debt stewardship.
Every intervention produced a temporary rally followed by a return to,
or a breach of, the prior yield level within weeks, itself evidence that this
is structural rather than cyclical.
I
want to credit GIC before I criticise it.
Its 2025/26 annual report confirmed increased allocations to “commodities,
gold, and infrastructure” specifically “to enhance inflation resilience,” and
it overhauled its entire investment framework this year, replacing a 65/35
equities-bonds reference portfolio it had run for thirteen years with a
Strategic Portfolio built around growth, uncertain rate navigation, and
inflation resilience. GIC’s stated
reasoning names “a changing world order, rising fiscal risks, and advances in
artificial intelligence,” the same three forces I have argued throughout this
paper.
Here
is the contradiction I believe needs resolving.
GIC’s equities allocation rose to 51 per cent in the year to March 2025,
up from 46 per cent the year before, even as its report acknowledged high
valuations “provide a challenging backdrop for forward returns.” GIC has confirmed it “invested selectively”
in AI, in companies it believes have “enduring value beyond short-term market
enthusiasm,” an acknowledgement that it distinguishes durable value from
speculative enthusiasm and has chosen to increase exposure regardless. Global SWF, an independent tracker, has noted
GIC’s reporting “has grown more opaque, now lacking specific asset class mix
disclosures” that earlier reports provided.
I cannot verify what share of the 51 per cent equity allocation carries
AI concentration, nor how the real assets increase compares in scale to the
simultaneous equity increase. I believe
GIC and MAS should resolve this contradiction directly, and should, at minimum,
internally restore the granular disclosure that Global SWF notes is no longer
available.
Part Five: Why China’s Chip Constraints Handed It the Efficiency
Advantage
US
chip sanctions, intended to slow Chinese AI progress, instead forced a
different kind of innovation. DeepSeek
released its R1 model in January 2025, matching OpenAI’s o1 on multiple
benchmarks at a reported training cost under US$6 million, using export-controlled
H800 chips. Nvidia lost approximately
US$589 billion in market capitalisation on 27th January 2025, the
largest single-day value loss in stock market history.
DeepSeek’s
V3 model cost US$5.58 million to train, a 5.5 per cent fraction of GPT-4’s
reported cost. By 2026, this pattern
spread across Alibaba’s Qwen, Moonshot AI’s Kimi, and Zhipu AI’s GLM. GLM-5, released in February 2026, became the
first frontier-scale model trained entirely on Huawei’s Ascend 910B chips,
without a single Nvidia GPU. DeepSeek
founder Liang Wen Feng described the underlying constraint in a July 2024
interview: “We have to consume twice the computing power to achieve the same
results ... Our goal is to continuously
close these gaps.” His team closed them
through architecture, sparse Mixture-of-Experts routing, FP8 low-precision
training, aggressive context caching, cutting inference costs by 90 to 97 per
cent against comparable Western models, not through acquiring more
hardware. In a world where US data
centre electricity demand is projected to reach 325 to 580 terawatt-hours by
2028, an architecture requiring a tenth of the compute for equivalent output is
the actual efficiency frontier, not a footnote to it.
Singapore’s
National AI Strategy update, unveiled by Minister for Digital Development and
Information Josephine Teo Li Min at ATxSummit on 20th May 2026,
names “efficient AI computing” as a stated priority. Yet Singapore committed over S$1 billion in
AI research funding, S$150 million to the Enterprise Compute Initiative, and
S$37 billion under RIE2030, while OpenAI committed S$300 million to its first
overseas Applied AI Lab here and Nvidia opened its second Asia-Pacific research
lab here. There is a stated commitment
to efficiency, alongside deepening operational dependence on the same Nvidia
supply chain already exposed to the rare earth export controls. Singapore’s published strategy does not yet
specify how efficiency translates into a concrete hedge against this dependency,
as opposed to simply meaning lower costs within an unchanged hardware
relationship. I have also looked at
AiRTS Pte. Ltd., a Singapore-founded
company building a patented enterprise product called GenAI Twin, as an example
of genuine domestic AI intellectual property.
AiRTS operates at the application layer, addressing hallucination and
output consistency in deployed models, not at the foundational training layer
where DeepSeek’s efficiency gains are.
It is evidence of a capable domestic ecosystem. It is not a substitute for the hardware
diversification I believe the national compute strategy needs.
Part Six: The Domestic Precedent
Singapore
imposed a de facto data centre moratorium from 2019 to 2022, after data centres
reached 7 per cent of national electricity consumption, projected to climb to
12 per cent by 2030 without intervention.
Rather than simply lifting the ban, Singapore rebuilt the approval
framework around efficiency as the primary gate. The Data Centre Call for Application 2,
launched in December 2025, requires every approved facility to achieve a Power
Usage Effectiveness of 1.25 at full load, at least 50 per cent green power
sourcing, and liquid cooling for at least 60 per cent of IT load, standards
that significantly outpace regional benchmarks in Hong Kong, Japan, and South
Korea.
This
standard has not repelled capital.
Amazon Web Services pledged S$12 billion under this framework. Google committed US$5 billion. Jurong Island will host a 700-megawatt
low-carbon data centre park. Singtel’s
Nxera unit is building a 58-megawatt facility backed by a US$476 million green
loan. Keppel plans to double its power
capacity by 2030.
I
believe any Singapore-linked data centre or compute cluster development outside
our borders, in Johor, in Indonesia, or in any regional partnership our capital
participates in, should be held to the same PUE and green energy standard we
already enforce domestically. A
Singapore-backed facility built to a looser regional standard carries higher
exposure to both rare earth cost inflation and water and power constraints, for
a lower return on the same capital, than a facility built to our proven
standard. The domestic policy already
exists. I believe we should simply apply
it consistently to where our capital builds next.
Part Seven: Unmanaged Data Centre Growth is a Quantified American Risk
PJM
Interconnection’s capacity auction clearing price rose from US$28.92 per
megawatt-day in the 2024/25 delivery year to US$329.17 for 2026/27, a factor of
more than ten. Data centres accounted
for 63 per cent of that increase, translating to roughly US$9.3 billion passed
to ratepayers. New Jersey residents saw
bills jump 17 to 20 per cent in a single year.
Virginia’s Dominion Energy implemented its first base-rate increase
since 1992, adding US$8.51 a month to a typical household bill, in a state
where data centres already account for 40 per cent of total electricity
consumption. Utilities requested US$18.6
billion in rate increases in the first half of 2026 alone, on pace to exceed
2025’s record US$29 billion. Seven major
technology companies signed a voluntary Ratepayer Protection Pledge in March
2026, committing to cover their grid costs. The pledge carries no legal weight and cannot
override existing tariff structures.
I
believe Singapore should present this specific, quantified American contrast,
not the general principle of sustainability, in every regional and
international forum where data centre policy is discussed. Our DC-CFA2 framework mandates efficiency
before approval is granted. America’s
experience shows precisely what happens without that mandatory gate.
Part Eight: The Hollowed-Out Middle Class is a Financial Risk to Any
AI-Exposure Model
The
top 1 per cent of US households held 31.7 per cent of national wealth in the
third quarter of 2025, the highest share since the Federal Reserve began
tracking this in 1989. That 1 per cent
held US$55 trillion, close to the combined wealth of the bottom 90 per
cent. Wage growth for higher-income
households reached 3 per cent in December 2025, against 1.1 per cent for
lower-income households. Moody’s
Analytics found the top 10 per cent of households drove 49.2 to 49.7 per cent
of consumer spending by mid-2025, though this figure has been disputed; the
Bureau of Economic Analysis’s data show the top 10 per cent by disposable
income responsible for only about 20 per cent of spending across 2004 to
2022. I believe this dispute itself is
evidence worth acting on: if credentialed economists working from the same
Federal Reserve data cannot agree on how concentrated spending has become, the
aggregate consumer spending figure is less reliable as a signal of broad
economic health than most headline coverage suggests. The wealth concentration data carries no
comparable dispute, sourced directly from the Federal Reserve’s Distributional
Financial Accounts: the top 1 per cent held 29.2 per cent of aggregate wealth
in Q4 2025, against 5.3 per cent for the entire bottom half.
I
believe any Singapore-based fund or institution modelling US consumer demand
resilience should explicitly disaggregate spending data by wealth decile rather
than relying on the aggregate figure, since that aggregate materially
overstates broad-based economic health in an economy this concentrated, and
since the same equity concentration directly transmits into this same consumer
spending pattern through the wealth effect.
My Conclusion
I
have traced one argument across eight parts.
A circular capital loop, an absent fiscal cushion, a structural yield
problem, and a hollowed consumer base sit underneath an American AI bubble our sovereign
funds are increasing exposure to, even as our central bank’s risk framework
does not yet cover the specific scenario that exposure creates. China’s constrained path produced a
structurally more efficient alternative.
I believe Singapore’s institutions should reassess this exposure with
the data, not with the assumption that scale alone protects against a
correction nobody, including Temasek Holdings’ and GIC’s leadership, claims to
be able to time.
Terence Nunis | Executive Chairman, Equinox Zenith | Author,
The 1% Playbook: The Billionaire Cheat Code