Singapore · AI-Enabled Proprietary Investment

An edge built on depth, vision and conviction, driven by data

Apex Edge Capital Singapore is a proprietary principal, AI-enabled investment company. It concentrates its own capital behind a few high-conviction sectors — identified through macro and micro research and a long-range industry view — with one aim: to outperform the market by at least 50%. Fundamentals first; evidence always.

Depth over breadth
A few high-conviction sectors, not a long watchlist. Each is chosen top-down from the macro regime and bottom-up from micro and industry-structure research, and held only while the fundamentals hold. We would rather own the right idea deeply than many ideas loosely.
Risk first
Survival before return. Every position begins with what it can lose and where the exit is, so no single outcome is ever decisive. Concentration is deliberate but bounded, and leverage is never taken into binary events — the point is to stay in the game long enough for good ideas to compound.
Patience as edge
Independent capital can wait. We hold through drawdowns while the thesis is intact, exit when it breaks rather than on noise, and sit in cash when conviction is absent. Time horizon, used with discipline, is itself a source of advantage.
Latest view
Circular by design — what the AI build-out shares with every revolution before it
Our latest Insight · 28 July 2026 — read it, then browse the full column.
→
About

Human conviction, machine depth.

Apex Edge Capital Singapore is a single-principal, AI-enabled proprietary fund. Independence is its advantage: it answers only to its own conviction — free to hold through volatility, to wait for the right price, and to concentrate where the case is strongest rather than diversify for its own sake. The aim is straightforward — to outperform the broad market by owning the right ideas, not the most ideas — and an agentic-AI research and risk workforce lets a single principal pursue it with institutional depth.

Its focus is the structural shift reshaping technology and the wider economy — the build-out of artificial intelligence and the layers beneath it, from power and the physical foundations of computing to the businesses built on top. The question it returns to is where durable value is actually created as that landscape evolves: which layer holds the advantage, why it holds, and how long it can last. That is the question the portfolio is built to answer.

Apex Edge Capital Singapore
Joanna Tang, Founder & Director
Joanna Tang — Founder & Director

Twenty years across wealth management, consumer banking, payments and fintech — in C-suite and board roles spanning APAC and the United States.

Joanna Tang founded Apex Edge Capital Singapore to invest her own capital with the independence and rigour she spent a career building. She has led businesses through full P&L ownership, complex institutional partnerships and digital-platform transformation — consistently in the most regulated corners of finance.

Most recently she was Managing Director and Board Director at FNZ, the global wealth-management technology firm, running its institutional business across Asia ex-Japan and serving as CEO of its licensed dealing-and-custody entity. Earlier she was Chief Executive Officer and Board Director of Lu International, the Ping An Group's licensed digital-wealth business across ASEAN, and held senior roles at Visa and Standard Chartered Bank, following analytics and risk leadership at Barclaycard and Citi in New York.

Her foundations are quantitative — a Master's in Statistics from the State University of New York at Stony Brook, an engineering degree from Beijing University of Aeronautics and Astronautics in China, and an MBA from the University of Chicago Booth School of Business — and she is a MAS Registered Representative and Chartered FinTech Professional. At Apex Edge Capital Singapore she brings that blend of operating leadership, institutional discipline and a data-driven mind to the companies shaping the next technological cycle.

Approach

Macro, micro, math — fundamentals first.

Every idea runs through the same sequence. Macro sets the weather — the regime, the cycle, the direction of policy and liquidity. Micro finds where value is actually captured within a sector: who holds the durable advantage, and why. Math tests both against the data. Fundamentals lead the judgment; evidence confirms it; and agentic AI carries the research and risk work at a scale a single desk could not otherwise reach.

Macrosets the weather
→
Microfinds where value is captured
→
Mathtests it against the data
→
Convictionthe decision, human
01

Concentration with conviction

Depth over breadth. Conviction sectors are chosen top-down from the macro regime — the cycle, policy and liquidity that decide which tide is running — and bottom-up from micro and industry-structure research that identifies who inside a sector actually holds the durable advantage. A few such ideas are researched thoroughly, sized to matter, and revisited as the facts change. Fundamentals lead: a thesis must stand on business economics — real demand, real pricing power, real cash — before it earns a position. Narrative and momentum are inputs to understand, never reasons to own.

02

Survival before return

Risk is considered first and explicitly. Every position starts with its downside and its exit defined, so that being wrong is survivable and no single outcome can be decisive. Concentration is deliberate but bounded; correlated bets are counted as one; and leverage is never carried into binary events whose outcome cannot be researched. Capital preservation is not caution for its own sake — it is what keeps the book in the game long enough for the good ideas to compound.

03

Patience as advantage

Independent capital can wait — and waiting well is an edge in itself. Positions are held through drawdowns while the thesis is intact, and exited when it breaks, not when the tape is merely noisy. Entries are taken on dislocations rather than chased, and cash is held, without apology, when nothing meets the bar. In a market that rewards activity, the discipline to do less — but at the right moments — is where much of the advantage lives.

04

Evidence, not hype

Data and AI serve the judgment, not the other way round. Every thesis is tested against the numbers — unit economics, valuation, the base rates for companies like this — before capital follows, and claims that cannot be evidenced are treated as unproven. An agentic-AI research and risk workforce runs this analysis continuously and at scale, surfacing what matters and stress-testing what feels comfortable. The tools sharpen the decision; the decision, grounded in fundamentals, stays human.

Insights

Weekly investment views.

A weekly point of view on the markets Apex Edge Capital Singapore follows most closely — the AI value chain, technology, and the macro backdrop that frames them. Written analysis to begin; video to follow.

The creditor with a redeem button — who is really funding the deficit

America’s debt pile has found an unlikely new buyer. It isn’t a central bank, and it doesn’t hold for decades. Tokenisation was meant to erode the dollar. It is buying Treasury bills instead.

Read →

Circular by design — what the AI build-out shares with every revolution before it

Critics call today’s AI spending ‘circular’ and hear a bubble. History says circularity is how capital-intensive revolutions have always financed themselves — the question is whether the loop is healthy.

Read →

When wide spreads matter — and when they don’t

Spreads are near-serene. But wide spreads don’t cause crashes — a wide spread widening fast, with a systemic shock and no backstop, does. A guide to reading the gauge.

Read →

The long end is talking

The 30-year yield is back at pre-2008 levels and the term premium has returned. The bond market isn’t forecasting a boom — it’s charging rent for uncertainty.

Read →

AI’s real bottleneck is electricity

The AI bottleneck is moving from chips to electrons. You can make more chips; you can’t, on the same timeline, make more electricity — or the grid to carry it.

Read →

Where the rent sits: reading the AI value chain

Revenue in a technology build-out appears everywhere at once — but durable profit concentrates in a few physical bottlenecks. A framework for telling the two apart.

Read →

Is this 2000 again?

Every chip rally meets the same question: is this 2000 again? The better question is which part of 2000 — and why real earnings change the answer.

Read →

Video commentary — coming soon.

Short video views will join the written pieces here in due course.

10 June 2026 AI Value Chain

Where the rent sits: reading the AI value chain

In any technology build-out, the question that decides returns is not who grows fastest — it is which layer keeps the profit once the growth arrives.

Every large build-out in technology produces the same optical illusion. Revenue appears everywhere at once — across chips, systems, power, networks and software — and it becomes tempting to treat broad exposure to the theme as the same thing as owning its economics. It rarely is. Revenue share and profit share are different maps, and they seldom line up.

A more useful way to read a fast-moving industry is as a value system: the full chain of layers that carry a product from raw inputs to the end customer. In the current artificial-intelligence build-out that chain runs, roughly, from electrical power and data-centre capacity, through advanced memory, optical interconnect and leading-edge fabrication, up to the compute systems, the models trained on them, and finally the applications people actually use. Growth touches all of it. Durable profit does not.

Where the rent sits

Economic rent — profit that persists rather than being competed away — tends to accumulate at whichever layer is at once the scarcest, the hardest to substitute, and negotiating against the weakest counterparties. In the present cycle several of those layers are stubbornly physical: the power to run data centres, the most advanced memory, the optics that move data between chips, and the handful of facilities able to manufacture at the leading edge. They are slow to build, capital-hungry and difficult to replicate — which is precisely why their pricing holds while more abundant layers commoditise.

Broad exposure to a theme is not the same as owning its economics.

Why durability decides

Identifying a bottleneck is only half the work. A layer is worth owning only if something keeps competitors from closing the gap — capital intensity, accumulated process know-how, an ecosystem that is expensive to leave. Absent that barrier, today’s bottleneck becomes tomorrow’s commodity and the rent migrates elsewhere along the chain. Bottlenecks move as an industry matures; the discipline is to keep asking where the constraint sits now, not where it sat a year ago.

None of this requires predicting which name wins the narrative. It requires locating where the economics actually settle, testing whether that position can be defended, and re-checking the answer as the chain evolves. That is unglamorous work. Over time, it is also where the returns tend to live.

The views above are general in nature and provided for information only. They do not constitute investment advice, an offer, or a recommendation to buy or sell any security.

28 July 2026 Markets & History

Circular by design — what the AI build-out shares with every revolution before it

Critics call today’s AI spending ‘circular’ and hear a bubble. History says circularity is how capital-intensive revolutions have always financed themselves — the question is not whether money loops, but whether the loop is healthy.

A familiar objection has attached itself to the artificial-intelligence build-out. The chipmaker invests in the model labs; the model labs spend that capital back on the chips; the cloud providers underwrite the demand that justifies the next order. Draw the arrows and you get a circle, and to a certain kind of sceptic a circle means a bubble — money chasing its own tail until the music stops.

It is a tidy story. It is also a misreading of how every industrial revolution has actually been built.

Circularity is not a defect of the AI cycle; it is a feature of capital-intensive transformation. General-purpose technologies — rail, electricity, the automobile, the internet — are too big and too uncertain to be pulled into existence by diffuse, arm’s-length demand. They need an anchor firm: a leader with the balance sheet, the credibility and the incentive to reinvest ahead of the market, set the standards, and drag an entire ecosystem up behind it. The capital loops because, early on, the leader is effectively financing its own future customers. That is not fraud. It is bootstrapping.

The right question, then, is not ‘is it circular?’ — it almost always is — but whether four conditions hold. Is the anchor financially healthy, funding the build from real profits rather than hope? Is the business pipeline smooth — is spending converting into shipped product, booked revenue, and repeat orders? Is there a genuine business case underneath, real end-use that would exist even if the loop were cut? And is there a credible path to future prosperity — durable demand, not a one-time land-grab? Where those hold, ‘circular’ is simply the plumbing of a revolution financing itself. Where they don’t, the circle is hollow, and it breaks.

The cautionary case first

Start with the sharpest warning, the one the sceptics reach for: the late-1990s telecom boom. Equipment makers like Lucent and Nortel lent their own customers the money to buy their gear — vendor financing aimed at demand that did not yet exist. That circle was genuinely hollow: the end-use hadn’t arrived, the balance sheets buckled, and the bust was brutal. Yet notice what survived. The fibre laid in the frenzy became the backbone of the modern internet economy — search, streaming, the cloud. Even a hollow circle left durable infrastructure behind. The lesson is not that circularity kills, but that circularity standing in for real demand kills — and even then, the useful capital often outlasts the crash.

Then the ones that built the world

Go back further and the pattern turns constructive. The American railroads ran on a tightly interlinked loop — Morgan’s capital underwriting lines that bought steel from mills the same financiers backed, with Carnegie’s works at the centre. Speculative excess produced real panics, in 1873 and 1893, that cleared the weakest players. But the rails endured, and the network they left behind carried a century of growth. The circularity funded genuine infrastructure with genuine demand behind it; the froth burned off, the value stayed.

Electricity tells it from the leader’s side. Edison, Westinghouse and General Electric did not wait for demand to appear — they manufactured it, funding adoption, standardising the grid, and elevating a whole ecosystem of utilities and appliance makers in their wake. The anchor firms spent to create the market, and the market repaid them for decades.

Which brings us to now. A chip champion such as Nvidia reinvesting to pull the AI stack into being is not an anomaly to be feared; it is the anchor-firm pattern, running to schedule. The discipline is to hold it to the four tests rather than to the shape of the diagram — to ask whether the profits are real, the orders are converting, the end-demand is compounding, and the prosperity on offer is durable. Judge the health of the loop, not the fact that it loops.

Every revolution worth the name has looked circular from the outside.

The circularity objection mistakes the mechanism for the malady. What separated the revolutions that made fortunes from the ones that made cautionary tales was never the geometry of the capital — it was whether real value was being created inside the circle.

The views above are general in nature and provided for information only. They do not constitute investment advice, an offer, or a recommendation to buy or sell any security.

22 May 2026 Semiconductors

Is this 2000 again?

Every semiconductor rally eventually meets the same question. It deserves a better answer than yes or no.

In the spring of 2000, the Nasdaq Composite touched 5,048 — a number it would not see again for fifteen years. Within thirty-one months it had fallen roughly 78%, and the Philadelphia Semiconductor Index, the market’s purest bet on chips, fell about 80% with it. Ever since, whenever semiconductors run hard, a familiar chorus starts up: is this 2000 again? The bubble-callers — who have reliably predicted twelve of the last three crashes — are back, and this time they have charts.

They are not obviously wrong to ask. But “is this 2000?” is the wrong question. The useful one is narrower: which part of 2000?

The bear case is real, and worth stating at full strength. A handful of chip names drive much of the index. Capital spending on AI runs into the hundreds of billions. The financing has taken on a circular quality that rhymes with past manias, and at moments in 2026 the SOX’s volatility has reached dot-com-era levels (Benzinga, June 2026). When the people who called the last crash start pointing at yours, check your homework.

But look at what actually broke in 2000, because it was not chips-the-technology. It was chips-the-valuation, resting on earnings that did not exist. Cisco, the bellwether of its day, changed hands at more than 100 times earnings; hundreds of dot-coms had no earnings at all — only a story and a burn rate. The bet was on revenue that had not arrived, and for many never would. When the story wobbled, there was nothing underneath to catch it.

Today’s leaders are a different animal. Whatever you make of their prices, the largest chip companies earn tens of billions in actual, taxable profit, sell every unit they can make, and generate cash rather than burn it. That does not make them cheap, and it certainly does not make them safe. It makes them the opposite of 2000’s problem. In 2000 the danger was paying for earnings that would never come. Today’s danger is paying for earnings that are real but extrapolated too confidently — assuming this pace of demand compounds in a straight line, in an industry whose defining trait for fifty years is that it never does.

In 2000 the market paid for earnings that never came. The modern risk is paying for real earnings, extrapolated too far.

Because semiconductors are cyclical — gloriously, violently cyclical. The SOX has had several 40–80% drawdowns since 1995 that had nothing to do with a bubble and everything to do with the ordinary rhythm of inventory, capacity and price. AI has bent that cycle, perhaps profoundly. It has not repealed it.

So where do we land? Not on “it’s 2000,” and not on “it’s different this time” — the four most expensive words in markets, closely followed by “it’s 2000 again.” The honest read: the build-out is real, the earnings are real, and the infrastructure being laid will outlast whatever correction eventually arrives — and one will. The task isn’t to guess the top but to separate the two things the crowd keeps conflating: the enduring value of the physical build-out, and the price stapled to it this quarter. The first is a decade-long story; the second is subject to the same cycle it always was. The tell we watch isn’t the valuation headline — those are always arguable — but the boring data underneath: inventories, lead times, capacity coming online, pricing. That’s where a chip cycle turns, long before anyone agrees whether it was a bubble.

Is this 2000 again? Probably not in the way the question implies. But if you remember one thing from that era, make it this: the internet was exactly as transformative as the optimists promised — and the Nasdaq still fell 78% on the way to proving them right.

The views above are general in nature and provided for information only. They do not constitute investment advice, an offer, or a recommendation to buy or sell any security.

24 June 2026 AI & Power

AI’s real bottleneck is electricity

The scarce input in artificial intelligence is quietly shifting from silicon to something older and much harder to build.

There is a number in the International Energy Agency’s work on AI that should reframe how you think about the boom. The IEA expects electricity used by data centres to roughly double by 2030, to around 945 terawatt-hours — slightly more than the entire annual electricity consumption of Japan today (IEA, Energy and AI, 2025). Not one data centre. Not one company. A large industrial nation’s worth of new demand, much of it to feed machines that think.

For two years the AI conversation has been about chips — who makes the fastest ones, who can buy enough. That bottleneck is real, but it is moving. Given money and time, you can make more chips. You cannot, on the same timeline, make more electricity — or the grid to carry it.

The optimists have a fair rebuttal. Chips keep getting more efficient — more computation per watt each generation — and models are run more cleverly. If efficiency improves fast enough, demand growth softens and the grid catches up. Efficiency has rescued many a doom forecast before.

The trouble is arithmetic. Efficiency per chip is rising, but the number of chips and the size of the models are rising faster — the classic rebound effect, where cheaper compute simply invites more of it. And even on a gentler demand curve, the binding constraint isn’t generation in the abstract. It’s dispatchable, reliable power in a specific place, wired to a specific data centre, now. In much of the US, the queue to connect a large new load to the grid runs to several years. You cannot download a substation.

Which is why the smartest money in AI has quietly gone shopping for power. In 2024 Microsoft signed a deal to help restart a reactor at Three Mile Island — yes, that Three Mile Island — to feed its data centres (Constellation Energy, 2024). Amazon and Google have moved on nuclear too, including next-generation small modular reactors. When the world’s most valuable software companies start behaving like utilities, they are telling you where the constraint has gone.

You can, with money and time, make more chips. You cannot, on the same timeline, make more electricity.

The economics follow the scarcity. In any build-out, durable profit collects at whatever link is hardest to replicate — and a permitted, powered, grid-connected site is far harder to conjure than another rack of servers. Power, transmission, cooling, the equipment that manages them: these unglamorous, capital-heavy, slow-to-build layers have a way of becoming the toll booth everyone else has to pay.

We wouldn’t bet against human ingenuity on efficiency — it has a long unbeaten record. But betting that efficiency alone dissolves an electricity bottleneck this large, this fast, is a hope, not a plan. The grounded view is that power is becoming the AI build-out’s true scarce input, and the tells worth watching are physical, not financial: grid interconnection queues, power-purchase agreements, the price of electricity near the big data-centre clusters, and how many software companies keep improbably turning into energy companies.

The AI story has always been about intelligence. It may turn out the binding constraint is the least intelligent thing in the chain — the humble electron, and the century-old infrastructure that moves it.

The views above are general in nature and provided for information only. They do not constitute investment advice, an offer, or a recommendation to buy or sell any security.

8 July 2026 Rates & Fiscal

The long end is talking

The most important conversation in markets isn’t on an earnings call. It’s in the 30-year Treasury — and it’s getting harder to ignore.

There’s an old line that the bond market is the one participant nobody can spin, because it’s simply a very large, very unsentimental creditor doing arithmetic. Lately that creditor has been raising its price. The US 30-year Treasury yield has pushed back above 5%, to levels last seen before the 2008 financial crisis (CNBC, 2026). For a generation raised to think of long bonds as the safe, boring corner of a portfolio, that’s a meaningful thing.

The reflex is to read a higher long yield as a growth or inflation story. Some of it is. But the more interesting driver is a phrase that spent a decade near zero and has now returned: the term premium — the extra yield investors demand simply for the risk of lending long, over and above where they expect short rates to go. When it rises, the bond market isn’t forecasting a boom. It’s charging rent for uncertainty.

Plenty of thoughtful people say the move is technical and temporary: heavy issuance to digest, some positioning, a term premium merely normalising after years of central-bank bond-buying artificially crushed it. On this view, 5% on a 30-year isn’t a crisis — it’s a homecoming.

Perhaps. But look at what the creditor is being asked to finance. The US is running deficits near 6% of GDP with unemployment low and no recession in sight — a peacetime, full-employment borrowing rate with almost no modern precedent outside war. The supply of long bonds is large and growing; the price-insensitive buyers of the past (foreign central banks, domestic banks) are less reliably there. When supply climbs and the marginal buyer wants more compensation, the price of duration rises. That isn’t a technical wrinkle. It’s a market repricing the cost of a government borrowing heavily, for longer, into an uncertain future.

A rising term premium isn’t the bond market forecasting a boom. It’s the bond market charging rent for uncertainty.

You can see officials feeling it in the plumbing — leaning on shorter-dated issuance and buybacks to manage the long end. Those are duration-management tools: they smooth the path, they don’t change the destination. The underlying question — who holds a rising pile of long-dated government debt, and at what yield — doesn’t disappear because the issuance calendar is rearranged.

We don’t read 5% long yields as imminent catastrophe, and we’re wary of the permanent-crisis crowd who’ve predicted a bond apocalypse every year for a decade. But we’d take the term premium seriously, because it prices something real: the market’s growing demand to be paid for lending long to a heavy borrower in an uncertain fiscal era. The tell isn’t the yield headline on any given day — those swing on data — but whether the term premium keeps grinding higher across the noise. If it does, the long end isn’t malfunctioning. It’s telling you the price of duration has structurally changed — and a great many asset prices are quietly calibrated to the old one.

The bond market, as ever, is talking. The only question is whether the rest of the market is listening.

The views above are general in nature and provided for information only. They do not constitute investment advice, an offer, or a recommendation to buy or sell any security.

22 July 2026 Credit

When wide spreads matter — and when they don’t

Credit spreads are the market’s smoke alarm. The trick is telling a fire from a low battery.

In mid-2026, the extra yield investors demand to hold risky corporate bonds over safe government ones — the high-yield spread — sat around 280 basis points, or 2.8 percentage points (ICE BofA). By the standards of credit history, that isn’t just calm; it’s nearly serene. Which, depending on your temperament, is either reassuring or precisely the problem.

There are two ways to read a tight spread, and they can’t both be right. One says: credit is relaxed, defaults are low, the coast is clear. The other says: complacency is the disease, and the tightest spreads come right before the fall. The second view carries a seductive wisdom. It is also, most of the time, wrong.

The worriers aren’t imagining things. Spreads this tight do mean investors are paid very little to take credit risk, leaving thin cushion if something breaks. And the record is full of calm credit markets that flipped to chaos with startling speed. “You are not being paid for risk” is a respectable reason to be watchful.

But here’s what the history actually shows, and it’s more useful than “tight equals danger.” Wide spreads, on their own, have a poor record of causing crashes. What causes crashes is a spread that is widening fast, driven by a systemic shock that policymakers cannot or will not contain. In 2008, high-yield spreads blew out to roughly 2,000 basis points — because a truly systemic event, the seizure of the banking system’s plumbing, had no immediate backstop. In March 2020, spreads spiked past 1,000 basis points in weeks — then collapsed almost as fast, because the Federal Reserve stepped in and, for the first time, backstopped corporate credit directly. Same instrument, opposite outcome, decided by what policy did.

Now set those against the false alarms. In 2011, 2015–16 and 2018, spreads widened meaningfully — and no cascade followed, because each shock was either contained by policy or isolated to one corner. The alarm flashed; the fire never caught.

Wide spreads don’t cause crashes. A wide spread, widening fast, with a systemic shock and no backstop — that causes crashes.

So the signal isn’t the level alone. It’s the combination: the level, the speed it’s moving, whether the shock behind it is systemic or contained, and whether a central bank stands ready to catch it. A spread of 280 tells you something true but limited — that credit is not, today, the market’s source of stress. It is not the all-clear, and it is not proof the fall is nigh. It tells you where the smoke isn’t.

We treat credit spreads as one of the most honest gauges in markets — a thermometer, not a crystal ball. Tight spreads today mean the credit system is calm, and calm is a fact, not a forecast. The permabear’s error is treating the low-battery chirp as proof of fire; the complacent one’s error is ripping the alarm off the wall. The disciplined middle watches the rate of change and the context — is it widening fast, is the shock systemic, is there a backstop? When those line up the wrong way, spreads stop being a thermometer and start being a warning. Until then, a serene credit market is quietly telling you the next stress will likely arrive from somewhere else first.

Which, if you sit with it, is the more uncomfortable thought. The next crisis rarely announces itself in the gauge everyone is already watching.

The views above are general in nature and provided for information only. They do not constitute investment advice, an offer, or a recommendation to buy or sell any security.

4 October 2026 Rates & Fiscal

The creditor with a redeem button

America’s debt pile has found an unlikely new buyer. It isn’t a central bank, and it doesn’t hold for decades.

Last month the 10-year Treasury yield touched 5.04%, its highest since 2007, and the Federal Reserve raised rates for the first time since 2023. The gross national debt crossed $40 trillion in August, having passed $38 trillion last October and $39 trillion in March — a trillion dollars every five months. Net interest ran to $970 billion in the 2025 fiscal year, roughly $150 billion more than the United States spent on its military.

Writing in Fortune, a former Senate budget staffer reached for Hemingway: bankruptcy happens gradually, then suddenly. It is a good line, and the arithmetic behind it is not in dispute.

But the number everyone stares at is the debt. The more interesting question sits on the other side of the ledger — not how much America owes, but who is holding it, and how quickly they can leave.

Start with the fear, because it is the wrong one. The International Monetary Fund’s reserve data for the first quarter of 2026 puts the dollar at 57.13% of allocated global reserves, down from 64.22% in early 2017 — a real erosion. It is also up from 56.42% in the final quarter of 2025.

More telling is who gained. Not the euro, which sits at 20.03% and has been going nowhere. Not the renminbi, at 1.99%. The dollar’s lost share has drifted to the Canadian and Australian dollars and a long tail of smaller names. This is fragmentation at the margin, not succession.

Tokenization was supposed to accelerate the exit. The opposite happened. As of June 2026, dollar-backed stablecoins were 97% of a $316 billion market. Euro stablecoins came to about $635 million — roughly two parts in a thousand. Putting the dollar on a blockchain turned out to be a way of exporting it to places that previously found it hard to hold one.

Here is the part that deserves more attention than it gets. Stablecoins are backed, and overwhelmingly they are backed by short-dated US government paper. By mid-2026 the sector held roughly $300 billion of Treasury bill exposure. Tether alone held over $141 billion — enough to place it among the twenty largest holders of US government debt on earth. Circle keeps about 80% of its reserves in bills.

Every dollar of stablecoin issued is, mechanically, a purchase of Treasuries.

So at the precise moment foreign official investors are taking a shrinking share of American debt, a new and entirely unplanned buyer has appeared at the short end. Washington has noticed, and the policy direction has been to encourage it rather than restrain it. On the narrow question the Fortune piece raises — will the world keep funding America? — the honest answer this year is yes, and from a direction nobody mapped.

America hasn’t lost a creditor. It has swapped part of a pension fund for part of a current account.

Which is where we would push back on the comfort. A foreign central bank is a patient holder. It buys for policy reasons, sits through drawdowns, and does not redeem because sentiment turned on a Tuesday. A stablecoin reserve is the opposite: redeemable on demand, by anyone, instantly — and concentrated, with two issuers accounting for over 80% of the market.

Bankers have a word for this distinction, and it is not a compliment to the new money. It is the difference between a term deposit and a current account. The balance sheet looks identical until the day it doesn’t.

We have already had the rehearsal. When Silicon Valley Bank failed in 2023, $3.3 billion of USDC reserves became temporarily inaccessible and the coin briefly lost its peg. Nothing systemic followed. But the mechanism was visible: a redemption wave becomes a forced sale of Treasury bills into whatever market happens to exist that morning.

None of this threatens the dollar’s reserve status. It was never going to. It changes something subtler — the duration, and the nerves, of the money funding the deficit.

The sterling precedent is instructive precisely because it is slow. The dollar overtook sterling in 1929 outside the Commonwealth, and the two ran at 45% and 51% in 1931. Sterling was still formally a reserve currency through the 1947 convertibility crisis, through Suez in 1956 costing £650 million of reserves, through the 14.3% devaluation of 1967. By 1952 it was already under 10% of Western European central bank reserves. The status was the last thing to go, not the first — every painful episode happened while Britain still held the title.

So we would spend less time watching reserve share, which lags by decades, and more on the composition and stickiness of the marginal buyer. The dollar’s position looks secure. Its funding base is quietly getting faster on its feet.

Which leaves a question we cannot answer. If a run on a stablecoin becomes a run on the bill market, is that a crypto problem or a fiscal one — and has anyone decided whose job it is?

The views above are general in nature and provided for information only. They do not constitute investment advice, an offer, or a recommendation to buy or sell any security.

Contact

Welcome inspiration and intellectual sparkles

Apex Edge Capital Singapore is a small, independent shop — and some of the best thinking comes from comparing notes with people who look at markets the same way. If you invest in or build around the AI value chain, technology or the macro backdrop, I’d genuinely enjoy trading ideas, debating a thesis, or simply learning from how you see it. No agenda — just good conversation between like-minded investors.

Principal
Joanna TangFounder & Director
Email
Based in
Singapore