Est. MMXV · Hong Kong
Our investment process is built on the belief that enduring excess returns come from owning hard-to-replicate assets and capabilities and taking well-founded views that differ from market consensus. We seek to identify structural changes before they are reflected in prices, recognizing that this requires patience, as the difference between being early and being wrong often only becomes clear over time.
We sized two positions around one idea: the coming decade would reward scarcity, and the two scarcities that mattered most were computational capacity and monetary supply.
NVIDIA was the compute leg. At a market capitalization of about $250 billion, a fraction of what it would later become, the market still valued it largely as a cyclical semiconductor and gaming name. We thought that framing was already stale. Data center had become its fastest-growing segment, workloads were parallelizing, and model sizes were climbing. The moat was never only the silicon. It was CUDA and the software layer that made the silicon expensive to leave, and that was the part the price ignored.
Bitcoin was the monetary leg. Trading near $9,000 and still treated as fringe, it offered the one property fiat could not: a supply schedule no policymaker could revise. Set against open-ended balance-sheet expansion, a credibly fixed supply looked less like a curiosity and more like an option on declining confidence in paper money.
The market saw a chip stock and a speculative token. We saw one trade expressed two ways.
The consensus word was “transitory.” We disagreed, and the disagreement was the position.
Monetary accommodation, direct fiscal transfers, near-zero policy rates, and M2 growth running north of 25% year over year had never arrived together outside of wartime. The relevant question was not whether prices would rise for a few quarters, which was close to certain, but whether the low-inflation regime that had anchored a decade of long-duration positioning had actually ended. We thought it had.
With the ten-year yield near 1.5% and real yields deeply negative, the asymmetry was clear. We rotated toward businesses that could raise prices without losing volume, assets with genuine scarcity, and cash flows geared to nominal rather than real growth. If the regime had turned, long-duration paper was the wrong place to be.
Uranium had spent a decade as an orphaned commodity, and that was the opportunity. Ten years after Fukushima, spot sat near $30 a pound, below the marginal cost of new production and far beneath the level needed to justify restarting idled mines. Supply had been quietly destroyed. Producers had shuttered capacity, high-cost mines had closed, and almost nothing new had been financed.
The demand side, meanwhile, was turning. Reactor build-out across Asia continued, Western utilities were drawing down inventories they would eventually have to replace, and nuclear was regaining political acceptance as a precondition for any credible decarbonization path. A commodity priced below the cost of its own replacement does not stay there for long.
We accumulated at those levels. The thesis needed no catalyst on any particular date. It needed only for the arithmetic of supply and demand to reassert itself, which in a market this small tends to happen violently rather than gradually.
Written weeks before the invasion of Ukraine, this was a bet on direction rather than event.
The investment world still ran on two post-Cold War assumptions: that globalization would keep deepening, and that the peace dividend was permanent. Both were fraying. We read the accumulating frictions across trade, technology, and supply chains as a system reorganizing around strategic competition rather than efficiency.
We expressed it through defense, adding TOKYO KEIKI (7721.T), and through gold. The thesis needed no single flashpoint. Governments that feel less safe spend more on defense, energy security, and industrial capacity, and they do it for years. That is a spending trend, not a trade.
As the market rushed the GPU, we went looking for what the GPU plugs into.
Every computing revolution eventually presents a physical invoice: power, cooling, networking, and metal. A data center is not just processors. It is megawatts, thermal management, and industrial supply chains that take years to build. That is where the bottleneck, and the pricing power, migrate once the marquee name is fully owned.
We bought Vertiv (VRT) at around $15 a share for the power and thermal layer sitting directly beneath the compute, and silver, then in the low-to-mid $20s per ounce, for its widening industrial footprint across electrification, solar, and electronics. Neither was an AI stock. Both were levered to AI’s electricity and hardware bill, which we expected to compound faster than the market assumed.
By mid-2024 several slow-moving forces stopped running in parallel and began to reinforce one another.
A decade of underinvestment in energy, mining, and heavy industry was colliding with rising electricity demand, supply-chain regionalization, larger defense budgets, and the power appetite of AI. After roughly two decades of flat U.S. electricity demand, utilities were revising load-growth forecasts higher for the first time in a generation. The arithmetic pointed one way: the world would have to spend far more on physical capacity than it had grown used to.
We expected Trump to win in November, and that was consistent with the thesis, but it was not the thesis. The capex story did not depend on the ballot. It rested on years of capital starvation running into a wall of future demand. On that view, commodities were not a cyclical trade to rent. They were the beneficiaries of a shortage the market had not yet decided to pay for.
Attention was still fixed on raw processing power. We think that focus is too narrow for where the cycle is heading.
Training frontier models does not only consume compute. It consumes high-bandwidth memory, enterprise storage, and a widening set of advanced components, and those inputs were not priced for the volumes the roadmap implies. History is unkind to the assumption that only the most visible supplier gets paid. As an ecosystem matures, value migrates outward into the adjacent infrastructure.
We took positions in Micron and in SanDisk, the latter freshly separated from Western Digital and trading again as a standalone flash pure play. Memory is not the story the market tells about AI. We think it is one of the inputs it ends up fighting over.
This is the clearest expectations gap we see across global markets, and we are expressing it as a pair.
Private credit has spent years enjoying cheap money, compressed spreads, and a wall of demand for yield, growing past $1.5 trillion under exactly the conditions in which underwriting quietly deteriorates and return expectations drift above what the paper can plausibly deliver. We are not calling a specific crack. We are saying the risk-adjusted return no longer compensates for what is being taken on, whatever the marketing says.
Chinese equities are the mirror image, priced for something close to despair after more than four years below their early-2021 highs. Valuation, a turning policy stance, and stabilizing fundamentals have skewed the distribution of outcomes to the upside, which is not how the consensus is positioned.
We are not trying to time either one. The trade is the gap between price and fundamentals: paying too much for private credit’s optimism, and being paid to hold China’s pessimism. When positioning gets that lopsided, the imbalance is the opportunity.