Investment Gems

The market keeps trying to force the AI buildout into an old template.

Some investors treat it like a software bubble. Others treat it like a datacenter land grab. Still others reduce it to a single question: is demand real?

That is the wrong framing.

Demand is real. Scarcity is real. The harder question, the one equity investors actually get paid for answering, is whether a company can convert real scarcity into per-share value before its capital structure consumes the upside.

That is the frame I keep coming back to:

Long the bottleneck, short the cost of capital.

In this cycle, many companies are clearly long the bottleneck. They control some combination of scarce power, GPU access, energized capacity, land, cooling, interconnect, logistics, or regulatory position. But many of those same companies are also implicitly short their own cost of capital. They need so much money, so quickly, to monetize the opportunity that the financing stack can become more important than the asset itself.

That is the central analytical mistake in today’s market. Investors keep treating relevance as return. They see scarcity and assume value accrues automatically. It does not. Value accrues to the owners of scarce assets only if they can finance expansion without turning the common stock into a perpetual funding mechanism.

This is why capital structure is no longer adjacent to the thesis. It is the thesis.

The AI buildout is an industrial bottleneck story first

For all the talk about agents, models, wrappers, and software flywheels, the market remains constrained by a much more basic set of realities:

• grid-ready power
• permitted sites
• substations and transmission access
• cooling architecture
• GPU supply, especially memory-constrained supply
• networking and integration
• reliable deployment and uptime

That is why so many conversations that begin in software end in infrastructure.

If compute were abundant, the economic center of gravity would already have shifted decisively to the platform layer. But compute is not abundant. Powered, operational, financeable compute remains scarce. That means the physical world is still setting the pace of the digital one.

This matters because scarcity changes the economics of businesses that would otherwise look mediocre. Commodity businesses are bad when supply is abundant. They can be excellent when capacity is rationed. In undersupplied markets, the “boring” owner of the bottleneck can earn extraordinary economics, at least for a period.

But only for a period. Which brings us back to financing.

Scarcity is not the same as shareholder return

A recurring error in markets is to mistake proof of relevance for proof of return.

Claude being rate-limited proves demand. It does not prove CRWV’s equity is undervalued.

A national trust charter for a crypto firm proves regulatory opening. It does not prove the stock is cheap.

A big OEM logo proves a solid-state battery company got in the room. It does not prove volume, margin, or manufacturability.

A large resource in the ground proves strategic relevance. It does not prove the common equity will compound.

The same mistake is now being made all over the AI capex complex. Investors are identifying the right bottlenecks and then skipping the hard part: who pays for the build, on what terms, with what protections, and at whose expense?

That is why “Long the Bottleneck, Short the Cost of Capital” is such a useful frame. It forces you to ask the only question that really matters:

Can the company turn scarcity into contracted cash flow before dilution, debt burden, or refinancing risk turns the asset into someone else’s return stream?

IREN: the purest expression of the framework

IREN is probably the cleanest case study because the company sits right on the fault line between industrial scarcity and capital market gravity.

The bull case is straightforward. IREN appears to control exactly the kinds of assets the market still cannot conjure on demand: power, sites, deployable capacity, and growing credibility as a serious infrastructure counterparty. If the management commentary we’ve been discussing is directionally right, IREN is not trying to sell a software dream. It is trying to own part of the scarce industrial substrate of AI.

That is the good part.

The harder part is that scarcity alone does not make the common stock work. IREN also needs enormous amounts of capital. The moment the company expanded the ATM, the stock ceased to be just an infrastructure story and became a capital structure story.

That does not automatically make the ATM bad. The lazy critique is to compare every large equity authorization to meme-stock dilution. That misses the point. AMC diluted to survive. IREN is trying to preserve the option to dilute in order to build. Those are not the same thing.

But the distinction only helps if the capital is raised into something financeable and accretive.

This is why I keep describing IREN as long the bottleneck, short its own cost of capital.

The company’s strategic opportunity is obvious: if there are effectively no idle production GPUs and if time-to-compute still matters more than perfect theoretical architecture, then real powered capacity has scarcity value. In that world, IREN does not need to become Nebius or Oracle Cloud to win. It does not need to own the full software layer. It just needs to own enough real capacity to matter while the market remains structurally short.

The risk is equally obvious: if it builds ahead of financeable demand, then the common equity becomes the shock absorber. Scarcity may remain real, but per-share value can still leak away.

So for IREN, the key variable is not whether demand exists. It is whether scarcity gets translated into:

• take-or-pay style contracts
• customer prepayments
• JV structures
• project-level debt
• GPU financing
• and enough non-equity capital that the ATM becomes a bridge rather than a lifestyle

If that happens, the ATM will look strategic in hindsight. If it does not, the ATM will become the story.

Oracle: even incumbents are discovering that balance sheet is now strategy

Oracle offers the same lesson from the other side of the market.

Unlike IREN, Oracle does not suffer from an early-stage credibility discount. It has an installed base, enterprise relationships, and a far deeper pool of legacy cash flow to harvest. In theory, that should make it one of the clear winners of AI infrastructure.

And yet the rumored behavior we discussed, labor cuts, asset-sale chatter, “bring your own chip” structures, and aggressive capex positioning, suggests something deeper: even a mature incumbent may need to cannibalize legacy cash flows and legacy organizational structure to stay in the race.

That is a profound signal.

It implies that AI is no longer just a product opportunity. It is a corporate reallocation event. Old margins, old divisions, old payrolls, and even old strategic identities are being mined to fund new infrastructure.

In other words, the org chart is becoming part of the financing stack.

That is why Oracle matters in this framework. It shows that capital structure pressure is not just a problem for smaller, more speculative names. It is becoming universal. The winners may not be the companies with the best AI demos. They may be the companies with the strongest ability to turn legacy cash flows into new capex without collapsing the economics of the legacy franchise in the process.

Oracle is not short relevance. It is short duration mismatch. Cash outlays happen now. Returns may take years. That gap is where the risk lives.

Strategy: capital structure itself can become the product

Strategy is not an AI infrastructure company, but it may be the clearest modern example of the underlying principle.

What Strategy has demonstrated is that in a scarcity narrative, the capital stack can stop being a passive funding tool and become an active operating advantage. It has taken one volatile asset, Bitcoin, and progressively sliced exposure into forms that appeal to different classes of capital.

Common equity captures one risk profile. Convertible and preferred-like structures capture others. The more the company can transform one underlying scarcity asset into multiple investable claims, the more it widens its funding base and deepens its strategic flexibility.

That is relevant for AI because it suggests something many investors still underestimate: the companies that win these capex races may not simply be the ones with the best assets. They may be the ones with the best financial architecture around those assets.

This is why Strategy belongs in the same conversation as IREN and Oracle even though the underlying asset is different. The transferable lesson is that once capital structure becomes part of the product, the company’s strategic degrees of freedom expand dramatically.

Why the lithium and QS analogies matter

The analogies to lithium and solid-state batteries are useful because they show how often investors get this wrong in every disruption cycle.

With lithium, the macro story can be completely right, EVs, grid storage, strategic materials, domestic supply security, and yet the wrong equities still disappoint. Why? Because resource scale, recycling potential, or national importance do not eliminate timeline risk, capex intensity, permitting drag, or dilution.

With solid-state batteries, the same mistake happens in a different form. Investors see customer logos and start mentally annualizing future market share. But logos do not equal commercial ramps. And commercial ramps do not equal attractive returns if manufacturability and cost remain unresolved.

The lesson is always the same: strategic relevance does not immunize equity holders from financing risk.

That is the lesson many AI investors still have not learned.

The market is still mispricing what kind of cycle this is

One reason this framework matters so much now is that the market has not fully decided whether the AI buildout is a normal capex cycle or something closer to strategic mobilization.

If this is merely a speculative build, then overcapitalized bottleneck owners will probably destroy value the way commodity players often do: overbuild, price collapse, disappointing returns, and equity issuance that socializes the mistake.

But if this is closer to industrial rearmament, if governments, hyperscalers, enterprises, and sovereign actors all treat compute as a strategic necessity, then the duration of spend could be longer than skeptics expect. In that world, scarcity persists longer, customer commitment deepens, and bottleneck owners have a larger window in which to turn relevance into real cash flow.

Even then, however, the capital structure question does not go away. It becomes more important. In a strategic buildout, the winners are not merely the owners of scarce assets. They are the owners of scarce assets with the cheapest and most flexible capital.

That is the non-consensus point. The real moat may not be the data center. It may be the ability to finance the data center.

What to own, and what to avoid

This framework leads to a fairly clear screen.

You want companies that have:

• genuine exposure to a real bottleneck
• evidence of customer urgency
• some form of financing diversity beyond common equity
• insider incentives aligned with limiting dilution
• a path from relevance to contracted monetization

You want to avoid companies that:

• are pre-selling strategic importance without financing clarity
• rely overwhelmingly on the common stock to fund buildout
• confuse memoranda of understanding with real demand
• are merchant builders in markets that can reprice quickly
• or require heroic assumptions about timing, margins, and capital availability

That is why so many current debates are actually debates about financing, disguised as debates about technology.

The “best” software team can still be a bad investment if it sits on the wrong capital stack.

The “boring” infrastructure owner can still be a great investment if it owns a bottleneck and finds a way to finance it without destroying per-share economics.

The conclusion

The AI buildout has produced a lot of bad categories.

People still ask whether it is a bubble or a revolution. That is too abstract.

People still ask whether demand is real. That is too late.

The better question is simpler and harder:

Who owns the bottleneck, and who can finance it without handing away the upside?

That is the investment problem.

IREN is interesting because it may own part of the bottleneck while remaining exposed to capital market gravity. Oracle is interesting because even incumbents are now liquidating pieces of the old order to finance the new one. Strategy is interesting because it shows how financial engineering can become a strategic operating advantage rather than a footnote.

The common thread is that scarcity is real, but scarcity alone is not enough. In this cycle, capital structure is the decisive variable between relevance and return.

That is why the right posture is not simply bullish AI, or bullish infrastructure, or bullish scarcity.

It is more precise than that.

Be long the bottleneck.
But never forget you may also be short the cost of capital.