← Writing

On the unit economics of inference.

A category priced like software will, in time, earn the margins of compute. The investment question is not whether this is true, but where in the stack the exceptions live.

For most of the last decade, software was understood to be the highest-quality business model available to public and private capital. Gross margins above eighty percent, dollar-based net retention reliably above one hundred, marginal cost of distribution rounding to zero. These properties were so reliable that the SaaS multiple became a stand-in for quality itself.

The current generation of artificial intelligence businesses is being financed at multiples that assume these properties. We think a meaningful fraction of them will not have them.

The reason is mechanical. Software's gross margin is a function of distribution: once written, a line of code costs nothing to deliver to the marginal customer. Inference is not software. Every query consumes compute, every token costs energy, and the cost is paid in real time at the moment of use. The marginal cost of an AI product is not zero. It is the cost of GPU-hours billed against a depreciating asset, plus the energy to run it.

This makes the economics of an AI product look structurally closer to those of a hosting business than to those of a SaaS business. The gross margin ceiling is not eighty percent. It is whatever margin the underlying compute provider chooses to leave on the table — which, given that the compute layer is itself competitive and capital-intensive, is unlikely to be generous over a full cycle.

The two arguments against this view.

There are two thoughtful counterarguments, and both deserve serious treatment.

The first is that inference cost will fall faster than usage rises, restoring software-like margins through deflation. This is partly true and partly misleading. Inference costs per token are falling sharply. But the average query is becoming dramatically more expensive: longer context windows, more reasoning, more tool calls, more multimodal inputs. The net effect on cost per useful unit of work has been ambiguous. Companies that planned for monotonic deflation have been surprised.

The second is that model differentiation creates moats: a company with a better proprietary model can charge a price untethered from compute cost. This is true for a small number of frontier labs. It is not true for the long tail of application-layer companies, which depend on access to models they do not control, on terms they do not set, with switching costs that are technical rather than contractual.

Where the exceptions live.

This does not mean AI is uninvestable. It means the investable surface is narrower than the financing volume implies, and the durable businesses are concentrated in particular structural positions.

The first position is owning the infrastructure layer itself — compute, data, orchestration, the tools that make inference reliable at scale. These businesses look more like cloud than software, but cloud has been an extraordinary category for capital. The margin ceiling is lower, but the demand curve is structural and the switching costs are real.

The second position is at the application layer, where the company owns proprietary data, distribution, or workflow integration that the model cannot replicate. The model is a commodity input. The moat is everything around it. In these cases the margin profile can look like software because the customer is paying for the data and the workflow, not the inference.

The third — and the smallest — is at the model layer itself, for the very small number of labs at the frontier. These are not venture investments in the conventional sense. They are infrastructure bets sized accordingly.

What this means for venture returns.

The implication for fund construction is uncomfortable for a category that has absorbed an enormous amount of capital. If the median application-layer AI company has the gross margin profile of a hosting business, the multiple at exit will reflect that, and the math of the fund will not work at the prices being paid in 2024 and 2025.

The most useful question to ask an AI company in 2026 is not what its product does. It is what its gross margin will look like at scale, and which of the three positions above explains that margin.

We have not invested behind the consensus answer to this question. We have invested where we believed a company occupied one of the three positions defensibly, and we have passed on a number of well-regarded companies where we believed the gross margin story did not survive scrutiny. Time will tell whether we are right. The mechanics, we think, are clear enough.

← Writing

Regulation as market structure.

In digital assets, the regulatory perimeter is not a constraint on the business. It is the business. Investors who treat it as risk-adjacent rather than value-determinative will misprice the category at both ends.

The cultural posture of the digital asset industry toward financial regulation has shifted twice in the last five years. The first shift was defensive — a period in which regulation was framed as an external threat to be litigated, lobbied against, or routed around. The second is more recent and more interesting. It is a posture of structural engagement. Serious operators in the category now read securities law, banking law, and prudential regulation the way an early-stage software founder reads developer documentation. As specification.

This is the correct posture, and the investment implication has not yet been fully priced in.

What changed.

For most of the last cycle, the working assumption in venture and growth investing in digital assets was that regulatory clarity would arrive eventually, that the timing was uncertain, and that companies which built useful things would be permitted to operate once the rules caught up. The strategic posture this produced was reasonable for its moment but is no longer well-calibrated to the present environment.

Three developments have changed the picture. First, regulatory clarity has arrived in a piecemeal but consequential form: through enforcement actions, through banking guidance, through the slow accumulation of precedent in court. The rules are not codified but they are observable. Second, the institutional capital that the category has long courted has finally begun to enter, and it has entered through a narrow door defined precisely by regulatory permission — exchange-traded products, qualified custody, registered intermediaries. Third, the platforms that have done well in this period are not the platforms that built the most novel technology. They are the platforms that read the perimeter accurately and positioned inside it early.

The investable consequence.

The consequence for capital allocation is that the regulatory posture of a digital asset business is no longer adjacent to its valuation. It is determinative of its valuation. A business that has thought rigorously about its securities-law surface, its custody architecture, its licensing footprint, and its banking relationships has, in our view, a fundamentally different risk profile than a business with identical product-market fit that has not.

This is not a defensive observation. It is offensive. The companies that read regulation as market structure are the companies that capture the institutional capital that defines this cycle. They are the companies that win the partnerships, the listings, and the integrations that require counterparty due diligence. They are the companies for which a regulator's letter does not constitute an existential event. The premium attached to this posture is large and, we think, durable.

What we underwrite.

In practice, this means a few things. We underwrite the regulatory architecture of a digital asset business with the same seriousness as its product. We ask whether the company has structured its tokens, its custody, its disclosures, and its market-making relationships in ways that survive scrutiny — not whether they survived scrutiny in 2021, but whether they survive it now.

We are skeptical of theses whose primary edge is regulatory arbitrage. Regulatory arbitrage is a real source of return, but it is not a defensible source of return. Arbitrages close. The businesses that survive their closure are the businesses that built something else underneath.

The investment question in digital assets is no longer whether a company has built something useful. It is whether the company has built something that can be owned, custodied, transferred, and reported on in a way that an institutional counterparty will accept.

This is a narrower investable universe than the category appears from a distance. It is also, in our view, a substantially better one. The companies that meet this bar are buying durable optionality on the institutionalization of an asset class. The companies that do not are buying optionality on the regulatory environment changing — which is a different bet, and one we are reluctant to underwrite.

One implication for stage.

A note on staging. Regulatory architecture is expensive to retrofit and inexpensive to design in early. The companies that thought about it at the seed stage trade at a premium at the growth stage that meaningfully exceeds the cost of having thought about it at the seed stage. This is a structural arbitrage available to early-stage capital that is alert to it, and it has shaped a number of our recent commitments.

← Writing

The replatforming thesis.

The last cycle of financial technology investing was a competitive bet against incumbents. The current cycle is a structural one. The difference is the most important variable in the category.

Between roughly 2015 and 2021, financial technology investing was organized around a simple competitive premise. Incumbent financial institutions were slow, their products were poorly designed, and their customer experiences were stuck somewhere in the late twentieth century. A well-capitalized challenger with better software and a focused product could win a vertical, scale to millions of users, and ultimately threaten the incumbent's franchise.

This thesis worked in some categories — consumer payments, neobanking in certain geographies, small-business software — and failed in others. The failures are instructive. The categories in which challengers did not displace incumbents were the categories in which the incumbent's advantage was not, in fact, competitive. It was structural. Balance-sheet capacity, regulatory permissioning, settlement access, payment-system membership, the cost of capital itself. These are not advantages that better software dissolves. They are advantages that better software is built on top of.

The cycle we believe is now beginning is different in kind. It is not a competitive cycle. It is a structural one. The replatforming thesis is the claim that several of the incumbent advantages that protected financial institutions through the last cycle are eroding — not because of better software, but because the underlying market structure is changing.

What erodes structural advantage.

Three forces are at work, and each deserves its own treatment in detail. They can be summarized briefly.

The first is the rebuilding of payment rails. Real-time settlement, instant cross-border transfer, and programmable money — whether on private rails like FedNow and the equivalent systems in other jurisdictions, or on public rails like stablecoins and tokenized deposits — are eroding the cost-of-funds and settlement-time advantages that defined large banks for decades. The advantage of a clearing relationship matters less when clearing is instant and accessible to a wider range of counterparties.

The second is the disaggregation of credit. The capital-markets ecosystem outside the regulated banking system has grown to a size and sophistication that materially affects who can fund what. Private credit, asset-based finance, and the increasing willingness of pension and insurance balance sheets to fund directly have created credit channels that do not depend on bank intermediation. The structural advantage of being a bank — being the only durable source of certain kinds of lending — is narrower than it was a decade ago.

The third is data. Open banking, open finance, and the broader trend toward consumer-permissioned data access are dismantling the data moats that protected incumbents. A customer's transaction history is no longer the bank's proprietary asset. It is, increasingly, the customer's, and increasingly portable.

None of these forces is complete. None is irreversible. But the direction of each is consistent, and the cumulative effect is meaningful. The incumbents' structural moat is not gone. It is, in specific and measurable ways, narrower than it was.

What this means for capital allocation.

The investable consequence is that the category of fintech companies worth backing has shifted. The last cycle's winners were companies that competed against incumbents on user experience. The current cycle's winners, we think, will be the companies that operate at the seams where structural advantage is eroding — and that capture economics that were previously locked inside the incumbent stack.

The most useful question to ask a fintech company in 2026 is not what it does better than the incumbent. It is which incumbent advantage has structurally weakened, and whether the company is positioned to capture the economic surface that is being released.

This is a narrower set of theses than the last cycle's. It does not include consumer challenger banks competing on UX. It does not include yet another vertical SaaS company with an embedded payments component. It includes companies that are restructuring credit channels, rebuilding payment infrastructure, capturing data flows that incumbents previously owned, and providing the operating systems that financial institutions themselves now need to compete in a faster, more open environment.

An asymmetry worth pricing.

One final point. The replatforming thesis implies an asymmetry that is, we think, underappreciated. If the structural moat erodes, the most valuable position in the resulting landscape is not the challenger. It is the infrastructure provider — the company that provides the rails, the connections, and the orchestration that the new ecosystem requires. These companies tend to look unglamorous from the outside. They sell to financial institutions rather than to consumers. They speak the language of compliance and integration rather than growth and engagement. They compound slowly.

They are also, in our view, the companies most likely to define the next decade in financial technology. We have allocated accordingly.

← Writing

On concentration.

Coverage is the easier discipline. Concentration is the more honest one. The difference shows up in the second decade of a firm, not the first.

A firm's portfolio construction is the most legible statement it makes about how it thinks. Everything else — the website, the thesis, the language of the deal memos — describes a posture. The portfolio is the posture revealed. A firm that holds fifty positions has decided, whether explicitly or by default, that the variance of any single position matters less than the average. A firm that holds eight has decided the opposite. These are different businesses entirely.

We hold few positions. This essay is an attempt to say why, plainly, in a way that does not depend on the rhetoric of conviction.

What coverage costs.

The case for coverage is well-rehearsed and not wrong. Early-stage outcomes are power-law distributed, the best companies are difficult to identify ex ante, and a fund that holds enough positions can absorb a great many failures so long as a small number of winners return the fund. This is the canonical argument for venture portfolio construction, and the math works. We do not dispute it.

What the math obscures is that coverage has costs which are not paid in dollars and which compound over time. The first is attention. A partner who is responsible for a portfolio of forty companies cannot, by simple arithmetic, know any one of them well. The second is selection. A firm that needs to deploy capital across a wide surface of opportunities cannot, in practice, pass on the median company in its pipeline; the rate of pass is set by the rate of deployment, not by the quality of the opportunity. The third is reputation. A firm whose name appears on too many cap tables eventually stops being a signal at all.

None of these costs is fatal. All of them are real. A firm that pursues a coverage strategy is implicitly betting that the benefits of diversification outweigh these costs, and that bet is sometimes right. We think it is more often right at scale — for firms that have institutionalized the workflow and the pipeline — than it is for firms our size. We may be wrong. The evidence will take a decade to assemble.

What concentration requires.

The argument for concentration is not the inverse of the argument for coverage. It is not "fewer companies, better outcomes." That formulation is wishful. The argument for concentration is structural: a firm that holds few positions can underwrite each of them in a way that a firm with many cannot, and can therefore make a different kind of decision at the moment of investment.

What does this different kind of decision look like? It looks like the willingness to pass on a company that almost everyone agrees is excellent, because something in the underwriting does not survive scrutiny. It looks like the willingness to write a check into a category that is unfashionable, because the question being underwritten is not "is this fashionable" but "is this true." It looks like the willingness to wait. A concentrated portfolio is, mechanically, a more patient portfolio. There is less capital to deploy and less pressure to deploy it.

The cost of concentration is the company you missed. The cost of coverage is the company you should have passed on and didn't.

We have made both kinds of mistakes. We expect to continue making them. The question is which kind of mistake the firm is structurally inclined toward, and whether the structure of the firm rewards or punishes the correction of those mistakes. In a concentrated portfolio, the correction is fast: a bad position is visible, and the next decision is not contaminated by twenty other positions making their own demands on attention. In a coverage portfolio, the correction is slower, because the signal-to-noise ratio is lower and the mistake can hide inside the average.

The decision discipline.

Concentration is sometimes described as a temperament. We think this is a category error. It is a discipline, and like all disciplines it can be practiced badly. A concentrated firm that holds eight positions chosen carelessly has none of the benefits and all of the costs of concentration. The discipline is not in the number of positions. It is in the quality of the underwriting that the small number permits, and in the willingness to decline opportunities that do not clear the bar that the small number requires.

The hardest part of the discipline is not the conviction. It is the patience. A concentrated firm spends most of its time not investing. This is uncomfortable. It feels unproductive. It is the opposite of how the venture industry is structured to feel, and the structural pressure to deploy — from limited partners, from a competitive pipeline, from the natural human desire to do something rather than nothing — is constant. The discipline of concentration is, in practice, the discipline of accepting that pressure and declining to act on it.

What we are buying.

A concentrated portfolio is a bet on the quality of underwriting over the quantity of opportunities. It is, more specifically, a bet that the firm's capacity to think carefully about a small number of things is greater than its capacity to think superficially about a large number of them. This is a bet about ourselves, not about the market. We think it is the right bet for the firm we are trying to build. We acknowledge it may not be the right bet for any other firm, and we do not advocate it as a general principle.

The portfolio is the posture revealed. Ours is what we have chosen to hold.