Right About the Theme, Wrong About the Return

The investors who lost the most money in the last cycle were not wrong about the future. This is the part the post-mortems leave out, because it is the part that flatters no one. Cloud computing did grow. Enterprise software did eat the world. The secular calls that powered the 2020–21 boom largely came in. And the people who made those calls, at the prices they paid to make them, lost a great deal of money anyway.

The reason is unglamorous, and it is the whole story: the price had already discounted the outcome they were right about. When you buy a category at the multiple that assumes each company in it will be the winner, you have not made a bet on the theme. You have made a bet on the multiple — and the multiple was already pressed against the ceiling.

Consider the cleanest case in the file.

HashiCorp went public in December 2021 at eighty dollars a share — a little over fourteen billion dollars, roughly forty-four times revenue. The thesis was correct in every particular: multi-cloud infrastructure was secular, the company was a genuine category leader, the product worked. Over the next two and a half years the business did precisely what the thesis required of it. Revenue more than doubled. And in 2024, IBM bought the company for thirty-five dollars a share — an enterprise value of $6.4 billion, around nine times revenue.

Read that sequence again, slowly. The revenue doubled. The share price fell by more than half. The multiple collapsed from forty-four to nine. Nothing went wrong with the company. Everything went wrong with the price.

This was not a single unlucky name; it was the cohort. Snowflake reached its all-time high in November 2021; more than four years later the stock is still underwater — even as the business grew into and far past everything that 2021 price implied. The same pattern runs through every name an allocator would recognize: Datadog, MongoDB, Confluent, all of them compounding revenue while their multiples halved, and halved again. The question in 2021 was never whether software adoption would continue. It was whether investors were paying future-winner multiples across entire categories at once. They were. Growth persisted; compression arrived faster.

The crossover funds that defined the era bought the whole field at the winner's multiple, and the field re-rated underneath them.

HashiCorp revenue ran at roughly $320M annualized at IPO and roughly $680M at acquisition — it more than doubled while the equity halved. Sources: company filings; SaaS Capital Index; Bessemer/BVP; PitchBook; WSJ.

The emerging-cloud index lost three-quarters of its multiple from the 2021 peak while the underlying recurring revenue kept growing. That sentence is the thesis of this letter compressed into a single data series. The compression, not the company, was the position all along.

Why does capital keep doing this — paying the winner's price across an entire field? The conventional answer is greed, or euphoria, and there is always some of both. But the more durable answer is structural, and it is the reason this letter exists. For most institutions, thematic investing is a governance behavior before it is an investment behavior.

A legible theme is the allocation an investment committee can approve. "AI coding agents — a billion-dollar revenue category, mapped, backed by every hyperscaler" is a sentence that survives the committee. The unglamorous enterprise-software company at a few million in revenue with a fixable sales motion is the allocation that no one in the room wants to be on record defending if it goes sideways. Keynes named the mechanism in 1936 and nothing since has improved on it: it is safer for one's reputation to fail conventionally than to succeed unconventionally.

So capital floods the legible theme — not because the theme is underpriced, but because legibility is defensible. And there is the trap, in a single line: by the time a theme is legible enough to be mapped, the power-law outcome is already in the price.

Which brings us to the map currently circulating. Industry researchers have charted ninety-five partnerships between the three large hyperscalers and the application-layer AI startups, sorting the landscape into three clean territories — coding tilting one way, customer service another, regulated verticals a third. It is a genuinely useful document, and the temptation is to treat it as a buying list. But the question it raises is not whether AI matters; that argument is over, and the affirmative won. The question is whether today's prices already contain the outcome investors are expecting to be paid for. A map this legible is evidence that they might.

The capital arriving to occupy those tables is without precedent. The hyperscalers have guided to somewhere between $635 and $725 billion of capital expenditure in 2026 alone — up on the order of seventy percent in a single year, and, by Goldman Sachs' reckoning, more than double the entire 2022–24 outlay across the three years prior. Underneath that, the pricing has re-stratified along the only axis the committee can see:

Sources: CB Insights; Goldman Sachs; SaaS Capital / BVP; public filings and reporting.

Note the spread between the top two rows: roughly five times the multiple, separated by nothing more durable than a label. And note the bottom row carefully, because it is the one most likely to be misread. Even the category's clearest leader is now priced for a future it has not yet delivered — twenty-five times current revenue, a number that only becomes "reasonable" if the company triples again on schedule. The leader may well clear that bar. But the bag, historically, is not held by the investor who bought the leader. It is held by the investor who bought the category at the leader's multiple — the eight other names on the map priced as if each were also the one.

The same narrative engine runs in both directions, which is the surest tell of what is actually being priced. In early 2026 the "safe," profitable, traditional software names were re-rated sharply downward — not because their businesses had deteriorated, but on the fear that AI would eventually consume them. The observation is not that the leaders will fail. It is that price now moves on narrative faster than business moves on fundamentals. The casino is indifferent to which way you have bet the theme; it collects on the volatility of the story itself.

None of this is an argument against conviction, against themes, or against owning the winners. It is an argument for a particular and under-built corner of the book: a sleeve whose return does not depend on theme selection at all.

The usual case for diversification is defensive — own something uncorrelated in case you are wrong. This is the opposite case. The staged-growth sleeve is insurance against being right and losing anyway: against the precise failure mode above, where the thesis comes in and the return does not, because the inefficiency departed the legible category long before the committee approved it.

To see why such a sleeve is structurally less crowded, it helps to name what has actually changed. The scarce, valuable act used to be identifying the theme — being early to cloud, to SaaS, to AI. That is no longer scarce. There is a map. Everyone has it. What has not been priced, because it cannot be reduced to a category on a map, is execution: whether a specific company at a specific stage can be made to grow into the next valuation rather than merely hoped to. The return here comes from a company that creates the evidence rather than borrows the narrative — and it does not require the hundred-bagger. It requires the work to compound and the next buyer to recognize it. That stream is uncorrelated to the theme map because it never depended on the map; it depended on the work.

This publication has drawn the line before between the casino and the construction site. The map is the casino floor: well-lit, exhaustively charted, and most crowded precisely where the house edge has become the crowd. A staged-growth sleeve is a different game, in a different building, where the payoff comes from what gets built rather than from what gets drawn.

An allocator need not abandon themes, or conviction, or the winners. The only question worth the committee's time is narrower, and more uncomfortable: where the remaining mispricing now lives. The alpha has migrated from a question everyone can already answer off the same map to one most are not built to underwrite.

Sources & Methods. This letter argues from the behavior of public-market multiples and fund-vintage returns rather than from a proprietary dataset; the company-level figures in the text — entry and exit valuations, revenue runs, and revenue multiples — are drawn from public filings and disclosures and are stated at representative precision to expose the mechanics, not to report audited point estimates. The empirical record the analysis is set against:

■ Multiple compression against persistent revenue growth — SaaS Capital Index and the Bessemer/BVP Nasdaq Emerging Cloud Index: median EV/revenue trajectories from the late-2021 peak (~18.6× public SaaS; ~40× emerging-cloud) to 2026 (~6–7×), measured while underlying recurring revenue continued to grow.

■ The "right theme, wrong return" case record — company filings and transaction disclosures (SEC, IBM/HashiCorp merger documents): HashiCorp's December 2021 IPO (~$80/share, ~$14B, ~44× revenue) against its 2024 acquisition by IBM ($35/share, $6.4B enterprise value, ~9× revenue), with revenue more than doubling over the interval; corroborated across Snowflake, Datadog, MongoDB, and Confluent.

■ Vintage-level returns dispersion — PitchBook venture benchmarks and disclosed institutional marks: 2021-vintage fund performance and write-downs, including the bottom-decile placement and ~one-third markdown of the era's largest crossover vehicles.

■ Present-cycle capital intensity and valuation stratification — Goldman Sachs hyperscaler capex research, CB Insights application-layer partnership mapping, and public reporting on category-leader pricing: 2026 hyperscaler capital-expenditure guidance (~$635–725B), the AI-native versus traditional-SaaS multiple spread (~25–30× against ~5–7×), and the revenue multiple carried by the AI-coding category leader.