# Market Sizing

**Compute for Equity — investor memo**
*Status: strategy memo. Figures are illustrative estimates triangulated from public sources;
they are directional, not audited. Prepared for the data room, June 2026.*

---

## 1. The market we are actually in

It is tempting to size "the GPU market" or "the AI market" and wave at a trillion-dollar number.
We will not do that, because it would not be honest about what we monetize. Compute for Equity is a
**clearing layer.** We do not sell GPUs and we do not invest a fund. We take a thin fee on the
*flow* of value as compute and energy are exchanged for equity. So the right way to size us is to
size the **flow we can clear**, then apply a realistic take rate. We build the number from the
bottom up and sanity-check it from the top down.

The three assets we make interoperable are each large and growing, but the relevant market is their
*intersection*: the subset of compute spend that an AI company would rather settle in equity than in
cash, brokered through a regulated venue. That intersection is small today because the instrument to
clear it barely exists — which is precisely the opportunity.

## 2. Top-down anchors (the context)

A few public reference points frame the opportunity:

- **AI infrastructure spend** is on a multi-hundred-billion-dollar annual trajectory, with leading
  estimates placing the cumulative AI infrastructure financing gap in the hundreds of billions —
  the figure we surface on the site, **~US$490B**, sits in that range. Whatever the exact number,
  the direction is not in dispute: compute is the dominant cost of building AI, and it is being
  financed with increasingly exotic structures.
- **The GPU rental / neocloud market** (CoreWeave, Lambda, Together, and the hyperscalers' rental
  lines) is already tens of billions of dollars annually and projected to grow toward the hundreds
  of billions by the early 2030s.
- **GPU-backed debt** has gone from zero to a recognized asset class in under three years: CoreWeave
  alone has raised facilities of \$2.3B, \$7.5B and \$8.5B collateralized by GPUs and their customer
  contracts — the last achieving an investment-grade rating. The financialization of compute is
  already happening; it is simply happening as *debt between two parties*, not as a *cleared,
  equity-linked, multi-sided market*.
- **Venture-as-compute** is live too: a16z's Oxygen program rents GPUs to portfolio companies below
  market in exchange for equity, with a stash reported in the tens of thousands of chips. The
  behaviour we intermediate — compute traded for equity — is already happening bilaterally inside
  single funds.

The conclusion from the top-down view: every adjacent structure (GPU debt, VC-as-compute, GPU
rental) validates demand for exchanging compute and capital — and every one of them is *bilateral*.
None is a neutral, regulated, multi-sided clearing venue. That gap is the white space.

## 3. Bottom-up TAM / SAM / SOM

We size the cleared flow directly.

**TAM — total compute-for-equity flow that could exist.**
Take global AI training-and-inference compute spend by venture-funded AI companies. A meaningful
share of that spend is, in principle, "equity-fundable" — i.e., the company would prefer to pay in
equity rather than burn cash, particularly for large, lumpy training runs. If global
venture-AI compute spend is on the order of **US$80–120B/yr** and even **15–25%** of it is
equity-fundable, the addressable *flow* is roughly **US$12–30B/yr** of compute that could be
settled as equity. That is the TAM for the value we clear — not the GPU market, the *convertible
slice* of it. As AI compute spend grows toward the hundreds of billions, this slice grows with it.

**SAM — the slice we can serve given regulation and geography.**
We serve professional clients, through an ADGM-regulated venue, anchored on energy-rich Gulf supply.
Realistically the early serviceable market is the flow that (a) involves a participant willing to
transact through a regulated venue, and (b) touches the supply we can originate. We estimate the
serviceable flow at roughly **US$2–5B/yr** within the first several years — the regulated,
relationship-reachable subset of the convertible slice.

**SOM — what we actually clear, near term.**
Phase 1 is ~10 hand-brokered deals. At an illustrative **\$1–5M per Compute-SAFE**, that is
**\$10–50M of cleared flow** in the first phase — tiny as a market share, decisive as a proof. As
Phase 2 opens the platform to a vetted cohort, we model cleared flow growing toward **\$200–500M/yr**,
i.e. roughly **5–15%** of the SAM. SOM is deliberately conservative because liquidity is hand-built,
not assumed.

## 4. Revenue model on the flow

We are capital-light by design; revenue scales with cleared volume, not with our own balance sheet.
The fee stack, illustratively:

- **Origination fee:** ~1–2% of the Compute-SAFE principal at structuring.
- **Clearing & settlement fee:** ~0.5–1% on settlement.
- **Custody fee:** a thin annual basis-point charge on assets under custody.
- **Secondary-market take rate (Phase 3):** ~0.25–0.5% per transaction once a secondary book exists.
- **Treasury float:** modest yield on segregated balances under conservative policy.

Blended, this lands around a **2–4% effective take rate on cleared primary flow**, plus recurring
custody and secondary revenue that compounds as the book grows. Apply that to the SOM:

- **Phase 1** (~\$10–50M cleared): roughly **\$0.3–1.5M** of fee revenue — enough to validate unit
  economics, not to be a business yet.
- **Phase 2** (~\$200–500M cleared): roughly **\$6–18M/yr** — a real, capital-light revenue line.
- **Phase 3** (SAM penetration deepening + secondary): the model scales toward
  **\$30–60M/yr** as cleared flow approaches the upper SAM and secondary volume turns over the book
  multiple times.

These are illustrative and tied to the financial model in the data room (`model.json`), which
carries the Bear/Base/Bull build. The point is structural, not the decimal: **a 2–4% take on a
multi-billion-dollar cleared flow is a venture-scale business, achieved without us ever owning a GPU
or raising a fund.**

## 5. Why the take rate is defensible

Thin take rates are only defensible when the venue provides something the parties cannot easily
replicate bilaterally. We provide four: a **neutral price** (we own no compute, so our Oracle quote
is trustable), a **standard instrument** (the Compute-SAFE, so deals are repeatable and tradable),
**regulated custody** (so institutional and sovereign capital can participate at all), and
**liquidity** (a secondary market no single counterparty pair can offer). Each justifies a slice of
the take. Strip any one of them and you are back to a private bilateral deal — which is exactly the
status quo we improve on.

## 6. Sensitivities and what would change the number

- **AI compute spend trajectory.** If AI infrastructure spend grows faster than consensus (plausible
  given current capex), every tier scales up. If it stalls, the convertible slice shrinks but does
  not vanish — dilution-sensitivity is structural to startups in any market.
- **Equity-fundable share.** Our 15–25% assumption is the swing variable. The a16z Oxygen precedent
  suggests real appetite; wider adoption pushes the share up.
- **Regulatory timeline.** A faster path to a full FSP pulls Phase 2 revenue forward; a slower one
  pushes it out. The RegLab on-ramp caps the downside by letting Phase 1 revenue start regardless.
- **Take-rate compression.** As the asset class matures, fees compress — but custody and secondary
  revenue (recurring, book-driven) grow to offset primary compression. The mix shifts from
  transactional to recurring as we scale, which is the healthy direction.

## 7. Comparable venues and why clearing layers earn premium multiples

The right comparables for valuation are not GPU clouds or VC funds — they are **clearing houses,
exchanges, and marketplace infrastructure.** Public market operators (ICE, CME, Nasdaq) and
financial-infrastructure businesses trade at premium multiples precisely because they are
capital-light, take-rate businesses with network effects and regulatory moats — the exact shape we
are building. A GPU cloud is a capital-intensive, depreciating-asset business that trades on
revenue multiples weighed down by capex; a fund earns carry but does not compound as
infrastructure. A *venue* compounds: each marginal deal costs us almost nothing to clear, and each
adds to the liquidity that attracts the next. That is why a clearing layer with a 2–4% take on a
growing multi-billion-dollar flow is worth more, per dollar of revenue, than the GPU business
underneath it. We monetize the *flow*, not the *iron* — and the flow is where the durable margin is.

A useful sanity check: GPU-backed debt went from \$0 to tens of billions in under three years as a
*bilateral* market. A *cleared, multi-sided, equity-linked* market for the same underlying — with a
secondary — is a structurally larger and stickier prize, because it serves more participants and
trades more than once. The debt market's velocity is the proof of appetite; the clearing layer
captures more of it.

## 8. The honest bottom line

We are not claiming a trillion-dollar TAM. We are claiming something more underwritable: a
**multi-billion-dollar annual flow of compute that AI companies would rather settle in equity**, a
**2–4% take** on clearing it, and a **regulated, supply-anchored position** that lets us capture a
growing share of it without owning the underlying assets. The adjacent markets — GPU debt,
VC-as-compute, GPU rental — already prove the behaviour at scale. We are building the neutral venue
that turns that behaviour into a cleared market. Start at \$10–50M of flow hand-built in Phase 1,
compound toward hundreds of millions as the platform opens, and the take rate does the rest.

---
*Human TODOs before external circulation: reconcile every figure here against the latest
`model.json`; cite specific third-party market reports for the top-down anchors; have finance sign
off on the take-rate assumptions.*