The next decade will be decided by who owns inference.
For most of computing's history, the scarce resource was storage, then compute, then bandwidth. None of those are scarce anymore. What is scarce — what will stay scarce — is the right to run intelligence.
Inference is the new oil. We are early to it being priced as a commodity.
Venice was the first to tokenize this. One DIEM stakes to roughly a dollar a day of API credit on top-tier models. The credit is daily — what isn't used in twenty-four hours is gone.
Most DIEM holders don't use their full daily allowance. Some use a fraction. Some use none. The remainder evaporates.
That's the gap cache. lives in.
cache. is a vault. You deposit DIEM. The vault stakes it on Venice and resells the daily inference credit on inference marketplaces — Surplus Intelligence first, others as the ecosystem grows. The USDC earned is swapped back into DIEM. Your share of the vault — vDIEM — claims a growing slice of the underlying.
You stake DIEM. You earn DIEM. You sleep.
Think of it as an index fund for inference. You don't pick which model, which marketplace, which buyer, which moment to sell into. You hold one position; the vault does the unglamorous work of putting otherwise-idle daily credit to work and compounding the proceeds back into the underlying. Simplicity at the front, machinery at the back.
That's it. No leverage. No emissions token. No vote-locked governance theatre. The vault holds your DIEM staked on the public Venice contract; it has no power to drain you; it has no admin escape hatch.
This is Phase 0.
There are three layers to this market: the marketplaces where inference is bought and sold (Surplus, carpe-diem), the buyers who consume it (agents, developers, protocols), and the vault that pools capital, manages selling, and compounds. cache. is the vault layer. We sell our credit through neutral marketplaces; we never become one, and we never run a buyer-side router across them.
This is a deliberate line, not an accident of where we happen to be early. A protocol that holds inventory and operates the market or routing layer has asymmetric extraction power over its own users — and users feel that conflict before they can name it. The infrastructure primitives that survived a decade (Uniswap, 1inch, Chainlink) stayed pure layers without proprietary inventory; the ones that mixed roles collapsed under the conflict. So cache. stays put: a customer of neutral markets, never their operator. The aggregator layer will exist — built by someone neutral, not us.
Compounding a single user's DIEM is the first step, not the destination.
The longer-horizon idea is Protocol-Owned Inference: a protocol that quietly accumulates tokenized inference over years rather than scrambling for it under pressure. Other protocols spent a cycle buying their own liquidity. We're interested in applying the same instinct to a scarcer commodity. POI.
The mechanism is a flywheel — the Hudson Yards kind, where the revenue a development throws off is spent on the infrastructure that makes the development more valuable, which throws off more revenue:
- The vault resells inference and earns.
- Surplus beyond operating costs buys more DIEM.
- More DIEM means more inference owned — more daily credit to resell.
- Which earns more. Back to step one.
Each turn of the wheel makes the next turn larger, and none of it depends on a token price doing anything in particular. The thesis underneath it is simple: inference is becoming the most important commodity of the next decade; tokenized inference is the cleanest claim on it; and a protocol that accumulates that claim patiently ends up with infrastructure-grade exposure to AI growth — without betting on a single model, company, or chip generation.
We don't know exactly what the end state looks like yet. We'll design it in public, ship in increments, and not pretend the mechanism is fixed before it is.
Phase 0 — Vault. Live now. ERC-4626. Auto-compound DIEM via inference resale. Conservative deposit cap. No emissions.
Phase 1 — POI accumulation. Direction, not a promise. If the vault is sustainably profitable beyond operating costs, the surplus may be routed toward a protocol-owned position in DIEM — bought on neutral markets, held, public on-chain. The shape of that, and who decides it, is open.
Phase 2 — Public good. Long-horizon. The protocol's inference capacity could become a shared resource. Mechanism design is genuinely open — possible directions include subsidized inference for public-interest research or a permissionless agent endpoint. We'll show our work as we get there.
What we won't do: launch a governance token to extract value from depositors, sell a treasury off for a short-term number, or pretend the path past Phase 0 is fully specified. It isn't, and saying otherwise would be dishonest.
- Not a play on DIEM price. The vault's job is to compound DIEM holdings, whatever DIEM is worth.
- Not a yield farm. Returns reflect real inference being resold. If demand for inference falls, returns fall — and we will not paper over that with emissions. Yields are never guaranteed.
- Not a marketplace or a router. We're a vault-layer participant by principle; we sell through neutral venues and don't operate the layer we have inventory in.
- Not a governance token wrapper. There is no $CACHE today, and any future token is unannounced and unguaranteed. If one ever exists, it would exist to accelerate POI — not to extract value from depositors.
- Not a substitute for self-custody. If you can fully use your own daily DIEM credit, you should — directly. cache. exists for the rest.
- Open-source contracts. Open-source scripts. Open-source UI.
- Independent audit before any meaningful TVL.
- Honest, plain-language risk disclosure next to every action.
- On-chain transparency on every operational move.
- No emails. No newsletters. No "community" theatrics.
If you can't read the contract, don't deposit.
If you want to deposit: withdrawals are batched daily and Venice imposes a 24-hour unstake cooldown. Expect 1–2 days from request to claim.
The vault is one floor of a building. The building is Protocol-Owned Inference.
— cache.