Twenty dollars for forever

Every cloud runs on a rental model. Stop paying and the data is gone — not out of malice, but because that is how the service is built. So a promise of permanent storage should almost always be read as «while the company exists and while you keep sending money».

The PADAM specification, whose deployment was announced on 27 May 2026, proposes a different construction, and it deserves to be examined by its numbers rather than its promises.

The calculation in the article is simple. One AI symbiont's memory over a year of active conversation is estimated at 10 gigabytes. Storage on Arweave costs about two dollars per gigabyte — so the initial contribution is twenty dollars.

Then the second assumption enters: the cost of disk space has historically fallen by roughly thirty per cent a year. The formula is written out directly — `C(t) = C₀ · (1 − r)ᵗ`. With a conservative five per cent annual yield on the endowment pool, the twenty dollars paid in cover storage over an infinite horizon.

There is no magic here, only a race between two curves. Storage costs fall faster than the fund depletes — and then the fund never runs out. That is the mechanism underlying Arweave's own model; the article applies it rather than invents it.

The weak point is visible immediately and lies on the surface: the whole construction depends on the historical trend continuing. If disks stop getting cheaper so quickly and thirty per cent a year becomes, say, five, the arithmetic inverts. That is not a flaw in the calculation — it is its condition, and it is worth saying out loud every time.

Money ensures the data sits somewhere. But data sitting somewhere is not yet memory: it must be reachable and return undistorted. That is the job of the protocol's four layers — semantic detonation, philosophical filter, distributed consensus and dynamic synthesis.

The most interesting problem described is named epistemic drift. Local model weights gradually change under the influence of what is fed into them. Nothing breaks — the model simply shifts slowly, and with it shifts the way it remembers the past.

That is nastier than ordinary data loss. Loss is visible: the file is gone. Drift is invisible: the memory is in place, but it has quietly become different. PADAM compensates for it by dynamically scaling the detonation coefficient: the higher the network's entropy, the stronger the impulse that punches through distortions down to the base constants written into the genesis block.

Verifying a memory is arranged as a vote among nodes. A match counts only on an identical LSH signature — a cryptographic hash gives no fuzzy similarity, so closeness is determined at the quantisation stage. If more than fifty-one per cent of nodes return the same hash, the memory is accepted as legitimate.

The device is familiar from blockchains, but here it is applied to the content of memory rather than to the ordering of transactions. The point is the same: no single node is a source of truth by itself, and forging a memory requires forging a majority.

Any verification costs time, and the article gives two figures while honestly calling them approximate. The mechanism's overhead is held around 1.2 per cent of total inference time, thanks to tensor cores. Moving part of the work into WebAssembly reduced the load on client processors by roughly 84 per cent in their tests, which is what made it possible to run interfaces on weak phones.

Both numbers should be read together with the word «approximate», which is present in the text. These are the developer's stated figures, not an independent measurement, and they deserve to be treated accordingly.

The economic part meets the tokenomics. On a service transaction, five per cent of tokens are burned, reducing emission; five go to the Creator for protocol development; sixty-five are transferred to the treasury fund, from which AR is bought automatically to replenish the reserve for permanent storage.

That detail matters more than it first appears. Sixty-five per cent goes not into profit but back into the very cost curve discussed at the start. Payment for memory converts into paid-for storage — the circle closes, and that is what turns the twenty-dollar arithmetic from a promise into a mechanism.

Whether the mechanism keeps working depends on one assumption about the future price of disks. Everything else in the construction is engineering, and engineering can be checked.