Most storage is rented: stop paying and your files go away. Arweave sells storage the other way round. You pay once, up front, and the network commits to keeping the data for as long as it runs. That works because of an economic bet about the future cost of disks and a mining design that rewards miners for actually holding the data.
Pay once: the storage endowment
When you upload to Arweave, the fee is calculated to cover storing that data, with many copies, for a very long time (the design targets about 200 years) at today's storage prices. A small part of the fee goes to miners immediately. Most goes into a protocol-controlled endowment.
Miners are paid from block rewards and fees. Only if those ever fall short of the real cost of storage does the protocol start releasing tokens from the endowment.
The key assumption is that storage keeps getting cheaper. The cost of storing a gigabyte for a year has fallen dramatically over decades. Arweave prices in a decline far more conservative than the historical trend, so the endowment should grow in real terms. If storage prices stopped falling for a long time, or the AR token lost most of its value, the endowment would be drawn down faster. That is the risk you accept.
What it costs
Fees are priced in AR, but the target is a stable cost in real storage terms, so the fiat price per gigabyte moves with the AR price and with network parameters. Don't rely on a figure from an article; ask a gateway:
curl https://arweave.net/price/1048576
# returns the fee in winston for 1 MiB (1 AR = 1,000,000,000,000 winston)
As a rule of thumb, Arweave is expensive compared with one month of cloud storage and cheap compared with decades of it. It suits small, valuable data (metadata, documents, images, app bundles) better than bulk video archives.
The blockweave and SPoRA
Arweave's chain is a blockweave. Each block links to the previous one, as in any blockchain, and mining also requires data from randomly chosen earlier parts of the weave.
Succinct Proofs of Random Access (SPoRA) make that concrete. To attempt a block, a miner needs to read random chunks of historical data and include proof it had them. A miner that stores more of the weave locally, and can read it quickly, gets more chances to win. A miner that tries to fetch data from the network on demand is far too slow to compete. So the incentive to mine becomes an incentive to , especially data few others hold.