It's 14 tps.

It's scalable because it doesn't have ever-mounting costs. If you took Bitcoin and gave it 100 MB blocks, that'd be up to 5.2 TB of data added per year to the history that some subset of users would have to completely store and all full nodes would have to completely download. Altcoins get away with using large max block sizes and saying that they're "scalable" because their blocks are usually empty, so they don't actually have to face the costs.

Grin has a 15 MB bitcoin-equivalent max block size, and while this increases the bandwidth which nodes require (from an absolute minimum of 53 kbps on Bitcoin to 200 kbps on grin - note however that the network cannot function if most nodes have only this minimum), it largely does not increase the history which nodes must keep track of, since mimblewimble uniquely allows the entire network to forget old data.

Maybe in the future there will be a willingness to increase the minimum required bandwidth on grin substantially as global Internet speeds increase. At first glance 200 kbps seems conservative for grin, and since grin does not aim to be a store of value, it can afford to be somewhat reckless anyway. But as long as your network relies on every network participant taking a look at every transaction (ie. on-chain transactions in all decentralized block chain systems), you have no hope of competing with the throughput of eg. Visa, since you have to support ordinary people with less-than-optimized Internet connections running nodes. Global broadcast just doesn't scale well, which is why we're not all tapped into the same Ethernet network. LN can get Visa-level transaction volumes because it keeps most transactions local to the participants.