Anyone can create a custom Plasma chain with a custom consensus mechanism. The creator decides on the details, so it could be created with one node or many. In practice this means there will be various clients that you can choose from or configure. It's up to the creator or interested parties to build the client or pick and configure an existing one. Those running the client maintain the child chain. Validator nodes could be just a few nodes or in the extreme case just one.
This works because the root blockchain Plasma smart contract enforces the state in the child chain, and because of the incentives of bonding and slashing in the Plasma protocol.
Here is how it works based on my understanding of the Plasma white paper. The root chain only penalizes the child chain if there is proof of fraud. The fraud proofs are in the form of state transitions on the child chain. This allows for deposit and withdrawal of funds into the child chain keeping account balances on the child chain, and the balance in the root chain is represented by a smart contract enforcing the child chain state transitions via fraud proofs.
Fraud proofs allow for an interactive protocol for fund withdrawals and exit of the child chain. In the case of a dispute about the new transaction state the disputing party attests to their version of the correct state of the new block in the process of requesting a withdrawal. Then anyone on the network can submit bonded proof to the root contract which will execute the proof. The bonded proof contains information about the previous state, and using the state-transition rules in the root contract actors on the child-chain can prove the new state is invalid. When this happens the child-chain is reverted to the previous block and the cheating block producer's deposit is slashed.