Understanding brain activity requires large-scale electrophysiology recording with a single-cell resolution. Over the past decades, the number of electrodes allowed for simultaneous recording doubled approximately every 7 years. Large data from high-density electrodes stress every part of the recording system, including data transmission, storage, and processing, and demands high power. Developing an energy-efficient method to compress the data with minimal distortion is necessary.Spikes are the sparse information carrier in the neural network. The proposed method, focused on spikes, consists of two parts: 1) An analog-to-digital converter compares the extracellular signal between different timestamps. It outputs the difference as 2-bit ternary pulse trains (delta modulation), potentially relieving the computational load for the following stages. 2) An event-driven serializer packs pulse trains into packets suitable for wireless transmission. Leveraging the sparsity of spikes and spatial redundancy of high-density electrodes, the proposed method achieves a compression ratio of >11.4X while consuming <1% of energy compared to state-of-the-art appliances.Spike sorting results based on reconstructed signals show that all spike waveform features are well-preserved. The encoded data can decode motor prediction tasks (data from neuroBench) using a neuron-inspired ‘spiking neural network.’ The accuracy is similar to GPU-server-based decoding results while consuming extremely low energy and having low decoding latency.Energy efficiency is crucial for battery life and dissipates less heat to surrounding tissue. Together with the tissue-coupled transdural data telemetry and ultrasound powering method developed in the group, our intracranial BCI device can be free-floating and easily scalable to record multiple brain areas.