The Ledger
Celestia’s Voyager Averages 21.2M TPS in Five-Hour Test
Celestia Labs says its new Voyager payment architecture averaged 21.2 million transactions per second in a five-hour test across 100 million accounts. The single-node benchmark leaves state roots, machine-to-machine replication and high-volume status queries for future work.

Voyager averaged 21.2 million transactions per second over five hours, according to a Celestia Labs benchmark published Oct. 29. The payment-focused blockchain architecture processed random transfers among 100 million accounts, peaked at 21.6 million TPS and ran execution on a single CPU core. The result measures one node—not a complete payment network operating at that rate.
What Celestia’s five-hour test measured
The benchmark covered about 400 billion transfers. Celestia says memory stayed near 48 gigabytes throughout the run, while the execution core operated at 96% to 100% utilisation. The journal and the Fibre publisher each moved about 4.8 gigabytes of data per second, including full transactions and signatures.
Those figures describe sustained execution under a specific workload: random balance transfers between a very large set of accounts. Voyager builds payment logic into the system instead of running it on a general-purpose virtual machine, and stores balances in a flat table rather than a Merkle tree. That design avoids recomputing a state root on every block, but also means the test does not establish performance for general-purpose smart contracts.
Celestia’s authors describe the test’s limits directly:
“The benchmark measures a single Voyager node. It doesn’t yet compute state roots, replicate data to other machines, or serve real-time transaction status queries at this volume, though it does produce receipts for every payment.”
Nick White and Achilleas Sarantaris, Celestia
Separate threads keep payments moving
Voyager divides the work among independent processing threads. An ingress layer checks signatures and batches payments before sending them to the core node. A ring buffer then feeds batches, in order, to parallel stages: the executor updates balances, the journal records activity, and the Fibre publisher and Stargaze storage service handle older batches.
The system also applies backpressure: when its queue fills, it holds back incoming transactions rather than letting work accumulate without limit. Celestia says the executor fetches payment data ahead of time, reducing waits on other threads. The benchmark therefore tests a specialised payment pipeline, not a chain asked to run arbitrary programs and compute a cryptographic state commitment at the same speed.
Agent payments are the proposed use
Celestia is pitching Voyager for software agents that pay per request for content, data or services. The company argues that conventional subscriptions may not suit an agent that needs a resource once, while card payments can be too slow or costly for tiny charges. Those are Celestia’s reasons for pursuing usage-based payments; the benchmark does not demonstrate customer demand or commercial use.
Voyager is designed to publish every transaction to Celestia Fibre so others can replay the data and check how payments were applied. Fibre went live on Celestia’s mainnet on Oct. 19, according to the announcement. Celestia says Voyager is its first execution layer built to use Fibre’s throughput.
Replication and verification remain ahead
The announcement says Voyager already saves the information needed to recover after a process crash and produces a receipt for each payment. Replication to other machines, state roots and real-time status queries are still pending. Those functions matter to users and services that need to verify balances, check whether a payment completed and recover from machine failures.
Celestia says it is taking Voyager toward production and plans to publish more technical details in the coming weeks. The next test will need to show how the system performs with replicated state and verification enabled—not only how quickly one node can execute transfers.
Sources
- Original source · Blog post Celestia Blog blog.celestia.org