The test wasn’t a narrow lab simulation. From start to finish, Celestia operated the full pipeline—encoding new blobs, distributing and storing the resulting pieces, gathering validator signatures, and finally committing the data onchain. That full pipeline is what separates this benchmark from an earlier demonstration of 1 Tb/s that Fibre showed off in a partial test earlier this year. According to Celestia, a prospective customer with unusually demanding throughput needs prompted the company to find out exactly how much capacity Fibre could sustain under real conditions.
During its strongest stretches, the network performed even better than the overall average. The best 60-second window averaged 3.69 Tb/s, while the best 30-second window hit 4.27 Tb/s. Celestia notes these figures only count blob data confirmed onchain — they exclude recovery pieces generated during encoding and the traffic involved in distributing pieces to validators, both of which would push real throughput higher.
Key Architectural Innovations Enable High Throughput
The core design choice behind Fibre’s performance is a simple separation: blob data goes directly to validators, while Celestia’s chain only records a commitment to that data plus the validators’ signatures. That split is what allows the chain to scale, since it never has to process the full weight of the underlying data — only a proof that it exists and was verified.
In practice, a client encodes a blob and sends each validator its own share of pieces. Validators check their pieces, store them, and sign to confirm receipt. Once validators representing two-thirds of voting power have signed, the client submits those signatures to Celestia along with the blob commitment.
Getting there required months of engineering. Since March, Celestia’s team has been cutting repeated verification work, reusing memory, and eliminating unnecessary data copies to let Fibre handle more uploads simultaneously. One of the biggest wins came from encoding itself. Before a blob ships, Fibre splits it into pieces and adds recovery pieces using Reed-Solomon coding, so the original data can be rebuilt even if some pieces are lost in transit. Because this happens for every blob, inefficiencies compound quickly.
Modern ARM chips — including the AWS Graviton machines used in testing — can process multiple values within a single instruction, but the Reed-Solomon coding Fibre relied on was still working through data one value at a time. Using NEON, ARM’s vector instruction set, Celestia’s engineers built vectorized kernels and fine-tuned the workload to take advantage of them. As a result, the median time to encode a 2 GiB blob dropped from 6.8 seconds to 0.9 seconds, a gain of roughly 7.5x.
Once encoding and distribution sped up, the chain itself became the limiting factor. In tests, uploads needed roughly 1 second to encode and 0.5 seconds to distribute pieces, yet confirmation could take up to 14 seconds, since data arrived faster than the chain could handle it. To address this, Celestia started caching and reusing successful signature checks rather than rerunning them, parallelized verification, and streamlined how validators process PayForFibre (PFF) transactions, the onchain messages that log each Fibre upload, which reduced the time to validate a block proposal from scratch—with no cache—from 10.35 seconds to 1.12 seconds.
Storage and Configuration Enhancements Support Scalability
Raising block capacity and switching storage infrastructure closed the remaining gap between encoding speed and chain confirmation. By increasing per-block capacity for Fibre submissions from 200 to 2,000 and shortening block time to 1 second, Celestia lowered the average confirmation wait to 2.2 seconds in the final benchmark.
Storage needed its own fix. Celestia initially picked network-optimized AWS instances for their bandwidth, but their attached EBS disks couldn’t write data as fast as it was arriving. The solution was shifting blob piece storage to AWS S3, the cloud provider’s object storage service. Because S3 performance started to decline at around 3,500 uploads per second, Celestia bundled 16 Fibre pieces per object and distributed writes across hash-based key groups and several buckets to cut down the overall request count. These storage and configuration changes, layered on top of the encoding and chain-processing fixes, are what ultimately cleared the way for the full-scale 3.07 Tb/s run.
Benchmark Differences and Future Scaling Plans
Several conditions in this test won’t carry over directly to Celestia’s mainnet launch. The benchmark used 2 GiB blobs, while mainnet will start with a 128 MiB limit. It also ran with 1-second blocks and a cap of 2,000 Fibre submissions per block, compared with the current limit of 200. The memory and concurrency configurations were calibrated for the specific AWS machines involved in testing, and the test run relied on an experimental performance branch of Fibre’s production pipeline, so several of the optimizations outlined here remain under refinement ahead of a wider rollout.
That distinction matters for anyone trying to gauge what Fibre will actually deliver on day one versus what it’s capable of under tuned conditions. Celestia frames the benchmark as a demonstration of ceiling potential, not a guarantee of out-of-the-box mainnet performance.
Why this matters going forward: Celestia argues that demand for blockspace is about to outgrow what existing blockchain systems can offer, driven by AI agents making payments on users’ behalf and a broader shift of global finance activity onchain. According to the company, at 3 Tb/s Fibre would be capable of serving an AI agent for every person on the planet, with each one submitting a transaction every four seconds. Celestia’s stated goal is to scale Fibre’s blockspace capacity to 3 Tb/s and beyond as application demand grows, rolling out mainnet capacity matched to early usage and expanding from there.
The work was led by Vlad Krinitsyn, with contributions from Rachid Chami, Preston Evans, Hlib Kanunnikov, Alex Kiss, Rene Lubov, and Rootul Patel.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.
