Celestia indicates that its Fibre technology achieved an average throughput of 3.07 Tb per second in a complete end-to-end benchmark test covering 120 validators. This result takes the discussion about the throughput of Celestia Fibre from a theoretical level to a stage that is closer to a viable blueprint. The company disclosed this figure in a blog post published on October 1, 2026, describing it as proof that the network is capable of handling far more data than most current blockchains.

Key Points

• Within a 143-second load window, across 120 validators, the average throughput was 3.07 Tb per second.
• This data rate is roughly equivalent to nearly 2 billion transactions per second.
• After using the ARM NEON instruction to vectorize the Reed-Solomon code, the encoding speed increased by 7.5 times.
• In testing, the block size submitted by Fibre has been increased from 200 to 2,000, and the block time has been reduced to 1 second.
• Celestia indicates that the benchmark test used settings that included 2 GiB and blob. These configurations will not be available when the mainnet goes live.

Celestia Fibre Achieves record throughput in benchmark tests

The most critical figure is 3.07 Tb. This value remains stable within a 143-second load window, with 120 validators each operating on independent AWS machines. Celestia indicates that this is sufficient to support a traffic volume of nearly 2 billion transactions per second. The company compares this to processing the entire transaction volume for the year 2025 in just over two minutes.

This test was not a narrow-sense laboratory simulation. From start to finish, Celestia carried out the entire process: encoding the new blob, distributing and storing the generated data fragments, collecting verifiers' signatures, and finally submitting the data to the blockchain. Celestia states that it is precisely this complete process that distinguishes this benchmark test from some of the tests demonstrated earlier this year involving Tb. According to Celestia, a potential customer with exceptionally high throughput requirements prompted the company to determine how much capacity Fibre can actually handle under real conditions.

During the periods of peak performance, the network’s efficiency even exceeded the overall average. The average throughput for the best 60-second window was 3.69 Tb /s, and for the best 30-second window, it reached 4.27 Tb /s. Celestia notes that these figures only account for the blob data confirmed on the chain, excluding the recovery fragments generated during the encoding process, as well as the traffic involved in distributing these fragments to verifiers; if these components were included, the actual throughput would be even higher.

Key architectural innovations support high throughput.

The core design choice behind the performance of Fibre is quite simple: blob The data is sent directly to the verifiers, while on the chain, only commitments to this data and the verifiers' signatures are recorded. It is this separation that allows the chain to scale, as it does not need to handle the entire burden of the underlying data; it only needs to process one piece of evidence that proves the data exists and has been verified.

Achieving this requires several months of engineering work. Since March, the Celestia team has been working to reduce duplicate verifications, reuse memory, and eliminate unnecessary data copies in order to allow Fibre to process more uploads simultaneously. One of the biggest improvements came from the coding itself. Before sending a blob, it is broken down into multiple segments, and Reed-Solomon coding is used to add the segments back together, so that even if some segments are lost during transmission, the original data can be reconstructed. Since this process is performed on each blob, efficiency losses can accumulate quickly.

Modern ARM chips, including the AWS Graviton machines used in testing, are capable of processing multiple values in a single instruction. However, the Reed-Solomon coding that Fibre relies on still processes data individually for each value. With the help of ARM's vector instruction set NEON, Celestia engineers built vectorized cores and optimized the workload to take advantage of these capabilities. As a result, the median time to encode a 2 GiB blob has been reduced from 6.8 seconds to 0.9 seconds, an improvement of about 7.5 times.

With the improvement in encoding and distribution speeds, the blockchain itself has become a bottleneck. During testing, encoding took approximately 1 second and distribution took 0.5 seconds, but the confirmation process could last up to 14 seconds, as the data arrived faster than the blockchain could process it. To address this issue, Celestia began to cache and reuse successful signature checks instead of executing them repeatedly; it also parallelized the verification process and simplified the way verifiers handled PayForFibre ( PFF ) transactions. PFF is used to record each message uploaded to the blockchain via Fibre. These changes reduced the time required to verify a block proposal from 10.35 seconds (in the absence of caching) to 1.12 seconds.

Storage and configuration optimization support scalability.

Increasing the block capacity and switching to a new storage infrastructure has bridged the remaining gap between encoding speed and on-chain confirmation times. Celestia increased the number of Fibre transactions that each block can accommodate from 200 to 2,000, and reduced the block time to 1 second, ultimately lowering the average confirmation wait time to 2.2 seconds.

Storage also needs to be repaired separately. Celestia Initially, network-optimized AWS instances were chosen for bandwidth reasons, but the EBS disks mounted to them could not write data at the same speed as it was being received. The solution was to migrate the blob fragment storage to AWS S3, which is the object storage service provided by the cloud service provider. Since the performance of S3 began to decline at around 3,500 uploads per second, Celestia packaged 16 Fibre fragments into one object and distributed the writing across multiple buckets using hash-based keys to reduce the total number of requests. These changes in storage and configuration, along with repairs to coding and chain processing, ultimately led to this successful 3.07 Tb /s run.

Benchmark Test Differences and Future Scaling Plans

Several conditions from this test will not be directly carried over to the launch of the Celestia mainnet. The benchmark test used 2 GiB blob, while the mainnet will initially adopt a limit of 128 MiB. The test also utilized a block interval of 1 second and a maximum of 2,000 Fibre submissions per block, whereas the current limit is 200. The memory and concurrency configurations were calibrated for the specific AWS machines used in the test, and this run relied on an experimental performance branch of the Fibre production process. Therefore, several optimizations mentioned in this document are still being fine-tuned before broader deployment.

This difference is important for anyone trying to determine what Fibre can actually provide on its first day, as well as what level it can achieve under optimized conditions. Celestia positions this benchmark test as a demonstration of its upper limit capabilities, rather than a guaranteed performance for the mainnet out of the box.

Celestia believes that as AI agents make payments on behalf of users and global financial activities are increasingly integrated onto blockchain networks, the demand for block space will soon exceed what existing blockchain systems can provide. The company states that at a level of 3 Tb /s, there is enough Fibre to provide a AI agent for everyone on Earth, allowing each agent to submit a transaction every 4 seconds. Celestia's goal is to expand the block space capacity of Fibre to 3 Tb /s or more as application demands grow, and to gradually introduce mainnet capacity based on early usage needs, with further expansions to follow.

This work is led by Vlad Krinitsyn, and participated in by Rachid Chami, Preston Evans, Hlib Kanunnikov, Alex Kiss, Rene Lubov, and Rootul Patel.

This article was generated with the assistance of artificial intelligence and has been reviewed by an editorial team.