500 confidential transactions per second, measured end-to-end

Measured end-to-end on Arbitrum over 15,000 confidential transfers (July 2026): 500 cTPS sustained, 750 cTPS peak, at $0.14 per million transfers.

This post explains what those numbers mean, how we measured them, and why we insist on the end-to-end qualifier.

What "end-to-end" means, and why it matters

Throughput claims in confidential computing are rarely comparable, because vendors measure different things. Some count how fast a cryptographic kernel executes on a benchmark rig, with no chain involved. Some count compute throughput on specialized hardware and leave settlement out entirely. An end-to-end figure counts something stricter: a confidential transfer initiated by a client, processed by the privacy network, and settled in a block on a live public chain. It includes every real cost: encryption, MPC computation, consensus, settlement.

That is the number an engineer building a payments product actually needs, and as far as we know, no other confidential-transaction system has published one measured on a live chain.

The setup

In July 2026 we ran 15,000 confidential ERC-20 transfers through Bubble on Arbitrum. The MPC nodes ran on ordinary cloud CPUs: no GPUs, no accelerators, no exotic instances. Transfers were dispatched by parallel client accounts simulating end users, and every transfer settled on the live chain. Methodology and raw results are available on request.

The results

• 500 cTPS sustained across the full run
• 750 cTPS peak
• $0.14 per million transfers in compute cost, which is why we round it to "essentially free"

Separately from the end-to-end run, we profiled the compute layer alone: a single commodity CPU core sustains an analytical ceiling of 3,400 cTPS. Parallelized across cores, that projects past 30,000 cTPS, meaning the cryptography is not the bottleneck; the chain is.

For scale

The fastest published figure we know of for FHE-based confidential transfers is 1,040 tx/s, achieved on a node of eight NVIDIA H100 GPUs, and described by its authors as compute throughput rather than settlement on a live chain. The published production figure for CPU-based FHE is roughly 20 tx/s. A GPU node of that class costs on the order of 10,000 dollars a month before the first transfer; our run used hardware you can rent for a fraction of that, and the per-transfer compute cost rounds to zero.

The comparison is not a victory lap over FHE, which is a serious research direction. It is a statement about what garbled-circuit MPC makes possible today, on hardware institutions already run, at costs that disappear into the noise of any payments business.

The trajectory

Our first public test, in November 2024, sustained 53 cTPS on low-end AWS instances, and we set a target of 1,000 cTPS. Since then the stack went to production as the privacy layer of COTI mainnet, where it has processed more than 125,000,000 confidential transactions since March 2025. The July 2026 measurement puts sustained end-to-end throughput at 500 cTPS with the compute ceiling far above it. The 2024 numbers are preserved here; the current ones live on the technology page and stay updated.

If your product needs confidential transfers at production volume, talk to our team. We will share the methodology and the raw data.