On July 27 Zama published a benchmark that landed in every infrastructure feed within a day: 1,040 confidential transfers per second. The company's founder, Rand Hindi, framed it as a 5,000 times gain over four years, from 0.2 per second at a 2022 demo, and put the cost of a single encrypted token movement at $0.000004. Both figures are accurate. Neither one tells you what happens when you send an encrypted USDT transfer on Ethereum tonight.
That gap is worth walking through slowly, because fully homomorphic encryption is the rare corner of crypto where the marketing number and the user number are separated by several orders of magnitude, and where the company itself has been fairly open about why.
Start with what an encrypted token movement involves
A normal ERC-20 transfer publishes the amount. Anyone can read it. A confidential transfer under the ERC-7984 standard keeps the balance and the amount encrypted the whole way through, and the network computes on the ciphertext without ever decrypting it. That is what fully homomorphic encryption buys you, and it is why Zama describes ERC-7984 as the encrypted counterpart of ERC-20.
The problem is arithmetic. Adding two encrypted numbers is not one operation, it is thousands, and each one drags along a step called bootstrapping that resets the noise built up inside the ciphertext. Zama's own explainer from February put FHE work at roughly 100 times slower than the plaintext equivalent. So a confidential transfer is not a transfer with a privacy flag set. It is a transfer plus a zero-knowledge proof that the sender knows the encrypted amount, plus two or three separate decryption operations handled off-chain by a threshold key network.
Every one of those pieces has a price. That matters later.
The 1,040 number describes one machine, and Zama says so in the post
The benchmark ran on a single protocol node holding eight NVIDIA H100 GPUs. Version 1.7 reached 1,040 confidential transfers per second against 400 in version 1.6, a 2.6 times jump that came mostly from a rewritten mixed-radix FFT built for GPU registers. Latency barely moved, 51.5 milliseconds against 54.7. A single H100 handled 47,900 bootstraps per second, and the full eight-card node reached 383,200.
Then the post says the part almost nobody quoted. The 1,040 figure is confidential compute throughput on one GPU node, not end to end settlement on a live chain. Zama's engineers go further and argue FHE is no longer the limiting factor at all. Block times and consensus are.
Read that alongside the headlines and the shape of the story changes. A lab measured how fast its encryption engine can chew through ciphertext. It did not measure how many encrypted transfers Ethereum will actually settle in a second, and the company never claimed it did.
Meanwhile the live protocol has been running on CPUs at about 20 a second
Zama's mainnet went live on Ethereum on December 30, 2025, with the first confidential USDT transfer. It has been processing real encrypted volume ever since. By February the protocol had passed $121 million in what Zama calls Total Value Shielded, the metric it introduced alongside its token, and the sealed-bid auction that distributed that token cleared at roughly $0.05 after raising somewhere between $118 million and $121 million.
The throughput on that live system is in Zama's litepaper, and it is not 1,040. The document describes a climb from 0.2 transactions per second to over 20 per second on CPU, with the GPU migration slated for the end of 2026 to deliver 500 to 1,000 per chain, and purpose-built FHE ASICs after that promising six figures per server. So the benchmark published in July is roughly the top of the range the litepaper penciled in for a deployment that has not shipped yet.
One secondary detail got mangled in the retelling. Some coverage of Hindi's remarks described a mainnet activation planned for the end of the year, which reads as though the protocol is not live. What is planned for the end of the year is the GPU path. The chain has been settling encrypted transfers for seven months.
Twenty per second sounds thin until you check what Zama argues it covers, which is enough headroom to encrypt all of Ethereum at current usage. The 500 to 1,000 range is the number aimed at Ethereum's fragmented L2 economy and most of what Solana does.
Four cost figures, four different things being measured
This is where the reporting got loosest. Three separate per-transfer prices for a confidential transfer have circulated in the past fortnight, and they differ by a factor of two hundred thousand at the extremes because they measure different layers.
| Figure | Where it comes from | What it actually covers |
|---|---|---|
| $0.000004 | Hindi, quoted July 28 | FHE compute for one ERC-7984 transfer |
| Under $0.00005 | Zama benchmark post, July 27 | Compute summed across all protocol nodes |
| $0.008 | Zama litepaper | What a user pays with the 99% subscription discount |
| Up to $0.80 | Zama litepaper | What a user pays at pay-as-you-go rates |
The litepaper breaks the user-facing figure down further: $0.005 to $0.50 for verifying the zero-knowledge proof, plus $0.003 to $0.30 for decrypting two or three values. Encryption is the cheap part of a confidential transfer. Proving and decrypting are what you pay for, and no GPU cluster makes a zero-knowledge verification on Ethereum free.
How to read the next throughput claim
FHE is going to keep producing numbers like this, because the field is improving genuinely fast and every improvement is publishable. A few questions separate a lab result from a live capability.
- Is the figure compute throughput or settled transactions? Those are different measurements and the gap between them is where most of the confusion lives.
- One node or the whole network? A per-node number cannot be added up across nodes in a threshold system where every node processes the same work.
- Which hardware path, and is it deployed? A GPU benchmark on a protocol running on CPUs describes next year.
- Does the quoted cost include proof verification and decryption, or only the encryption step?
- Is the figure self-reported? Zama's benchmark is, and no independent methodology has been published alongside it.
None of that makes the work less real. Going from 0.2 to over 20 in production, with $121 million shielded on a public chain, is a harder result than most privacy projects have managed, and it sits in a different category from the sealed anonymity sets of Zcash's shielded pool. Cryptographic overhead is also the same wall the post-quantum research at the Ethereum Foundation keeps running into, and it does not come down through announcements.
Frequently asked questions
Is Zama's mainnet live right now?
Yes. It launched on Ethereum on December 30, 2025, and executed the first encrypted USDT transfer that day. The thing scheduled for late 2026 is the GPU upgrade to the node software.
So what does a confidential transfer cost me today?
Between roughly $0.008 and $0.80, according to Zama's litepaper, depending on whether you are on a subscription or paying per use. The sub-cent figures in the headlines are compute cost, not the fee.
Will the 1,040 per second number apply to Ethereum?
Not directly. Zama's post states that chain consensus, not encryption, now sets the ceiling on settlement, so mainnet capacity will be governed by block space rather than by how fast the H100s run.
Which chains support ERC-7984 confidential transfers?
Ethereum mainnet today, with EVM chains and a Solana integration on the roadmap. Zama's framing is that no chain needs a protocol upgrade to support the standard, since the heavy work happens in coprocessors off-chain.
The next checkpoint is concrete enough to diary. Zama has committed to the GPU path landing on production nodes by the end of 2026, so between now and then there are two numbers worth tracking: whether Total Value Shielded keeps climbing from $121 million, and whether the per-transfer fee a user actually pays moves off the $0.008 to $0.80 band once the hardware changes. If throughput rises tenfold and the fee does not budge, the bottleneck was never the encryption.
