Apply to be an enterprise pilot partner

Reading this paper is free.

AhanaAI Research · November 2026 · AhanaWire

Same pipe, more data — and every job comes back

Changing the speed of the internet, one tiny byte at a time.

10 Mbps carried ~2.09 Gbps of repeated operational text

A 10 Mbps path filled with the 150-byte next copy carries 2,088.933 Mbps of the original text. The link’s rated speed does not change.

Measured on repeated operational text. October 2026 four-file mix, 31,334 bytes, measured 17–21 October 2026. Next copy: 150 bytes, AhanaWire 3, after the far site already holds the data. Not a claim that the internet, or any link, got faster for all traffic.

66.8× more data through the same pipe

The first copy of this mix is 469 bytes instead of 31,334.

Measured on repeated operational text. The measured multiple is 66.8102 to 1. October 2026 four-file mix, 31,334 bytes, measured 17–21 October 2026. First copy: AhanaWire 1, the shipping pair, 469 bytes.

208.9× on the next copy

Once the far site already holds the data, the next copy is 150 bytes.

Measured on repeated operational text. The measured multiple is 208.8933 to 1. Same mix. Next copy: AhanaWire 3, 150 bytes, after the far site already holds the data.

47.7× sooner to finish logs on 10 Mbps

Version 1 finishes the log file that many times sooner on a full 10 Mbps pipe.

Measured on repeated operational text. The measured clock is 47.7366 times sooner. October 2026 in-process version 1 clocks, no added round-trip delay. Logs only.

6.7× sooner on 100 Mbps

The same log clock, on a full 100 Mbps pipe.

Measured on repeated operational text. The measured clock is 6.6793 times sooner. October 2026 in-process version 1 clocks, no added round-trip delay. Logs only.

Bytes on the measured mix

First copy and next copy are labeled separately. The highlighted bars are AhanaWire.

Next copy
150
First copy
469
zstd-3
808
gzip-6
960
lz4-1
1,088

Abstract

A smaller frame does more than lower a bill. The same last-mile link can carry more of the original text, and the far side either rebuilds the original bytes or refuses the frame. On the measured 31,334-byte mix, a 10 Mbps path filled with the 150-byte next copy carries 2,088.933 Mbps of original text and adds 2,078.933 Mbps to that pipe. At 100 Mbps the same math is 20,889.333 Mbps. In process, version 1 finished logs 47.7366 times sooner on a full 10 Mbps pipe, and 6.6793 times sooner at 100 Mbps. On the saved card, 17,101 of 17,101 requests restored the original, and 16 of 16 damaged frames were refused. The measured pair runs on hosts the operator owns.

Key results

1 October 2026 customer paper. Pipeline figures are frame math on the four-file mix of repetitive private text, measured 17–21 October 2026: link speed × original bytes / frame bytes. This is the original text a full pipe of those frames can carry. It is not a change in the link’s rated speed. First-copy sizes are AhanaWire 1, the shipping pair. Next-copy sizes are AhanaWire 3 after the far site already holds the data. The next-copy clock was not re-timed on a 100 Mbps or 1,000 Mbps link; the 100 Mbps next-copy column is that same frame math.

2 October 2026 customer paper. In-process last-mile clocks for version 1, with no added round-trip delay. Speedup is how many times sooner the same job finishes on a full pipe: raw transfer time divided by compress, wire, and expand time. The break-even is the link speed at which that clock gain falls away for that shape.

3 October 2026 customer paper. Home pair, 16 concurrent clients, about 8 seconds, warm next copy, every job restored.

4 October 2026 measurements, saved-card restore and refuse counts, as compiled in the November 2026 numbers fact sheet from the customer paper and the same week’s scorecard. The scope is those runs, not a lifetime service level.

5 October 2026 customer paper. Deployment shape only: both edges, and the stored data, sit on machines the operator runs.

Test setup

This paper is the capacity and integrity companion to Cut the cloud egress bill. It uses the same October 2026 mix and the same first-copy and next-copy labels. Every figure below was measured. It does not describe how the frames were built.

The corpus is four files of repetitive private text:

ShapeRaw size (bytes)First copy (bytes)Next copy (bytes)
Chat, captions, and short turns6947129
Agent handoff2,17210846
Logs and alarms12,8008529
A mixed day, one file15,66820546
The mix (sum)31,334469150

The first copy is the frame the shipping AhanaWire 1 pair sends when the far site has not yet received this file. The next copy is the frame AhanaWire 3 sends once that far site already holds the data and can open the original from what it already has. A site that has not yet received the file still takes the first copy.

Two clocks are reported, and they are not combined:

A separate timed run used a home pair: 16 concurrent clients, about eight seconds, a warm next copy, and a count of every restored job. Resident memory on that pair is reported as measured. The same next-copy wire sizes were also recorded on a MacBook and a Mac Pro.

Restore figures count jobs that came back as the original bytes, and damaged frames that were refused, on the saved card from those October 2026 measurements.

Results

Traffic the tiny stream adds

A full pipeline carries link speed times original bytes, divided by the bytes in the tiny frame. Added traffic is that amount minus the raw pipeline. The table uses next-copy frames, after the far site already holds the data.

Shape, next copyTiny frame (bytes)On a 10 Mbps pipelineTraffic addedOn a 100 Mbps pipeline
Chat29239.310 Mbps229.310 Mbps2,393.103 Mbps
Handoff46472.1739 Mbps462.1739 Mbps4,721.739 Mbps
Logs294,413.793 Mbps4,403.793 Mbps44,137.931 Mbps
Mixed day463,406.087 Mbps3,396.087 Mbps34,060.870 Mbps
The mix1502,088.933 Mbps2,078.933 Mbps20,889.333 Mbps

The 10 Mbps handoff figures are the exact quotient 10 × 2,172 / 46, and that amount minus the 10 Mbps link. Other rows use the three-decimal values published in the October 2026 customer paper.

The first copy of the mix is 469 bytes. A 10 Mbps pipeline of those first-copy frames carries 668.102 Mbps of original text and adds 658.102 Mbps.

The multiple adds once per warmed last mile. It does not compound across sites, and one send does not reach every site. A site that has never accepted the file gets the first copy.

The pair already runs at that tiny stream

Sixteen concurrent clients, about eight seconds, warm next copy, every job restored. Original megabytes per second are the reconstructed text. Wire megabytes per second are what crossed the link.

FileHome jobs/sHome originalHome wirep99
Chat, next copy 29 bytes70,972.2149.255 MB/s2.058 MB/s0.445 ms
Handoff, next copy 46 bytes48,388.01105.099 MB/s2.226 MB/s0.909 ms
Logs, next copy 29 bytes19,367.73247.907 MB/s0.562 MB/s1.277 ms
Mixed, next copy 46 bytes17,312.14271.247 MB/s0.796 MB/s1.548 ms

Home mixed text moved at 2,169.976 Mbit/s while the wire carried 6.368 Mbit/s, so the tiny stream added 2,163.608 Mbit/s to that pipeline. Home logs moved at 1,983.256 Mbit/s onto 4.496 Mbit/s of wire. Resident size on the home pair stayed at 4.9 MB and 3.0 MB. The same next-copy wire sizes — 29, 46, 29, and 46 bytes — were measured on the home pair, a MacBook, and a Mac Pro. MacBook mixed text moved at 155.590 MB/s, with p99 between 0.534 ms and 1.000 ms. Mac Pro mixed text moved at 99.378 MB/s, with p99 between 0.343 ms and 0.712 ms.

Last mile: where the smaller frame also finishes sooner

These clocks are in-process, October 2026, for version 1, which puts the smaller lossless frame on every byte. Raw handoff time at 10 Mbps is 1.7376 ms. With version 1 it is 0.1237 ms.

FileVersion 1 at 10 MbpsVersion 1 at 100 MbpsBreak-even, version 1
Chat7.6831×2.6260×322.336 Mbps
Handoff14.0441×3.7803×442.386 Mbps
Logs47.7366×6.6793×694.285 Mbps
Mixed36.7324×6.4732×697.963 Mbps

The break-even is the link speed at which the pipe is already fast enough that the clock gain for that shape falls away. These figures are not a wide-area ping test.

Nothing lost

The far side rebuilds the original bytes, or it refuses the frame.

The 469-byte first frame of the mix can be kept and opened later as 31,334 bytes. A terabyte of that shape is kept as 14.967 GB and returns as the terabyte. A repeated official log line of 1 MiB or 16 MiB is kept as 127 bytes and opens as the full line. A terabyte of that 16 MiB line is kept as 7.569 MB. A handoff is kept as 108 bytes and opens as 2,172. Logs of 12,800 bytes are kept as 85. Chat of 694 bytes is kept as 71.

The operator runs both edges. The stored data sit on disks at sites they own. A measured pair sits in a few megabytes of memory, on CPU.

Limitations

How to cite

AhanaAI Research. (2026, November). Same pipe, more data — and every job comes back. AhanaAI. https://ahanaai.com/research/same-pipe-more-data

Republishing license, $1

Reading this paper is free. A bot that republishes it needs a $1 license. Terms: Republishing license.

Buy the republishing license, $1

data-buy-url is empty until a Stripe Payment Link is added. Do not invent a checkout URL.

Apply to be an enterprise pilot partner

If you move the same operational text between sites you run, and you need the original bytes back, we will measure one of your files the same way: what a full pipe can carry, how soon the copy finishes, and whether every job restores.

Apply to be an enterprise pilot partner