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.
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
- A full pipe carries link speed times original bytes, divided by the bytes in the tiny frame. Added traffic is that amount minus the raw pipeline.1
- On the next copy of the measured mix (150 bytes, after the far site already holds the data), a 10 Mbps pipeline carries 2,088.933 Mbps of original text and adds 2,078.933 Mbps. On a 100 Mbps pipeline it carries 20,889.333 Mbps.1
- On the first copy of the same mix (469 bytes), a 10 Mbps pipeline carries 668.102 Mbps of original text and adds 658.102 Mbps.1
- In process, version 1 at 10 Mbps finishes chat 7.6831 times sooner, a handoff 14.0441 times sooner, logs 47.7366 times sooner, and a mixed day 36.7324 times sooner. At 100 Mbps those clocks are 2.6260, 3.7803, 6.6793, and 6.4732. Version 1 break-evens are 322.336, 442.386, 694.285, and 697.963 Mbps.2
- A measured home pair, 16 clients, about eight seconds, warm next copy, moved the mixed day at 271.247 MB/s of original text while the wire carried 0.796 MB/s — 2,169.976 Mbit/s of text on a 6.368 Mbit/s stream. Every job in that run restored. Resident size stayed at 4.9 MB and 3.0 MB, on CPU.3
- On the saved card, 17,101 of 17,101 requests restored the original bytes, and 16 of 16 damaged frames were refused. Every measured product arm restored the original.4
- The operator runs both edges. The store sits on disks at sites they own.5
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:
| Shape | Raw size (bytes) | First copy (bytes) | Next copy (bytes) |
| Chat, captions, and short turns | 694 | 71 | 29 |
| Agent handoff | 2,172 | 108 | 46 |
| Logs and alarms | 12,800 | 85 | 29 |
| A mixed day, one file | 15,668 | 205 | 46 |
| The mix (sum) | 31,334 | 469 | 150 |
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:
- Pipeline capacity is frame math: a full pipe of a given frame size, at 10 Mbps or 100 Mbps. It answers how much original text that pipe can carry, not how fast a network interface is rated.
- Last-mile speedup is an in-process bench for version 1. It answers how many times sooner the same job finishes on a full pipe, with no added round-trip delay.
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 copy | Tiny frame (bytes) | On a 10 Mbps pipeline | Traffic added | On a 100 Mbps pipeline |
| Chat | 29 | 239.310 Mbps | 229.310 Mbps | 2,393.103 Mbps |
| Handoff | 46 | 472.1739 Mbps | 462.1739 Mbps | 4,721.739 Mbps |
| Logs | 29 | 4,413.793 Mbps | 4,403.793 Mbps | 44,137.931 Mbps |
| Mixed day | 46 | 3,406.087 Mbps | 3,396.087 Mbps | 34,060.870 Mbps |
| The mix | 150 | 2,088.933 Mbps | 2,078.933 Mbps | 20,889.333 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.
| File | Home jobs/s | Home original | Home wire | p99 |
| Chat, next copy 29 bytes | 70,972.21 | 49.255 MB/s | 2.058 MB/s | 0.445 ms |
| Handoff, next copy 46 bytes | 48,388.01 | 105.099 MB/s | 2.226 MB/s | 0.909 ms |
| Logs, next copy 29 bytes | 19,367.73 | 247.907 MB/s | 0.562 MB/s | 1.277 ms |
| Mixed, next copy 46 bytes | 17,312.14 | 271.247 MB/s | 0.796 MB/s | 1.548 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.
| File | Version 1 at 10 Mbps | Version 1 at 100 Mbps | Break-even, version 1 |
| Chat | 7.6831× | 2.6260× | 322.336 Mbps |
| Handoff | 14.0441× | 3.7803× | 442.386 Mbps |
| Logs | 47.7366× | 6.6793× | 694.285 Mbps |
| Mixed | 36.7324× | 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.
- Every measured product arm restored the original bytes.
- On the saved card, 17,101 of 17,101 requests restored the original.
- Sixteen of 16 damaged frames on that card were refused.
- In the home-pair runs above, every job restored.
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
- Pipeline megabits per second are frame math filling a link with the measured frames. They are not a claim that a 10 Mbps network interface has been upgraded, and they are not a multi-site fabric test. The 100 Mbps next-copy column was not re-timed on a 100 Mbps link.
- The multiple adds once per warmed last mile. It does not compound. A site that has never accepted the file still receives the first copy (469 bytes on this mix, not 150).
- Last-mile speedups are in-process version 1 clocks with no added round-trip delay. They should not be quoted as wide-area latency.
- The home-pair speed table is one timed run on the machines named above. It should not be combined with other clocks from other benches.
- Restore and refuse counts cover the measured October 2026 runs. They are not a contractual durability or uptime promise.
- This paper does not report a result on already-compressed media, on random bytes, or on traffic that does not resemble the measured shapes.
- No organization is named as a customer, pilot, or case study.
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