Cellular Data Cost per Device for Streaming LLM Audio

Concept render

Raw 16-bit PCM at 16 kHz uses 1.92 MB per audio minute, and at 24 kHz it uses 2.88 MB. Opus at 16 to 20 kbit/s uses 0.12 to 0.15 MB. At $0.03 per MB (Hologram list price, October 31, 2026), that is $0.058 against $0.0045 per minute. Measure overhead on your own link.

A voice device that streams to a cloud model pays twice: once to the model provider, and once to the carrier for every byte that crosses the cellular link. The carrier bill depends on the audio format, on which direction carries it, and on how many minutes each device talks. This guide gives you the arithmetic and a worked example at one public list price. It contains no Duobay price and no measured result.

How many bytes does a minute of voice use?

The conversion is MB per minute = kbit/s × 60 / 8 / 1000, in decimal megabytes. Cost per minute is that figure times the price per MB. The price used here is the Hologram self-service list price of $0.03 per MB (Hologram pricing page, as of October 31, 2026). Plan prices, tiers and allowances differ between carriers, so treat it as one dated reference point.

Opus at 20 kbit/s is about 13 times cheaper per minute than 16 kHz PCM (0.0576 divided by 0.0045 is 12.8). The RFC lists 16 to 20 kbit/s for wideband speech with 20 ms frames, and the codec supports 6 to 510 kbit/s (RFC 7587, accessed October 31, 2026). The last row adds about 40 bytes of IP, UDP and RTP headers per 20 ms packet. That is 50 packets per second times 40 bytes times 8 bits, or 16 kbit/s. The 40-byte figure is our estimate from standard header sizes, not a measurement, and a TCP, TLS or WebSocket path adds different overhead that no vendor publishes.

Which direction costs more?

The two ends of the link use different formats. Per the Google Live API guide, input is raw 16-bit PCM at 16 kHz and output always uses 24 kHz (Google Live API guide, as of October 31, 2026). The OpenAI Realtime page says to send 24 kHz mono audio when you use PCM (OpenAI Realtime guide, as of October 31, 2026). Those pages describe the model API, not your cellular link. A gateway that you run between the device and the model can choose the format that crosses the cellular link. Whether the model endpoints accept Opus directly is not something we verified.

Worked example with assumed minutes

Assume a device uses 5 minutes of audio up and 5 minutes of audio down per day, for 30 days. Those minutes are an illustration, not a typical value. The table below uses the per-minute figures above and $0.03 per MB, without any packet overhead.

Recheck of scenario A: 5 × 1.92 = 9.6 MB up, 5 × 2.88 = 14.4 MB down, 24 MB per day, 720 MB per 30 days, 720 × 0.03 = $21.60. Scenario B changes only the downlink: 5 × 0.15 = 0.75 MB. Add the Hologram charge of $1 per month per SIM to each line, and the $3 one-time cost of the SIM card.

What is the formula for your device?

monthly_data_cost = 30 × (U × MBmin_up + V × MBmin_down) × O × P + S

MBmin = kbit/s × 60 / 8 / 1000        (decimal megabytes per audio minute)

U, V     minutes of audio up and down, per device per day
O        overhead multiplier for headers, TLS, framing, keepalives (≥ 1)
P        price per MB, in dollars
S        monthly charge per SIM, in dollars

With O = 1, S = 0 and the scenario A inputs, the formula gives 30 × (5 × 1.92 + 5 × 2.88) × 0.03 = $21.60, which matches the table. Use the rows below to see which inputs are sourced and which are yours.

The overhead multiplier O is the number most likely to surprise you, so run the formula with several values of O and with double the minutes. If the answer flips, the decision depends on a figure you have not measured yet. For related tracker work on low-power cellular links, see LTE and GNSS trackers.

What are the common cost traps?

These follow from the published formats, prices and documentation above. They are engineering reasoning about how the numbers behave. None describes an incident at any company.

• Downlink dominates when the reply is raw PCM. The Live API documentation says output is always 24 kHz 16-bit PCM. At 2.88 MB per minute against 1.92 MB for the 16 kHz upload, a reply that is not re-encoded costs more per minute than the question did. Scenario B above shows the effect of re-encoding.
• Always-on streaming multiplies everything. There are 1,440 minutes in a day. Streaming Opus at 20 kbit/s around the clock is 1,440 × 0.15 = 216 MB per day, or 6,480 MB over 30 days, which is $194.40 at $0.03 per MB. Raw 16 kHz PCM around the clock is 1,440 × 1.92 = 2,764.8 MB per day, or 82,944 MB over 30 days, about $2,488. Voice activity detection or push-to-talk cuts the minutes in the formula directly.
• Forcing constant bitrate at a very low rate. The Opus documentation warns that hard CBR can cause noticeable quality degradation for LPC and hybrid modes at very low bitrates, and that VBR is the default. Saving bytes by forcing CBR low can cost you recognition quality, so test quality when you set the rate.
• Counting only the audio payload. The bill counts every byte on the link, so the codec bitrate understates it. Packet headers add a roughly estimated 16 kbit/s at 20 ms packets, which is close to the whole Opus payload. TLS, WebSocket and keepalive overhead is not published, so keep O as a variable until you measure it.
• Billing is per KB after the session closes. Hologram documents that usage is charged per KB after the SIM stops transmitting, when a data session closes. A long-lived session can therefore show up in your balance late. This is an engineering inference, so check how your own carrier reports usage.
• Overage is not on the price page. The Hologram documentation says usage beyond an included allowance is charged at an overage rate, and the rate is not published there. It also says included plans and pools are for qualifying organizations. Do not model overage with a guessed rate. Ask the carrier in writing.
• Long replies. Downlink minutes scale with how long the model talks. A product that lets the assistant give long answers raises V directly. Cap reply length in the prompt and in the session settings, then verify it on logs.

What should you measure on your own board?

The published sources do not give bytes per minute on a real LTE-M, NB-IoT or Cat-1 link, and they do not give the overhead of your protocol stack. We have not measured these for you, and no such result appears in this article. Measure them before you commit to a plan.

Fix every pass line before you run the test. A threshold chosen after you see the result tends to land just under the result.

What should you ask before you commit?

Take this list to your team, and to any supplier who proposes a design or a data plan.

• Which audio format does each direction use on the cellular link, and who converts it?
• Does a gateway re-encode the model reply to Opus before it reaches the device, and where does that gateway run?
• What is our plan for U and V, and is it based on logged prototype sessions?
• Is voice activity detection or push-to-talk in the design, and what does it do to U?
• What was O on a real LTE-M, NB-IoT or Cat-1 link, measured with packet captures?
• What does the carrier charge per MB and per SIM, as of which date, and how is a megabyte defined?
• What happens when usage passes the included allowance, and is the overage rate in writing?
• How long after a session closes does the usage appear on the bill?
• What is the cap on a device that streams by mistake, and how does the firmware stop it?
• Which numbers in the model are measured on our own hardware, and which are still assumptions?

If you want help turning these formulas into a measured data budget for your device, tell us about your device. You can also read how our engineering gates handle connectivity and cloud choices.

Sources

All pages accessed October 31, 2026. Prices and documentation change, so re-read them before you decide.

• IETF, RFC 7587, RTP payload format for Opus: https://www.rfc-editor.org/rfc/rfc7587.html (accessed 2026-10-31)
• Opus codec, encoder controls: https://opus-codec.org/docs/opus_api-1.5/group__opus__encoderctls.html (accessed 2026-10-31)
• Google, Gemini Live API guide: https://ai.google.dev/gemini-api/docs/live-guide (accessed 2026-10-31)
• OpenAI, Realtime conversations guide: https://developers.openai.com/api/docs/guides/realtime-conversations (accessed 2026-10-31)
• Hologram, pricing: https://hologram.io/pricing/ (accessed 2026-10-31)
• Hologram, data plans overview: https://docs.hologram.io/guides/billing/data-plans-overview (accessed 2026-10-31)