Non-Terrestrial Networks: 3GPP Satellites, Delay Budgets, and Direct-to-Device

Satellite and cellular have coexisted for decades as separate industries with separate handsets, separate spectrum, and separate business models. What changed with 3GPP Release 17 is that the satellite became part of the 3GPP system — same NR protocol stack, same core network, and increasingly the same device.

The engineering behind that shift is more interesting than the headline. NR was designed around assumptions that satellites violate comprehensively: cells that stay still, round-trip delays under a millisecond, and Doppler shifts you can ignore.

---

The terminology

Three terms recur constantly and are worth fixing early:

Service link — satellite to UE. This is the radio link that replaces the terrestrial cell.

Feeder link — satellite to ground gateway. The backhaul segment.

NTN payload — what the satellite actually does with the signal, which is the architectural fork covered below.

---

Orbits and what they cost you

LEO gives usable latency and manageable link budget — which is why direct-to-device work concentrates there. The cost is that a satellite is overhead for perhaps 5–10 minutes, so you need a large constellation and near-constant handover.

GEO gives one satellite covering roughly a third of the planet with no tracking needed. The cost is delay: a ~540 ms round trip through the gateway is a hard architectural constraint, not a tuning problem.

MEO sits between and sees less use in 3GPP NTN work.

---

What long delay breaks

NR's timing assumptions are built into procedures at multiple layers. Satellite delays break several of them, and Release 17 had to address each.

Timing advance

A terrestrial cell has a maximum timing advance corresponding to tens of kilometres. NTN needs to accommodate propagation delays orders of magnitude larger.

The solution has two parts. Ephemeris broadcast — the satellite's position and velocity are broadcast in system information. UE-side pre-compensation — a GNSS-equipped UE knows where it is, knows where the satellite is, computes its own delay, and pre-compensates before transmitting.

That GNSS dependency is a real design assumption. Release 17 NTN broadly assumes the UE has position knowledge, which is fine for handsets and IoT trackers and less fine for anything else.

HARQ

The HARQ round trip becomes the dominant latency term. With up to 16 processes and a 540 ms GEO round trip, the link stalls waiting for feedback — the processes simply run out.

Two mechanisms address it:

Extended HARQ RTT timers, scaled to the actual delay so the UE doesn't consider a transmission failed while the acknowledgement is still in flight.

HARQ feedback disabling per process. Some processes run without feedback at all, relying on RLC ARQ or blind repetition instead. This is a genuine departure from terrestrial NR — you're trading the HARQ combining gain for throughput that isn't blocked on a half-second round trip.

Doppler

A LEO satellite at 7.5 km/s produces substantial Doppler shift. At S-band frequencies around 2 GHz, the shift reaches roughly ±50 kHz, with the rate of change also significant as the satellite passes overhead. At Ka-band it scales up by an order of magnitude.

Again the answer is pre-compensation: the network compensates on the downlink using known ephemeris, and the UE pre-compensates its uplink frequency. The residual is what the receiver's frequency tracking has to handle.

---

Moving cells

Terrestrial mobility assumes the UE moves and cells stay put. In LEO NTN, the satellite moves at 7.5 km/s and the UE may be stationary — mobility management inverted.

3GPP defines two beam behaviours:

Earth-fixed beams — the satellite steers its beams to keep them pointed at a fixed ground area as it passes. The cell stays put on the ground; the satellite serving it changes. This keeps tracking areas meaningful and looks more like terrestrial mobility.

Earth-moving beams — beams sweep across the ground with the satellite. Simpler for the satellite, harder for the network, because cell coverage areas continuously change.

Earth-fixed is generally preferred where the payload can steer, because it preserves the assumption that a tracking area corresponds to a place.

There's also feeder link switchover — as the satellite moves out of a gateway's view, the feeder link transfers to another gateway. That's a network-side handover with no terrestrial equivalent, and it has to happen without disrupting active sessions.

---

Transparent vs regenerative payloads

The architectural fork, and the one that determines what the satellite actually is.

Transparent (bent pipe)

The satellite is an RF repeater. It receives, amplifies, frequency-converts, and retransmits. The gNB is on the ground, at the gateway.

The satellite carries no protocol intelligence, which makes it simpler, lighter, cheaper, and easier to upgrade — you change the gNB software on the ground and the constellation gains features without a launch.

The cost is that every packet traverses the full UE-satellite-gateway path, and the feeder link is in every delay budget.

Regenerative

The gNB, or part of it, is on board. The satellite demodulates, decodes, and processes.

This shortens the service-link delay because the radio protocol terminates at the satellite rather than the ground. With inter-satellite links, traffic can route between satellites without touching a gateway at all — which is what makes coverage over oceans and poles feasible.

The costs are real: more power, more mass, more thermal load, more complex hardware, and a software upgrade path that requires uploading to orbit. Radiation tolerance requirements make the processing hardware substantially more expensive.

A regenerative payload can also host a split gNB — DU on the satellite, CU on the ground — which is a neat application of the CU/DU split to a genuinely different constraint.

Release 17 focused on transparent payloads. Regenerative is where later releases and most commercial ambition are heading.

---

Direct-to-device

The application generating the most commercial interest, and the one with the most confusion around what's actually being delivered.

Direct-to-device (D2D) means an ordinary handset — no dish, no special terminal — connecting to a satellite. The link budget is brutal: a phone antenna has near-zero gain, limited transmit power, and no ability to point.

Making it work requires large satellite antennas, low frequencies (L- and S-band, where 3GPP defined FR1-NTN bands including n255 and n256), very low data rates, and substantial processing gain.

That last point sets expectations correctly. D2D is a messaging and emergency-connectivity service, not a broadband one. Early commercial offerings deliver text, location sharing, and emergency contact. Voice is at the edge of feasibility. Video is not on the table.

Two parallel tracks exist and get conflated:

IoT-NTN — NB-IoT and eMTC over satellite, standardised alongside NR-NTN in Release 17. Arguably the more commercially mature path, because IoT traffic profiles suit the constraints: tiny payloads, delay-tolerant, low duty cycle. Asset tracking across oceans and remote sensor telemetry are natural fits.

NR-NTN — full NR over satellite, targeting higher capability at correspondingly higher cost.

Some commercial direct-to-device services also use non-3GPP or LTE-based approaches with spectrum leased from terrestrial operators. Those are a different technical path to a similar user-facing outcome, and they're often what's meant when a handset vendor announces satellite messaging.

---

Where NTN actually fits

Coverage extension — oceans, deserts, mountains, polar regions, and anywhere terrestrial deployment can't be justified economically.

Emergency communications — service when terrestrial infrastructure is damaged or absent. The strongest public-interest case, and a regulatory driver in several markets.

IoT asset tracking — containers, vehicles, and equipment moving through areas with no terrestrial coverage.

Backhaul — connecting remote terrestrial cells where fibre or microwave isn't viable. The oldest satellite application, now integrated into the 3GPP architecture rather than sitting outside it.

Maritime and aviation — moving platforms that have always needed satellite and can now use 3GPP-standardised equipment.

The framing that matters: NTN complements terrestrial networks, it doesn't compete with them. A LEO constellation's total capacity spread over its coverage area is a tiny fraction of what a single dense terrestrial network delivers. The value is reaching places terrestrial can't, not serving places it already does.

---

The takeaway

Service link = satellite↔UE. Feeder link = satellite↔gateway.

LEO for latency and link budget, GEO for coverage with a ~540 ms round trip.

Long delay forced real protocol changes: extended timing advance with ephemeris broadcast, extended HARQ timers, and per-process HARQ feedback disabling.

GNSS-based pre-compensation for both timing and Doppler is a core Release 17 assumption.

Transparent = gNB on the ground. Regenerative = gNB in orbit, enabling inter-satellite routing.

Direct-to-device is messaging and emergency service, not broadband. IoT-NTN is the more mature path.

NTN is a good example of a specification adapting to constraints it was never designed for — and the interesting parts are exactly the places where terrestrial assumptions had to be dismantled.

---

Further reading

• 3GPP TR 38.821 — Solutions for NR to support non-terrestrial networks
• 3GPP TS 38.300 — NG-RAN overall description, NTN aspects
• 3GPP TS 38.331 — RRC specification, ephemeris and NTN configuration
• 3GPP TR 36.763 — Study on NB-IoT/eMTC support for NTN
• The gNB explained — the CU/DU split referenced above