Short answer: an open-source voice stack (Asterisk, FreeSWITCH, or jambonz) is still the right answer when media must stay on your own infrastructure or when your volume is high enough that per-minute margin pays for the engineers who run the stack. Managed CPaaS wins when the team that would operate the stack does not exist, and in 2026 the decision moved again: AI voice agents made the LLM/STT/TTS orchestration layer the hard part, and no open-source stack ships that layer managed. The middle path, keeping your open-source media layer and hanging a managed platform for channels and agents, is now a documented, shipped option rather than a whiteboard sketch.
The quick answer by team shape
| Your shape | Likely winner | Why |
|---|---|---|
| Staffed telephony team, high minutes, own carrier contracts | Open-source stack | Zero per-minute margin; the engineer cost is already on the payroll |
| Regulated media residency (media must not leave your network) | Open-source stack, or hybrid with on-prem media | Sovereign media is the one requirement no managed platform can outsource |
| Product team without a telephony rota | Managed CPaaS | The on-call burden of a media server is a product tax paid nightly |
| Team running jambonz or FreeSWITCH that wants AI agents and more channels | Hybrid: your media layer plus a managed platform for channels and agents | You keep the stack you operate; the platform carries the parts you would otherwise build |
The rest of this post is the reasoning behind those four rows.
What you actually own when you own the stack
"Free" telephony software is free the way a free puppy is free. Running Asterisk, FreeSWITCH, or jambonz in production means owning six things, and each one has a real cost line:
- The media server. RTP paths, codec negotiation, DTMF handling, recording storage, and capacity planning for concurrent calls. A media server that runs out of channels at peak drops calls, and the paging for that is yours.
- The dialplan and routing logic. Every route decision, failover chain, and number mapping is code you wrote and you maintain. FreeSWITCH and Asterisk express this differently (XML dialplan vs
extensions.confvs jambonz's webhook-driven applications), but the ownership is identical. - SIP trunk contracts and number ownership. You negotiate the carrier contract, hold the DIDs, manage the interconnect, and answer for quality on the PSTN leg. Wholesale termination is cheap per minute and expensive per incident.
- STT/TTS plumbing. Speech vendors become your contracts, your key rotation, your latency budget, and your failover logic. In an AI voice agent, the conversation quality ceiling is set by this plumbing.
- The compliance scope. If you card-pay on a call, the media server is in your PCI scope. Recording retention, consent capture, and audit evidence are your storage and your policy engine, configured per tenant or per line of business by your own team.
- On-call for the stack. Media servers fail at 3 a.m. The team that owns the stack owns that page.
Asterisk brings the deepest PBX ecosystem and the largest pool of engineers who have touched it. FreeSWITCH is the media engine under a generation of commercial voice products and scales better as a pure switching fabric. jambonz is the youngest of the three and the most relevant to this comparison in 2026: it is a cloud-native, webhook-driven voice stack whose application model looks like a CPaaS programming model, which is exactly why it is the open-source layer teams keep when they adopt a managed platform for everything above it.
Managed CPaaS inverts the ownership list. The platform owns the media servers, the routing, the carrier relationships, the speech plumbing, and the on-call rota; you own configuration, your application logic, and a usage bill. The bill is the price of transferring six ownership lines to a vendor that runs them for every tenant, not just yours.
The 2026 inflection: AI voice agents moved the hard part
The original case for open-source telephony was that the switching layer was the hard part, and owning it captured the margin. In 2026 the switching layer is the solved part. The hard part of a voice product is now the agent layer: LLM orchestration with turn-taking and barge-in, STT/TTS vendor management across regions, prompt and tool-call evaluation, call recording and transcript storage wired into your compliance posture, and the channel surface around the call (SMS follow-ups, WhatsApp, email summaries).
No open-source voice stack ships that layer managed. You can bolt each piece onto jambonz or FreeSWITCH with webhooks, and teams do, but every bolted piece is another item on the ownership list above, written into a newer and faster-moving problem space than dialplan scripting. The calculus that justified owning the stack in 2016 (media switching is scarce and expensive) does not apply to a switching layer that is now commodity while the orchestration layer above it is where the scarce engineering goes.
This is also where the middle path became real. A managed platform that accepts your SIP trunk keeps your open-source stack in the loop as the media layer while the platform carries the parts you do not want to build: Devotel Orbit provisions per-trunk SIP credentials for BYO trunks with worked provisioning examples, publishes jambonz bridge worked examples for exactly this topology, and accepts BYO SMPP binds so your existing messaging carriers stay yours too. In that shape, jambonz or FreeSWITCH keeps doing what it does well (media, on your infrastructure, under your control) and the platform carries the channel breadth, the agent runtime, and the PSTN relationships. The BYO carrier guide walks the tradeoff between keeping your carrier contract and letting the platform be carrier of record.
TCO math: own stack vs managed, same methodology
The SIP trunk TCO post decomposes a managed trunk bill into (channels on primary trunk × per-channel fee) + (failover channels × per-channel fee) + (talk-minutes × blended per-minute rate). The same decomposition works for the ownership decision, with one extra term on the open-source side:
Open-source TCO = (engineer share of salary) + (compute and storage) + (wholesale carrier minutes) + (your on-call cost, priced as engineer time)
Worked on the same 50-seat deployment the TCO post uses (10 primary channels, 5 failover channels, 7,000 talk-minutes across three destination tiers):
| Line item | Managed CPaaS (illustrative) | Open-source stack (illustrative) |
|---|---|---|
| Trunk channels (15) | $30.00 | $0 (your trunks, your carrier) |
| US/CA minutes (5,500) | $77.00 at blended retail per-minute rates | ~$38.50 at ~$0.007/min wholesale |
| Tier-2/3 minutes (1,500) | $67.50 | ~$18.00 at wholesale |
| Compute and recording storage | included | ~$150.00 |
| Engineer time to run the stack (0.25 FTE) | $0 | $2,000–$4,000 |
| Monthly total | ~$174.50 plus platform value-add | ~$2,200–$4,200 |
Read the two columns honestly. At 7,000 minutes the open-source stack's per-minute advantage (~$70/month) is real and it grows linearly with volume, but it sits under an engineer line that does not scale with volume. The crossover is a volume and staffing question, not a philosophy question:
- Minutes without a telephony hire: managed wins at every volume. Hiring 0.25 FTE to save $70/month at this volume is not a saving.
- Minutes with the engineer already on payroll (a team that runs the stack for other reasons): open-source wins from roughly the point where the per-minute delta exceeds the incremental compute line, and the margin keeps widening at tens of thousands of minutes.
- AI agent volume: add the agent layer and the managed column absorbs STT/TTS orchestration pricing while the open-source column absorbs another engineering line. The agent layer is where most 2026 deployments tip, because its cost is engineering time on the open-source side and usage pricing on the managed side.
Substitute your own rate card and salary numbers; the pricing page publishes the managed side and your carrier contract publishes the wholesale side. The CPaaS pricing models TCO post runs the same decomposition on the messaging side, where per-message billing replaces per-minute billing.
Decision matrix: team shape, volume, compliance load
| Dimension | Open-source wins | Managed wins | Hybrid wins |
|---|---|---|---|
| Team shape | Telephony engineers on payroll who want to own the stack | Product teams without a voice rota | Teams with a media stack but no agent/channel team |
| Volume | Tens of thousands of minutes and growing, with staffed ops | Anything below the point where per-minute margin pays an engineer | High media volume you already carry, with platform-billed channels above it |
| Compliance load | Teams that must keep media on-prem or in a sovereign region, with their own policy tooling | Teams that want recording retention, consent, and tenant-level caps as configuration rather than projects | On-prem media plus platform-managed evidence and retention controls |
| AI voice agents | Rarely, outside research teams | Default | When the media layer must stay yours but the agent layer should not |
| Carrier posture | Own contracts, own DIDs, own interconnect incidents | Carrier of record, one bill, platform-owned carrier health | BYO trunk: your carrier and numbers, platform credentials and routing surface |
The compliance column deserves a word of care. Open-source gives you control of the media path, which is a genuine regulatory asset in some regimes; it also gives you the obligation to build retention, consent, and tenant scoping yourself, tenant by tenant. A managed platform hands those over as configuration your own team sets, which is control without construction. Map your actual regulatory requirement to one of those two shapes before the cost math, because the compliance shape constrains the cost math, not the other way round.
When open-source genuinely wins
Three cases, stated plainly, because a comparison that cannot name its loser names nothing:
- Sovereign or on-prem media. If the regulation says the audio may not leave your network or your jurisdiction, a managed media plane is disqualified before pricing starts. Open-source stacks (or a hybrid that keeps media on your jambonz/FreeSWITCH box) are the only shape that complies.
- Zero per-minute margin at scale, with the staff already hired. A contact-center operator moving tens of millions of minutes a month through FreeSWITCH, with the team that has run it for years, is capturing margin no managed bill will ever match. The engineer line is sunk; the per-minute delta is pure margin.
- Deep media-path customization. Custom codecs, bespoke signal processing, research pipelines: when the media path itself is the product, owning it is the product.
Outside those three, the honest 2026 read is that the open-source decision is a staffing decision wearing a cost costume, and the hybrid shape exists precisely for teams whose answer to "do we still want to own the media plane" is yes while their answer to "do we want to build the agent and channel layer" is no.
Where Devotel Orbit fits
One labeled section, same posture as the stack-dismantling series: Devotel Orbit is the managed side of every row above, including the hybrid row. As carrier of record it terminates and originates on its own wholesale softswitch behind one usage-based bill drawn from the same wallet as SMS, WhatsApp, email, and agents; as a BYO surface it provisions SIP credentials against your trunk, takes your jambonz or FreeSWITCH box as the media layer via the documented bridge path, and keeps your SMPP carriers bound to your account. The voice alternatives hub collects the head-to-head pages, and the Orbit vs Signalwire comparison runs the same ownership-list method against the nearest managed incumbent for buyers who have already ruled open-source out.
Frequently asked questions
Is jambonz a CPaaS competitor or a CPaaS component?
Both, depending on the team. Run alone, jambonz is a self-operated voice platform and competes with managed CPaaS on the ownership list in this post. Bridged to a managed platform for channels and agents, it is a component: the media layer in a hybrid topology. The 2026 shift toward AI voice agents pushed more jambonz deployments into the second shape, because the agent layer is the part nobody runs jambonz to build.
Can we keep our Asterisk PBX and still add AI voice agents?
Yes, through the same BYO trunk shape. Your Asterisk box keeps the dialplan and the media; the platform side carries the agent runtime, speech orchestration, and the channel follow-ups. The bridge is a SIP trunk with provisioned credentials, not a rip-and-replace.
At what volume does owning FreeSWITCH beat a managed bill?
There is no universal number, because the engineer line dominates the equation. With a telephony engineer already on payroll for other duties, the per-minute delta turns positive in the tens of thousands of minutes. Without one, the managed bill is cheaper at any volume a mid-market team realistically carries, because 0.25 FTE priced at market rates outruns the minute savings.
Does open-source telephony reduce our compliance scope?
It moves the scope rather than reducing it. Owning the media server puts media storage, recording retention, and call-data handling inside your own infrastructure and your own policy tooling, which is an asset under sovereignty requirements and a burden otherwise. Managed platforms hand those controls over as tenant-level configuration your team sets and audits.
What does the hybrid middle path cost?
Your existing stack cost (already sunk if you run it) plus the managed platform's per-channel and per-minute lines for what the platform carries. There is no platform fee for the bridge itself; the trunk channels and the minutes and the agent usage are the bill, on the same usage-based pricing as any other tenant.
The takeaway
Own the stack when media sovereignty, sunk engineering, or scale-margin says so; buy the managed bill when the team that would run the stack does not exist; and in 2026, take the middle path when the media plane is yours but the agent layer should not be. The ownership list, not the headline rate, is the decision instrument: write down who owns the media server, the dialplan, the trunk contracts, the speech plumbing, the compliance scope, and the 3 a.m. page, and the winner names itself. The SIP trunk TCO math supplies the managed-side arithmetic, the BYO carrier guide supplies the hybrid topology, and the voice alternatives hub supplies the head-to-heads for the managed shortlist.
Published 28 October 2026.