Everyone else sells you an API. Sinch sells you the network underneath it.

The CPaaS market has stopped being a messaging-gateway business and become an orchestration business: globally distributed, intelligent, and increasingly judged on transport quality rather than API ergonomics. The 60%-to-95% adoption curve between 2025 and 2029 is the demand side of that shift, and it consolidates the market into an oligopoly where the durable differentiator is who owns physical infrastructure.

There are two latency budgets, and conflating them loses deals. End-to-end conversational turn latency — speech in to speech out — needs to land near 800 milliseconds to feel human. Per-hop network and media latency needs to stay under roughly 100 milliseconds for an AI agent to interact naturally over a telephone network at all. Model inference consumes most of the turn budget, which means the transport layer has almost none to spare.

Where the 800 milliseconds actually goes is the part worth memorizing, because it reframes the whole conversation. Capture and voice activity detection take 10 to 30 milliseconds. Speech to text takes 60 to 120. Model inference takes most of what remains. By the time the network is asked for its share, there is almost nothing left — which is why the transport layer has to be measured in tens of milliseconds rather than hundreds, and why an extra broker hop is not a rounding error but the difference between a conversation and an awkward pause.

Scale is the other half of the credibility argument, and the numbers are worth carrying verbatim: more than 150,000 businesses on the platform, including eight of the ten largest technology companies in the world; over 600 direct mobile-network-operator connections; more than 250 billion messages a year into over 190 countries; and 140 million-plus phone numbers reachable programmatically. Thirteen messaging channels sit behind one integration layer.

There is also a governance number that reframes AI conversations entirely: industry data puts 74% of enterprise AI customer-communication deployments as having hit operational problems or been rolled back — unmasked PII, hallucinated outputs, or escalation handoffs that dropped context. The question a buyer is really asking is not whether AI works. It is whether it can be governed.

The financial picture explains the product roadmap. In Q3 2025 Sinch reported net sales of 6.7 billion SEK against a market forecast it narrowly missed, with EPS of -0.01 USD. Underneath flat top-line revenue, the margin story executed: 5% organic gross-profit growth, 8% organic adjusted EBITDA growth to 915 million SEK, and a 14% adjusted EBITDA margin.

That is a deliberate mix shift away from commodity SMS transit toward higher-margin, next-generation channels — RCS, email, verification, and AI orchestration. When a prospect asks why the roadmap emphasizes RCS and silent authentication over cheaper bulk SMS, this is the answer: those are the channels where owning the network converts into margin instead of being competed to zero.

Management has put numbers on where that goes: a 2027 target of 8 to 10% annual gross profit growth with an adjusted EBITDA margin of 20 to 22% of gross profit. Analysts read that as the deliberate rotation away from commodity SMS toward higher-margin surfaces — RCS, email, verification. The customer list is the credibility layer for it: Nespresso, Visa, Nordstrom, and roughly eight of the ten largest technology companies in the world, with Gartner and Forrester placing Sinch in the leading tier for CPaaS.

The restaurant analogy has four floors, not two. The API surface is the menu and waiter. Beneath it sits the orchestration layer — routing logic, channel abstraction, template and identity handling, compliance rules per jurisdiction. Beneath that sits carrier interconnect: the commercial and technical agreements that let traffic enter an operator's network. And at the bottom sits the physical network itself: switches, signaling, and IP peering that actually move packets.

Every CPaaS vendor sells you floors one and two. The strategic question is how many of the lower floors they own versus rent. Vendors that own only the top two floors buy floors three and four as transit — which means their unit economics, latency, and quality ceiling are all set by someone else's pricing and someone else's network.

Transcoding is the clearest example of what the orchestration layer earns you. You submit one payload; Sinch renders it per channel — a rich card in Apple Messages for Business, an approved template in WhatsApp, or a graceful downgrade to plain SMS text with a link when the destination supports nothing richer. When a designer needs pixel control, there is an explicit bypass to send a channel-exact layout. Underneath it, fallback chains are an ordered list of channels — try RCS, then SMS — executed inside Sinch rather than in your retry logic, with validation that every channel in the chain is actually configured before the request is accepted.

Three terms from the research paper matter more than the rest, because they name the mechanism rather than the marketing. Origination and termination describe where traffic enters and leaves a network: a vendor that both originates and terminates on infrastructure it owns never hands your traffic to a third party. Transit is the opposite — buying passage across someone else's network, which introduces the three costs below. And cost of goods sold is where network ownership shows up on a balance sheet: owned routes carry a structurally lower COGS than leased capacity, which is why discounting on owned routes does not destroy margin.

CAMARA is the newer one to learn. It is the open standard set that exposes carrier network capabilities — number verification, SIM-swap signals, device location — as programmable APIs. It is the standards backbone behind silent authentication, and it points at a future where the network authenticates the user and typed codes disappear.

The deep-dive vocabulary is here too: transcoding and fallback chains on the messaging side, drain rate and ingress queue when throughput comes up, MSISDN channels and webhooks when someone asks how inbound works, rate centers and on-net versus off-net when the conversation turns to voice economics, and barge-in, voice activity detection, and speech-to-text when it turns to AI. Learn the ten you will actually say out loud; know where to find the rest.

The precise technical claim is this: Sinch originates and terminates traffic directly on its own tier-1 network. That is not the same as having good routes. It means the packet's first and last hops are both on infrastructure Sinch controls, so there is no third-party transit segment where behavior becomes unobservable.

Third-party transit imposes three distinct costs, and they are worth naming separately on a call. Latency: every transit segment adds queueing and re-serialization delay. Margin degradation: every intermediary prices its segment, compounding markup on traffic the vendor cannot reprice. Quality loss: packet routing decisions, retry behavior, and delivery reporting are all made by a network the vendor does not operate and cannot audit — which is exactly how synthetic delivery receipts get generated in the first place.

Direct interconnect also changes what telemetry is available. When you own both endpoints, packet loss, congestion, and route health are first-party measurements rather than inferred from a partner's reporting, which is what makes machine-learning-driven dynamic rerouting possible at all.

The reason a central network exists at all is arithmetic. Connecting a hundred carriers directly to one another requires nearly five thousand physical links — a number that never gets built. That is the problem Neutral Tandem was founded to solve and the asset Sinch now owns as Inteliquent: carriers connect once, to the tandem, and the tandem reaches everyone else. An aggregator without that hub is buying transit across it.

The Inteliquent acquisition is the economic moat, not a feature line. Owning a tier-1 network outright gives Sinch control over three things a reseller cannot touch: direct carrier connection terms, packet routing efficiency, and cost of goods sold. Competitors lease carrier network access from underlying providers — historically, in some cases, from Sinch and Inteliquent themselves. A vendor reselling your network cannot out-engineer your latency.

Voice Relay and WebRTC are the mechanism that turns that ownership into AI capability. Voice Relay is a low-latency media bridge that sits between the AI model and the voice backbone, keeping media handling inside the owned network instead of traversing broker hops. WebRTC carries real-time audio into browser and in-app contexts on the same path. Together they are what makes the sub-100ms per-hop network budget reachable, leaving the remaining turn budget for model inference.

The physical scale is the argument. Roughly 300 billion voice minutes a year, more than 12,800 on-net rate centers, a 100 gigabit IP backbone linking over 30 geo-diverse points of presence, and 99.999% uptime. About 95% of US voice traffic stays on-net — never handed to a transit carrier, so no interconnect fee, no extra hops, no borrowed packet loss. Alongside SIP trunking that replaces legacy PRI lines with programmatic access to global numbers, E911, CNAM, and directory listings, the network also performs the unglamorous work of interconnect: normalizing each carrier's SIP dialect and hiding internal topology behind session border controllers.

For AI voice, Voice Relay streams call audio over a WebSocket directly to a model and handles barge-in — when a human interrupts, it detects the voice activity, halts speech synthesis, and forwards the new audio. That matters because the turn budget is already spent elsewhere: 10 to 30 milliseconds for voice activity detection, 60 to 120 for speech to text, and the remainder for inference. The transport layer has no slack to give back.

The routing topologies are worth knowing by name, because a technical buyer will ask which one fits their estate: App-to-App over WebRTC, App-to-Phone out to the PSTN, Phone-to-App inbound, App-to-SIP, and App-to-Conference. App-to-SIP is the one that answers the legacy question — it terminates straight into an on-premise PBX, so a customer with existing phone infrastructure adds programmable voice alongside it instead of replacing it.

The Conversation API is a unification layer, not a multiplexer. It normalizes identity across channels — phone number, app-scoped ID, social handle — into one contact profile, then adapts per-channel template rules, approval workflows, and identity models behind a single call. That abstraction is what makes localized regulatory compliance a platform capability rather than an integration project for each market.

On the email side, Mailgun is the developer transactional surface with burst characteristics measured in tens of millions of messages per five-minute window, Mailjet is the marketing composition surface, and Email on Acid is pre-send rendering QA. High throughput matters less than sustained throughput under burst: transactional email arrives in spikes tied to product events, and burst capacity is what keeps the 95% placement figure from degrading exactly when it matters.

Concrete numbers, because engineers ask for them. The Conversation API runs in regional data centers — US East, EU Ireland, and Brazil — and you choose the region for data residency. Sustained throughput is capped at 800 requests per second per project; the ingress queue holds 500,000 accepted messages per application; the default drain rate to carriers is 20 per second, negotiable upward. Authentication is an OAuth 2.0 bearer token valid for one hour, minted from a Key ID and Key Secret where the secret is displayed exactly once. Inbound events arrive by webhook, and routing rules can send replies, delivery reports, and read receipts to different endpoints.

On email, Mailgun's Rapid Fire SLA is the number to quote: up to 72 million requests per hour, with 99% attempted delivery of 15 million messages inside five minutes. That is burst capacity backed by pre-warmed IP pools and queue management — precisely what a traditional SMTP relay throttles at the exact moment a password reset or fraud alert needs to land. And when the security reviewer joins the call, the certifications are the answer: SSAE-16 SOC I and SOC II, ISO 27001, and full HIPAA and GDPR compliance — which is what lets a healthcare or financial-services buyer put transactional email in scope at all.

The webhook surface is richer than a delivery callback, and the payload names are worth carrying into a technical conversation. Inbound messages arrive as message events; delivery and read state arrive separately; a composing_event payload drives a typing indicator in the customer's UI; a read_message_event payload triggers a read receipt. That is what makes a two-way conversation feel like a chat app rather than a send queue with replies bolted on.

On email the burst figure is the one to quote: Mailgun's SLA covers surges of up to 25 million emails every five minutes, with real-time list validation to clean recipients before send, automated IP warm-up on new sending infrastructure, and deliverability analytics over both the REST API and plain SMTP. The compliance stack underneath it is SOC Types I and II, ISO 27001, HIPAA, and GDPR.

The quieter product in the portfolio is programmable fax, and it wins deals in regulated sectors precisely because nobody else bothers. A HIPAA-compliant Fax API replaces on-premise fax hardware, which is how a telemedicine platform sends records, lab results, and referrals in the same interface as its voice, SMS, and video — Spruce Health being the reference. Voice control runs on SVAML for TTS, recording, DTMF capture, IVR trees, and number masking.

Silent Mobile Auth is the leading edge of a standards movement, not a proprietary trick. Under CAMARA-aligned network identity APIs, the carrier network itself becomes the authenticator: the platform validates a cryptographic SIM token over a direct carrier data session and returns a verdict. No code is generated, transmitted, displayed, or typed — so there is nothing to phish, intercept, or socially engineer.

The strategic implication is worth saying out loud to a security buyer: the paradigm shift underway renders traditional user-input authentication obsolete. Typed one-time passcodes become the legacy fallback rather than the default, and the attack surface that AIT fraud monetizes shrinks with it. This capability requires direct carrier data-session access — which is a network position, not a software feature.

Underneath the pitch there is a ladder of four modalities, ordered by friction and cost. SMS OTP is the universal baseline and the most interceptable. Voice OTP reads the code aloud for accessibility or where SMS is unreliable. Flashcall places a call that drops before it connects — the SDK reads and matches the caller ID, so there is no PIN in transit and no termination charge. And silent network authentication asks the operator to confirm that the SIM on the data connection matches the claimed number, with nothing shown to the user at all.

The standards layer is what makes the top rung buildable. The GSMA Open Gateway initiative is pushing 90-plus network APIs toward CAMARA specifications, and Sinch acts as the single integration point across hundreds of carriers rather than making you negotiate operator by operator.

Number Lookup deserves its own sentence, because it is a spend-control tool as much as a fraud tool. Run the target list through it in bulk before a mass campaign: the API returns whether each number is active, whether it is a mobile handset or a landline, and which carrier hosts it. Landlines and dead numbers get dropped before a single message is submitted, so budget is not spent on destinations that were never deliverable — and the same lookup feeds pre-routing and SIM-swap risk checks.

Read that ladder as a routing strategy rather than a feature list. Configure the Verification API to attempt the cheapest modality that can succeed first — Flash Call, or a WhatsApp OTP where per-attempt tariffs undercut A2P SMS — and fall back to SMS only when it must. The saving is simply one minus the alternative cost over the SMS cost, and on Flash Call that reaches up to 50% of a standard SMS OTP tariff while also being faster, because there is no PIN to read and type.

Two details close technical deals here. The whole authentication lifecycle is managed server-side — PIN creation, storage, expiry, validation, and delivery retries — so none of that state lives in your application. And the API integrates out of the box with existing CIAM platforms including Okta and Auth0, which means multi-factor authentication drops into the access flow an enterprise already runs instead of asking them to rebuild it.

Read the acquisition list as a stack diagram rather than a shopping list. SAP Digital Interconnect bought carrier interconnect depth. Inteliquent bought the physical network layer and the CLEC license that makes operating it legal in all 50 states. Pathwire bought a whole channel — the email lifecycle, from composition through QA through delivery telemetry. MessageMedia bought segment reach; Wavy, ACL Mobile, and MessengerPeople bought regional channel depth where local operator relationships decide deliverability.

Each purchase acquired a layer that cannot be built in a sprint: a carrier license, an interconnect footprint, a deliverability reputation. That is the structural reason a software-only competitor cannot close the gap with engineering velocity — the missing pieces are regulatory and physical, not architectural.

Inteliquent is the one worth understanding properly, because it started life as Neutral Tandem — a company built on an arithmetic problem. Connecting a hundred carriers to each other directly needs nearly five thousand physical links, which nobody builds. So carriers connect once, to a central tandem hub, and the hub transits everyone's traffic to everyone else. Sinch bought the hub. It is now an all-IP SIP network rather than legacy TDM switching, and the newest layer sits above all of it: an AI orchestration surface where Sinch Skills builds no-code agents and the Model Context Protocol lets an AI agent call Sinch messaging as a tool with governance and audit trails attached.

Governance is what makes the AI layer sellable, and it is structural rather than procedural. Smart Conversations runs on the ingestion pipeline itself: it classifies intent and sentiment in real time, filters abusive language before it reaches a live-agent console, and masks personally identifiable information before any payload is written to a log. Chatlayer keeps a stateful integration with the agent platform so an escalation carries the transcript and the extracted entities across.

The developer-facing piece of that is Sinch Skills — structured knowledge modules installed with npx skills add sinch/skills that give an AI coding assistant the platform's real rules: application key and secret versus project-level OAuth 2.0 tokens, the mandatory caller ID on a text-to-speech callout, and how asynchronous delivery callbacks actually behave. It exists because those three mistakes account for most hallucinated Sinch integrations.

The honest comparison: both Sinch and Twilio are reliable at enterprise scale in North America, and both give developers a clean surface to send an SMS or a WhatsApp message. The divergence is one layer down. Twilio leases carrier network access from underlying providers; Sinch owns the tier-1 Inteliquent network outright. That ownership converts directly into control over carrier connection terms, packet routing efficiency, and cost of goods sold.

Be candid about the shared weakness too, because it builds credibility. Any platform processing hundreds of billions of interactions a year can be slow to respond on small deployments — that is true on both sides. What Sinch offers at the enterprise tier is dedicated account management and SLA structures customized for multinational compliance, which is a different product from self-serve support.

Be honest about the trade the reviews describe. Twilio is widely rated easier to start with — better docs, a smoother self-serve on-ramp for a developer with a credit card. Sinch reviews skew the other way at small deployment sizes, where support responsiveness can lag, and skew strongly positive at enterprise scale where named account management, custom compliance work, and underwritten SLAs are the product. Sell into the scale where that is an advantage, and do not pretend it is a hobby-project platform.

Three commercial levers come straight out of network ownership. High-throughput SLAs are underwritable because Sinch operates the circuits the commitment depends on — you cannot credibly guarantee throughput on capacity you rent per message. Volume-optimized pricing works because owned-route COGS is structurally lower, so discounting at volume compresses margin instead of erasing it. And localized regulatory compliance is customizable per market because compliance rules live in the orchestration layer that Sinch controls end to end.

For a procurement conversation, that reframes the negotiation. You are not asking for a discount on a resold unit rate; you are consolidating spend onto a vendor whose cost base moves with your volume, under one master commitment across regions, with regulatory fees passed through at cost.

The account model is a commercial lever in its own right. Account, then Project, then Subproject — and a Subproject keeps usage metrics fully isolated while inheriting user management from its parent. For an ISV or reseller that is per-client cost tracking without per-client contracts, and when a client churns you delete the Subproject.

Above the raw APIs, Sinch Engage sells as tiered SaaS with published numbers, and knowing them is the difference between an answer and a follow-up email: Basics at $49 a month, Conversations at $99, Pro at $249, and Advanced at $799, with premium channels such as WhatsApp Business and Instagram carrying a $10-per-month add-on on top. Sinch Chat bills separately on monthly active users with the first thousand free — a low-friction entry point for a CX team that has no developers to spare.

Rate limits are commercial too. The Conversation API default drain rate to carriers is 20 messages per second and the ingress queue holds 500,000 per application; both move through commercial agreement, not a support ticket. Sustained traffic above 800 requests per second returns 429s. Knowing those numbers lets you size a deal honestly instead of discovering the ceiling in production.

Read these eight as architecture evidence, not logos. Doctolib's 72-hour RCS launch is only possible because RCS Upscale converts an existing SMS flow into verified RCS rather than requiring a new integration — the abstraction layer doing its job. HSBC's 10,000 whitelisted templates in China are a carrier-relationship asset, not a software feature; a vendor buying transit has nobody to call when a regulator changes the rules.

The commerce cases prove the payload layer. Familhao's 23-point renewal lift came from order_details and order_status payloads authorizing a Pix or card payment inside the thread, and Clarins' 35x redirection lift came from multi-card carousels with embedded CTA buttons. Both are the same mechanism: one standardized JSON payload transcoded into channel-native interactivity, with a server-side fallback chain underneath in case the handset cannot render it.

The verification cases show the ladder in production. Nets runs two-way SMS confirmation across 8 billion annual card transactions; Triodos runs 250,000 SMS OTPs a month across 170,000 accounts. Both sit on the first rung. The strategic conversation is what happens when those move up to flashcall and silent network authentication over CAMARA — same outcome, no code in transit, and nothing to phish.

The sharpest single sentence in the research: Sinch originates and terminates traffic directly on its own tier-1 network, bypassing the latency, margin degradation, and quality loss inherently associated with third-party transit. If you can say that sentence naturally and then explain each of the three costs, you have the entire architectural argument in about twenty seconds.

The forward-looking half is just as short. Conversational AI needs sub-100ms network latency to hold a natural phone conversation, and Voice Relay plus WebRTC put that inside reach. CAMARA-aligned network identity is turning telecom networks into silent authenticators. Both trends reward whoever owns the physical layer — which is the whole thesis.

The technical closer, if the room wants it: one Conversation API call transcodes into every channel with a server-side fallback chain behind it, verification climbs a ladder from typed OTP up to silent network authentication over CAMARA, and voice runs on the tandem network that the rest of the industry interconnects through — roughly 95% of it on-net.

Three lines to carry out of here. One: every competitor sells an API; Sinch sells the API and the network beneath it, originating and terminating on infrastructure it owns. Two: that ownership is what produces sub-500ms delivery, carrier-native receipts, Level-A attestation, 95% inbox placement, structurally lower cost of goods sold, and a media path fast enough for real-time voice AI. Three: you never have to win the argument in a slide — you have to win one high-volume route in a split test.

The forward view, in one line: as CPaaS converges with network identity standards like CAMARA and with agentic AI orchestration, the vendors that own physical infrastructure become the default layer and the resellers become a pricing tier.

And if someone asks for the deep-technical version, four numbers carry it: 800 requests per second of Conversation API ingress, 72 million email requests per hour on the Rapid Fire SLA, 300 billion voice minutes a year across 12,800+ on-net rate centers at 99.999% uptime, and a sub-100ms media path into an AI model. Every one of those is a consequence of owning the layer underneath the API.