Why innovating down the stack is easier than innovating up — and what that means for anyone whose position is now being rented rather than bought.

Between March and July this year, four companies from three different industries announced the same product.

On 10 March at Enterprise Connect, Salesforce launched Agentforce Contact Center — voice, digital channels, CRM data and AI agents in a single system, generally available on day one. On 6 May at SIGNAL, Twilio shipped a new conversation layer: Orchestrator, Memory, Intelligence and Agent Connect, all GA, all built to coordinate interactions, hold context and connect human and AI agents. In June at NiCE World, NiCE moved Cognigy’s agentic AI into the native core of its platform and wrapped a workforce suite around human and AI agents together. And in July, at its analyst summit in Croatia, Infobip presented AgentOS as the control layer where agents, data, channels and customer intent decide what happens next.

Four announcements. One architecture diagram. Channels at the bottom, data and context in the middle, agents on top, orchestration governing the lot.

I came off a quarter on the road in June convinced that there is no third door: every midsize player in this market has to buy its way up the stack or make itself worth buying. What the first half of 2026 added is the part that makes the choice harder. Not everyone is climbing in the same direction, and the direction turns out to matter enormously.

Then July happened

Two announcements in the last month made the picture considerably less comfortable.

The first was AWS, in stages and without much noise. RCS for Business went live on End User Messaging at the end of March. In late May it went from two countries to twenty-two. In June it added rich cards, carousels, webviews and interactive buttons — and, more consequentially, RCS Conversation pricing: one flat rate for unlimited messages inside a 24-hour session, across twenty-one countries. Next to it sits a one-time-passcode product requiring no number registration and no carrier approval. You call the API.

The second was OpenAI. GPT-Live landed on 8 July as a full-duplex voice model. Two weeks later came Presence: a managed platform for production voice and chat agents with policies, permissions, escalation rules, simulation testing and post-launch governance, delivered by forward-deployed engineers and systems integrators rather than sold self-serve. The proof point is OpenAI’s own support line, which it says resolves three quarters of inbound issues without a human. Early customers named: BBVA in Mexico, SoftBank, the Australian insurer IAG.

Neither company is building CPaaS. Both are renting it — plugging into carrier connections, number ranges, RCS agent registrations and termination that somebody else spent fifteen years assembling — and selling the result to enterprises under their own brand, at their own price, on their own commercial model.

A good share of what sits underneath is wholesale supply from the very players now competing with it. That has been true for years, and technically the two worlds have already merged: plenty of wholesale messaging and voice infrastructure has run in AWS regions for a long time. What is new is the buyer. It used to be a developer with an API key. Now it is a platform with a go-to-market.

The WhatsApp move

Look closely at how AWS priced RCS and you will recognise it immediately, because it is not a new model at all. One flat charge covering a 24-hour window of messaging is exactly how Meta priced WhatsApp Business for years. Same shape, same window, same commercial promise: stop counting messages, buy a conversation.

Which makes this a straight competitive attack, and a well-aimed one. WhatsApp is Meta’s rail. RCS is the telco rail, with Google carrying it. AWS has no rail of its own and has just made the telco one commercially interchangeable with Meta’s — for enterprises who buy on the shape of the invoice rather than the technology underneath.

The timing is the part that gives the game away. Meta retired conversation-based pricing on 1 July 2025 and moved WhatsApp template messaging to per-message billing, because as the channel matured Meta wanted to monetise volume. AWS has now adopted, in the middle of 2026, the model Meta abandoned a year earlier.

That is not a coincidence and it is not naivety. It is the difference between a company that needs messaging margin and a company that does not.

Innovating down is easier than innovating up

Here is the thing the four keynotes have in common and nobody said out loud.

I picked this up from Steve Jobs years ago and it has held up better than almost any other framework I use. Innovating down the stack is far easier than innovating up it, because when you go down you already know what you need. Apple knew precisely what silicon to build, because Apple was the largest and most demanding user of that silicon. A chipmaker trying to climb up into devices has no such advantage — it is guessing at a customer it has never been.

Apply that to this market and the asymmetry is glaring.

Salesforce, NiCE, ServiceNow and now OpenAI are coming down. They already own the workflow, the record, the case, the interaction. They know exactly what they are missing and exactly why they want it. Look at what Salesforce actually built: not a Genesys clone, not feature parity with NiCE or Five9, but the core of a next-generation contact centre. That is not a weak product. That is a product built by people who know where the market is going and are only acquiring what they need to get there — undivided ownership of the voice channel, so a call becomes a conversation, and a conversation becomes intent, signal and action inside the system they already control. That is the whole thesis of the thing, and it explains why the feature gap does not embarrass them.

And watch what that descent costs them, because it is not cheap and they are paying it anyway. On 15 June Salesforce agreed to buy Fin — the company formerly called Intercom — for around $3.6 billion in cash, with more than thirty thousand customers, expected to close in Q4 of its fiscal 2027. The number that explains the price is the resolution rate: Fin resolves roughly three quarters of support volume end to end, against about 62 percent for Salesforce’s own help agent. Salesforce did not buy a roadmap. It bought a working resolution rate — because it knows exactly what it needs and precisely what that is worth to it.

CPaaS is going up. Into a suite fight, against buyers it has not historically sold to, in a category where it holds the weakest brand and the longest sales cycles — and, critically, without the advantage of being its own customer. Nobody at a messaging platform has ever run a global service organisation. They are guessing at the buyer, exactly as the chipmaker guesses at the device.

That asymmetry, more than any architecture choice, is why the CPaaS versions of this story read thinner even when the engineering is better.

You cannot disrupt a business you depend on

And there is a second constraint sitting underneath the first, which is older and harsher.

Nobody disruptively innovates a business they rely on to make payroll. Not because they lack the ideas — the people inside these companies see it perfectly clearly — but because the board cannot sign off on destroying the revenue line that funds the attempt. This is not a failure of imagination. It is arithmetic.

So the disruption arrives from outside, as it always has. AWS can price a 24-hour messaging session flat because messaging margin is a rounding error inside Amazon. It can spend a couple of billion learning a market for the strategic option value and barely notice. The AI-voice startups can take risks no listed platform can contemplate, because they have no legacy revenue, no installed base, no reputation and no channel conflict to protect. Creative destruction has never come from the incumbent. That is the definition of the term.

Which is why the sensible expectation for the CPaaS category is not self-disruption. It is what Twilio has actually done: build a credible AI platform on top of the infrastructure position and defend the margin. That is a strong, rational move. It is not a disruptive one, and it should not be sold as one.

The irony is the sharpest thing in this whole story. Fifteen years ago CPaaS did precisely this to the telcos — took a wholesale product the operators were selling by the unit, repackaged it, repriced it, wrapped it in a developer experience, and captured a margin the operators never saw coming. The operators could not respond, because they depended on the very revenue being repriced.

That move is now being run on CPaaS, from above, by people for whom the channel is a cost line rather than a business.

Where the descent stops

Except the descent does stop somewhere, and the place it stops is the most interesting thing on the diagram.

You can buy your way down through applications, through orchestration, through channels. Then you hit interconnect: IPX, direct operator relationships, number ranges, regulatory licensing, the accumulated contract estate across hundreds of markets. That does not yield to capital or to good engineering. It yields to twenty years of relationships and compliance work.

Which is precisely where the telco wholesalers live. Infobip spent a decade and a half building direct operator connections across the world, and bought its way further down in 2022 with Peerless Network — roughly $200 million for a US voice network reaching almost every American across forty-nine states. Proximus Global sits on IPX, carrying an enormous share of the international traffic and messaging that moves between networks. National operators hold the regional version of the same asset. When AWS lights up RCS in twenty-two countries, it is not because Amazon built twenty-two sets of carrier relationships. Somebody sold it access to theirs.

So the picture is split in a way the architecture diagrams obscure. Technologically the two stacks converged years ago and largely run in the same clouds. Operationally, legally and regulatorily they have not converged at all: the telco players still do the telco part, the IT players host it — and the IT players are unambiguously in front on the AI.

That split matters more now, because the other moat just evaporated.

For fifteen years the best CPaaS platform was the one with the best documentation, the cleanest API, the fastest time to first message. Developer experience was a real moat — it took years to build and it was genuinely hard to copy. It is dissolving in front of us. When AI writes the integration, documentation quality stops being a differentiator, and a three-month development project becomes a fifteen-minute prompt for the raw code. The idea that software is by itself a defensible position — the idea that carried the entire SaaS era — is evaporating faster than most business plans account for.

Watch how the category is responding, because it is more candid about the diagnosis than the keynotes are. On 4 August, Sinch shipped Agent Tools: IDE extensions, a simulator, an MCP server and structured product knowledge, aimed squarely at Claude Code, Cursor, GitHub Copilot and ChatGPT Desktop. The stated logic is that developers now work inside AI-assisted environments and expect discovery, implementation and testing to happen there.

Read that as what it is. The customer for developer experience is no longer a developer. It is a coding agent — and a coding agent develops no loyalty to your documentation, your portal or your brand. It reads whichever API definition is best structured that morning. Shipping the tools is the right move. It is also the clearest evidence yet that the thing they defended for fifteen years has stopped being a moat and become table stakes.

What survives is physical and regulatory: fibre, cable, spectrum, interconnect, numbering, licences, jurisdiction. Musk can put a constellation in orbit and it still will not replace the terrestrial estate. The moat did not disappear. It moved back down to the layer this industry has spent a decade trying to climb away from.

And the regulatory half of that just got heavier. On 2 August the European Commission’s AI Office gained formal enforcement powers over general-purpose AI models, and new transparency obligations took effect: a chatbot has to identify itself as one, AI-generated content has to carry machine-readable marks, with exposure up to €15 million or three percent of worldwide turnover. Every voice and chat agent in this piece now operates inside that. Compliance cost is not evenly distributed either — it is a line item for Amazon and a project for everyone else.

The market is scoring the billing unit

Meanwhile the market has been quietly settling a different argument.

On the architecture diagram Salesforce holds the strongest position anywhere in the CX stack. The market took more than forty percent off it in the first half of 2026, after a twenty percent fall the year before.

That is not a verdict on the architecture. It is a verdict on the billing unit. Enterprise software price-to-sales multiples compressed from roughly 5.6 times at the end of 2025 to about 4.2 times by mid-March, with more than $280 billion wiped from software stocks in a single February session. Salesforce reported $1.2 billion of Agentforce ARR growing over 200 percent and got sold anyway, because if agents do the work you stop buying licences for the humans who used to.

Salesforce knows it, and did something about it. On 8 June it agreed to buy m3ter, a London metering and rating platform built for consumption-based billing, closing the deal on 1 July and folding it into Agentforce Revenue Management so enterprises can bill usage and outcome-based pricing natively. Salesforce’s own framing was blunt: most enterprises are still running billing systems built for seat-based pricing while their products have gone usage-heavy. That is a company trying to re-base its own unit of account before the market finishes marking it down.

It was not alone. Stripe absorbed Metronome. Adyen bought Orb for $335 million on 11 June. Three different kinds of acquirer — CRM, payments infrastructure, payments processing — concluded within weeks of each other that metering was worth owning rather than integrating. Nobody buys metering infrastructure in a market where the billing unit is settled.

Now look at Twilio, which the same market had written off two years ago. It reported Q2 on 6 August: record revenue of $1.5 billion, up 22 percent reported and 17 percent organic, earnings well ahead of consensus, record free cash flow — and full-year growth guidance raised from 14–15 percent to 18–18.5 percent. The stock jumped around 16 percent on the print.

Twilio did not out-architect Salesforce. Twilio bills for consumption. An agent that makes ten thousand calls is ten thousand calls. An agent that replaces a service rep is minus one seat. Management pointed at demand from enterprises and AI-native companies alike, which is the same thing said twice: both are buying volume, and agentic workloads generate volume rather than eliminating it.

In an agentic market, your exposure is set by your unit of account, not your position on the stack. Which is exactly why AWS could publish that flat session price and Meta could not keep it.

And nobody knows what any of it should cost

Now the honest part, because I have watched a lot of people in this industry pretend otherwise.

Nobody knows where the pricing lands. Not the CPaaS platforms, not the CCaaS incumbents, not the labs.

Everything downstream of a token price is unstable while the token price is unstable, and the token price is a moving target. Inference costs have fallen repeatedly and will fall again. Open weights are now a live commercial and political question rather than a research one: on 24 July, Jensen Huang made his first-ever post on X to share a letter from twenty-five companies arguing against restricting open-weight models, and the signatory list doubled to around fifty within a day — with Amazon and Anthropic conspicuously absent. The argument the letter makes is not primarily technical. It is about competition, control over your own data and deployments, and sovereignty.

That is the same argument this industry has been having about where inference runs, under whose jurisdiction, on whose hardware. It is now the centre of AI policy, and it will move the cost base of every conversational product on the market.

You cannot responsibly price the layer above a foundation that is repricing this fast. The research programme I run through CPaaSAA is digging into the revenue models — our network API study lands around our Amsterdam conference in September — and there are pricing specialists doing serious work on the question. But the jury is out, and anyone claiming otherwise is selling something.

The one CPaaS going the other way

Which makes Telnyx worth watching more closely than its size suggests, because it is the only significant player in this category that read the Jobs asymmetry correctly and went down instead of up.

Telnyx owns a carrier network across forty-plus countries. It also now owns its GPUs — a fleet in the thousands — colocated directly alongside its telephony points of presence, so speech recognition, model inference and speech synthesis run on the same network carrying the call. In July it lit up sovereign GPU capacity in Dubai, with inference and storage staying in-region. Its CEO’s argument is blunt and correct: training is a bounded capital event, inference is a permanent operating cost, and if you do not own the compute and the network under it, the unit economics will not work.

That is innovating down the stack from the one position where the Jobs advantage applies — you know exactly what you need, because you are your own most demanding customer. It is also the only version of this story where the sovereignty argument is backed by owned assets rather than a slide, and notably the only one where somebody can quote a firm per-minute price for an AI voice agent without flinching. Owning the stack is what makes the number sayable.

Three motions, one enterprise

Meanwhile the demand side is more crowded than any of these diagrams admit. Three different motions are converging on the same confused enterprise buyer.

The horizontal platforms — Salesforce, OpenAI, Infobip — sell a platform plus deployment engineers and integrators to make it real. The CCaaS incumbents are re-basing the seat licence onto the AI agent before the seats disappear. And the vertical point solutions attack from the opposite end: no platform, no legacy, one workflow in one industry solved end to end, built on top of somebody else’s CPaaS.

That third group is where the money is going. Gradium, out of Paris, closed a $100 million seed in July with Nvidia joining. ElevenLabs raised at $11 billion in February. Parloa took $350 million in January. None of them carry legacy revenue, channel conflict or a margin structure to defend — and all of them need numbers, termination and delivery from the layer underneath.

The enterprise, meanwhile, knows it needs AI, does not know where to start, and needs the result certified, secured and live in weeks rather than quarters. Everyone above is selling to that anxiety. Everyone below is supplying the plumbing for it.

So who can actually manoeuvre

Two things are true at once about Salesforce. It lost close to half its value in eighteen months, and it remains one of the largest and best-capitalised software companies on earth, with more strategic optionality than anyone else in this fight. A derating is not a death sentence. It is a change in the cost of capital.

A CPaaS platform with compressing gross margin, rising regulatory cost and a shrinking multiple does not have that latitude. It cannot buy its way out of very much. Sinch’s Q2 in July showed the squeeze precisely: six percent organic revenue growth against two percent organic gross profit growth, and an API platform where revenue rose nine percent while gross profit went slightly backwards. Growth in the right products, margin going the wrong way. And it is worth looking hard at what does get bought when the cheque is written: Infobip’s acquisition this year was email infrastructure. Profitable, sensible, immediately accretive — and firmly today’s business rather than tomorrow’s.

That is the third-door problem in its sharpest form. The choice was always buy up or be bought. What the first half of 2026 clarified is that the ability to exercise either option is not evenly distributed, and it is being redistributed by a market that is repricing everyone faster than any of them can reposition.

The moat moved. The pricing is unresolved. The capital is relocating. And the layer that turns out to be hardest to replicate is the one everybody else has decided to rent rather than buy.

Renting is a decision the renter gets to revisit — on their timetable, not yours.

There is a version of this industry that saw all of that coming, because it ran the same play itself. CPaaS was the outsider once. It looked at a wholesale product being sold by the unit, saw a business the incumbents could not defend without cannibalising themselves, and took the margin. Everything now being done to CPaaS is that move, repeated by people standing one layer higher with considerably more capital and nothing at all to lose.

Knowing how the play works is not the same as being able to stop it. But it is a much better starting position than being surprised by it.

Rob Kurver writes on cloud communications, AI and telecom transformation at thenextcloud.co.