The Voice AI Stack:
what's commoditized,
and what isn't

Every layer of the voice stack is being commoditized from the bottom up. The question worth asking is not whether a layer will fall, but how far the water has already risen — because that line tells you exactly where it is still rational to spend engineering effort, and where you are paying rent on something that has already become a dropdown.

As of July 2026 the waterline sits precisely on turn-taking. Below it — carriage, transport, speech recognition, the good-enough language model, orchestration, baseline speech synthesis — you find three or more credible vendors priced within a whisker of each other, an open-source floor underneath them, and switching costs measured in dropdown clicks. Above it — expressive voice, evaluation harnesses, production ops, regulated telephony, domain depth — you find price set by value rather than cost, and switching measured in quarters.

Two things follow immediately. First, the layers below the line are where a buyer should be ruthless about price and a builder should refuse to spend a single sprint. Second, and less obvious: the line is moving upward at a knowable rate, which means a moat you are building today at layer n has a shelf life you can estimate. §10 puts dates on it.

The ordering above is not arbitrary and it is not simply "cheap things first." Look at what the fallen layers have in common and a single property does most of the explanatory work:

Run the stack against it. Speech recognition has word error rate: a single number, publicly leaderboarded, reproducible on shared audio. It commoditized hardest and fastest. Text-to-speech has mean opinion score, which is a survey, not a measurement — so TTS split cleanly in two: a commodity floor where "clear and intelligible" is effectively binary, and a durable premium tier for expressiveness, where the metric is still a human shrug. Turn-taking only recently acquired numbers — false-barge-in rate, endpoint latency, semantic end-of-turn accuracy — and it is commoditizing right now, in public, at exactly the pace the rule predicts. Evaluation has no agreed scalar at all: nobody can hand you a number for "was this call good" that a second vendor would reproduce. It is the most expensive add-on on any price list. And compliance is not measured, it is attested — audits and contracts, not benchmarks — which is why it never rides a cost curve at all.

This is my framework rather than a vendor's, so treat it as a hypothesis with a prediction attached. The prediction is uncomfortably specific and appears in §10: the eval layer's price collapses within months of the first credible public benchmark for conversational quality, not gradually before it. If you want to know when your eval moat expires, watch for the leaderboard, not the competitors.

"Commoditized" is usually asserted. It can be measured. If a layer has truly cleared, the credible suppliers will have converged: the ratio between the cheapest and dearest vendor doing the same job collapses toward one. If a layer hasn't cleared, that ratio stays wide, because buyers are still paying for something price alone doesn't capture.

The obstacle is that these layers are quoted in incompatible units — per minute, per hour, per million characters, per million tokens. So everything below is normalized to one dollar figure per minute of live conversation M, using a single consistent model of what a minute contains: four assistant turns, roughly 2,000 input tokens and 60 output tokens per turn, and about 265 characters of synthesized speech (an agent that holds roughly a third of a 150-words-per-minute conversation). Those assumptions are stated so you can break them; §06 shows where they bend.

The chart makes the argument better than the prose can. The three tightest bands — telephony at 1.8×, baseline speech synthesis at 2.3×, English speech recognition at 3.0× — are exactly the layers below the waterline. The three widest — orchestration at 5.5× above a free floor, speech-to-speech at 7.1×, and the full range of speech synthesis at 11× — are exactly the layers where the argument about value is still live.

That inversion is worth sitting with, because it generalizes. Platform margin does not live where prices look high. It lives where a buyer has stopped comparing — and buyers stop comparing precisely once a layer feels commoditized. The moment you accept that a layer is a commodity is the moment you stop auditing its price.

One minute of inbound support, built three ways, every figure drawn from list prices fetched on the same day. The first column is what the components cost. The second and third are what you can buy the same minute for.

Cross-check the middle bar against what the market actually charges for a whole bundled minute and it holds: an Indian platform sells an all-inclusive connected minute at ₹6 (about $0.10), another lands between $0.07 and $0.15, a third publishes a headline band of $0.07–0.31, and an enterprise contract bottoms out around $0.07 at volume O. The market clears at roughly ten cents against a two-cent component floor.

There is a second reading of the same chart that matters more for anyone selling minutes. Notice that a whole bundled retail minute at $0.10 costs less than two and a half times the premium voice line alone in the third bar. Platform pricing is a packaging decision far more than it is a cost floor — which is exactly why it is vulnerable.

Everything above rests on these numbers, so here they are unconverted, in the units the vendors publish them in. Every row is a list price from the vendor's own pricing page on a single day.

Two forces keep the commodity layers from collapsing all the way to zero. Both are structural rather than commercial, which means no amount of vendor competition removes them.

The context tax — and an honest disagreement with my own model

My model puts a good-enough brain at roughly $0.0009 per minute. One platform's own price list puts the same class of model at $0.081 per minute — about six times higher P. I am not going to paper over that gap, because both readings are actionable and I cannot yet distinguish them from public data.

Either platforms mark the commodity brain up several-fold, or a real production voice context is three to six times heavier than a naive estimate — long system prompts, retrieved knowledge-base chunks, and a full transcript re-sent on every single turn. My instinct is that the truth is mostly the second: a voice agent re-uploads its entire conversation with each reply, so token consumption grows quadratically with call length while the conversation only grows linearly. Either way the lesson is identical and it is the single most useful operational takeaway in this piece:

The audio tax — visible in the vendors' own price lists

Speech-to-speech models promise to delete the pipeline: no separate recognition, no separate synthesis, no seam. The reason they cannot simply undercut a cascaded stack is printed in the price list of the company that makes both.

Stack that tax on top of the fact that a speech-to-speech model re-sends accumulating audio history and the economics resolve cleanly. The cheapest realtime audio model lands at about $0.018/min — genuine parity with a cascaded pipeline. The most expensive lands at about $0.128/min, roughly ten times a cascaded build and a seven-fold gap between two suppliers of the same capability M. One vendor has collapsed speech-to-speech pricing sevenfold below the other; that is price competition arriving, not a commodity forming.

So the prediction here is narrow and, I think, safe: speech-to-speech wins on naturalness well before it wins on cost, and cascaded pipelines stay the production default for anything touching money or compliance — because a cascaded stack hands you an inspectable transcript at every hop, and a native audio model does not. Expect blended architectures, not a switchover.

Here is the most useful thing on any voice platform's pricing page: the add-ons. Each one is a line item the vendor believes it can charge for separately, which makes the price list a direct readout of what the market still treats as scarce. Measured against the two-cent commodity minute, the readout is stark.

Read the top bar again. Evaluating a call is priced at nearly five times the cost of running it, and roughly twice the platform's own voice-engine fee. In almost every other part of software, testing is the cheap part — you run the test suite a thousand times because each run costs nothing. In voice AI the relationship is inverted, and that inversion is the single clearest statement anyone in this industry has made about where the scarcity is.

The rest of the margin is sold as rent rather than usage, which tells you it is access being priced, not compute:

Strip out everything the price sheet has settled, and four things are left standing. None of them is a model. Two of them are not even technology.

Notice what these four have in common: not one of them gets cheaper when inference gets cheaper. Three of them get more valuable, because a world of cheap capable agents is a world with more calls to evaluate, more regulations to satisfy, and more systems of record to write into.

There is a failure mode I keep running into when large product organizations build voice agents internally, and it is the exact inverse of what you would expect. In-house teams are usually accused of reinventing infrastructure. What they actually do is subtler and more expensive: they build the un-commoditized layers rather well, and then dramatically under-buy the commodity ones.

The archetype looks like this. The team has invested real effort in a scenario harness — synthetic test audio, an adversarial test model, generated conversation scenarios — and in an operations cockpit that beats commercial platforms outright on the parts that matter after launch: product insights, repeat-caller detection, per-line-of-business summarization. Those are precisely the layers above the waterline. That is money spent correctly.

And then the commodity layers sit frozen at whatever was wired in during the first sprint:

The trap, stated plainly: a fixed default at a commodity layer feels like a decision that was made. It usually isn't — it is a decision that was never revisited after the layer beneath it collapsed in price. The audit is worth running annually, and it takes an afternoon.

A commoditization map is only worth writing if it makes claims that can later be checked and found wrong. So here is the waterline as a moving object, with dates attached and — more importantly — with the specific observation that would prove each prediction false.

A price sheet dated to a single day is only useful if it is equally explicit about what isn't on it. The following was not verified in this pass and is deliberately absent rather than guessed at. Several of these are load-bearing.