Last week I did something I hadn't done since 2020: I sat at my kitchen table at one in the morning and divided a very large number by a very small one. The large number was $1.7 billion โ the annual cloud cost Amazon is reportedly absorbing for Alexa+. The small number was 1.6 cents, my back-of-envelope estimate of what a single voice exchange costs to run. I wanted to know how many conversations that buys.
The answer โ roughly 100 billion a year, or 280 million a day โ told me something the headline didn't. It told me the product is not being used heavily. It's being used broadly, thinly, and expensively. That is a completely different diagnosis from "AI is too expensive," and I recognised it instantly, because I have shipped that exact product before. In 2020 I put my whole savings into an unaudited yield farm and watched it drain in forty-eight hours. The failure wasn't the exploit. The failure was that nobody, me included, had ever divided the cost by the usage.
The facts, as far as they can be established: Alexa+ is free for Prime members at $139 a year, and $19.99 a month for everyone else. The reported $1.7 billion is a cloud cost line, and inference is served largely through Amazon Bedrock, which means Anthropic's Claude models sit somewhere in the stack. Moonraker, the agentic browser project, is reportedly being considered for scale-back. There is no named source, no year-over-year baseline, no active user count, and โ this matters enormously โ no revenue side at all. No subscription conversion rate, no voice commerce commission, no advertising attach.
So I did what I always do with incomplete inputs: I treated the number as a signal node rather than a fact. The signal is not that Amazon has a cost problem. The signal is that the largest cloud provider on earth has now put a specific dollar figure on its tolerance for subsidising a consumer generative AI product. That sentence has never been written down in public before. It has enormous implications for an industry โ mine โ that has been subsidising consumer products with token emissions for four years and calling it growth.
Here's the arithmetic, and I'll flag every assumption, because the assumptions are the argument. A typical Alexa+ exchange โ system prompt, a few turns of context, the ASR transcript, a tool response โ runs about 3,000 input tokens and 500 output tokens. At mid-tier retail pricing of $3 per million input and $15 per million output, that's roughly 1.6 cents per exchange. Divide $1.7 billion by 1.6 cents and you get about 106 billion exchanges a year, or 290 million a day.
Now the interesting part. If Alexa+ has 100 million monthly actives, that's 2.9 exchanges per user per day. If it has 50 million, it's 5.8. Either way, this is not a heavy-use assistant. The $1.7 billion is not buying depth of engagement; it is buying breadth of coverage. Amazon is paying to keep hundreds of millions of lightly-touched devices inside its voice perimeter, not to serve power users. The per-user cost is brutal at the low end: at 100 million actives, each user burns $17 a year, or 12% of a Prime membership. At 50 million, it's $34, or 24%. That's the number I would want on my desk before deciding anything.
Voice makes this structurally worse than chat, and almost nobody prices it correctly. The pipeline is ASR, then LLM, then TTS, serialised, inside a sub-second latency budget. That budget forbids large batch scheduling โ precisely the mechanism that makes text inference cheap. So you pay a premium per token for a request whose informational value is lower. "Turn on the lamp" does not deserve 1.6 cents. The voice assistant is the only consumer AI product where the unit cost is high and the unit value is near zero at the same time.
Which brings me to the question I actually care about: is there a routing layer? If Amazon has built intent classification that sends simple requests to a small on-device model or to Nova Micro and escalates only genuine conversation to Claude, then $1.7 billion is buying a defensible experience. If every request hits the same mid-tier model, routing is missing and there is 50% to 70% of headroom sitting on the table. The report doesn't say. But a cost figure that large, leaked in that shape, is weak evidence that routing is immature โ you don't leak a number that big if you've already solved the easy half of the problem.
The silicon lever is the one that moves the decimal point. Amazon is the only hyperscaler shipping both a training chip and an inference chip at scale โ Trainium2 and Inferentia2. Industry experience puts custom inference silicon at 40% to 60% below comparable GPU cost once you account for energy and the absence of a margin stack. Move half of Alexa+ inference onto Inferentia and $1.7 billion becomes $1.0 to $1.2 billion without touching a single feature. Which reframes the Moonraker rumour: this looks less like retreat and more like buying eighteen months of budget discipline while the architecture catches up.
Moonraker itself deserves a colder read than it's getting. Agentic browser tasks โ multi-step planning, tool calls, long-context maintenance โ have end-to-end success rates somewhere in the 30% to 60% band, and success decays roughly exponentially with step count. That is not a consumer-grade reliability number. It is entirely possible the project is being slowed because it doesn't work well enough, not because it costs too much, and those two explanations carry opposite strategic meanings. One says Amazon is cheap. The other says Amazon is honest.
Now let me be blunt about why a crypto person should care, because the parallel is not decorative. Prime bundling and token subsidisation are the same instrument. Both mask negative unit economics behind an asset the user already holds. Amazon hides inference cost inside a $139 membership; a DePIN inference network hides it inside an emissions schedule. In both cases the product looks free, the growth looks organic, and the invoice is paid by someone who never sees it. We didn't build business models in the last cycle. We built subsidies with tickers attached and called the TVL a moat.
Truth in blockchain isn't in the throughput chart. It's in the invoice. When I audited protocol treasuries in 2022, the number I always asked for โ and almost never got โ was fully-loaded cost per unit of real usage, denominated in something other than the protocol's own token. Amazon just published theirs by accident: $1.7 billion for roughly 100 billion units. That's a benchmark every decentralised compute network should be forced to answer against, and I suspect most would rather not.
Here's where I'll break with the consensus take that big tech's cost discipline is a tailwind for decentralised AI. I think it's the opposite. The moment centralised inference gets 50% to 70% cheaper through custom silicon and routing, a token-subsidised network running at 30% utilisation cannot compete on price โ and price is the only axis most of them compete on. The DePIN inference projects that survive won't be the ones with the best emissions curve. They'll be the ones holding a cost advantage that exists for a physical reason: stranded energy, idle GPU arbitrage, data that legally cannot leave a jurisdiction, or a censorship-resistance requirement someone will actually pay for.
There's a second-order point the coverage missed entirely. Cutting a browser agent is a security improvement, not just a cost one. An agent that clicks, fills forms and moves money converts a prompt injection from an embarrassing output into a drained account. If Moonraker is paused, the attack surface pauses with it โ and nobody writing about this has framed the decision that way, because the cost narrative is more legible than the risk narrative.
So the number I'll be tracking isn't $1.7 billion. It's the ratio between inference cost and real usage, published by anyone, in a currency they didn't mint. When a decentralised network can state that number and defend it, the conversation changes. Until then, we're all just dividing big numbers by small ones at one in the morning, hoping the answer looks like a business.