Twenty dollars buys 66.7 million tokens. Another eighty buys 2.5 billion. Same platform, same model family, same billing page. The marginal price of a token falls by a factor of 9.4 between two adjacent rows of a table that most coverage read as a straightforward tier ladder.
That asymmetry is not a rounding artifact. It is the entire architecture of the home AI agent market in late 2026, compressed into one column of arithmetic. Read the curve and you learn what the product actually is. Read the headline and you learn what the marketing department wants you to believe it is.
Three events closed the free-tier era inside a five-week window. On October 12, 2026, xAI gated GrokBot's agent functionality behind X Subscription Premium at $30 per month, leaving only a stripped base model accessible without payment. On October 13, Apple announced HomeView at $299 and buried the consequential clause in the footnotes: server-side Apple Intelligence features carry daily usage limits, and expanded access will be available for a fee in the future. Meta, running since September, never pretended otherwise — Meta Muse meters tokens on a published curve, from a 100-million-per-week free ceiling up to $100 per month for 3 billion.
The consensus reading is that compute became too expensive to give away. That reading is incomplete, and the gap between it and the actual mechanics is where the money is.
Start with the curve itself, because the curve is the confession.
Meta's free tier allows 100 million tokens per week. Annualized, that is 5.2 billion tokens. Monthly, it is 433.3 million. The $20 tier allows 500 million per month. The incremental capacity purchased by that first $20 is 66.7 million tokens — a marginal rate of $0.30 per million. The next $80 purchases 2.5 billion tokens, a marginal rate of $0.032 per million. The ratio between the two marginal prices is 9.4x.
The first paid tier is not a capacity purchase. It is a toll on functionality. A user who upgrades from free to $20 buys a 15.4% increase in the token ceiling and a 100% increase in what the platform is willing to let them do. The tokens are not the product. The removal of the throttle is the product. The tokens are the unit of account the throttle is denominated in.
This matters because it inverts the standard unit-economics story. If the $20 tier were priced against compute cost, the marginal rate would flatten or decline gradually across the ladder. Instead it cliffs downward by nearly an order of magnitude at the second step. That shape is not a cost curve. It is a conversion funnel with a price tag bolted on, calibrated to extract the maximum from users who have already demonstrated intent by clicking upgrade.
Now price the compute honestly, because the cap only makes sense against a real cost floor.
A 70B-class dense model served with continuous batching on an 8xH100 node runs roughly $2.50 per GPU-hour at prevailing cloud rates, or $20 per node-hour. Aggregate throughput at high batch sizes lands in the neighborhood of 15,000 output tokens per second, which is 54 million tokens per hour. Divide the node cost by the throughput and the gross compute cost is approximately $0.37 per million output tokens. That figure is a back-of-envelope estimate and it moves with quantization, caching, context length, and model architecture, but it is the right order of magnitude.
Run the Muse ceiling against that number. 500 million tokens at $0.37 per million is $185 of gross compute against $20 of revenue. Full consumption of the published cap produces a negative gross margin of roughly 825%.
A cap that would bankrupt the seller if every buyer reached it is not a capacity guarantee. It is a ceiling priced for the tail and marketed to the median.
So what does the median look like? An agent workload is not a chat turn. A single task involving tool calls, multi-step reasoning, and retry loops burns 50,000 to 200,000 tokens. A light user running five such tasks per day at 80,000 tokens each consumes 400,000 tokens daily, or roughly 12 million per month. At $0.37 per million, that is $4.44 of gross compute against a $20 subscription — a 78% gross margin.
The published 500-million ceiling sits at 41x median consumption. The free tier's 433-million ceiling sits at 36x. Neither number describes expected usage. Both describe a marketing surface area large enough to make the subscription feel like a bargain while the fair-use clause, written in a document nobody reads, absorbs the variance.
That structure — a generous-looking cap, a heavy-tailed distribution, and a discretionary enforcement mechanism — is an insurance contract. The platform collects a premium from the median and pays out to the tail, and the tail is defined by terms the platform controls unilaterally. It is a sound business model. It is also a model in which the buyer cannot compute their own exposure, because the meter reading is never disclosed.
xAI's structure is more aggressive, and the aggressiveness is structural rather than numerical.
GrokBot's agent layer costs $30 per month, or $360 annually, and ships with no published token ceiling. That omission is the entire contract. A metered product with a published cap has a bounded liability. A flat-fee product with an unpublished cap has an unbounded one, because the definition of acceptable use sits entirely with the operator. An unpriced meter attached to a flat fee is not a feature. It is a reservation of rights.
The contrast with Meta is instructive and rarely drawn. Meta sells capacity — auditable, denominated, comparable across tiers. xAI sells access — discretionary, undifferentiated, and revocable by policy update. Capacity pricing competes on unit economics. Access pricing competes on switching cost. The second is far more durable, and far worse for the buyer.
Apple's structure is the most extractive of the three precisely because it is the quietest.
The $299 HomeView is a hardware sale with an embedded call option on future revenue. Apple has stated that daily usage limits apply to server-side features including Siri AI, photo editing, and Image Playground, and that expanded access will be available for a fee. No price has been disclosed. No date has been disclosed. What has been disclosed is the shape of the contract: a device purchased today, with full functionality contingent on a contract that has not yet been written.
This is an incomplete contract in the strict sense. The buyer pays the full consideration up front and receives a bundle whose most valuable components are subject to future unilateral pricing. Once the $299 is sunk, the switching cost is the hardware itself. The negotiating position at renewal time is structurally weaker than the position at purchase time, and the platform knows it.
Model the option value, illustratively. Ten million subscribers at $5 per month is $600 million in annual recurring revenue attached to hardware that has already been sold and depreciated. Apple does not need to announce the price to capture it. It needs only to wait until the installed base has no alternative.
Now stack the three together, because the compounding is the part most consumers have not modeled.
GrokBot at $30 and Muse at $20 is $50 per month, or $600 annually, before a single overage charge. Add HomeView hardware at $299 amortized across a three-year replacement cycle, and the effective annual cost rises by roughly $100. Add Apple's undisclosed fee at any plausible level and the conservative total lands between $700 and $900 per year per household, for a set of capabilities that were free, partially free, or bundled eighteen months ago.
Three rentiers, three meters, three sticky layers. The agent runtime bills monthly. The model API bills by consumption. The device ecosystem bills at purchase and again at the point of maximum lock-in. Each layer is designed to be individually tolerable and collectively inescapable.
This is the consumer-side mirror of the enterprise compute-landlord thesis, and the mirror is accurate. Enterprises are being squeezed by cloud cost growth that outpaces workload growth. Households are being positioned as tenants of their own living rooms — buying the hardware, renting the intelligence, and paying a metered rate on top. The business model is tidy. The residence is expensive.
Which brings the analysis to the question the crypto side of this market keeps answering incorrectly.
The reflexive response is decentralized compute. DePIN GPU markets, tokenized inference, spot capacity cleared on-chain. The pitch is that permissionless supply undercuts hyperscaler pricing and returns the margin to the user.
That pitch competes on the wrong unit. The consumer is not buying GPU-hours. The consumer is buying a bundle of deterministic first-token latency, tool-calling reliability, update cadence, and a support surface. Spot GPU markets deliver volatile pricing and variable availability — the precise inverse of what an agent runtime requires. An agent that stalls on a cold start is not a cheaper agent. It is a broken product.
The moat is not the GPU. The moat is the billing relationship and the device. Decentralized capacity can undercut the compute line item by 40% and still lose, because the compute line item is not what the customer is paying for.
Where blockchain actually fits is one layer up, and it is a much narrower wedge than the industry wants to admit.
The real defect in the Muse curve, the GrokBot omission, and the Apple footnote is not that compute is centralized. It is that the meter is unilateral. A subscriber cannot audit their own consumption. They cannot verify that 500 million tokens were consumed rather than 800 million, cannot inspect the counting methodology, cannot challenge a fair-use determination, cannot port a usage history to a competitor. Every one of those gaps is a settlement problem, not a compute problem, and settlement is the one thing a blockchain is genuinely good at.
Concretely: payment channels and stablecoin rails can settle sub-cent metered consumption without the 2.9% plus thirty cents that card networks extract from a $20 charge. Operator-signed consumption receipts can be published per billing cycle, giving the subscriber a verifiable record of what they were charged for and why. Escrowed prepayment can convert a discretionary flat fee into a bounded, refundable liability. None of that requires decentralized inference. All of it requires an auditable meter.
The obvious objection is verifiable inference — zero-knowledge proofs that the model ran as claimed. I spent four months in 2025 auditing the circuit design of a STARK-based ZK-Rollup for technical due diligence, and the binding constraint was proof generation time, not proof size. That was for a rollup circuit operating on transaction batches. An inference circuit for a 70B-parameter model with a 100,000-token context is several orders of magnitude heavier. Proving overhead at that scale is not a $20-per-month product. It is not a $200-per-month product either.
The viable design is attestation-of-metering rather than proof-of-inference. A trusted execution environment or an operator-signed receipt attests to consumption counts, settled on-chain for auditability. The guarantee is weaker — you trust the enclave, not the math — but the cost difference is measured in orders of magnitude, and the consumer actually needs the weaker guarantee. Nobody is disputing whether the model produced a paragraph. They are disputing whether they were billed for 500 million tokens or 800 million.
There is a familiar pattern here, and it is worth naming because the industry keeps repeating it.
Proof layers get built ahead of demand. In the rollup era, the argument was that every Layer 2 needed dedicated data availability infrastructure. The arithmetic never supported it — the overwhelming majority of rollups never generated enough data volume to justify a dedicated DA layer, and the ones that did were a short list. The infrastructure was revolutionary in design and underutilized in practice, because supply was built against a projected demand curve that arrived late and smaller than modeled.
Verifiable inference is on the same trajectory. Proving capacity is being built against an imagined future in which every agent interaction demands cryptographic verification. At consumer price points, that future is not near. What consumers actually demand is a receipt. The gap between a receipt and a proof is the gap between a product and a research program.
Which leaves local inference as the only structurally non-rentier option on the table, and it deserves an honest accounting rather than a dismissal.
An 8B-class open-weight model runs acceptably on a laptop with unified memory. A 30B-class model at aggressive quantization runs on a well-configured desktop. For a meaningful class of agent tasks — file organization, calendar management, document summarization, local search — the capability gap between a local model and a hosted frontier model is smaller than the price gap suggests. The local path costs electricity and nothing else, and it requires no token, no chain, and no subscription.
So the claim that there is no free path to a functional home agent is overstated. What closed in October 2026 was the free path to a maintained home agent. The subscription does not buy intelligence. It buys the update cadence, the tool-calling reliability, and the absence of operational burden. That distinction matters, because it reframes the entire pricing debate. The $20 and the $30 are not charges for model capability. They are charges for the labor of keeping the thing working.
And that reframe dissolves most of the revolutionary framing attached to decentralized alternatives. A permissionless inference network does not reduce the labor of maintaining an agent. It relocates it, and usually increases it, because now the operator is also responsible for node selection, availability monitoring, and payment channel management.
The counterintuitive conclusion is that the free tier was never a product.
It was price discovery. Across 2025 and 2026, every consumer AI platform ran a large-scale measurement exercise: give the agent away, instrument the usage distribution, find the median, find the tail, find the elasticity of conversion against the throttle. The October closures were not a cost event. They were the end of a measurement period. The pricing tables published in September and October are the output of that measurement, and the 9.4x marginal cliff inside Meta's ladder is what the data looked like when it came back.
This is the part that most post-mortems miss. The narrative is that compute got expensive. The evidence is that compute was always expensive and the platforms needed two years of telemetry to determine exactly where to place the meter. Once placed, the meter became the product.
For anyone building on the crypto side of this market, the implication is a hard reprioritization. Decentralized inference competes with hyperscalers on a unit the consumer does not purchase. Verifiable inference solves a trust problem the consumer does not have at a price the consumer will not pay. Auditable metering solves the problem the consumer actually has — an unauditable bill — at a cost that fits inside the existing subscription.
The question for 2027 is not which model is best, and it is not which network has the most GPU capacity. It is who audits the meter. If metering remains unilateral, the home agent becomes a utility with a discretionary rate schedule, and the household pays whatever the schedule says because the hardware is already on the wall. If metering becomes auditable, the subscription becomes a contract in the enforceable sense rather than the rhetorical one.
Apple has not announced its fee. xAI has not published its cap. Meta has published a curve whose first step costs nine times what its second step costs. None of those three facts is a coincidence, and all three resolve the same way: the meter is the contract, and right now only one party can read it.
