The press release arrived with a strange symmetry.
The summary and the body were the same paragraph, repeated verbatim, as though the writer had run out of facts before running out of formatting. Phase โ a name I could not find in any founder database, audit registry, or grant ledger I had access to โ had "launched Proxy," a product that lets smartphones deliver AI connectivity. The stated purpose: lower the cost and barrier of artificial intelligence, accelerate its adoption, decentralize the infrastructure beneath it.
That was the whole text. No chain. No token. No team. No funding round. No benchmark. No audit. No date. Not even a device compatibility list, which is a peculiar omission for a product whose entire premise is the phone in your hand.
In 2017, auditing a reentrancy flaw for a DAO successor in Zurich, I learned that the most important part of a document is usually the part it declines to write. My report on that 500 ETH vulnerability was rejected by the frontend team for being "too academic." The lesson I took was not that engineers are philistines. It was that a document's silence is structured, and structure is information. An announcement this empty is not a failure of communication. It is a communication.
So let us read it the way I read code I do not trust: slowly, and against the grain.
The Long Line of Borrowed Hardware
The idea that ordinary people's devices can be welded into a compute network is not new. It is, in fact, one of the oldest promises in the industry, and its graveyard is well-marked.
Golem shipped in 2016 with a thesis almost identical in shape: idle CPUs everywhere, a market to match supply with demand, a token to price the clearing. It was elegant. It largely did not happen, not because the cryptography failed but because the unit of work was never fungible enough to price. A CPU-minute is not a CPU-minute when every machine is a different architecture running a different thermal envelope under a different operator's tolerance for noise.
Filecoin solved a version of this by making the unit brutally simple โ a sealed sector, a proof, a reward. Helium solved it by making the unit physical and auditable โ a radio that either witnessed a packet or did not. Render solved it by narrowing to one workload with one measurable output โ frames. The survivors in this lineage all share a trait: their output could be verified without asking the supplier to be honest.
Then came the 2023โ2025 convergence, when DePIN and AI fused into a single narrative. io.net aggregated GPUs. Akash sold container hours. Bittensor turned model training into a market. The vocabulary shifted from "storage" to "intelligence," and the valuations followed the vocabulary, as they always do.
Each cycle the resource changes. CPU, then disk, then bandwidth, then GPU, and now โ apparently โ the phone. What never changes is the two questions that decide whether the network is real: how do you verify the work, and what does the work actually cost the person supplying it.
Phase's announcement answers neither. But the vocabulary it chose answers something else.
Four Architectures, One Word
"Smartphones deliver AI connectivity" can describe at least four mutually exclusive systems, each with a different cost structure, a different trust model, and a different regulator.
One: the thin-client proxy. The phone is a routing endpoint. It forwards a request to remote inference and returns the answer. The phone contributes zero compute. What is being sold is access, packaging, and UX โ a brokerage between AI demand and AI supply. This is the reading the product's own name supports.
Two: distributed inference. A model is sharded across many devices, each computing a slice. For autoregressive language models this is close to unworkable โ the intermediate activations between layers are large, the round-trip latency across consumer mobile networks is unpredictable, and the failure of one node stalls the chain. It becomes plausible only for embarrassingly parallel workloads: embedding generation, image transforms, batch classification.
Three: on-device inference with a network layer. The phone runs the model locally; the network distributes and versions models, meters usage, and settles payment. Compute stays on the device. This is the most technically honest version and the least dramatic.
Four: federated contribution. Devices contribute data or gradient updates rather than inference. Different economics entirely, and a privacy regime that would make most general counsel reach for the door.
The announcement collapses all four into a single phrase. That is not a small ambiguity. It is the difference between a CDN and a data processor, between a fee business and a securities question.
So let me take the charitable technical reading โ the one where phones genuinely do inference โ and price it honestly, because the arithmetic is where this narrative always dies.
The Model Has to Fit in the Palm
A modern flagship phone carries 8 to 12 gigabytes of RAM, of which perhaps five to six are usable after the operating system takes its cut. A seven-billion-parameter model at 4-bit quantization occupies roughly 3.5 GB for the weights alone, before the KV cache, before the runtime, before the browser tab that the user refuses to close.
This means a phone can comfortably host small models โ one to three billion parameters โ and cannot host frontier models at all. Quantization and distillation, the two techniques that make mobile inference viable, are also the two techniques that strip capability. A 3B model does classification, retrieval, embedding, speech, summarization, and on-device personalization beautifully. It does not do reasoning at the level that justifies the phrase "democratizing AI."
That gap matters, because the value proposition of a smartphone compute network is not general intelligence. It is privacy-preserving, low-latency, small-model inference โ a real market, a growing market, and a market roughly one order of magnitude smaller than the one the token narrative usually sells.
Which brings us to the part of the deck that never appears in the deck.
The Battery Is the Ledger
A phone's battery holds around 17 watt-hours. Sustained AI inference on a modern system-on-chip draws three to five watts, and thermal throttling arrives within minutes because a phone has no fan โ only a glass back and the ambient temperature of a pocket.
Run three hours of sustained inference and you have consumed roughly 12 watt-hours. That is about seventy percent of a full charge, every day, for a device whose owner did not buy it to be a server.
The electricity is not the problem. At residential rates, 12 watt-hours costs about a third of a cent. The problem is cycle life. A lithium-ion cell holds its rated capacity for perhaps 500 to 800 full cycles before degrading to eighty percent. Seventy percent of a charge is 0.7 of a cycle. Do that daily for a year and you have spent roughly 255 cycles on somebody else's inference workload. Do it for two years and you have consumed the battery's designed life and are looking at a replacement bill of sixty to a hundred dollars, or a phone that dies at four in the afternoon.
Amortized, a participating device is donating something like thirty to forty-five dollars a year in accelerated hardware depreciation. That is the true marginal cost of the node, and it never appears in any DePIN model I have read.
Which produces the sentence I want on the record: decentralized inference does not eliminate the cost of AI; it relocates that cost into the battery cycle of a stranger.
If the network pays the device owner less than the depreciation it causes โ and at token prices set by emission schedules rather than by demand, it almost certainly will โ then the network is not cheaper than the cloud. It is cheaper to the buyer because the subsidy is being extracted from hardware that the supplier does not account for on a balance sheet. That is not efficiency. That is an unpriced externality wearing a token.
I modeled a version of this in 2020, in a paper on yield farming that predicted token incentives would centralize governance rather than distribute it. Fifty thousand people read it. The market ignored it until the crash. The curve here is the same curve, drawn on different axes: emission-subsidized supply floods in, the unit price of the resource collapses, only mercenary nodes remain, quality degrades, and the demand that was supposed to arrive never does because it was never the point.
Who Verifies the Phone
Now the harder question, the one that separates a whitepaper from a network.
A phone completes an inference task and returns a result. How do you know the work was done correctly, on the claimed hardware, at the claimed cost?
You have three options. Redundant execution โ run the job on two or three devices and compare, which triples the cost and halves the margin. Optimistic verification โ accept the result, allow a challenge window, require challengers to re-execute, which demands stake and demands that a phone be able to reproduce the computation deterministically, which floating-point arithmetic across heterogeneous chips does not generously allow. Or hardware attestation โ have the device prove its identity and its integrity through a trusted execution environment.
The third option is the only one that works at mobile scale. And on mobile, the trusted execution environments are Apple's Secure Enclave, Android's StrongBox, and ARM's TrustZone. There are approximately three vendors on earth.
So the trust anchor of a "decentralized" smartphone compute network is a certificate chain rooted in two or three corporate enclaves, governed by terms of service that the network operator does not write and cannot amend. That is a legitimate architecture. It is also the precise opposite of the sovereignty story that gets told at conferences.
I have spent enough time inside governance forums to have a private taxonomy for this. Protocols that preach decentralization while their root of trust sits in a vendor's firmware are not lying, exactly. They are performing a kind of institutional theatre in which the word does the work the architecture cannot. Identity is a protocol; soul is the private key โ and when the private key lives inside a vendor-controlled enclave, the soul is on loan.
There is a further operational truth that any node operator learns within a month. Phones leave Wi-Fi. Phones thermal-throttle. Phones get picked up, pocketed, and put on a charger at 2 a.m. Node churn in a mobile network is not ten percent; it is a continuous, adversarial, geographically correlated churn. Every distributed job must be checkpointed against the assumption that thirty percent of its workers will vanish mid-computation. That overhead is real, it is paid in latency, and it is the reason Helium's coverage maps were always more beautiful than its throughput.
None of this makes the idea wrong. It makes the idea narrow. There are workloads where this works: batch embedding jobs, image preprocessing, on-device personalization where the data never leaves the handset, latency-tolerant classification at the edge of a network that has poor backhaul. Those are real. They are also unglamorous, and unglamorous does not raise a round in a bull market.
The Name Is the Confession
Here is where I stop being generous.
A proxy is not a participant. A proxy is a stand-in โ something that acts on your behalf while you are absent. In networking, a proxy forwards without understanding. In law, a proxy votes with delegated authority and carries a fiduciary shadow. In finance, a proxy is a document that discloses who really controls the thing you think you own.
Phase named its product after the act of substitution. If the architecture were compute, the vocabulary would have been a grid, a fabric, a mesh, a pool โ the nouns of supply. Instead we got the noun of brokerage.
The product is a routing relationship, not an infrastructure revolution. That is not a criticism. Brokerages are excellent businesses with real margins. But a brokerage regulated as an intermediary has obligations โ data processing duties, consumer protection duties, anti-money-laundering duties โ that a "decentralized physical infrastructure network" is structurally designed to disclaim.
And there is a second confession buried in the silence. In a bull market, a launch with no team, no token, and no funding is either extraordinary discipline or deliberate sequencing. My experience runs toward the second. The teams I have seen do real work tend to over-describe, because they are proud of the difficult parts and because the difficult parts are the only things they can prove.
What arrives after an announcement like this is predictable in structure if not in timing: a node app, a points system, a waitlist, an airdrop rumour, and finally a token whose emissions schedule is calibrated to the narrative rather than to demand. Watch for the points system. Points are the polite fiction a project uses when it wants token economics without token disclosure.
When the pool empties, only the intent remains. Strip the vocabulary away and the intent here is legible: convert the world's most abundant, most personally intimate, most poorly accounted-for hardware into a supply curve, then sell access to that supply curve before anyone asks what it costs the supplier.
The most contrarian reading I can offer is also the kindest. Perhaps Phase genuinely intends only architecture number one โ a thin-client proxy that never touches the user's silicon, never needs attestation, never touches the battery. A modest, honest, profitable routing layer. If so, the failure is not technical. It is that a modest, honest routing layer would not have been covered, would not have been aggregated, and would not have been read by anyone. The inflation is in the telling, not in the code โ and that is the more common sin.
What to Watch, and What Would Change My Mind
Ignore the token. Tokens are lagging indicators of narrative, not leading indicators of capability.
Watch the telemetry, because telemetry is the only thing a project cannot fake for long. Three falsifiable signals, in order of evidentiary weight.
First, sustained-power and thermal benchmarks disaggregated by device class. Not peak TOPS. Sustained watts over ninety minutes, with throttle curves, on a mid-range Android phone that a real person in a real market actually owns. Any team that has done the engineering has these numbers. Any team that has not will describe "efficiency" instead.
Second, a named verification mechanism with the trust anchor disclosed in plain language. If the answer is hardware attestation, say so and say whose enclave. If the answer is redundant execution, publish the multiplier and the resulting unit economics. Silence here is not caution; it is an admission that the question has not been solved.
Third, a unit-economics disclosure that compares payout per device-day against modelled battery depreciation. This is the disclosure no DePIN project makes, and it is the one that would prove the network is a business rather than a subsidy. If Phase publishes it, I will write about them again โ differently.
If none of the three appears within six months, the narrative will have outlived the product, which in this industry is not a scandal but a business model.
The audit is not a check; it is a confession. The question was never whether a phone can run a small model. It can. The question is who pays for the battery, and whether the answer is written down anywhere that a supplier could read before saying yes.
We have spent a decade tokenizing resources people did not know they were selling. The next narrative will not be about intelligence on chain. It will be about the invoice โ the one sent to a stranger's hardware, in a currency of cycles and heat, arriving silently in the second year of ownership. Somebody is going to build the disclosure standard for that. I would like to know who, before the market decides it does not want one.