# Qubic: The Long Version (QC, Qubic Church)

## Qubic: The Long Version — Chapter 8. Murdoch’s Gas in Action — UPoW

Source: https://qubic.church/books/qubic-the-long-version/08-murdochs-gas-in-action-upow
Author: QC · Publisher: Qubic Church · 2026

### The Gas No One Collected into a Pipe

What, exactly, do six hundred and seventy-six computors do for the entire week? What work earns them their place in the next epoch — and why is it “useful” rather than, as in Bitcoin, arbitrary?

This chapter is the answer to that question. And it’s also a return to the very metaphor I carefully introduced in chapter one and promised one day to bring to bear. Let’s go back to Coalbrookdale of 1792.

William Murdoch is standing by a furnace in which coal is being charred for the production of coke. The process by then has been in use for eighty years already: since 1709, since the time of Abraham Darby, the iron-foundry industry of England has rested on this very transformation of coal into coke. From the furnaces, every day, dense black smoke pours out — a byproduct of the charring, inconvenient, sometimes dangerous, always released into the atmosphere as waste. That’s how it had been done all this time: coal yields coke, smoke flies into the sky, rain washes the soot off facades, and no one complains, because coke means cast iron and cast iron means the empire. Murdoch looks at the smoke differently. He runs a pipe from the furnace to a lamp in his own workshop, purifies the gas, lights it. It burns. It burns evenly and for a long time. It works.

Over the next fifteen years this idea will grow into the lighting of London with gas. By the 1820s an entire gas industry will have grown up — a new economy founded on the byproduct of an old one. Charring coal now produces two valuable outputs simultaneously. The first is heavy, tangible — coke. The second is invisible and powerful — gas.

In chapter one I wrote one phrase which has hung over the book ever since as a promise. I called Bitcoin mining the Murdoch’s gas of 2009 that no one has yet collected into a pipe. The process is running: computer cycles are being burned by the billions, power stations are operating, chips are heating up. For the sake of what? Of one thing — finding a digest whose first N bits are zeros. Found — the block is credited, the reward issued. All the rest of the computational byproduct gets thrown into the atmosphere. Goes off as heat. Into nothing.

In 2013 one engineer looked at this picture and saw exactly what Murdoch saw in 1792. Not “how to make a more efficient Bitcoin.” But “what if we collect the byproduct.” In 2022, on the basis of this idea, the Qubic network started up. By 2024 the UPoW mechanic, which this chapter is about, had stabilized.

Three questions to be answered. First: what exactly do Qubic miners solve every tick? Second: why does the result of this work simultaneously serve consensus and something else — and what, exactly, is that “else”? Third: does it work today as designed — or only halfway?

Let’s start with the mechanics.

### One Task, Several Outputs

A Bitcoin miner solves one task. He varies the value of the nonce field in the block header, computes sha256 of the header, looks at the resulting digest, checks: are the first N bits zeros? If not — change the nonce, compute again. If yes — the block is found, transmitted to the network, the reward taken. That’s it. The whole Bitcoin-mining industry, tens of billions of dollars in hardware, hundreds of terawatt-hours per year — is brute-force search for a coincidence: a digest with the required number of leading zeros.

What’s happening inside? Each sha256 computation passes through sixty-four rounds of internal transformations, and at each intermediate step some state results — bits, register states. By construction of the function, these intermediate states mean nothing outside sha256. They aren’t data, they aren’t a signal, they don’t contain any recoverable structure. They are computational byproduct that exists for exactly as long as it takes to produce the next one. The chip computes them, the chip forgets them. At the output — a digest. One output, one value, one purpose.

And this, it’s important to say, is not a bug. It is the right design of the function for its purpose. A hash function whose task is to be a cryptographically strong transformation must not leave useful traces in its intermediate states. If sha256 produced something useful in the course of computation, that would make it vulnerable: predictability of byproducts contradicts the requirement of indistinguishability from a random function. Bitcoin isn’t “badly made.” Bitcoin is made correctly — for what it does. But for what Qubic does, a function of a different design is needed.

Here is how the task is described by Come-from-Beyond — the creator of the system, with whom the reader has already met in chapter four. The quote is public, from his post of July 23, 2024:

“Qubic miner task is similar to Bitcoin hashing and can be reduced to finding such digest that first N bits are zeros, we set the difficulty manually (dev team chooses it every epoch) by deciding the value of N. Each computor simulates the ANN (which we have represented as a hash function during the reduction above) and generates the 'digest’ which is verified to have first N bits set to zeros.”

Read the phrase slowly. Qubic miner task is similar to Bitcoin hashing — the Qubic miner’s task is similar to Bitcoin hashing. And can be reduced to finding such digest that first N bits are zeros — and reduces to searching for a digest whose first N bits are zeros. The same task. Not similar, not analogous — formally reducible to the same one. But further: Each computor simulates the ANN (which we have represented as a hash function during the reduction above). Each computor simulates an artificial neural network — which through a formal reduction is represented as a hash function. This is the structural difference. The digest whose first N bits are zeros is generated not by sha256 but by running data through a neural net. At the input — some parameters, inside — a forward pass through layers of the ANN (matrix multiplications, nonlinear activations, aggregations), at the output — a sequence of bits which is checked for leading zeros.

I’ll underline: this is not an analogy and not a metaphor. The quote says explicitly: “which we have represented as a hash function during the reduction above” — represented as a hash function through a formal reduction. That is, it is mathematically established that the run through the ANN behaves, in the relevant properties, as a hash function, and therefore the task “find an input for which the first N output bits are zeros” has the same meaning as in Bitcoin. The difference is that the very process producing the digest now leaves traces.

The intermediate states of the ANN simulation — neuron activations, weight distributions on trained layers, the structure of how the data passed through the network — these are not empty bits inside sha256. These are data which, in the most honest sense, are neural-network data. They can be used for training. For evolutionary selection of architectures. For everything that makes up modern machine learning.

One bit of work by a Qubic miner produces three outputs simultaneously. The first output is consensus proof. A digest found with the first N zeros is a confirmation that the computor performed a valid simulation. This is the equivalent of “found a block” in Bitcoin: a proof of work which the network can verify and credit. On this output the certification of ticks is built. This is what makes UPoW a proof-of-work in the strict sense — there is work, there is its proof, the work cannot be faked more cheaply than performed.

The second output is AI training data. The intermediate states of the ANN simulation. The concrete values on which the work was done. This data is a valuable byproduct, reserved in the project for a consumer named Aigarth — a separate component of the system that uses the obtained results for evolutionary selection of neural-network architectures. I’ll talk about Aigarth in detail only in chapter eleven; here one thing matters: UPoW produces the data, Aigarth consumes it. As Come-from-Beyond himself put it on May 13, 2024: “Miners do useful PoW, not Aigarth”. Producer and consumer are two different nodes, and the separation between them needs to be kept in mind.

The third output is score. Score is the structured memory of the network which, by the results of the epoch, determines who stays, who leaves, who arrives. The content of the score can now be unfolded.

Score is the aggregated number of valid digests found by a computor over an epoch. Each digest is one successfully performed ANN simulation, each simulation is one unit of useful work. The epoch’s score is the sum of these units over the week. The network doesn’t just count votes or ticks; the network counts work of a specific type. The more valid simulations a computor has done, the higher its score. The higher the score, the more confidently it stays among the 676 for the next week, not falling into the 225 rotation slots.

What in chapter seven was called “the network’s memory” has now received content. The memory is a tally of how many times in a week you’ve produced precisely the kind of work the network considers grounds for keeping you in your place. Not money paid, not proof of stake, not social reputation — but a count of valid neural-network simulations. Architecturally this too is a rarely encountered solution: “the one who stays in the game is the one who has done more work by its rules.”

Now we can come back to the parallel that hung over chapter one. Bitcoin is the Darby phase. Coal burns, coke is given, smoke goes off. Compute is burned, consensus is given, everything else gets thrown out. The scheme is complete and working; it has no hole — it just has one output. Qubic is the Murdoch phase. The same process rebuilt so that the same input produces several outputs at once. Not “an improved Bitcoin PoW” — a different industrial regime of the same computational combustion. Just as coking in 1792 was not “better burning of coal” but a phase from which the chemistry of byproducts began in England.

In 1792 no one in Coalbrookdale thought “ugh, what bad coking, before we were doing it wrong.” Coking was being done correctly even before that — for the sake of coke, which is needed for cast iron, which is needed for the empire. Murdoch simply saw that the process had a potential second output that until then had been ignored. Bitcoin PoW is analogous: it does precisely what it is required to do; its function by construction has no useful byproducts, and that is the right design for its task. UPoW isn’t better. UPoW is about something else.

And here I have to pause. Before unfolding what “the system has three outputs, but one of them is so far not used at full strength” means, strict order needs to be brought to one place where the community regularly gets confused. Three different things that get called by similar words but do different jobs need to be separated. This is a short section, but without it the next, central section of the chapter won’t work as intended. Three Layers of Work: What Makes Qubic Qubic A reader familiar with crypto may quite reasonably think: “useful PoW is no news. Flux rents out compute for scientific computations. Gridcoin paid for BOINC tasks back in 2013. Old hat. What’s the sensation?”

The answer is that none of that is UPoW. That’s PoUW. A one-letter difference, but a structural one.

PoUW — Proof of Useful Work. In existing implementations (Flux, early Gridcoin, a number of academic projects) useful work exists on top of the consensus mechanism. Inside the network classical PoW or PoS is running — for security; in parallel with it the network provides infrastructure through which external clients rent compute for scientific calculations or rendering. Two independent layers, joined by one token. Network nodes can simultaneously secure the chain and perform useful tasks — but “securing the chain” and “useful tasks” are two separate jobs, each with its own purpose. UPoW — Useful Proof-of-Work. In Qubic, useful work is itself the consensus mechanism. Not “in addition to” and not “on top of.” The very same ANN simulation that Aigarth needs is what produces the digest that certifies the computor as an active participant in the network. One process, simultaneously doing both jobs. There aren’t two layers that need to be combined.

Come-from-Beyond on April 3, 2024, formulated this in a short phrase: “UPoW is not equal to PoUW”. Not the same thing. I’m putting this into a separate section deliberately, because substitution of one for the other is the most frequent mistake in discussions of Qubic. It’s easy to get confused, because the letter order is almost identical, and the functional difference — fundamental.

Now the second separation that needs to be made here. UPoW is one thing. External mining, in which Qubic also engages, is another.

In parallel with UPoW, two more streams of computational activity run on the network. The first is external mining. First, from May 2025 to April 2026, the network mined Monero. From February 2026, a DOGE add-on was launched.

It’s important to understand what external mining is structurally. It is not part of UPoW. It is performed by the same computors but with a different task, on different infrastructure, with a different result. DOGE mining gives the network revenue (an inflow of coins which are then converted into QU through the buyback mechanism and burned — these are details for chapter nine). UPoW gives the network consensus. These are two different jobs which a computor can perform in parallel, but which are architecturally not the same thing.

Come-from-Beyond formulated this on July 15, 2025, directly: “Too much misconception about #Qubic #UPoW on X. Mining of $XMR isn’t a part of it, mining for #Aigarth is. Finding blocks for #Monero looks useful but it doesn’t influence certification of validators directly.” And separately, in the same thread: “Monero mining isn’t UPoW”.

Here it’s important to go slowly. External mining “looks useful” — looks useful indeed. Very much so. It brings revenue to the network, it provides hashrate, in 2025 it even temporarily made Qubic the largest participant in the Monero network. All of that is real outcomes. But it does not directly influence the certification of validators. A computor doesn’t get a place in the next epoch for having mined a Monero block. A computor gets its place for having produced enough valid digests — for UPoW. There’s empirical proof of this, and it’s strong. In Discord posts of April 17, 2026, Come-from-Beyond described the observed picture: out of 676 computors, around 20 do only UPoW, without XMR and without DOGE. They receive at the same time about 80% of the maximum revenue of the epoch. That means the bonus from external mining makes up around 20% of full income. This isn’t an after-the-fact estimate; it’s a direct measurement on a live network.

From this follows a statement worth saying aloud. Roughly 80% of the network’s value is produced by UPoW. Roughly 20% by external mining. And — the critical test: switch off DOGE — the network keeps working. Switch off UPoW — the network is dead. UPoW is the foundation. External mining is an annex.

Here chapter three closes. There I wrote: the mining process must itself produce decentralization. Bitcoin doesn’t do that — mining produces concentration through ASIC economies of scale, and two pools today control most of the hashrate. Qubic does it differently. UPoW produces score. Score determines who stays among the 676 in the next epoch. The process of producing digests and the process of producing the network’s composition are literally one and the same process. Not “we mine, and in parallel update the composition through some other procedure,” but “we mine — and by that very fact update the composition.” The cycle is closed on itself: work generates the right to keep working; the right to keep working is determined by how much work you have done.

This also structurally closes another promise. In chapter three I formulated Axelrod’s fourth condition: a win in the game cannot be converted into the right to change the game itself. In the architecture of UPoW this condition is fulfilled like this: a high score gives you the right to stay among the 676 for the next round of the game. It doesn’t give you increased weight of vote, doesn’t give you an enlarged share of consensus, doesn’t give you influence on parameters. Only — the right to keep playing. Won more — you stay. The rules of the game — you don’t change. A clean implementation of what Axelrod, describing the tournament, formulated as a necessary condition for the resilience of cooperation.

And a brief historical remark that fits here. On September 23, 2021, seven months before the launch of mainnet, Come-from-Beyond named in a public tweet three architectural properties of the future network: “Useful PoW, unlike #Bitcoin / 100% certainty of transaction outcome / Feeless transfers like in #IOTA”. Useful PoW — unlike Bitcoin. One hundred percent certainty of transaction outcome. Feeless transfers, as in IOTA. UPoW was named by name even before the network existed. Not post-hoc rationalization (“we built something, and now we’re explaining what it was”); a declaration stated before the start, verifiable afterward.

A separate clarification for the attentive reader. Qubic as an engineering project has a long prehistory. In chapter five I mentioned Jinn — a research effort from the early 2010s, where the concept of useful work was first thought through as a possibility. UPoW is its realization a decade later, in another context and on a different basis. From idea to working network — a decade passed. That’s a normal pace for a nontrivial infrastructural idea.

Now, with the separation of three layers in hand, I can speak directly about what I warned of in the first section of the chapter. About that gap between built and running at full strength which the system’s creator named publicly himself.

### Honest Disclosure: “Useless PoW” and Why It’s Here

By this point the reader may suspect that UPoW is too beautiful to work. He’s half right. The idea works. The implementation isn’t yet at full strength. The system’s author said this himself, twice, publicly — and the second time, in July 2025, used a word that is hard to forget.

The first quote, verbatim. On July 17, 2025, Come-from-Beyond wrote on Discord: “On serious note, current mining params are useless, currently UPoW means Useless Proof-of-Work”. And in the same thread, in a separate phrase: “just like our current UPoW”. It wasn’t critics who wrote this — the system’s creator wrote it himself, in front of witnesses, on a public platform, in a channel where the engineers and the token holders are gathered.

The word “Useless” isn’t accidental. The usual marketing register allows “not yet at full strength,” “remains to be polished,” “in the process of optimization.” Come-fromBeyond chose none of those soft formulations. He chose a word which on any reading sounds like a statement of fact: at the moment — useless. And with a second move he reinforced it: just like our current UPoW.

And this isn’t the first such statement. On December 7, 2023, a year and a half earlier, the same thought had already been voiced — a bit more softly, but with the same structure. “Right now our Useful PoW is Half-useful, because Aigarth cannot process all the provided solutions”. Right now our Useful PoW is Half-useful, because Aigarth cannot process all the provided solutions. The problem has been recorded not since last year. It has been known to the system’s creator since at least the end of 2023. What, exactly, do these two quotes mean? I’ll work through it slowly, because this is the structural center of the chapter, and on the precision of the wording depends whether the reader will understand it correctly.

The infrastructure of UPoW works fully on all of its design goals for consensus. Computors simulate the ANN. Digests are validated. Score is assigned. Rotation of 225 places happens every Wednesday. The network is decentralized, secured, functioning. If you look at Qubic today as a blockchain — everything is in order, the consensus layer does exactly what it promised to do, and that first of UPoW’s three outputs (proof for consensus) is produced in full.

The problem is not in the first output. The problem is in the second. AI training data is produced every tick — but not all of it is consumed. Aigarth, designed as the consumer of this data, cannot yet process the full volume of solutions the network supplies it. The network generates more than Aigarth eats. The difference is precisely what Come-from-Beyond calls half-useful and useless.

The reasons, in his own words, are two. The first is the throughput limit of Aigarth. The consumer is so far smaller than the producer; this is an engineering scaling task on which work is in progress. The second is the parameters of the task itself. The ANN that the computors simulate is chosen and tuned so as to provide validity for consensus. For Aigarth’s needs it isn’t optimized — it may be of the wrong architecture, the wrong sizes, with the wrong properties to be maximally useful for evolutionary selection of useful neural networks. This is an architectural design problem solved only by changing parameters and — ultimately — by the deploy of Aigarth in its production form. The parallel back to coal here suggests itself. Murdoch ran the first pipe in 1792. But gas lighting for the first few years worked only in his own workshop in Redruth. London began to be lit by gas — really, with street lighting of the streets — in 1807. The gasworks industry, in earnest, as an industry, took shape by the early 1820s. Between “we can collect the gas” and “gas lights London” — fifteen years passed. Between “gas lights London” and “gas powers industry” — another ten. That’s a normal pace for an infrastructural shift.

In the case of UPoW: mainnet started up on April 13, 2022. UPoW as a mechanic stabilized in 2024. The planned deploy of Aigarth — April 13, 2027. Between “the mechanic works” and “the mechanic is used at full strength” the project has reserved approximately five years. That’s a working pace.

Now an important question: why is this said publicly rather than swept under the rug?

Come-from-Beyond could have stayed silent. He could have waited until Aigarth caught up, and spoken about UPoW only in the past tense: “built,” “works.” The press materials accompanying any technological project would easily allow that. None of Qubic’s critics had brought “Useless PoW” to the public in the system creator’s own formulation — because the formulation itself was created by him. He could have kept it to himself.

He didn’t. And this is no random habit. In chapter four I described a principle Comefrom-Beyond has been following for a second decade: don’t sell what isn’t there yet. Hiding “Useless PoW” would have meant violating that rule in the very place where it gets tested most palpably — in the central mechanic of the central chapter of the entire project. To say it aloud is to assert the rule more strongly than any pr message could formulate it. In a world where ninety percent of token projects hide engineering problems behind marketing until the moment of collapse, putting one’s own main weakness out in public is not weakness and not an admission of defeat. It is discipline. What is Phase 3 and what does DOGE have to do with it?

First: the XMR-mining code is removed from the core. Not switched off by a flag that could be flipped back in a minute — removed completely, so that for a return a new release with restoration of the code would be required. This is a surgical architectural commitment. The Monero chapter is closed.

Second: Phase 3 of compute distribution is activated. The new proportion is roughly 80% to Aigarth training, 20% to DOGE mining. In previous phases a substantial part of compute went to XMR mining; that part is now freed and given to Aigarth. For the first time since the launch of mainnet, Aigarth has access to 100% of the compute reserved for AI work.

Phase 3 isn’t “the moment when UPoW became fully useful.” It’s the moment when the gap between built and used began to close through scaling of the consumer. How quickly it will close, and whether it will close by the planned date, is a question for Part Five of the book. There I will be returning to questions of Aigarth’s architecture and to its deploy.

From Phase 3 to the planned deploy of Aigarth — 356 days. Fifty-one weeks. Fiftyone epochs. In the terms of time introduced earlier — fifty-one rounds of the iterated game.

I’ll close this section with an honest frame.

UPoW in April 2026 is infrastructure that one engineer began to think through in 2013, spent a decade implementing, and launched on mainnet in 2022. It works against all its design goals except one: the AI component so far uses part of what is produced. This is a crack between built and utilized. It exists. It has been recorded by the creator publicly, twice, with an interval of a year and a half.

This is the honest state of UPoW as of April 2026. Neither less, nor more. Under the Hood: Numbers, Revenue, and One Architectural Bonus Now to the gears. The mechanics of UPoW are expressed not only in words; behind every formulation of the previous sections stand specific constants which the team sets in code and which can be looked up in the Qubic repository. This section is for the reader who wants to know exactly what the working architecture is assembled from. I’ll start with the parameter N from the quote of July 23, 2024. In the code of the Qubic core this parameter is called ADDITION_SOLUTION_THRESHOLD_DEFAULT, and it determines how many leading zeros there have to be in a digest for the solution to count as valid.

I’ll underline an important nuance. In Bitcoin, difficulty is self-regulated by the algorithm: every 2016 blocks the network looks at how quickly solutions are being found and adjusts difficulty so that the average interval stays around ten minutes. In Qubic, difficulty is set explicitly by the development team every epoch. This is an engineering decision that makes sense. The hardware base of Qubic is the computors, and their composition for the next week is known in advance; there is no need to imitate blind self-regulation when one can calibrate the parameter for the current hardware directly. It’s the same straightforwardness we encountered in chapter seven in epoch transitions: not “it’ll work itself out,” but “the team made a decision.”

The reference implementation of the miner from Qubic is open-source and is called Qiner. The technical reader can go to the repository and look at exactly how the ANN simulation is built at the level of code — not from words but from sources. It isn’t a black box; it’s a public engineering implementation of what was discussed in sections 2– 3 as a structural reduction. I won’t be retelling the code here — that’s material for a separate breakdown — but the very fact of its openness matters: the claims about the mechanics are independently verifiable.

Now to the money. How exactly does the network pay for a computor’s work? The revenue formula, recorded in PR #844 of the Qubic core and active in Phase 3: revenue = IPC × M × (S² + B·E) / DIVISOR I’ll work through what each symbol means.

IPC — Issuance Per Computor, the issuance to one computor per epoch. The constant equals 1,479,289,940 QU; this is the unit of a trillion new QU issued in an epoch, divided by 676. The upper ceiling a computor can reach for an ideal week. S — a scaling constant, equals 1024. All factors are normalized to the range [0, S]. B — the weight of the bonus for external mining, equals 256.

M — the mandatory component, takes values from zero to S. Assembled from two parts: tx factor (weight 85%) — how actively the computor leads ticks, and oracle factor (weight 15%) — to what extent it participates in the commit/reveal cycle of oracles. This is work the network considers obligatory. The component multiplies the rest of the formula. Didn’t perform the mandatory work — everything else gets multiplied by a smaller number, down to zero.

E — mining factor, also in the range [0, S]. In Phase 3 this is only DOGE shares. This is work the network considers optional. The component is added in via the term B·E, doesn’t multiply the base.

I’ll run through the ranges. With full mandatory work (M = S) and zero external mining (E = 0) the formula gives revenue making up about 80% of IPC. With full mandatory work and full mining (E = S) — exactly IPC, 100%. If a computor fails the mandatory work — the mandatory component collapses, and regardless of how much DOGE it has mined, it gets little.

The construction directly corresponds to what I described in the third section in words. Work on which consensus depends — TX relay, oracles, UPoW — is rigidly built into the multiplicative layer of the formula. Work on which only the network’s revenue depends — DOGE mining — is built in as a weak, additive bonus. Switch off DOGE — you’ll lose around 20%. Switch off the mandatory part — you’ll lose everything. And here let’s note one thing about UPoW’s second output, which in the fourth section was discussed as a bottleneck. AI training data isn’t paid for directly by the formula. A computor doesn’t receive revenue for producing intermediate ANN states useful to Aigarth; it receives revenue for having found a valid digest — the first output — and for having performed the mandatory network work. The second output is currently pure byproduct, written into the atmosphere awaiting a consumer. Or, to flip it, the economy pays for two outputs out of three: consensus and the rotational score (through the mandatory part and through the fact of finding digests). The third, the AI component, is paid for on the credit of expectation.

There’s one more property of UPoW I’ll mention separately, because it isn’t earned by design but by structure. Bitcoin PoW through the sha256 function is theoretically vulnerable to quantum acceleration: Grover’s algorithm gives a quadratic speedup for the preimage-search problem, and at the appearance of sufficiently large quantum computers, Bitcoin mining will find itself in a new reality — not broken, but rebuilt. UPoW is built differently. The task “find an input to an ANN for which the first N output bits are zeros” doesn’t reduce to any of the problems for which quantum algorithms with speedup are known. There is no quantum algorithm performing a forward pass through an arbitrary neural network faster than the classical one.

Come-from-Beyond on November 10, 2025, formulated this in a single phrase: “Unlike PoW, our UPoW isn’t vulnerable to QC”. Unlike PoW, our UPoW isn’t vulnerable to quantum computers.

Here the tone with which I say this matters. This is not a primary design goal. Quantum resistance was not on the list of tasks for the sake of which Qubic was conceived in 2013. It’s a side property of the architecture. But such is the nature of many elegant engineering solutions: when a thing is made correctly for one reason, it turns out to be correct also for reasons its creators didn’t have in mind. Murdoch collected the gas for the sake of lighting his workshop, not for the sake of having chemical works run on the same gas a hundred years later. As it happened, the collected gas turned out to be useful both there and here.

The last thing I’ll say in this section is a short quote I’ll be coming back to in Part Five of the book. On November 20, 2023, Come-from-Beyond described Qubic’s functional map in a single sentence: “Aigarth uses Qubic’s UPoW, also Qubic has everything that an AI would need: Smart contracts as brain, Oracles as sensors to outer world, Outsourced computations as Manipulators to change that outer world.”

Smart contracts as brain, oracles as sensors, outsourced computations as manipulators — and UPoW as the source of energy for everything. On this map UPoW isn’t “a feature of consensus” but the energy subsystem of an AGI substrate. Like the heart in a body: it pumps, it doesn’t think, but without it the rest doesn’t work. Unfolding this map is the job of Part Five; here it’s needed only as a reminder of what, in the end, all this machinery is for.

### Parsons, the Turbine, and the Twice-Collected Gas

There is one more figure worth calling into this chapter to close it. Not Murdoch — with Murdoch we’ve already worked, and we’ll be working again in the finale. Charles Algernon Parsons, England, 1884.

Before Parsons, the steam turbine worked in a single stage: high-pressure steam struck a large bladed wheel, transferred to it part of its energy, and exited the other side — still with an enormous reserve of energy, which simply went off into the surrounding space. The efficiency of the first turbines was around 20%. Four-fifths of the steam’s energy — out the chimney, literally.

Parsons built a multi-stage turbine. Steam passed not through one wheel but through a series — gradually, giving up at each stage part of its energy and moving on to the next. The first stage took what it was strong enough to take. The second — what passed by the first. The third — what passed by the two. And so on, in a cascade, until the steam exiting the last stage carried with it only residual pressure. Efficiency, thanks to this construction, rose from 20% to 50–60% and higher.

The principle Parsons applied to steam is formulated abstractly like this: don’t collect everything in one operation, but build a cascade of operations, each of which collects what the previous didn’t. Multi-stage extraction. Extraction in several stages. The idea is applicable to anything that constitutes a process with surplus energy at the output.

UPoW is multi-stage extraction applied to compute.

Bitcoin PoW is built as a single-stage turbine: compute goes in, a digest comes out, all the rest of the computational energy is lost as heat. One stage, one output. UPoW rebuilds the same process into a cascade:

The first stage — compute goes into the form of an ANN simulation and at the output yields a digest with N zeros. Consensus proof. This is the output Bitcoin also has. The second stage — the intermediate states of the simulation, which in Bitcoin were thrown out by construction, here remain as data. AI training data. An output Bitcoin doesn’t have.

The third stage — the aggregate of valid digests over an epoch is folded into a score, which determines the rotation of the network’s composition. Score. An output Bitcoin doesn’t have.

One input — three outputs. The engineering form of what Murdoch did with coal, applied to computational work 217 years later.

I’ll underline one thing so the parallel doesn’t read as metaphor. Parsons and Murdoch worked on one type of task. Not on one substance — they worked with steam and with coal respectively — but on one form of question: how to rebuild an existing process so that it produces several outputs instead of one. Neither one of them invented a new process. They rebuilt an existing one. UPoW is one more point in this series. Not “the invention of a new PoW” but the application of an old principle (multi-stage extraction) to a new substance (compute instead of steam or coal).

This chapter is called “Murdoch’s Gas in Action” — not “UPoW Explained.” The title is chosen this way precisely because, 217 years after Murdoch ran the first pipe from furnace to lamp, the same engineering idea is being applied to the digital analog of that very gas.

The Gas Is Collected. The Pipe Is Laid.

Three sections of compressed mechanics behind us. Now a brief return to the overall view.

Qubic miners solve the same task as Bitcoin miners: find a digest whose first N bits are zeros. But their digest is generated by simulation of a neural network, represented as a hash function through a formal reduction. One job at the input — three outputs at the output: proof for consensus, data for potential AI training, score for rotation of the network’s composition. The infrastructure works against all design goals. The second output — the AI component — is so far not used at full strength; the system’s creator himself spoke this aloud twice publicly, first in December 2023, then in July 2025. Phase 3 was activated: the XMR-mining code is removed, Aigarth for the first time gained access to 100% of the compute reserved for it. To the planned deploy of Aigarth — 356 days.

Chapter six described the space of this network — 676 voices. Chapter seven described its time — weeks of epochs. This one described its energy — UPoW and what it produces. And economics always begins with energy.

From Murdoch in 1792 to Parsons in 1884, 92 years passed. From Bitcoin in 2009 to Qubic in 2022 — 13 years. Industrial cycles have compressed from centuries to years, but their internal logic hasn’t changed: first the process, then the gathering of the byproduct, then the multi-stage optimization. We, the readers of this chapter in 2026, stand at a point which by analogy with coal corresponds to somewhere around 1805– 1810: the gas is collected, the first lamps are burning, but the great gas industry — is still ahead.

The KEY chapter of the first half of the book doesn’t end on a triumphant note precisely because triumph isn’t yet earned — in 2027 a test lies ahead, and now — work. Just as in 1792 Murdoch lit a lamp, but the streets of London didn’t yet know they would soon be lit by gas.

The next chapter is about the economics that keeps this machine running. We’ve described what computors do. We haven’t discussed how the network pays for this work — what exactly happens with the QU token every Wednesday, why issuance and burning go in parallel, why Qubic didn’t choose the Bitcoin model of fixed supply. The breathing token: every epoch a trillion new QU come into being, a substantial part is burned through the buyback mechanism, and supply pulses in time with the weeks of epochs.

Murdoch’s gas is collected. The pipe is laid. Now let’s figure out who is paying for its operation — and what the price of gas is in a world where gas makes not light but intelligence.