Bitcoin mining · economics · critique

The "Physics" That Forgot the Facts

A recent article tries to make the case that one particular mining stack comes out ahead over a 50–100-year horizon. While the ideas behind it are worth understanding on their own, we don't believe that the numbers and assumptions built on top of them hold up.

The article has the vocabulary down, and at first glance it can seem persuasive. Look closer though, and it starts to fall apart. It opens with things that are genuinely fixed (SHA-256, the 21 million cap, the halving schedule) and uses these to frame a set of tables that project a mining setup's earnings all the way out to the year 2126, ranked by a "100-Year Score".

No method is shown anywhere. We rebuilt its tables every realistic way we could come up with. Section 02 shows the only model that fit, along with what that reveals about the second fifty years. We'll grant the author his own stale rates, the growth his own tables imply and every generous rounding, and show the numbers still don't hold up to the barest scrutiny.

The mistakes are at least instructive, so we'll explain the mechanics as we go (the halving, a miner's share of the network, pool and transaction fees, stale shares) and let the errors show themselves. His article gets the trivia right and the conclusions wrong.

Every one of our figures comes from a single share calculation, using stated assumptions that vary by figure and are set out in full at the end. We've used inputs anyone can check: the device's rated hashrate, the current network hashrate, and Bitcoin's issuance schedule. Where we had to assume something, we've used whatever is most generous to the author's own conclusions, so if anything the corrections here are on the conservative side.

01 The halving

The coins come early, then it's fees

Miners get paid in two ways: a block subsidy, which is newly issued coins, and the transaction fees in the block. Today the subsidy is by far the larger of the two, but it doesn't stay constant forever. Every 210,000 blocks, which works out to roughly four years, the subsidy is cut in half. It started out at 50 BTC per block back in 2009 and sits at 3.125 BTC today. From there it drops to 1.5625 at the next halving, then 0.78125, and so on, until the last tiny fraction of a coin is issued somewhere around 2140.

The schedule is brutally front-loaded. Each epoch issues half as much as the one before, so the coins come early and then thin out quickly. By around 2040, somewhere near 99% of all the bitcoin there will ever be will have already been mined. The whole century from 2040 to 2140 only issues the last little sliver.

There's an easy way to picture just how little. Because each epoch pays half of the one before it, everything still left to mine after any halving adds up to no more than that single epoch which has just passed. So from the very next halving, barely two years away, everything mined across almost the entire hundred years the author projects will add up to no more than this current epoch pays on its own.

Spot the flaw

In the article's tables, every 100-year figure is exactly double the 50-year one: 515,000 becomes 1,030,000 sats, 4,288 becomes 8,576 BTC. An even split needs the second fifty years to pay out as much as the first, and the halving rules that out: the first fifty carry nearly all of the remaining subsidy, the second fifty almost nothing besides fees.

Table 2a · small scale

Table 2b · 5 EH/s

Table 2c · ~60 EH/s

The author's own yield tables, exactly as published. Click to enlarge.

Here's how the schedule actually decreases for a miner, using the article's own example device, a Bitaxe Gamma 601 at 1.2 TH/s, on his own implied assumptions. Each bar is a four-year halving epoch, each less than half the size of the one before: the halving halves the payout, and the network growth his tables imply shrinks the miner's share on top. The last bar rolls up everything from 2048 to the end of issuance and still barely registers:

So if we rebuild the article's small-scale table (Table 2a) with the halving properly applied, and the network growing as his own tables imply, two things stand out. The totals fall to roughly a tenth of what's claimed, and, more importantly, the jump from year 50 to year 100 effectively disappears:

There is one plausible objection to this: perhaps the flat century is a bet on transaction fees rising to take the subsidy's place. But a century-long fee forecast would be guesswork, six Bitcoin lifetimes of it, and guesswork is not "immutable physics".

More pertinently, if a fee bet is in his tables, we could find no trace. Rebuild them, as we do next, and the method needs no fee income at all. What it implies for the second fifty years undoes the whole projection.

02 Network share

A miner can't reliably earn more than its share of the network

What a miner earns isn't really a property of the machine on its own. On average it's a share: the miner's hashrate divided by the total network hashrate, times whatever the network is paying out. Any single block comes down to luck, but over years, let alone a hundred, the real total settles to almost exactly that share. So a miner who controls a millionth of the world's hashpower earns roughly a millionth of the coins. A pool, a gateway or a template policy can slightly nudge the edges, but none of those change the underlying share of hashpower.

As of writing, the network runs at roughly 873 EH/s, and a Bitaxe Gamma 601 adds 1.2 TH/s to that. Divide one by the other and the device is about 0.0000001374% of the network, which at today's block subsidy works out to somewhere in the region of 60 satoshis a day. That comes straight from public statistics, and independent profitability estimators (miningnow, asicminervalue) land in the same range.

The share also puts a cap on what a miner can ever earn. There are only about 950,000 BTC of subsidy left to be mined, and over its whole life a miner can collect only its share of that fixed pot. Transaction fees sit on top of the pot: a miner only collects its share of whatever fees the network earns, not more. Efficiencies can help claw back the tiny fraction lost due to stale work, but nothing outside of extraordinary luck lifts a miner above that.

That share is the tool for checking the article's two bigger yield tables, 5 EH/s (2b) and ~60 EH/s (2c), and since no method is given, we set out to rebuild them. The only model that fit is one which mining profitability calculators have offered for years: let difficulty keep climbing. With the halving respected, no fee income at all, and difficulty growing at about 4.8% a year, the model reproduces both big tables' fifty-year figures almost exactly: 4,288 BTC for the 5 EH/s table and 51,500 for the 60 EH/s. No workings are shown, so we can't know this is what the author did. But nothing else we tried, simpler or more involved, landed us on his figures. And there is nothing wrong with them: they are consistent with the halving and with conservative growth, and on their own they are entirely defensible.

Run the same model past year fifty, and the projection collapses. The reason isn't the growth rate we picked, it's the halving: nearly all of the remaining subsidy is issued in the first fifty years, so the second fifty add less than one bitcoin for the 60 EH/s operation; the article projects another 51,500, because every hundred-year figure in every table is simply the fifty-year figure doubled.

Where do the extra 51,500 BTC in the author's projections come from?

The Bitaxe table is the odd one out. Per unit of hashrate its figures run exactly five times too high. The likeliest explanation is an honest slip. Five Bitaxe Gammas make 6 TH/s: run 6 TH/s through the same method and, rounded to the nearest thousand, out comes 515,000. But the slip sits in the one table most readers at home can act on, and with the doubling stacked on top, the claim ends up ten times too optimistic.

There's more wrong with these tables than the totals. Each lists a stale rate, a yield, a "Diff vs Best" and a "% Loss", yet nowhere does the article say how any of them relate, and on their own terms the columns don't appear to agree. The clearest sign: the same 0.25% legacy pool is marked 0.40% loss in one table, 0.54% in another and 0.49% in the third. One input, three different outputs, with nothing explaining the gap.

03 The "100-Year Score"

Precise scores, with no method shown

The first of the article's numbered tables ranks six mining setups by a "100-Year Score", presented under the heading "Top 6 Mining Stack Rankings" and above its own caption, which we quote exactly: "Rankings by Long-Term Profitability (50–100 Year Horizon)". The stack it promotes lands at 9.8 out of 10, the next is rated at 9.4, on down to legacy Stratum at 4.1. Such precise values carry the look of a research model's output. But no model is disclosed, so nobody can verify the results for themselves.

The article's ranking table, as published.

A 9.8 without a model to back it amounts to little more than a personal preference. The two top scores, 9.8 and 9.4, belong to the exact setups the article goes on to recommend, and the lowest belong to the competition, described with labels like "higher stale" and "spam tax" rather than any measurable statistics.

04 Stale shares

A tiny penalty, and a temporary one

A stale share is one which a miner finds a moment too late, after the network has already moved on to the next block. It earns nothing, so lower latency and locally built templates, which cut stale work, are a genuine benefit of the tooling the article promotes. We'll grant that. The real question is how big the effect is. The article puts legacy setups at 0.25% stale and local-template setups at roughly zero, but that 0.25% doesn't come from any independent measurement which we can find. It's the exact example Ocean's own DATUM marketing uses ("from even 0.25% before to nearly 0% with DATUM"), and it's close to the Stratum V2 project's own figure of about 99.8% share acceptance improving towards 100% with local job declaration. Both are best-case vendor numbers rather than evidence about how a given pool actually performs.

Even at face value, the effect is tiny. A 0.25% stale rate means keeping 99.75% of revenue, and as we saw earlier, the article's own tables dock the legacy pool more than that without ever saying why. Over the corrected lifetime of our little Bitaxe, the gap between the stack it promotes and the legacy pool it ranks lowest comes to about 260 satoshis in subsidy rewards, spread across a hundred years. That's roughly what a miner earns from the fee on a single low-fee-paying transaction included in a block.

The bigger problem here is the strange assumption underlying the whole comparison: that a "legacy" pool stays legacy for a hundred years. It won't. The low-latency, local-template tooling the article is promoting is open source and simple to adopt, and any pool actually losing a quarter of a per cent to stale work would close that gap quickly, because nothing is stopping it. It's an invented disadvantage stacked on top of the doubling: a hundred-year comparison in which no competitor ever upgrades.

There is also one comparison the article's own logic demands but never makes. The stacks it favours earn part of their edge by filtering "spam", which means declining the fees those transactions pay. In the fourteen months to February 2024, inscriptions alone paid miners more than 6,000 BTC, roughly one and a half per cent of everything miners earned in that time, six times the entire stale-rate penalty the article hangs on its worst-ranked pools. He presents a ranking based on lost revenue but ignores the biggest loss to his top pick.

05 The pool fee

Not physics, and already changed once

The article leans on a real perk. Miners who run DATUM, an open-source tool for building block templates locally, get a 50% discount on Ocean's pool fee, paying 1% instead of the standard 2%. That discount is genuine. But the projection treats it as a fixed constant for a century, and it isn't.

DATUM, the protocol, is fee-agnostic. All it does is build templates locally and coordinate the reward split with the pool, and in its own documentation it "is not specific to a pooled reward system". The fee and the discount aren't defined by DATUM at all. They're set by the pool operator as a matter of commercial policy, and Ocean's own materials frame DATUM as the technology and the operator as the one controlling pricing. A different pool could run the identical protocol and charge something completely different.

Ocean has already exercised that control. The pool launched in November 2023 with 0% fees, then in mid-November 2024 introduced the current structure of 2%, or 1% with DATUM. That was its first fee change, roughly one year after launch. So a primary input to a 100-year projection is a private company's revocable pricing decision, one that has already been changed once in the pool's roughly two-and-a-half years of existence.

Nothing in Bitcoin's "immutable physics" stops Ocean from raising the base fee, trimming the discount or restructuring its payout system tomorrow, and a century guarantees many such decisions will be made if the company survives that long.

06 The spam filter

What the filter actually costs

The article lists spam, "inscriptions and arbitrary data", among the things that "act as silent taxes" on its chosen metric, hashes per joule, and its closing line declares that over decades the winner is the miner who "wastes the fewest joules on stale shares, fees, spam, or bandwidth". This is worth spelling out: the article's efficiency metric rests on it, and it misunderstands what a mining machine does.

An ASIC never hashes transactions. It hashes the block header, 80 bytes, fixed, whatever the block carries: a timestamp, a nonce, and one 32-byte digest standing in for everything else. A block stuffed with inscriptions, a block of plain payments and an empty block hand the machine an 80-byte header each, at the same energy per attempt. The contents are hashed by the node's own processor as it folds them into that digest while building the template: a fraction of a second of ordinary computing, the same work byte for byte whatever the bytes represent. The ASIC then grinds that header trillions of times a second, changing only a nonce, never touching the contents. As a matter of fact, a block of ordinary payments can work the node harder than one stuffed with inscriptions: thousands of small transactions mean thousands of signatures to check, where a handful of bulky inscriptions carry almost none. Either way it is a moment's work for an ordinary processor, and none of it is done by the miner.

More to the point though, Bitcoin already has a mechanism for junk, and it's the same one the author seems to revere. Block space is scarce and fixed. As adoption grows and that space is contested, we can expect that people paying real money to move real value will outbid low-value data for it. The fee market is the filter, and it makes "spam" uneconomic to sustain. The inscription wave of 2023 and the Runes craze that followed the 2024 halving both all but burned themselves out, no rule change required.

A transaction fee is a miner's income, not its cost. A miner who screens valid transactions out of its blocks simply earns less than one who leaves them in, and the fees it declines are picked up by the next operator who isn't so fussy. When block space is contested, which is exactly the future the article projects, filtering valid transactions is a standing decision to earn less than the competition. Somehow the author scores filtering as an advantage in his rankings, and the cost of it never appears in his projections.

The article name-drops BIP-110 in the same breath as Stratum V2, as though the two belonged to one modernisation effort. They don't. Stratum V2 is a transport protocol, it changes how miners talk to pools. BIP-110 is an extremely contentious soft fork to Bitcoin's consensus rules.

Does it even work? A determined data-stuffer just routes around the filters. Peter Todd embedded the entire BIP-110 text into a BIP-110-compliant transaction, and a separate demonstration stored a 60 KB image using none of the filtered paths. The reference Ordinals implementation has an approved BIP-110-compatible inscription envelope queued to merge the instant the rule activates, a companion parser already reads it, and a public readiness list tracks the wider ecosystem. The fork hasn't even activated yet and it's already circumvented.

And even if it did work, that would be worse. A protocol that can retroactively rule valid transactions unacceptable and soft-fork them away is a protocol that no serious builder can rely on. Why commit to a vault, an inheritance scheme or a Layer-2 rail if tomorrow's rules might brand the outputs they depend on as "spam"? That uncertainty discourages engineers from building the long-horizon monetary use cases the author says he wants.

The lack of support for the proposal is clear: at the time of writing, signalling sits under 1%, far below the 55% that would trigger early activation. But low support won't stop it. BIP-110 is written to lock in regardless: after a mandatory flag day, nodes running it reject any block that doesn't enforce the new rules. Attempting to force a contested change onto an unwilling network in this way is exactly what could result in a chain split.

So the article projects a hundred-year, fee-funded future while nodding to a hostile proposal that neither stops "spam" nor lets anyone reliably build the monetary economy which that future depends on. Wanting less junk on-chain is reasonable. Forking the protocol into permanent uncertainty to combat it ineffectively is a reckless gamble.

07 In fairness

What the article does get right

A good-faith reading has to credit the parts that hold up, and there are some. The entities are real and mostly described accurately: Ocean is a real pool, DATUM is real open-source software, Bitcoin Knots is a real node implementation, and the 50% DATUM fee discount is genuine (with the caveats above about who controls it). The history is right, too: Stratum V1 dates to around 2012, and Stratum V2 was indeed pioneered by the Braiins team in 2019. The article's report that seven major pools joined the Stratum V2 Working Group on 7 May 2026 checks out. It's worth noting who they are though: they include Foundry and AntPool, the pair the article's own ranking parks second from bottom, at 5.9 out of 10. So the pool operators it rates worst are among those driving one of the very standards it rates best. The AtlasPool detail checks out too: the pool announced a migration to Knots across five of its nine endpoints in late May 2026.

The one table we'd leave almost untouched is the Stratum V1-vs-V2 comparison. Its headline figures (roughly 60–70% less bandwidth, encryption, job negotiation, "up to 7.4%" efficiency) come straight from the Stratum V2 project's own published materials. The same goes for the "95% bandwidth reduction" credited to DATUM, which traces to Ocean's own pitch for remote sites, a drop from roughly 200 GB to 8 GB per day for a 1 EH/s operation. These are vendor benchmarks, and "up to 7.4%" is a best-case scenario rather than a typical one, but they are at least traceable to published sources.

Worked backwards, the fifty-year figures in the 5 EH/s and 60 EH/s tables deserve real credit: they fit a halving-aware subsidy model with difficulty growing about five per cent a year, almost to the coin. If that is what was done, it is sober modelling, conservative even. The trouble is what happened to those numbers afterwards: doubled into a hundred-year column that the same model says should add less than a single bitcoin.

Local template construction, whether via DATUM or Stratum V2's job declaration, is a real improvement for decentralisation and it does reduce stale work. The intuition that efficiency and latency matter to margins is sound. On fees, the article misses half the picture: the "spam" it wants filtered pays transaction fees that become a miner's main income once the subsidy is mostly gone. Filtering transactions out cuts margins rather than protecting them, so it is the one "edge" we can't credit even in part.

The problem is that these true things are used to prop up projections that don't survive scrutiny.

The bottom line

Strip away the tables and the article's remaining claims are modest and partly true: run efficient, low-latency, sovereign infrastructure and a miner can shave fractions of a per cent off its losses. Absolutely worth doing. But that's not what he's selling us on. He's selling a hundred-year yield forecast, with no method to back it. The hundred-year columns are just the fifty-year columns doubled. The three tables can't agree on one set of assumptions. The headline ranking is scored on nothing anyone can check. And the two biggest levers, self-imposed filters and a discount one company can repeal, go conveniently ignored for one hundred years.