Bitcoin's Network Effect Is Measurable. The Number Everyone Quotes Is Not.

Almost every bullish argument for Bitcoin eventually arrives at the same sentence: it benefits from network effects. The sentence is true. It is also doing a great deal of unexamined work, because the same idea is used to explain why Bitcoin is useful and to calculate what it should be worth, and those are not the same claim.

So we went back to the research, then to the data. The research turns out to be fifty years old, well developed and genuinely contested. The data turns out to support the effect clearly and the arithmetic around it far less clearly.

Where the Idea Comes From

The concept predates the internet by two decades.

In 1974, the Bell Labs economist Jeffrey Rohlfs published a paper on the demand for communications services with an observation that seems obvious once stated and was not obvious at the time: the demand of each user depends on how many other users there are. From that he derived the start-up problem. Below a certain number of participants, a network has no stable equilibrium other than zero. It does not grow slowly. It collapses.

In 1985, Michael Katz and Carl Shapiro split the effect into two mechanisms that are still the standard vocabulary. A direct network effect means more users make the network itself better. An indirect one means more users attract more complementary products, which then attract more users.

In the same year, Paul David published his study of the QWERTY keyboard and gave the field its most quoted example of path dependence: an arrangement that persists because it was adopted early, not because it is optimal.

In 1989, W. Brian Arthur formalised what that implies for competing technologies. Where returns increase with adoption, the outcome is not determined by which technology is better. It is determined by small early events that get amplified. The winner is partly an accident of history.

Read as a set, those four papers describe Bitcoin's competitive position uncomfortably well, including the part most Bitcoin arguments leave out. Arthur's conclusion is that the first mover's advantage is real and that it is luck rather than merit.

How Researchers Try to Prove It

Direction is one thing. Magnitude is another, and this is where the field splits.

The n² claim. The engineer Robert Metcalfe sketched the familiar version in 1980 while selling Ethernet hardware. A network of n nodes has n(n−1)/2 possible connections, which grows roughly as n². Two details usually get lost. Metcalfe counted compatible communicating devices, not users. And the label "Metcalfe's law" was attached by the writer George Gilder in a 1993 Forbes column, not by Metcalfe. Metcalfe's own formulation also carries a linear cost term and an affinity constant per user, and in his 2013 retrospective he concedes there may be "diseconomies of network scale" where that constant turns down and swamps the n².

The n log n counter-claim. In 2006, Bob Briscoe, Andrew Odlyzko and Benjamin Tilly published "Metcalfe's Law is Wrong" in IEEE Spectrum. Their objection is empirical rather than mathematical: "connections are not all used with the same intensity. In fact, in large networks … most are not used at all." Weighting connections by how much they are actually used gives n log n. Concretely, in their own example, a network of 100,000 members worth one million dollars would be worth four million at 200,000 members under Metcalfe, and 2.1 million under theirs.

Their historical evidence is the strongest part of the paper. The Bell System and independent telephone companies competed in the same American neighbourhoods for two decades without interconnecting. CompuServe, Prodigy, AOL and MCIMail ran separate email systems until the mid-1990s. If value really scaled with the square of users, those firms left enormous sums on the table for twenty years.

The counter to the counter. In 2014, Leo Van Hove showed in NETNOMICS that this specific interconnection argument does not hold up logically. Two networks refusing to merge is compatible with n² value under plausible assumptions about who captures the gain.

The first real test. In 2015, Xing-Zhou Zhang, Jing-Jie Liu and Zhi-Wei Xu did what nobody had done in the previous thirty years: they checked. Using Tencent from 2003 to 2014 and Facebook from 2004 to 2014, they fitted four competing models. Metcalfe's n² won, and not narrowly. Measured as root mean square deviation in billions of dollars, Metcalfe scored 0.12 for Tencent and 0.64 for Facebook, against 1.51 and 1.45 for Odlyzko's n log n.

So the score is roughly this. The theory is disputed, the one large empirical test favours n², and the dispute has never been settled for money as opposed to messaging.

The Pattern Outside Bitcoin

Before turning to Bitcoin, it is worth being precise about what the effect looks like when you can measure it end to end. Three cases with numbers rather than anecdotes.

Telephones in the United States. From 10 percent of households in 1903 to 41 percent by 1929, then down to 31 percent by 1934 as the Great Depression removed a quarter of the network. Recovery to 62 percent by 1950, a plateau at 94 to 95 percent from the 1990s to 2002, then a collapse to 41.7 percent by 2018 as mobile phones replaced it. A full life cycle, including the fall.

The internet and mobile worldwide. Internet use went from 15.6 percent of the world population in 2005 to 73.6 percent in 2025. Mobile subscriptions went from 0.21 per 100 people in 1990 to 111.5 in 2025. The ITU counted roughly 6 billion people online in 2025.

Messaging. WhatsApp reported more than 450 million monthly users at the February 2014 acquisition, one billion in February 2016, two billion in February 2020 and more than three billion in April 2025. That final figure is about half of everyone on Earth who is online.

And the concentration that follows. In the third quarter of 2010, Gartner counted seven smartphone platforms with measurable sales, led by Symbian at 36.6 percent, then Android at 25.5 percent, iOS at 16.7 percent and BlackBerry at 14.8 percent. In August 2026, StatCounter measures Android at 67.61 percent of mobile browsing and iOS at 32.36 percent, leaving 0.03 percent for everything else. The two figures measure different things, units sold against devices in use, and the market leader of 2010 is on neither board.

The mechanics behind all of this are covered in our knowledge article on what the network effect is and why Bitcoin has one. What follows is the part that has not been written down carefully anywhere we could find.

What Bitcoin's Own Data Shows

We pulled sixteen years of Bitcoin network data from the Coin Metrics community API, taking the 1 January value of every year from 2011 to 2026 plus 1 September 2026, and we are publishing the series so the calculation can be repeated.

The first finding needs no statistics at all.

Date	Addresses with a balance	Active addresses	Price in dollars
1 Jan 2011	65,813	1,071	0.30
1 Jan 2015	3,926,220	145,765	314.78
1 Jan 2019	22,211,324	433,715	3,808.12
1 Jan 2023	43,266,583	719,716	16,606.75
1 Jan 2026	55,237,400	570,904	88,684.22
1 Sep 2026	56,789,267	677,321	77,407.70

Addresses holding a balance grew by a factor of 863 in under sixteen years, and the count fell in exactly one of those years. Between January 2018 and January 2019 it dropped by 4,431,375, or 16.6 percent, while the price fell 71.7 percent. That is the only down year in the series. The 2022 collapse did not produce one: the price fell 65.1 percent and the address count still rose 9.2 percent.

The 2018 exception is worth understanding rather than hiding. A mania creates addresses that hold almost nothing, and when it ends those addresses are emptied or swept together. What survived the 2018 drawdown was the part of the network that was not there for the mania, and that part has risen every year since.

Active addresses did nothing of the kind. They peaked above one million in January 2021, fell by a third, recovered partially and now sit below the 2021 level. Usage tracks the price cycle. Accumulated participation does not.

That divergence is the cleanest on-chain evidence we know of that adoption and price are separate variables, and it is the reason we treat the two as separate questions throughout.

Running the Regression Ourselves

The standard way to test Metcalfe against real data is to regress the logarithm of value on the logarithm of users. The slope of that line is the exponent. If it comes out near 2, Metcalfe holds. If it comes out near 1, value simply tracks membership.

On our seventeen data points, using addresses with a balance as the measure of network size:

β = 2.069, standard error 0.101, R² = 0.966.

Using daily active addresses instead: β = 2.091, standard error 0.189, R² = 0.891.

Taken alone, that is a striking result. Metcalfe's n², proposed for Ethernet hardware in 1980, appears to describe Bitcoin's market capitalisation almost exactly.

Except that Wheatley, Sornette and colleagues ran the same test in a peer-reviewed paper in 2019, using daily data through 2018 and active users, and got β = 1.69 with a standard error of 0.0076 and R² of 0.95. Their own comment on it is unambiguous: "the calibrated value of the slope, β = 1.69, with standard error 0.0076, is clearly far from Metcalfe's value 2."

Two competent calculations on the same asset, two different answers. So we tested how stable ours actually is.

The Exponent Moves

Window	Addresses with a balance	Active addresses
2011 to 2026	2.069 (R² 0.966)	2.091 (R² 0.891)
2013 to 2026	2.111 (R² 0.920)	2.643 (R² 0.780)
2015 to 2026	2.425 (R² 0.932)	3.227 (R² 0.621)
excluding the 2018 and 2021 peaks	2.058 (R² 0.965)	2.129 (R² 0.890)

Drop the first four years and switch the user measure, and the exponent runs from 2.07 to 3.23. Add the published result of 1.69 and the full range across defensible specifications is 1.69 to 3.23.

What that spread means in practice is easiest to see by asking what a doubling of the network is worth under each answer. At an exponent of 1.69, doubling the network multiplies value by 3.23. At 2.0, by 4.00. At 2.07, by 4.20. At 3.23, by 9.38.

Two honest caveats belong here, and they cut against our own result. First, both series trend upward over time, and regressing one trending series on another produces high R² values even where no causal link exists. A 0.966 is not proof of anything. Second, addresses are not people. One person can hold a thousand addresses and one exchange address can hold a million customers. Every study in this field, including this one, inherits that problem.

Where the Effect Is Not Showing Up

There is one part of the Bitcoin network where the numbers point the other way, and it does not usually appear in network effect arguments.

The Lightning Network's total capacity was roughly 5,718 BTC at the start of 2023. On 30 August 2026 it was 3,753.8 BTC across 32,518 channels and 16,230 nodes. Measured in bitcoin, capacity and channel count have both been falling for years. Measured in dollars the picture is mixed, because the price rose over much of that period.

There are reasonable explanations. Channels have become more efficient, liquidity is managed more actively, and larger nodes have consolidated smaller ones. Those explanations may well be correct. They are also exactly the kind of explanation that gets offered whenever a favoured metric goes the wrong way, so we are flagging the number rather than the excuse.

Our Reading

Our position, stated as ours: Bitcoin's network effect is real, measurable and has not broken once in fifteen years. The exponent people use to convert it into a value is not a property of Bitcoin. It is a property of the window you choose and the users you count.

That distinction matters because the two claims get used interchangeably. "Bitcoin has a network effect" is a statement about usefulness, and the data supports it about as well as this kind of data can. "Bitcoin should therefore be worth X" is a statement that requires a stable exponent, and there isn't one. Our own regression makes that point against our own result: we found the number closest to the popular version, and then found that moving the start date by four years pushes it to 3.23.

The strongest case against us is not that Metcalfe is nonsense. It is that Metcalfe is a good approximation and Bitcoin happens to fit it. Zhang and colleagues showed exactly that for two social networks with a decade of revenue data, and our full-sample estimate of 2.07 is entirely consistent with it. If you believe that, the instability we found is just noise from short sub-samples, and the honest answer is that seventeen annual observations cannot settle it.

What would change our mind: if the exponent held near a single value across different time windows, different user measures and different data providers, the objection would collapse and the number would be worth taking seriously as an input. That test is repeatable by anyone with the series we used.

What we will not do is put a price on the other side of the equals sign. The relationship between adoption and price is the subject of a separate field of modelling, which we examined in our analysis of Bitcoin price models, and it fails for the same structural reason found here: the parameters move.

For the question of what all of this means, rather than what it measures, there is a separate essay on why Bitcoin's starting position cannot be copied. And for the comparison that keeps coming up whenever the store-of-value case is made, our knowledge article on Bitcoin versus gold sets out the two side by side.