Binance published a notice on 13 October. At 07:00 UTC on 10 November 2026 it closes Binance Pool, Cloud Mining, Pool Savings and the Transaction Accelerator, along with the mining-pool API, the Stratum endpoints and the Telegram hashrate alerts. The pages stay up for fourteen days, until 07:00 UTC on 24 November, so people can read their payout history, and then they go. The stated reason is an ongoing review of the product suite.
Every wire story carried the same number. Binance Pool was a pool that “once handled 2.3% of the network”. That figure is wrong in both directions at once, and the way it is wrong is the most interesting thing about the story.
TL;DR
- The 2.3% in every headline is a six-year lifetime average, not a peak. Binance Pool found 22,189 of Bitcoin’s 971,539 blocks, which is 2.2839%. Its actual weekly peak was 13.886% in November 2022, and its share in the week to 12 October 2026 was 0.37%.
- The migration already happened and nobody logged it. The share fell below 12% in February 2023, below 8% in October 2023 and below 4% in June 2024. The 10 November shutdown moves whatever is left of 0.37%.
- Concentration went down, not up. Over the same window the Herfindahl index of Bitcoin pool share fell from 1,674 to 1,592, and the number of pools above 1% rose from nine to ten. The three largest pools absorbed 1.14 points of Binance’s 13.52.
- 126 named pools have already left Bitcoin and 15 of them found more than 5,000 blocks each. BTC Guild peaked at 41.81%, GHash.io at 39.33%. Binance Pool is the 127th exit and the first with a press release.
- The property people actually fear is the one no block records. Nothing in a serialised Bitcoin block tells you who assembled the template. The only pool whose block construction is visibly different is OCEAN, and that is visible only because it pays miners directly in the coinbase.
- Satoshie’s randomness layer is more concentrated than Bitcoin mining, and it matters less. In the 24 hours to 08:37 UTC today, four addresses delivered all 1,815 Chainlink VRF fulfilments on Base. That is a Herfindahl of 2,508 against mining’s 1,592. Those four decide when you get a number. They cannot decide which.
Three numbers that are all called “Binance Pool’s hashrate”
Start with the one that is checkable. mempool.space attributes 22,189 blocks to Binance Pool out of 971,539 mined in Bitcoin’s history. That is 2.2839%, and it is the number sitting on the pool’s own page under “block share, all time”. It is an average over six and a half years of a share that started at nothing, climbed into double digits, and came back down. It is not a peak and it is not a current reading. It got into the headlines because it is the figure the explorer shows you first.
The peak is on the same site, one endpoint away. In the week to 21 November 2022, Binance Pool’s share of blocks was 13.886%. Its yearly averages run 6.35% in 2020, 10.25% in 2021, 11.62% in 2022, 8.69% in 2023, 3.31% in 2024, 1.94% in 2025 and 1.98% so far in 2026. So the headline number understates what Binance Pool once was by a factor of six, and overstates what is leaving on 10 November by a factor of six. One number, wrong twice, in opposite directions.
The second figure in the coverage was a 65% decline in computing power since December. Here is what that depends on. Binance Pool found zero blocks in the last 24 hours. It found 8 in the last week, which at 0.745% of blocks against a network running about 1,012 EH/s implies roughly 7.5 EH/s. The weekly bucket for the week of 12 October implies 3.77 EH/s. Its December 2025 weekly readings were 23.52, 19.09, 38.00, 22.11 and 23.16 EH/s, averaging 25.18.
So the decline since December is 85%, or 70%, or 100%, depending entirely on which window you put in the denominator and the numerator. The published 65% is a defensible choice among several. It is also a reminder that a hashrate figure derived from block counts at sub-1% share is mostly variance: this pool’s own December readings swung from 19.09 to 38.00 EH/s in consecutive weeks without anything happening.
The move finished two years ago
The thing that makes 10 November a non-event is that Binance Pool’s hashrate left a long time ago, gradually, with no announcement at any point.
The last week in which it held 12% or more of blocks was 13 February 2023. The last week at 10% or more was 17 July 2023. The last at 8% was 30 October 2023, at 6% was 8 January 2024, and at 4% was 3 June 2024. By mid-2024 the pool was already a rounding error relative to what it had been, and it stayed in the 1.5% to 3% band for two further years before falling off in the last fortnight.
Which means the interesting question was never “where will the hashrate go on 10 November”. It is “where did 13.5 points of hashrate go between November 2022 and now, while nobody was writing it down”.
Where it went
Comparing weekly pool shares at the 21 November 2022 peak against the week to 12 October 2026 gives this.
| Pool | Nov 2022 | Oct 2026 | Change |
|---|---|---|---|
| SpiderPool | 0.00% | 6.47% | +6.47 |
| SECPOOL | 0.00% | 4.07% | +4.07 |
| MARA Pool | 1.60% | 5.36% | +3.76 |
| OCEAN | 0.00% | 2.40% | +2.40 |
| Luxor | 1.00% | 3.14% | +2.14 |
| F2Pool | 16.38% | 18.21% | +1.82 |
| ViaBTC | 9.29% | 9.61% | +0.32 |
| Foundry USA | 26.77% | 26.52% | -0.25 |
| AntPool | 19.38% | 18.95% | -0.43 |
| Braiins Pool | 4.70% | 1.48% | -3.22 |
| Binance Pool | 13.89% | 0.37% | -13.52 |
The three largest pools in the world, between them, picked up 1.14 points of the 13.52 Binance shed. The largest, Foundry, got none of it and is marginally smaller than it was. The share went to SpiderPool, SECPOOL, MARA, OCEAN and Luxor, four of which were either nonexistent or statistically invisible when the migration started.
The concentration measures follow from that. The Herfindahl index across named pools fell from 1,674 to 1,592. If you treat every unattributed block as a single hidden bloc, which is the most pessimistic reading available, it fell from 1,679 to 1,597. The count of pools holding at least 1% of weekly blocks rose from nine to ten. On every standard measure, the four years in which Bitcoin’s third-largest pool collapsed were four years in which Bitcoin mining got slightly less concentrated.
I tried to break that result
The obvious objection is that the new names are fronts. If SECPOOL and SpiderPool are back doors into an incumbent, the Herfindahl number is a decoration. So I pulled the last 100 blocks from every pool that found at least 20 blocks in the past year, 2,070 blocks across 21 pools, and ran three tests.
Coinbase payout addresses. Every address paid by the coinbase transaction of all 2,070 blocks, grouped by pool, then intersected pairwise. Across all 210 pool pairs, the number of shared payout addresses is zero.
Merged-mining jobs. Thirteen of the 21 pools merge-mine, writing the fabe6d6d marker and a 32-byte auxiliary block hash into the coinbase. If two pools took their auxiliary work from the same aggregator they would sometimes commit to the same auxiliary block. Across 1,197 auxiliary hashes there is exactly one repeat, and it is ViaBTC reusing its own stale auxiliary job across five of its own blocks. No cross-pool reuse at all.
Coinbase structure. This one found something, and it is weaker than it looks. AntPool, SECPOOL, Luxor, Poolin, BTC.com, ULTIMUSPOOL, WhitePool, RedRockPool, ViaBTC and Binance Pool itself all write a merged-mining merkle size of 16 and a nonce of zero, with near-identical byte layout and padding. F2Pool uses the same merkle size with its fish emoji stuffed into the nonce field. Foundry, MARA, OCEAN, Braiins, SBICrypto and NiceHash do not merge-mine at all, and SpiderPool does so in only 57 of its last 100 blocks.
That is a software family, not an ownership chart. Pools fork each other’s stratum implementations and have done for a decade. It is worth noting precisely because Binance Pool was already in it: hashrate moving from Binance Pool to SECPOOL does not change the template software at all, while hashrate moving to SpiderPool, MARA or OCEAN does.
None of these three tests proves independence. They are failures to find dependence, which is a weaker thing, and two pools can settle with each other off-chain in ways no coinbase will ever show. But the cheap version of the cynical story, that the share quietly went back to the same desks, does not survive contact with the block data.
This has happened 126 times already
mempool.space names 157 pools that have found at least one Bitcoin block. 31 of them found a block in the past year. The other 126 have stopped. Among pools that ever found more than 5,000 blocks, 15 of 26 are gone.
The list includes every pool that has ever frightened Bitcoin. BTC Guild peaked at 41.81% of weekly blocks in May 2013 and found its last block in July 2015. GHash.io, whose 2014 run at a majority of hashrate triggered the only genuine network-wide centralisation panic in Bitcoin’s history, peaked at 39.33% on this weekly measure in June 2014 and stopped in October 2016. BitFury peaked at 18.65%, BTC.TOP at 16.44%, Huobi Pool at 12.28%, Poolin at 18.81% and BTC.com at 28.95%. Poolin and BTC.com are technically still alive, at 0.28% and 0.92%.
Binance Pool will be the 127th name to leave, and as far as I can tell it is the first to leave with a notice, a timestamp and a fourteen-day grace period for your records. Every other departure was a chart that went flat.
The half nobody is measuring
Here is the thing that makes all of the above beside the point.
Nobody actually fears a pool’s share. What people fear is the thing the share is a proxy for: a pool operator decides which transactions go into the block and in what order, and at enough share that becomes a power over what the chain will accept. Share is the measurable stand-in. Template authorship is the property.
And template authorship is the one thing a block does not record. A block assembled by a pool operator and a block assembled by an individual miner under Stratum V2 job declaration serialise to the same bytes. There is no field, no flag, no version bit. I went looking for one across 2,070 blocks and there is nothing to find.
In May this blog published a post celebrating seven pools joining the Stratum V2 working group at roughly 75% of global hashrate, and told the gaming industry it should be embarrassed by comparison. Five months later I can report two things about that post. The five of those seven that actually run pools, Foundry, AntPool, F2Pool, SpiderPool and MARA, still total 75.51% of weekly blocks, so the headline figure held. And DMND, one of the seven named, has found one block in Bitcoin’s entire history.
There is a third thing in the table above that the May post would not have predicted. Braiins Pool, which wrote Stratum V2 and has pushed it harder than anyone, is the second-largest loser of the window: 4.70% down to 1.48%, a 69% fall in relative share while Binance Pool was collapsing beside it. Whatever moved 13.5 points of hashrate around this market over four years, an open standard for block construction was not it.
What I cannot report, from any public data, is whether a single block in that 75.51% was assembled by the miner rather than the pool. That is not a gap in my research. It is a gap in what blocks contain. The decentralisation shift everyone applauded is the only one on this list that cannot be audited from the chain, and I applauded it on a press release.
What the chain does show
One pool is visibly different, and only because of how it pays. OCEAN’s coinbase transactions pay between 9 and 64 addresses per block, 218 distinct addresses across its last 100 blocks, and mempool.space attributes 22 individually named miners inside those blocks. Across the other 20 pools and 1,970 blocks, every single coinbase paid either one or two addresses. Never three.
OCEAN is also the only pool whose blocks diverge measurably from what the public mempool predicted: its average match rate is 91.65%, more than five points below the next lowest, where every other large pool sits between 96.74% and 99.96%. ViaBTC does something adjacent but cosmetic, writing the individual miner’s username into the coinbase tag, 66 distinct usernames in its last 100 blocks, which is a label rather than a payment.
Be careful with what that proves. Paying miners directly in the coinbase is a payout decision, not a template decision, and a low match rate is consistent with miner-selected templates and with several duller explanations. The honest summary is narrower and worse: of the two properties people care about, the chain lets you measure the one that moved 13.5 points harmlessly, and does not let you measure the one that would actually matter.
We have the same shape, one layer up
Satoshie does not run on Bitcoin and no pool decides any of our draws. But the structure of the problem transfers exactly, and it would be dishonest to run this argument without pointing it at ourselves.
Our raffles and coinflips resolve on Base using Chainlink VRF. So this morning I pulled every log the VRF v2.5 coordinator 0xd5d517abe5cf79b7e95ec98db0f0277788aff634 emitted across Base blocks 52,588,062 to 52,632,061, the 44,000 blocks ending at 08:37:49 UTC today, roughly 24 hours. That window contains 1,815 RandomWordsRequested, 1,815 RandomWordsFulfilled and 1,815 L1GasFee events. Nothing unmatched in either direction.
Those 1,815 fulfilments arrived in 1,810 transactions sent by exactly four addresses, carrying 475, 472, 450 and 412 of them, and 1,809 of the 1,810 went through the same BatchVRFCoordinatorV2Plus contract at 0xd93077894a1b77e2395ae65572433a524ed5a0c5. The headcount is unchanged from the measurement I ran on 5 October, where a fifth address appeared once and never again.
Run the same index on that set and the answer is 2,508. Satoshie’s randomness delivery layer is 58% more concentrated than Bitcoin mining. Four addresses, no notice period, no press release if one stops.
The difference is not the number. It is what the number can reach. Last week I replayed a real fulfilment transaction against the coordinator with single bits flipped, and every edit was rejected on chain: a tampered proof fails the _uWitness check, a tampered seed returns NoCorrespondingRequest(), a tampered commitment returns IncorrectCommitment(). The proof is verified before the callback is allowed to run. So those four addresses control exactly one variable, which is when a number arrives, and the worst case is a stalled draw rather than a chosen winner.
Pool concentration is the other way round. The block is valid whoever assembled it, and the question of what went into it and in what order has no on-chain answer at all. We measured the same gap on Base and found 71% of transactions sitting behind a cheaper one in the same block.
The claim is narrow and I want to keep it narrow. Satoshie is not more decentralised than Bitcoin mining. We measured, and on the one index both systems share, we are considerably worse. The claim is that our fairness property is the kind you can check from chain data alone, and Bitcoin’s template-authorship property is not, and that difference does not show up in any concentration number either of us publishes.
Honest limits
Four, and the first is the biggest.
Pool attribution is a heuristic. Every figure here rests on mempool.space matching coinbase tags and known payout addresses to pool names. A pool that wants to be invisible can be. 220,690 of Bitcoin’s 971,539 blocks are attributed to nobody, although in the past year unattributed blocks were only 0.546% of the total.
Weekly hashrate at sub-1% share is noise. I used two different windows above and got 3.77 and 7.54 EH/s for the same pool on the same day. Any single-week EH/s figure for a small pool, including mine, should be read as an estimate with a wide interval.
100 blocks per pool is a sample. The address-overlap and merged-mining results cover the most recent 100 blocks each, not every block those pools ever found. A shared address six months ago would not appear.
Falling concentration is not the same as safety. A Herfindahl of 1,592 with one pool at 26.5% is still a market where one operator assembles a quarter of all blocks. The index went the right way. It did not go anywhere good.
Three questions worth asking
When a provider in your stack leaves, does anything in your system notice, or does somebody have to read a notice?
For each property you tell users they can verify, can they verify it from data the chain emits, or only from something the provider chooses to publish?
If your provider set shrank from four to three tomorrow without an announcement, how long before you knew, and would your users find out from you or from a chart?
The dial and the thing
A pressure gauge tells you the pressure in the pipe it is screwed into. It is accurate, it is cheap to read, and it is not the thing you care about. What you care about is whether the pipe holds, and the gauge only speaks to that if someone has worked out the relationship between the two and written it down.
Bitcoin has a very good gauge on pool share. It read 13.886% in 2022 and reads 0.37% today, and it has been quietly correct the whole way down while a hundred and twenty-six other needles dropped to zero without anybody filing a story. On 10 November one of them gets a press release. The needle it moves is already at the bottom of the dial.
The pipe is template authorship, and there is no gauge on it. That is the half of this story nobody is measuring, and it will still be unmeasured on 11 November.
Satoshie runs provably fair raffles and coinflips on Base, with every draw resolved by Chainlink VRF and every proof verified on chain before a winner is paid. Come and check one at satosh.ie.
📷 Photo by Miguel Ángel Padriñán Alba on Unsplash
