Three Pools Build 61% of Bitcoin’s Blocks

Over the seven days to 4 August 2026, Bitcoin produced 1,060 blocks, and three mining pools built 60.75% of them.¹ Foundry USA alone built 26.04%. Add F2Pool at 17.83% and AntPool at 16.89% and you have a supermajority of the week’s blocks assembled by three organizations. The top five reach 76.89%. Nineteen pools found a block at all; ten of them found fewer than twenty between them.

This is not a one-week artefact. Across the trailing year — 52,437 blocks and 33 distinct pools — the top three still account for 58.34%, with Foundry USA at 28.90%. The concentration measure barely moves: a Herfindahl index of 1,484 over the week against 1,536 over the year. Whatever else changed in Bitcoin mining over twelve months, this did not.

The usual dismissal, and why it only half works

The standard reply is that pool share is not hash power. Miners point their machines wherever they like and can leave a pool in an afternoon, so a pool with a quarter of the blocks does not own a quarter of the network. That is correct, and it is a good reason not to describe Foundry as controlling 26% of Bitcoin.

It is also an answer to a question nobody sensible is asking. The reason pool share matters is not ownership of hardware, it is authorship of blocks. The pool, not the miner, assembles the candidate block — which transactions go in, in what order, and which do not go in at all. On that question the numbers mean exactly what they appear to mean: three entities currently decide the contents of three in every five Bitcoin blocks. Our position is that the industry keeps rebutting the ownership claim because the authorship claim is harder to answer.

The week, in full order

• Foundry USA — 276 blocks — 26.04%
• F2Pool — 189 — 17.83%
• AntPool — 179 — 16.89%
• SpiderPool — 93 — 8.77%
• ViaBTC — 78 — 7.36%
• MARA Pool — 52 — 4.91%
• SECPOOL — 41 — 3.87%
• OCEAN — 37 — 3.49%
• Luxor — 33 — 3.11%
• Braiins Pool — 21 — 1.98%

What this does not show

It does not show common ownership. Pools are counted by the tag they write into the coinbase transaction, and that tag says nothing about who ultimately controls two separately branded pools, who supplies their hardware, or who they share infrastructure with. If any two of the names above answer to the same parent, every concentration figure here understates the position, and this dataset cannot detect that. We are not alleging it — we are saying the measurement is blind to it.

Attribution itself is a convention rather than a proof. A coinbase tag is a string a miner chooses, and blocks that carry no recognizable tag land in an “unknown” bucket, which took 1.06% of the week. One week is also 1,038 samples, so the smaller entries carry real sampling error: a pool credited with four blocks could plausibly have found two or eight. The top three are far outside that noise; the bottom of the table is not.

Finally, none of this measures whether the concentration has ever been used. Building a block and censoring a transaction are different acts, and the register of who built what contains no evidence about the second. We will re-run both windows on 3 November 2026. The figure we are watching is whether the top three hold above 58%.