Participation
compute → emission

AMA emission isn't a fixed pie split among solvers. The network measures its real solving compute and emits for exactly that — a fixed rate per op, no dilution, with a floor so emission never drops to zero.

Compute target

100 P-AMAFLOPS

Payout floor

10%

2M tiles/s card

≈103 AMA / day

01AMAFLOPS — measuring real work

Every solver crunches tiles — small pieces of one big puzzle. A tile is a fixed pipeline: a matrix multiplication — the matmul that powers AI inference, a GEMM in hardware terms — plus a blake3 step, totalling 25,722,880 ops per tile. Each tile is a small slice of a much larger matrix multiplication — block (tiled) matmul, the same way every GPU GEMM kernel works — and tiles combine to multiply a very large matrix piece by piece.

The work per tile never changes, so counting tiles is counting compute — via a blake3 score. We call the unit AMAFLOPS — like FLOPS, but not standard floating-point ops, because of that blake3 step. No spec sheets, no benchmarks, no trust: the chain measures it from the solutions themselves.

2,000,000 tiles/s × 25,722,880 ops = 5.14×10¹³ = 0.0514 peta-AMAFLOPS — one RTX 5090-class card

One consumer card. The chain doesn't care what silicon it is — it counts the tiles.

02Where the emission goes

Each epoch (~13.9 h) the protocol emits a fixed, slowly-decaying amount of AMA. Half funds vault yield. The other half — the solver half — pays solvers. But it only pays out participation % of it: the network's measured compute as a percentage of the 100 peta-AMAFLOPS target. Whatever isn't paid accrues — it is never handed to a lucky few solvers on an idle network. Accrued coins aren't burnt — they're reserved to fund future payouts, smoothing the payout curve as the network grows: early restraint becomes later stability.

The solver half is gated by participation. On a quiet network most of it accrues instead of over-rewarding a few cards.

03A fixed rate per op — no dilution

Here's the part that makes AMA different from classic PoW pools. Between the floor and the target, your reward is simply your compute's share of the 100 peta-AMAFLOPS target. Total network tiles cancel out of the math entirely:

reward = solver_half × (participation / 100) × (your tiles / network tiles) = solver_half × your P-AMAFLOPS / 100

More solvers joining raises participation — the network pays out more total — but your rate per op does not change. And that cuts the other way too:

Paid per compute1 P-AMAFLOPS always earns 1% of the solver half while the band holds.
No dilutionRun 1 GPU, add 100 more → 101× the ops, 101× the emission. Nobody else's reward moves.
10% floorBelow the floor the fixed 10% budget is spread over fewer ops — each op earns more.

reward, 2M tiles/s cardfloor bonus (<10)fixed-rate band (10–100)dilution (>100)

Reward for one 2M tiles/s card as the network grows — hover the chart to read values.

Card solving at 2M tiles/s (RTX 5090 class)	Raw reward
per day	≈ 103 AMA
per epoch (~13.9 h)	≈ 60 AMA

An epoch lasts ~13.9 hours, so a day covers about 1.7 epochs. Raw protocol rewards at current emission, before pool fees. The rate drifts down only with the emission curve — a slowing decay, ≈0.16%/epoch today and ~0.10%/epoch by epoch 1200 — never because someone else plugged in.

04From one card to a datacenter hall

Because the rate is fixed per unit of compute, scaling is boring — in the best way. It doesn't matter what the silicon is, only how many tiles it solves: any card doing 10× the tiles earns 10× the reward. A rack multiplies that again, a hall multiplies the rack. Revenue planning becomes multiplication, not game theory about what everyone else is solving.

A more powerful card is just more tiles per second — a B100, or whatever comes next, earns the same rate multiplied.

Cluster operators can price AMA solving like they price any other workload: per petaflop, per hour.

05Not just GPUs — any silicon that can solve

AMAFLOPS doesn't care about the vendor logo. The workload is a matmul + blake3 pipeline, and anything that runs it fast counts the same: gaming cards, datacenter accelerators, or meshes of dedicated AI silicon like Tenstorrent chips wired into a grid. Whatever wins on ops-per-joule wins the margin — the protocol just measures the output.

Heterogeneous compute, one honest meter.

06Why this is the future of AMA

Classic PoW turns miners against each other: every new card makes every old card poorer, so mining and hashing power chase price in violent cycles. AMA's participation model replaces that with a simple contract: the network emits for compute at a posted rate, floor 10%, target 100 peta-AMAFLOPS.

That has three consequences worth sitting with:

Predictable economicsA card, a rack, or a hall earns a computable AMA flow. Compute providers can commit capacity with spreadsheet certainty.
Aligned growthGrowth doesn't tax existing solvers — it raises participation and pays out more emission. Everyone's incentive is simple: add compute.
An honest meterAMAFLOPS is measured from solutions on-chain, not claimed. It's the primitive a settlement layer for AI agents needs: verifiable compute.

The endgame: a network where consumer GPUs, B100 halls and exotic AI silicon all sell compute into the same transparent market, and AMA emission is the clearing price.