igneum-network/igneum
1
0
Fork 0
Code Issues Pull requests Releases Activity
igneum/docs/analysis/mission/future.md

80 KiB
Raw Blame History

The last mission, lane 3: the far future, 2030 to 2036

Written 7 October 2026 by the future lane. Scope: the ten-year axes Horizon lane 7 did not cover, and the years 2030 to 2036 on the axes it did. Lane 7's rows to 2030 (docs/analysis/horizon/frontier.md 2.1 to 2.6) are taken as given and not repeated; where this lane disagrees, the row is named and the reason sourced. Every figure from memory is labelled approximate. Every figure from a source names it, with a URL in the sources list and the access date (all accessed 7 October 2026 unless stated). Times are UK. The arithmetic behind sections 2, 3, 4, 6 and 7 is model.py in this lane's scratch directory (scratchpad/mission-future/), reproduced in the tables.

The design this lane tests (horizon-2026-10.md section 1): an hourly random GPU program, a 256 MiB cache over a daily multi-GB dataset (2 GiB at genesis, doubling at years 4, 12, 28), latency-shadow work in a six-rung N ladder stepped by 90 percent miner signal, class v5 (dataset from chain state) as a candidate, miner-only finality with BLS vote keys weighted by 30 days of blocks, SP1 zkEVM proving by miners (the 20 percent pool), 100 IGN a block gliding to a 1 percent tail, no stake, no dev fund, no other chain in consensus.

0. Progress

Time (UK) State
08:2x Brief read: CLAUDE.md, frontier.md 2 and 6, algorithm.md 5, chip-model-v3.md 5, horizon-2026-10.md 1, finality-in-proof.md 4, 51-percent.md, bench-log rental entry
08:3x to 08:5x 50 web searches (the session's search budget ran out at 50; the rest went by direct fetch), 27 primary fetches
08:5x Model run (model.py): emission by year, chip thresholds, rental curve, heat credit, PQ bytes, wallet battery
08:5x to 09:00 Sections 1 to 11 written, 566 lines, copy-law check clean (no em or en dashes, ASCII only)
09:00 DONE. File handed to the coordinator. Nothing committed; nothing on the devnet touched; no build or measurement run

The tiers, used in every row: home miner with one 8, 12, 16, 24 or 32 GB card; a rig; a pool user; on Windows, Linux, macOS; NVIDIA, AMD, Apple.

1. GPU and memory roadmaps, 2027 to 2036

1.1 The roadmap, sourced

Item What is announced or reported Source Label
HBM4 Mass production at Samsung and SK hynix from February 2026; 2,048-bit interface, 8 Gbps a pin, about 2 TB/s a stack; 12-high 36 GB; SK hynix 16-high 48 GB from Q3 2026; Micron samples over 11 Gbps, 2.8 TB/s TrendForce 9 Jan 2026; EE Times CES 2026; Astute Group cited
HBM4 price About USD 550 a 36 GB stack, USD 15.3 per GB (factory gate) siliconanalysts.com/tools/hbm-analysis approximate (the site cites no primary)
HBM4E Late 2027 to 2028; 14 to 16 Gbps a pin, 3.6 to 4.0 TB/s a stack; 16-high; custom base dies on TSMC N3 (the GUC and TSMC "C-HBM4E" line, 12.8 GT/s by 2027); Samsung HBM4E samples May 2026 at 3.6 TB/s TrendForce 23 Dec 2024; Tom's Hardware (TSMC/GUC); TechTimes 30 May 2026 cited
HBM on a consumer card None announced. The Feynman datacentre architecture (2028) "supports HBM"; no consumer HBM part from NVIDIA, AMD or Intel on any roadmap found Wikipedia Feynman page; the 2027 to 2028 rumour set cited (absence)
GDDR8 No JEDEC standard. SK hynix's roadmap to 2031 lists "GDDR7-Next" for 2029 to 2031 Tom's Hardware (SK hynix roadmap); TechSpot cited
GDDR7 devices 2 GB ended at Micron (Sep 2026); 3 GB shipping at USD 60 to 70; 4 GB and 6 GB devices reported for 2027 to 2028 chip-model-v3 5.1; wccftech; club386 3 GB cited, 4 and 6 GB rumour
RTX 60 (Rubin GR20x) Late 2027 slipped to 2028; GDDR7; the 6090 reported at 512-bit with 32 GB or 48 GB TweakTown; wccftech; BigGo rumour
AMD RDNA 5 / UDNA Mid-2027 to 2028; GDDR7 at 36 Gbps; flagship "AT0" 154 CUs, 36 GB on 384-bit, 1.7 TB/s, 380 W; shares a chiplet design with the next Xbox; GDDR7 support landed in the Linux driver TweakTown; TechPowerUp; Tom's Hardware driver note rumour, driver patch cited
NVIDIA consumer chiplets Nothing reported; Rubin consumer parts described as monolithic the same rumour set approximate
Strix Halo class Strix Halo: 256-bit LPDDR5X-8000, 256 GB/s. Medusa Halo (2027 to 2028): LPDDR6 on 256-bit (461 GB/s) or 384-bit (691 GB/s) VideoCardz; hardware-corner.net rumour
Apple M5 Max: up to 128 GB unified, 614 GB/s (40-core GPU), 460 GB/s (32-core); M5 Ultra Mac Studio August 2026 Apple newsroom 3 Mar 2026 and Aug 2026 cited
LPDDR6 JESD209-6 published July 2025; 2 sub-channels a die, 12 DQ each, 4 CA each; activate timings not public JEDEC press release cited; timings unknown
DDR5 32 banks in 8 groups of 4 (x4/x8); JESD79-5D Nov 2025; tFAW a four-activate window JEDEC; DDR5 core datasheet cited
The latency floor "The latencies of three fundamental DRAM operations have not improved significantly in the past 18 years"; improvements "relatively stagnant for the last two decades" Lee et al. (arXiv 1604.08041); Chang et al. (arXiv 1805.03154) cited

1.2 What it means for the memory-latency-bound hash

The hash advances one dependent 4-byte read per memory latency; the rate is activates per tFAW window times channels, and energy is per activate (chip-model-v3 5.3). Pin speed does not move it. So the ten-year question is only: do channels per watt per dollar move, and does tRC or tFAW move. The sources say tRC has been flat for about twenty years, and no DRAM roadmap to 2031 (SK hynix) names a row-cycle improvement. HBM4 doubles channels per stack (lane 7, 2.3). HBM4E adds pin speed and a custom base die, not channels. GDDR7-Next is 2029 to 2031 and unspecified. Consumer HBM does not exist on any roadmap to 2028.

Year Flagship consumer memory (projected) Random reads per second, flagship (approximate) Mid-tier card (12 to 16 GB) Tier that wins or loses against 2026
2026 32 GB GDDR7, 512-bit, 16 devices 17.5 G measured (5090) 12 GB, 192-bit, 6 devices: about 6.5 G baseline
2028 36 to 48 GB GDDR7, 384 to 512-bit (RDNA 5 rumour, RTX 60 rumour) 16 to 21 G: the same, capacity adds no channels 16 to 18 GB on the same channel count (3 GB devices): the same rate nobody: a 2026 card keeps its rate against a 2028 card
2031 48 to 64 GB GDDR7-Next (SK hynix window) unknown; if channels per device rise to 8 (approximate guess), 1.5 to 2x 24 GB mid-tier on the same count the 2026 owner falls to 0.5 to 0.7x of the new card, the normal GPU cadence
2036 64 to 96 GB (extrapolation of lane 7's 1.167 a year); HBM on a halo consumer part possible but unannounced 2 to 4x of 2026 if a consumer HBM4-class part ships; otherwise 1.5 to 2x 32 GB mid-tier the 8 and 12 GB tiers are gone from the installed base (Steam trend, algorithm.md 5.6), not from the hash

Apple and APUs: the M5 Max's 614 GB/s is a 512-bit LPDDR5X bus (approximate); LPDDR activates are the same DRAM physics, so the Apple tier stays "0.78 uJ per hash, 21 W" class (algorithm.md 5.3a), the best per joule and the worst per dollar (USD 129 per MH/s). Medusa Halo on 384-bit LPDDR6 would be a 24 to 36 channel part (approximate: 2 sub-channels a die): a mid-tier card's read rate at laptop watts. Consequence per tier: Apple and APU miners stay the per-joule leaders and never the per-dollar ones; nothing in the hash changes that in ten years.

1.3 The cache, the dataset schedule and the ladder, re-read against the roadmap

Design item Roadmap fact Verdict Recommendation
256 MiB cache "above every GPU's on-die cache" (the spec's rule) 5090 L2 96 MB, GB202 128 MB; MI300X carries 256 MB Infinity Cache (datacentre, approximate from memory); RDNA 3's 7900 XTX 96 MB; no consumer part at 256 MB announced Safe to 2028. At risk from 2029 to 2031 if a consumer part ships a 256 MB last-level cache, which the datacentre already does Add a cache-size rung to the era-draw ladder beside N (256 to 512 MiB), stepped by the same 90 percent signal, in place of the fixed year-4 doubling alone; the trigger is a shipped consumer part with LLC at or over the cache size
Dataset 2 GiB, doubling at years 4, 12, 28 Card memory 32 GB now, 48 to 64 GB by 2030 (lane 7); one HBM3 stack holds 24 GB Right. The dataset is not a lever against the stored-dataset chip (chip-model-v3 5.7) and never binds a tier before year 12 (algorithm.md 5.6) Hold the schedule; the public card-lifetime sentence should carry the prover footprint, not the dataset (algorithm.md 5.6's one change)
N ladder rungs measured on 2026 cards The honest card's bind point moves with each generation: a 2028 card with the same read rate and 1.5x the ALUs binds at a higher N; HBM4 doubles the chip's rate per stack Rungs are a 2026 measurement. They are the right shape and the wrong numbers for 2029 Re-measure the rungs per card generation (the Steam top-10 cards each era) and publish the bind points; the signal mechanism already lets miners refuse a rung their cards cannot hold, so no genesis change, only a measurement duty written into the era-draw docs
C-HBM4E custom base die (2027 to 2028) The base die under the stack becomes a logic die on N3 that a customer designs (TSMC and GUC) This is the f = 1 chip's controller moved under the memory: the chip-model's "controller and PHY die beside it, 10 W, USD 50, plus a USD 200 one-stack interposer" row (5.3) collapses into the base die. Lane 7's 2.3 did not price it Disagreement with lane 7 row 2.3 "HBM4 one stack (2028)": the controller and interposer lines fall toward zero, so the HBM4 chip's dollars per MH/s fall below the USD 2.8 GDDR7 figure by 2028, approximate. The answer is unchanged in kind (N, the price per joule) and larger in degree; section 2 carries it

Per tier, section 1 in one line each: 8 GB (mines to year 12, never proves beside its miner); 12 GB (mines to year 28, loses mine-and-prove at year 4); 16 GB (mines and proves to year 12); 24 and 32 GB (unconstrained to 2036 on every roadmap found); rig (the rate per card is flat through 2028, so a 2026 rig is a 2028 rig); pool user (the pool's share tracks the installed base, which loses the 8 GB tier by 2030); Windows and Linux (no change); macOS (per-joule best, per-dollar worst, both for ten years); NVIDIA (channel count flat to 2028); AMD (RDNA 5 brings GDDR7, a 36 GB flagship: AMD's first competitive random-read part since the 9070 XT's 2.4 to 2.7 G); Apple (as macOS).

2. The ASIC maker's economics, 2026 to 2036

2.1 Inputs

Input Value Source
Mask set, total NRE: TSMC 28 nm USD 1 M, 1.8 M siliconanalysts.com/data/wafer-pricing (Sep 2026)
16 nm 1.8 M, 3.2 M same
7 nm 3.5 M, 5.5 M same
5 nm 6.5 M, 10 M same
3 nm 15 M, 22 M; design cost of a 3 nm chip USD 400 to 600 M all-in same; siliconanalysts tsmc-3nm-cost
2 nm masks USD 15 to 30 M; design cost quoted at USD 724 M semiwiki thread; siliconanalysts
A16, A14 no mask quote found; "capex per 1,000 wafers at A14 higher than N2" semiwiki A14 thread
Wafer, 300 mm 28 nm 3,000; 16 nm 5,500; 7 nm 9,500; 5 nm 20,000; 3 nm 20,000 (range to 27,000) siliconanalysts wafer-pricing
Leading-edge tapeout, all-in USD 30 M to 100 M+ at 3 to 5 nm; 5 to 30 M at 7 to 28 nm siliconanalysts tapeout guide, 1 Mar 2026
The f = 1 chip 166 MH/s, 78 W bare, 228 W with a 150 W shadow core at k = 1; USD 470 of memory, controller and board; USD 2.8 per MH/s chip-model-v3 5.4, lane 7 2.3
The honest card 5090 at class v4: 3.27 uJ, 431 W cap, 132 MH/s; USD 1,999 MSRP algorithm.md 5.3a
Emission (model) 100 IGN a block, 90-day ramp from 10 percent, monthly glide at 2.9 percent, tail 1 percent a year from year 11.4 tail-emission.md
Rental equilibrium hash joins until rent equals subsidy: 39, 156, 780 GH/s at USD 0.005, 0.02, 0.10 per IGN security-budget.md via 51-percent.md

Emission by year from the model (IGN, approximate): year 1 2.32 B, year 2 1.87 B, year 3 1.31 B, year 4 0.92 B, year 5 0.65 B, year 6 0.46 B, year 7 0.32 B, year 8 0.23 B, year 9 0.16 B, year 10 0.11 B; supply 4.18 B at the end of year 2, 7.07 B at year 5, 8.33 B at year 10.

2.2 The decision tree, written out

The maker chooses a target share s of the hash and a node. Revenue over the chip's two-year life is s x E2 x p, where E2 is the two-year emission and p the IGN price. The fleet needed is s/(1 - s) x H(p), and at the rental equilibrium H(p) = 7.8 M MH/s per USD of price, so the fleet's cost is also linear in p: 0.43 x 7.8 M x USD 4.64 per MH/s (the chip's two-year cost per MH/s with power at USD 0.05 per kWh, model) against two-year revenue per MH/s of 0.0117 x 17,520 = USD 205. The fleet term is 2 percent of revenue and drops out. The project pays when p is over p* = C_proj / (s x E2), and the market cap at which it pays is p* x supply. At s = 0.30 (an economic miner just under the veto third):

Node (project all-in) Years 1 to 2 (E2 = 4.18 B) Years 3 to 4 (2.24 B) Years 5 to 6 (1.10 B) Years 7 to 8 (0.55 B) Years 9 to 10 (0.27 B)
28 nm controller only, no shadow core (USD 5 M) p* 0.0040, cap USD 17 M 0.0075, 48 M 0.0151, 114 M 0.0306, 247 M 0.0620, 517 M
28 nm controller + N5 shadow core (USD 30 M) 0.0239, 100 M 0.0447, 287 M 0.0906, 682 M 0.1836, 1,481 M 0.3718, 3,099 M
N3 single die, controller + shadow + lanes (USD 60 M) 0.0478, 200 M 0.0895, 574 M 0.1813, 1,363 M 0.3672, 2,961 M 0.7437, 6,198 M
N2 (USD 150 M, 2026 quotes) 0.1195, 500 M 0.2237, 1,436 M 0.4532, 3,408 M 0.9179, 7,403 M 1.8592, 15,495 M
A16 / A14 (USD 250 M, extrapolated) 0.1992, 833 M 0.3729, 2,393 M 0.7553, 5,680 M 1.5298, 12,339 M 3.0987, 25,825 M

Reading. The stored-dataset chip without a shadow core pays at a USD 17 M market cap in the first two years, because its project is a 28 nm controller (chip-model-v3 5.6) and the chip's hour costs 56x less than rented hash. The class v4 shadow core forces an N5-class die (30 mm^2 at N = 100,000, lane 7) and lifts the bar 6x to USD 100 M. The glide lifts every bar about 2.3x per two years, so by years 9 to 10 the same N5 chip needs a USD 3.1 B cap. Nothing here needs N2 or A16: the chip is memory, and a leading node buys it nothing. So the "ASIC maker's economics at every node" collapses to two nodes, 28 nm and N5, and the N ladder is what moves between them.

Minimum volume to break even against buying cards (saving USD 2,131 a chip over two years against 5090s at the same hash, model): 28 nm bare 2,346 chips (0.39 TH/s); 28 nm + N5 shadow 14,077 chips (2.34 TH/s); N3 28,155 (4.67 TH/s); N2 70,387 (11.7 TH/s); A14 117,312 (19.5 TH/s). Against the rental-equilibrium hash (39 to 780 GH/s at today's three price inputs) the 2.34 TH/s break-even fleet is 3x to 60x the whole network: the chip only pays once the price is high enough for the network to be a few TH/s, which is the p* column above said another way.

2.3 The memory-controller chip against HBM4, per joule and per dollar, 2026 to 2036

Year Memory the chip buys Chip uJ per hash at N = 100,000, k = 1 (approximate) Honest 5090-class uJ Edge per joule Chip USD per MH/s What moves it
2026 GDDR7, 16 x 2 GB 1.37 3.27 2.4x 2.8 the measured row (algorithm.md 5.3a: 2.1x at the 5090's 431 W cap)
2028 HBM4, one stack 36 GB, C-HBM4E base die as the controller 1.33 (algorithm lane's correction of lane 7's 1.11) 3.3 (a 2028 card at the same read rate) 2.5x 2.0 to 2.5 (the interposer and controller rows fall into the base die; approximate) HBM4 price USD 550 a stack falls as HBM4E takes the premium (approximate)
2031 HBM4E or HBM5 (SK hynix lists HBM5 on the 2029 to 2031 window) 1.2 to 1.3 at the same N; 0.9 at N = 100,000 if activates per channel double again 3.0 to 3.3 2.5x to 3.5x 1.5 to 2.0 activate parallelism per stack; unsourced beyond HBM4
2036 the same class the per-joule edge is bounded below by N x 11 pJ x k, so at N = 330,000 and k = 1 the chip pays 3.6 uJ of program work whatever its memory 5.0 at 330,000 (the 5090 is compute-bound there) 1.3x 1.5 N and k only

Does the chip get cheaper or dearer per joule as HBM prices fall: dearer in dollars relative to the GPU through 2027 (lane 7 2.2, memory is 70 percent of its bill), cheaper from 2028 when the base die absorbs the controller, and flat per joule, because per joule is set by activates and by N. Per tier: the home 5090 owner stays inside 2.1x to 2.5x of the chip at N = 100,000 through 2031 and inside 1.3x at N = 330,000; the 4070-class 12 GB owner the same within 10 percent; the AMD 9070 XT owner sits at 5x to 8x behind the chip at every rung (its measured 10.6 uJ) and is the first tier a chip displaces; the Apple tier sits under the chip's per-joule line at every rung. The rig owner is a 5090 owner times eight. The pool user inherits the pool's card mix.

2.4 The FPGA route

AWS F2 (f2.6xlarge, VU47P, 16 GB HBM2, USD 1.98 an hour on demand, USD 0.66 spot) is the measurement the algorithm lane planned (algorithm.md 5.1); nothing has run. The tightened range stands: 2.3 to 2.9 G reads a second a card, 0.30x to 0.47x of the 5090 per watt, at USD 4,000 to 5,000 a card (approximate). Versal HBM and Agilex 7 M-series carry HBM2e (faster pins, the same tFAW), so they sit in the same row. An HBM4-based FPGA (none announced) would carry HBM4's channel count and the 0.3x to 0.5x row would become 0.6x to 1.0x (approximate, derived). Verdict: no FPGA displaces any tier to 2031; the F2 hour should still run, because the per-stack activate rate it measures is the input every row above rests on.

Recommendation for section 2: (1) the N ladder at genesis, as lane 2 and lane 7 said, with the bind-point re-measurement duty of 1.3; (2) write the p* table into the public threat model with the sentence "a stored-dataset chip pays at about USD 100 M of market cap in year 1 and about USD 700 M in year 5 (model, approximate)"; (3) no genesis parameter changes for N2 or A16, because the chip never needs them.

3. AI compute demand and the GPU supply

3.1 The numbers

Row Value Source Label
Datacentre GPUs shipped 2023 3.85 M units (NVIDIA 3.76 M, AMD 0.5 M, Intel 0.4 M) TechInsights via HPCwire, 10 Jun 2024 cited
Datacentre GPUs 2024, 2025 USD 123 B of GPUs and accelerators in 2024, USD 207 B in 2025 (Omdia); NVIDIA estimated 5.2 M Blackwell GPUs in 2025; GB200 cabinet forecasts cut to 25,000 to 35,000 (2.5 M GPUs) Omdia Aug 2025; Tom's Hardware cited, unit counts secondary
Installed datacentre fleet by end-2026 15 to 20 M Hopper and Blackwell class units (sum of the rows above) arithmetic approximate
Consumer AIB shipments Q2 2026 12.5 M units, +10 percent QoQ, +6.6 percent YoY; H1 2026 24.3 M; NVIDIA about 90 percent Jon Peddie Research Q2 2026 cited
Used H100 USD 18,000 to 22,000 in 2026; residual 40 to 75 percent at 36 months, 25 to 35 percent at 60+ months; A100 80 GB USD 12,000 to 18,000 mercatus-ai.com (verified 23 Jun 2026); intuitionlabs cited, secondary
H100 rental USD 1.49 (Vast.ai hosts) to 6.98 an hour; was over USD 7 in early 2024; spot about USD 1.00 cloudzero; spheron; shattered.io (2026) cited
Consumer card rental 5090 USD 0.21 to 0.44 an hour (Vast.ai, 6 Oct 2026); 4090 0.28 to 0.60 lane 7 2.4 cited
Igneum hash rental USD 0.0117 per MH/s-hour (RunPod community pods, 38 pods, 1,748 MH/s for USD 20.44 an hour); 8x 4090 rig USD 0.0129; a 5090 pod 98 to 128 MH/s for USD 0.41 to 0.74 bench-log, 6 Oct 2026 measured

3.2 The used-GPU flood, 2027 to 2030

Hopper fleets bought in 2023 to 2024 reach the 36-month residual cliff in 2026 to 2027 and the 60-month floor in 2028 to 2029. Can they mine Igneum: the hash is memory-latency-bound, so an H100 is its five HBM3 stacks (80 GB) and an A100 its five HBM2e stacks. At chip-model-v3's unmeasured HBM3 ceiling (10.7 G reads a second a stack) an H100 reads 53 G, 3x a 5090; at the JEDEC tFAW ceiling (2.3 G a stack) 11.5 G, 0.65x. Nobody has measured it; the fleet measured A4000, A5000, L4, 3090 and 4090 (prover-tiers-real-cards.md), not an HBM part. The honest statement: an H100 mines Igneum at 0.65x to 3x of a 5090 at 700 W, which is 0.3x to 1.4x per watt (approximate, both ends unmeasured).

Scenario Hash it adds Against the rental equilibrium (39 to 780 GH/s) What the ladder does
1 percent of a 1 M retired H100 fleet (10,000 cards) at 1x a 5090 1.3 TH/s 2x to 35x the whole network nothing: the shadow binds compute, and an H100 has 4x a 5090's ALUs per read; it holds every rung
10 percent 13 TH/s 17x to 330x the same
Rental at USD 1.49 an hour for 0.65x to 3x of a 5090 USD 0.0045 to 0.021 per MH/s-hour at or above today's 0.0117: no cheaper than consumer pods at list price n/a
Idle-time rental (the owner's marginal cost is power) USD 0.00016 to 0.0007 per MH/s-hour at USD 0.05 per kWh 17x to 70x under today's rent n/a

The ladder is a joule argument against a fixed-datapath chip; against a GPU with more ALUs it is silent. The used-fleet question is a price question only.

3.3 The ten-year rental curve and the 51 percent table

The H100 rate fell about 45 percent a year from early 2024 to 2026 (USD 7+ to about 2); consumer-card rent fell as purchase prices rose (lane 7 2.4). Three curves for USD per MH/s-hour from today's 0.0117 (model):

Year At -20 percent a year At -30 percent a year At -45 percent a year
2028 0.0075 0.0057 0.0035
2031 0.0038 0.0020 0.0006
2036 0.0013 0.0003 0.00003

What it does to 51-percent.md: nothing to the dollar cost of any attack, because every row there is priced at the rental equilibrium, where rent equals subsidy. A cheaper rent means more hash joins until rent equals subsidy again: at 2036's -30 percent curve the equilibrium hash is 35x today's at the same IGN price, and the attack costs the same twelve days of emission. What the curve changes is the home card's share of the subsidy, which falls 35x with the hash, and the supply ceiling of the rental market, which stays the real limit (the market gave zero pods when asked for twenty, bench-log). Update the 51-percent price basis each year from a measured rental, not from a curve.

3.4 Per tier: does a home card stay competitive against an idle datacentre card

Cost per MH/s-hour, electricity only, a 5090-class card at 3.27 uJ (model):

Owner and power price USD per MH/s-hour Margin under today's 0.0117 rent Margin if idle datacentre hash sets the rent (0.00016)
UK home, 26.32 p (Ofgem cap, Q4 2026) 0.00114 10x loses 7x
Germany home, EUR 0.387 0.00137 9x loses 9x
US home, 17.7 c (EIA, H2 2025) 0.00058 20x loses 4x
Texas industrial, 5 c (approximate) 0.00016 72x break-even
Paraguay, 4.4 to 6 c (ANDE crypto tariff, Decree 7824/2022, contracts end 31 Dec 2027) 0.00016 72x break-even
Iran, licensed, about 1 c 0.00003 358x wins

The home card is competitive while the marginal supplier of hash is a rented consumer pod at list price (today). It is not competitive the day the marginal supplier is an idle datacentre card on industrial power, and the used-fleet flood of 3.2 makes that day a price event, not a technology event. Per tier: 8 and 12 GB home cards lose first (their uJ is 1.3 to 1.7x worse than the 5090's); 24 and 32 GB cards last; a rig is a home card eight times, on the same power price; a pool user's payout tracks the pool's share, which falls with the home share; Windows and Linux are the same; macOS at 0.78 uJ (M5 Max) holds a 4x power edge over the 5090 and loses last of all the home tiers; NVIDIA and AMD as their uJ rows (algorithm.md 5.3a); Apple as macOS.

What Igneum should do: the reward rule that prices rented hash out (lane 7's 3.1, weight-aged keys paid more) is the only lever in the protocol; the earnings page should show the miner's own cost per MH/s-hour against the network's implied rent, so a UK miner sees the day the line crosses; and the H100 and A100 random-read rate should be measured on one rented card this month (two hours, under the measure lock), because every row of 3.2 rests on it.

4. Energy, heat and the home

4.1 Prices and forecasts

Region Household price 2026 Ten-year direction Source
UK 26.11 p per kWh (Jul to Sep 2026 cap), 26.32 p (Oct to Dec 2026) Cornwall Insight's wholesale path falls to GBP 83 per MWh by 2029 ("over GBP 40 above historic"); retail scenarios 22 to 42 p by 2030 Ofgem; Cornwall Insight (Jan 2024); solarpanelsforfactories (secondary)
EU EUR 28.96 per 100 kWh average H2 2025; Ireland 40.42, Germany 38.69, Belgium 34.99; Hungary 10.82, Malta 12.82, Bulgaria 13.55 The Commission's Electrification Action Plan (17 Jul 2026) targets electricity at most 2.5x gas for households by 2030 Eurostat 5 May 2026; Commission
US 17.7 c average H2 2025; "continue steady increase" AEO2025 reference: 13 c (2024, a different basis) to over 20 c by 2050 EIA
Cheapest mining regions Iran about 1 c (licensed); Ethiopia 2 to 5.3 c; Paraguay 4.4 to 6 c; Kazakhstan about 4 c; Nigeria 4.8 c hosted spark.money; oneminers; hashrateindex Paraguay (4 May 2026) secondary

4.2 Home mining as heating

Product or trial Facts Source
Heatbit Trio, Maxi, Maxi Pro USD 849 (10 TH/s, 400 W mining + 1,100 W resistive) to USD 1,499 (60 TH/s, 1,500 W); seasonal BTC USD 300 to 420 (Heatbit's own figure); Wired's review: mining covers 30 to 40 percent of electricity at 12 to 15 c per kWh miningboard.com; techbuzz (5 Apr 2026)
21energy, MintGreen, HotMine convector radiators 250 to 2,700 W; hydronic boilers for radiators and hot water miningboard.com
Qarnot "radiateur numerique" 100 RIVP social-housing flats in Paris 15e heated by compute radiators from 2013; Qarnot moved to boilers; the model is "the building pays the capital, heat is free" fr.wikipedia Qarnot; maisonapart

The economics per kWh (model): a 5090 at the 431 W cap is a 0.43 kW heater. Credit the heat at the gas price delivered through a 90 percent boiler, or at a heat pump's electricity (COP 3):

Region Electricity Heat credit, gas Heat credit, heat pump Effective price in the heating season
UK 26.3 p 7.0 p (27 percent) 8.8 p (33 percent) 17.5 to 19.3 p
Germany 38.7 c 13.3 c (34 percent) 12.9 c (33 percent) 25.4 to 25.8 c
US 17.7 c 5.6 c (31 percent) 5.9 c (33 percent) 11.8 to 12.1 c

A third off the power bill for the 1,500 to 2,000 heating hours a year (approximate), everywhere. It moves a UK home miner from 7x to 5x the Texas industrial price, not to parity. Consequence per tier: it matters most to the tiers with the worst uJ (8 and 12 GB, AMD) and least to Apple (21 W is not a heater). Recommendation: a heat mode in Ember (run the miner only while a room thermostat or schedule calls, power cap set to the room's load, the hourly program unchanged), which costs nothing in protocol and is the one feature home-mining heaters ship.

4.3 Grid balancing

ERCOT paid Riot USD 31.7 M in August 2023 (24.2 M curtailment credits plus 7.4 M demand response) and Riot booked USD 30.6 M of power-curtailment credits in Q3 2025, up 147 percent on the year (ABC13; Riot releases). The programmes name "large flexible customers"; a home GPU is not one. Home aggregators exist (the UK's supplier demand-flexibility sessions, US utility programmes; approximate, from memory) and pay per kWh shed against a baseline. A fleet of Ember miners is a shed-able load only if a signal reaches it. Recommendation: a curtailment input in Ember (a webhook, a schedule, a price threshold from a public spot feed) that pauses the miner and resumes it; the protocol sees a key that stops voting for an hour, which the LEAVE item of 0.3.16 already prices as nothing (vote-or-burn.md: no burn). Per tier: a pool user's pool must pass the signal down; a rig is the only home tier big enough to enrol directly.

4.4 The rules by region

Region Rule Source What it means for an Igneum home miner What Igneum builds
EU, MiCA White papers must state the consensus mechanism's climate impacts; CASPs must publish the sustainability indicators per asset (Delegated Regulation (EU) 2025/422; mandatory kWh a year, more above 500,000 kWh); Article 142 report on environmental impact and "minimum sustainability standards"; the 2022 Parliament vote dropped a PoW ban; the 2026 MiCA review consultation (reply by 31 Aug 2026) asks only how appropriate the disclosure regime is (Q53) and nothing on PoW MiCA; 2025/422; the consultation PDF (fetched) nothing on the miner; the exchanges that list IGN need the kWh figure publish Igneum's own 2025/422-format indicators from the hash-rate and the measured uJ, updated each era
Norway ban on new PoW mining data centres from autumn 2025; data-centre registry; existing sites run CoinDesk 23 Jun 2025 home mining untouched; no new hosting nothing
Sweden data-centre electricity tax relief removed July 2023 (SEK 0.006 to 0.36 per kWh); SEK 500 M of back-tax on nine firms 2024 to 2026 CoinDesk 14 Apr 2023; crypto.news home miners already paid full tax; hosting is dead nothing
Russia mining banned in 10 regions 1 Jan 2025 to 15 Mar 2031; Moscow region from 15 Aug 2026 to 2032; seasonal bans in Irkutsk, Buryatia, Zabaikalsky; no new regions in 2026 TASS; Cryptopolitan a region check at install; the 2024 law's 6,000 kWh a month household allowance (approximate, from memory) covers one card a region prompt in Ember with the banned list
Kazakhstan mining legal outside the AIFC (Nov 2025 law); "strategic mining" rules from 1 Aug 2026 with a share to the national reserve Caspian News; KuCoin licensed industrial; a home card is below any threshold found nothing
Paraguay ANDE crypto tariff USD 44.33 per MWh (Decree 7824/2022), lifted toward 5.1 to 6 c; every crypto contract ends 31 Dec 2027 hashrateindex 4 May 2026 the cheapest hydro region closes to new load in 2028 nothing
Iran licensed mining at a mining tariff; seasonal shutdowns in power shortages (approximate) ainfp.org; MEXC the cheapest power on earth and the least reliable nothing
China illegal; joint Notice 6 Feb 2026 reaffirmed; one report of Sichuan, Inner Mongolia, Xinjiang reopening from 1 Jan 2026 is uncorroborated lightspark; egw.news (single source) a Chinese home miner runs at legal risk the region prompt
US, federal SEC staff statement 20 Mar 2025: PoW mining, solo and pooled, is not a securities offering; DAME 30 percent excise proposed 2023, 2024, 2025 budgets, never enacted; EIA's emergency survey (Jan 2024) blocked in court (Mar 2024); CLARITY failed cloture 49 to 50 on 15 Sep 2026; GENIUS Act (stablecoins) 2025 Dechert; Blockworks; Fortune; Orrick 2 Oct 2026 nothing on the miner nothing
US, states Texas, Kentucky, Wyoming pro; New York's fossil-PoW moratorium (2022, approximate); California's DFAL licensing from 1 Jul 2026 reaches crypto businesses, not a home card sazmining; crypto.news a home miner is not a licensee anywhere found the region prompt

Per-region electricity-price prompt: the earnings page should default the price per kWh from a public table by country (the Eurostat, Ofgem and EIA rows above), let the miner edit it, and show the miner's own break-even hash share, because the number that ends a home miner is the bill, not the regulation.

5. Zero-knowledge proving cost, 2026 to 2036

5.1 The numbers today

Row Value Source Label
EF real-time proving target (10 Jul 2025) P99 under 10 s, capex under USD 100 K, under 10 kW, open source, 128 bits (100 accepted at first), proof under 300 KiB, no trusted setup; written for solo stakers "from home" on a 10 kW supply blog.ethereum.org cited
Ethproofs 2025 review (6 Dec 2025) USD 1.69 to under a penny a proof in nine months; 7 zkVMs, 15 provers, about 200,000 blocks; four teams at sub-10 s P99; single-GPU proving from 16 min to under 60 s; 2026 target sub-8 s P99 and kWh a proof as the metric hackmd willcorcoran cited
SP1 Hypercube 99.7 percent of blocks under 12 s on 16 x RTX 5090, cluster under USD 100 K; about USD 0.02 a block single-node Succinct blog; The Block cited
Pico Prism 99.9 percent under 12 s on 64 x 5090 Brevis cited
Cysic Venus 7.4 s a block on 24 GPUs (type unstated); ZK ASIC claims (1.33 M Keccak a second, 50x energy) with no shipping date bex.co 17 Apr 2026 cited, ASIC unverified
ZisK p99 9.62 s on 4 x 5090 GitHub comparative analysis, Sep 2026 secondary
Airbender 51 s average on one RTX 4090, under a cent bex.co Jan 2026 secondary
RISC Zero R0VM 2.0: 35 min to 44 s a block (Dec 2025) wavect; RISC Zero blog secondary
Jolt Lattice Jolt (9 Sep 2026): 2 to 3x faster, 65 to 80 KB proofs, 2 M cycles a second CPU, over 10 M with a GPU, post-quantum cryptobriefing cited
Cost a block on the tracker about USD 0.005 (Sep 2026) lane 7 2.6 secondary
Hardware provers Cysic (C1, ZK-Air, ZK-Pro), Fabric (VPU), Ingonyama (ICICLE; Accseal Leo ASIC in ICICLE v3), Irreducible (FPGA clusters, Binius), Supranational; "10 to 100x on MSM and NTT" claims h33.ai survey; Ingonyama; Cysic docs claims, no shipped benchmark on a zkVM block
Igneum's own 11 rented cards: shard beside the miner 10.7 s (5090) to 37.5 s (3060); alone 4.8 to 18.4 s; SP1 compressed verify 0.032 s on a Mac core prover-tiers-real-cards.md; bench-log 5 Oct measured

5.2 The trend line to 2036 and the 20 percent pool

Lane 7's 33x a year is software catching hardware and cannot hold. Hardware alone is 1.5x a year. The EF's own 2026 metric is kWh a proof, which is the right axis for a miner-prover: a shard that costs a joule of GPU time is paid from the pool and competes with the same joule in the lottery.

Year Block proof cost on the open market (USD, approximate) Hardware that proves an Ethereum block in real time What proves an Igneum shard (4.7 M cycles) under 10 s Does a home 12 GB card have a place
2026 0.005 to 0.02 16 x 5090 (SP1), 4 x 5090 (ZisK) 5090 beside its miner (10.7 s), every card alone yes, alone or hand-off (prover-tiers-real-cards.md)
2028 0.001 to 0.005 1 to 4 consumer cards; first ASIC or VPU boards if any ship (none has) every card from the 3060 at 3x a year (lane 7's table); 12 GB beside the miner at 3x, not at 1.5x yes alone; beside the miner only under 3x
2031 0.0003 to 0.001 one consumer card, or a prover ASIC at 10 to 50x per joule if the claims land a 12 GB card in 1 to 3 s yes, but a prover ASIC would take the external job market first (dollar-priced jobs go to the cheapest joule), and the pool second only if shards are open to anyone
2036 under 0.0001 commodity any card the pool's economics are the lottery's: whoever holds vote weight is drawn (sortition by weight), so the home card keeps its share of the pool whatever an ASIC does to the open market

The design's defence is already in place: the pool is drawn by vote weight (sortition), and vote weight is 30 days of blocks, which a prover ASIC does not have. A prover ASIC centralises the external job market, not the pool. What it would do to the pool is set the shard deadline: if the chain tightens the deadline toward ASIC times, home cards miss it. So the shard deadline must be a function of the measured fleet median (lane 7's rule), re-read each era, and the shard size must stay at a size a 12 GB card proves alone inside it (today 4.8 to 14.4 s). "Verified proving as the reward" (horizon-2026-10 item 1: the proof verified in consensus) is the piece that makes the pool unforgeable, and its activation is the decision owed.

Per tier: 8 GB (proves alone, never beside its miner; keeps its pool share by weight); 12 GB (the swing tier; the hand-off profile is the product); 16 GB and up (unconstrained); rig (proves beside mining on every card from 2028 at 1.5x); pool user (the pool operator's prover does it); Windows, Linux (the same); macOS (an M5 Max proves a shard, unmeasured against the rented table; Metal lane open); NVIDIA (every measurement is NVIDIA); AMD (no measured shard time; the 9070 XT has no SP1 GPU backend measured here, so the AMD tier proves on CPU or not at all until it is measured, which is the first owed number); Apple (as macOS).

5.3 Proof-system risk and the swap interface

Event Date What a malicious prover could do Source
SP1 v3.4.0 (LambdaClass, 3MI, Aligned) disclosed 26 Jan 2025 "generate valid SP1 proofs of incorrect execution of an arbitrary program": universal forgery, from two bugs plus a Plonky3 evaluation gap LambdaClass blog; Blockworks
RISC Zero zkVM 2.0.0 to 2.0.2 15 May 2025 a missing constraint in the rv32im circuit let any 3-register instruction be proven wrong; on-chain verifiers stopped by estop HackenProof
SP1 Hypercube JALR 20 May 2026 a completeness bug (prover crash), not soundness; the EF's audit found only 51 of 62 opcodes fully proven, four load instructions proven against wrong specifications zkevm.ethereum.foundation

Three soundness-class events in 18 months across the two leading zkVMs. On Igneum today a soundness bug is a light-client problem (spec 10.1; finality-in-proof.md 4.4): every full node executes natively and vetoes a record whose statement differs. It becomes a chain failure the day any of three things is true: (1) proof verification in consensus pays a shard on the proof alone (the 0.3.16 switch) and a forged proof carries a correct statement, which earns the pool and nothing else; (2) finality is carried inside the proof (lane 7's 3.3) and a forged proof carries a forged weight table, which a full node still vetoes, so the failure is confined to proof-only clients; (3) the chain ever lets a proof replace execution for full nodes, which the design forbids. So the chain-failure case is (3), and the rule is: never. The swap interface needs: a pinned verifier key per proof system with a version in the record; two independent zkVMs accepted in parallel (the EF's own "multiple provers" lesson), with a record valid only if its system is on the active list; a 95 percent class-change signal to retire a system, and a stop switch (RISC Zero's estop pattern) that any 1/3 of weight can flip to "execution only" inside an hour. Per tier: nothing a miner does; a pool's prover must build for two systems; the cost is two guest builds and two verifier keys.

6. Light clients on phones, 2026 to 2036

6.1 The state of the art

Client What it verifies Bytes and time Source
Ethereum sync-committee clients (Helios, Kevlar) 512-validator sync committee signatures; syncs in about 2 s, no storage; "lightweight enough to run on mobile" about 25 KB per two days (24,576 bytes of keys plus the signature, header and branch) a16z Helios; annotated spec
Mina a recursive SNARK of the whole chain a 22 KB chain; "a few milliseconds of processing" Mina docs via gate.com, iq.wiki
Celestia Lumina data-availability sampling in a browser or phone (Wasm, Rust) random shares a block; bandwidth not published for 2026 (approximate: tens of MB a day) celestiaorg/lumina; Eiger
Bitcoin BIP-157/158 (Neutrino: Breez, Blixt) compact block filters served by full nodes headers and filters in under 5 minutes Blixt; Lightning Labs
StarkWare's Bitcoin header proof every header since genesis in one STARK 1 MB, under 100 ms on a phone cryptotimes 11 Sep 2025
Zcash (Zashi/Zodl SDKs, lightwalletd; Tachyon-Lite) shielded sync through servers server-dependent (the anti-pattern) Zcash forum, GitHub
Kaspa wallets (Kaspium, kaspa-core) none: wallets talk to public nodes over wRPC n/a (approximate) coincodex; Zelcore
Phone verify times Groth16 about 5 ms; a Keccak circuit's verify 0.12 s on an iPhone 15 Pro (Mopro); a 45 KB STARK in 16 ms (laptop, approximate); Groth16 proving 2.9 to 3.1 s on a Galaxy S25 and iPhone 17 Pro Mopro benchmarks; provebench; shattered.io cited, mixed devices
WebGPU in Chrome 113, Edge 113, Safari 18, Firefox 141 n/a Wikipedia WebGPU

6.2 What Igneum's finality-in-proof needs from the phone

The client holds one proof and fetches the newest segment proof from any node (finality-in-proof.md 4.2); SP1's compressed verify is 0.032 s on a Mac core (bench-log 5 Oct), a Groth16 or Plonk wrapper is unmeasured. Model (a wallet updating every 10 minutes, 144 times a day, a phone big core at 3 W, LTE at about 1.5 J a MB, an 18 Wh battery):

Form the wallet verifies Bytes a day CPU Radio Battery a day
Groth16 or Plonk wrapper (about 1 KB with public values) at about 10 ms on a phone (approximate, 3x the Mac core) 0.14 MB 4 J under 1 J 0.007 percent
Compressed STARK at the EF's 300 KiB ceiling, about 100 ms on a phone 42 MB 43 J 63 J 0.16 percent

Either is nothing. The choice is the EF's: no trusted setup and under 300 KiB, which is the STARK row, and by 2030 a WASM-SIMD or WebGPU verifier on a phone should put the STARK verify under 30 ms (approximate projection from the 294x STARK-verifier speedup reported in 2026 and the WebGPU adoption row). The 2026 answer is the Plonk wrapper (no trusted setup, small) for the phone and the compressed STARK for nodes. Bytes a day are set by update frequency, and a wallet that updates on open rather than on a timer is under 1 MB a day in either form.

"Every miner is also a light-client server" has prior art: Bitcoin full nodes serving BIP-157 filters under the NODE_COMPACT_FILTERS service bit (every full node that opts in serves every light client), and Ethereum's Portal Network (every node serves a slice; "without having to trust or put extra strain on full nodes"). The anti-pattern is Zcash's lightwalletd, a few servers everyone trusts. Recommendation: the node serves igneum_getSegmentProofBytes over p2p and a rate-limited HTTP path, on by default in Ember, so the serving set is the miner set; the phone client dials three miners and takes the longest valid chain of proofs. Per tier: no cost a miner notices (one proof of a few hundred KB served on request); the pool node serves for its members; Windows, Linux, macOS, every vendor the same.

7. Post-quantum signatures and a proof-of-work chain

7.1 The standards and sizes

Scheme Standard Public key Signature Status Source
ML-DSA-44 FIPS 204 1,312 B 2,420 B final (Aug 2024) Cloudflare 9 Jul 2026; FIPS 204
ML-DSA-65 FIPS 204 1,952 B 3,309 B final encryptionconsulting
ML-DSA-87 FIPS 204 2,592 B 4,627 B final same
SLH-DSA-SHA2-128s / 128f FIPS 205 32 B 7,856 B / 17,088 B final Cloudflare
FN-DSA-512 (Falcon) FIPS 206 897 B 666 B draft; final expected late 2026 to early 2027; floating-point signing is "difficult to implement securely" encryptionconsulting; Cloudflare
ML-KEM FIPS 203 n/a n/a final NIST
BLS12-381 (today) n/a 48 B 48 B aggregate for 8,192 voters quantum-broken by Shor on the curve's discrete log ethereum.org PQ page

Timelines: NIST IR 8547 deprecates ECDSA, EdDSA, RSA and EC Diffie-Hellman after 2030 and disallows them after 2035 (final version confirmed). The Global Risk Institute's 2025 survey (26 experts, report page dated 9 Mar 2026): a cryptographically relevant quantum computer "quite possible (28 to 49 percent)" within 10 years and "likely (51 to 70 percent)" within 15. Resource estimates: ECC-256 at 1,200 to 1,450 logical qubits and 70 to 90 M Toffoli (Google, 2026), about 500,000 physical qubits against about 1 M for RSA-2048 (Gidney, May 2025); 835 logical qubits with 2^30.6 Toffoli (Luo et al., arXiv 2607.13816, Jul 2026). Bitcoin: BIP-360 merged into the BIPs repository 11 Feb 2026 as a draft (P2MR, formerly P2QRH), a companion BIP for ML-DSA and SLH-DSA, a testnet implementation March 2026, no activation. Ethereum: a PQ team formed January 2026; BLS to be replaced by leanXMSS (hash-based) with leanVM aggregating "3,000 bytes" of signature against BLS's 96 by "250x"; core PQ infrastructure targeted "by approximately 2029"; EIP-8141 account abstraction for user migration.

7.2 What quantum does to Igneum

Component Break Consequence Size of the fix
BLS vote keys (in every header) Shor on BLS12-381's G1 discrete log (a 255-bit subgroup over a 381-bit field; wider than secp256k1, so a later break by a year or two, approximate) every vote key is public; a CRQC forges two thirds of weight and certifies any chain. Catastrophic the migration below
Hash-to-curve none; it is a hash nothing nothing
ECDSA accounts in the EVM as Ethereum: exposed public keys after a first spend user funds; the same exposure as Ethereum's "0.1 percent dormant" row, smaller account abstraction with an ML-DSA verify precompile from genesis
The VRF (aggregator pick) if EC-based, forgeable; aggregation is not safety (every voter signs) liveness nuisance the same key succession
The class-group VDF (epoch seed, era draw) Shor computes the class-group order, which removes the sequentiality assumption of a Wesolowski-style VDF (approximate; from memory of the class-group VDF literature) an attacker with a CRQC grinds the hourly program seed a hash-chain fallback behind the same version byte; flagged, not sized
SP1 (STARK core, Groth16/Plonk wrapper) the STARK is hash-based and stands; the pairing wrapper falls the on-chain and phone verifier moves to the compressed STARK the wrapper choice of section 6

7.3 The cost in bytes, 8,192 voters, every voter signing every 30-second checkpoint (model)

Scheme Per checkpoint Per day (2,880 checkpoints) Averaged per block at 1 block/s Key in the header
BLS aggregate (today) 1.0 KB (48 B + a 1,024 B bitmap) 2.9 MB 34 B 48 B
ML-DSA-44 19,361 KB 57.1 GB 645 KB 1,312 B
ML-DSA-65 26,473 KB 78.1 GB 882 KB 1,952 B
FN-DSA-512 5,329 KB 15.7 GB 178 KB 897 B
SLH-DSA-128s 62,849 KB 185 GB 2,095 KB 32 B
SLH-DSA-128f 136,705 KB 403 GB 4,557 KB 32 B
ML-DSA-44 aggregated by a SNARK at the EF's 250x 77 KB 223 MB 2.6 KB 1,312 B, or a 32 B hash of it

A naive ML-DSA swap costs 57 GB a day of votes against a block stream of about 100 MB a day (approximate), 500x. It is not shippable without aggregation. The aggregator already exists in the design (VRF picks 8 aggregators a checkpoint): the migration is "aggregators carry a STARK of N verified PQ signatures", which is Ethereum's leanVM shape, and the per-checkpoint cost falls to the order of 100 KB.

7.4 The plan to pre-commit at genesis

  1. A sig_scheme version byte in the vote item and in the vote-key registration; genesis value 0 = BLS12-381.
  2. Key succession at registration: every vote key registers with a 32-byte hash of a successor public key (any scheme). A succeed item signed by the old key and the new key moves the 30-day weight to the successor without a reset. Miners generate the successor at first run; Ember stores it offline.
  3. A PQ verify precompile (ML-DSA-44 first, FN-DSA when FIPS 206 is final) in the EVM from genesis, so account abstraction can migrate users before 2030.
  4. The flip is a class change: 95 percent signal with a floor height, as the P2 mechanism; the aggregated-vote format ships first, the scheme flip second.
  5. The VDF carries the same version byte with a hash-chain fallback.
  6. The public sentence: "Igneum's vote keys and accounts can move to NIST post-quantum signatures by miner signal; the key succession is in every key from genesis."

Per tier at migration: a home miner on any card runs the updater and signs one succeed item from Ember; ML-DSA-44 signing is well under a millisecond on any CPU (approximate), every 30 s; the bandwidth cost is the aggregated form's, about 100 KB a checkpoint, which an 8 GB card's host handles; a rig's one key succeeds once; a pool's key is the operator's, and pool members do nothing; Windows, Linux, macOS, NVIDIA, AMD, Apple: nothing hardware-specific, since the signature runs on the CPU.

8. Regulation of mining and of coins with no issuer, 2026 to 2036

8.1 The texts

Regime What it says Source
MiCA Article 4(3) Title II (offers and white papers) does not apply where the crypto-asset is "automatically created as a reward for the maintenance of the distributed ledger or the validation of transactions" (point (b)); where there is no identifiable issuer "the obligation to produce a white paper does not apply to an issuer, as none exists"; a CASP that plays an active or promotional role may have to draw up one (Article 5 and the trading-platform duty under Title V) Osborne Clarke (4 Oct 2023); Conventus Law; Squire Patton Boggs
MiCA sustainability Article 66 and Delegated Regulation 2025/422: CASPs publish the consensus mechanism's energy (kWh a year mandatory; more above 500,000 kWh); the Commission's Article 142 report may propose "minimum sustainability standards"; the 2026 review consultation (reply by 31 Aug 2026) asks only how appropriate the regime is (Q53) the consultation PDF; carbon-ratings; Latham tracker
MiCA transition over on 1 Jul 2026; unlicensed service to EU clients is a breach ESMA statement Apr 2026
UK The Financial Services and Markets Act 2000 (Cryptoassets) Regulations 2026: regulated activities from 25 Oct 2027 (issuing qualifying stablecoins, safeguarding, operating a qualifying cryptoasset trading platform, dealing, arranging, arranging staking); gateway and savings window 30 Sep 2026 to 28 Feb 2027; a "qualifying cryptoasset" is the FCA perimeter term (the formal definition sits in the Regulations and was not retrievable here); mining and validating are not in the activity list found; the 2023 financial-promotions regime already covers promotions of qualifying cryptoassets to UK consumers FCA policy statements page; Lewis Silkin 30 Sep 2026; Skadden Jul 2026
US SEC Corporation Finance staff statement 20 Mar 2025: PoW mining, solo or pooled, is not an offer of securities (non-binding, facts and circumstances); CLARITY (CFTC over digital commodities) failed cloture 49 to 50 on 15 Sep 2026, a motion to reconsider preserved; GENIUS Act 2025 covers payment stablecoins; DAME 30 percent excise proposed three times, never enacted Dechert; Orrick 2 Oct 2026; Cointelegraph
FATF Recommendation 16 revised June 2025; implementation by end-2030; unhosted wallets still treated differently by country FATF best practices Jun 2025; Sumsub
DIFC (Igneum Labs LTD's seat) DFSA crypto-token regime amended 12 Jan 2026: firm-led suitability, no more Recognised Crypto Tokens list, controller approval at 30 and 50 percent; applies to firms providing financial services in crypto tokens; nothing on mining or software Dechert 13 Jan 2026; Arabian Business
Dubai outside DIFC VARA rulebook v2.1 (2026): exchange services, disclosures, token issuance; a federal licensing decision Feb 2026 TradingView/Coinpedia; Dechert

8.2 What it means for each actor

Actor EU UK US DIFC FATF
A pool operator paying members not a CASP for mining; a CASP if it holds members' coins or exchanges them (custody, exchange); the energy figure is the CASPs' duty, not the pool's not an activity listed, unless it safeguards or deals; from 25 Oct 2027 a pool that holds balances looks like safeguarding SEC: pooled mining not a security; FinCEN money-transmitter guidance on pools is the open point (approximate) not a financial service unless it custodies a pool that holds and sends on behalf of members is a VASP-shaped entity by end-2030
A home miner selling coins a seller, not an offeror; taxed as income then gains (country rules) the same; promotions rules do not reach a private sale income at receipt (IRS), then gains n/a the exchange it sells on carries the travel rule
The entity publishing the software (Igneum Labs LTD) no issuer under 4(3); the white paper is nobody's duty; publishing the 2025/422 indicators voluntarily lets CASPs list not an activity; public text must not be a financial promotion of a qualifying cryptoasset to UK consumers (the 2023 regime), which rules out "buy", "invest" and price talk the SEC statement covers mining, not the publisher; CLARITY's "mature blockchain" test is the thing to watch if it passes a software company in the DIFC with no financial service and no token sale sits outside the DFSA token regime on its face n/a
The testnet with no value nothing; no asset nothing nothing nothing nothing
The external proving market paid in dollars a job is a service; settlement in IGN on-chain with no operator custody is not a CASP activity; an operator that converts dollars to IGN is an exchange the same; dealing or arranging if an operator sits between money transmission if an operator holds funds a dollar-settled service is a financial service only if the entity holds or exchanges the same

8.3 What Igneum writes into its genesis and public text now (not legal advice; counsel is engaged)

  1. No issuer, no offeror, no sale: every coin is created as a block reward (MiCA Article 4(3)(b) language, verbatim in the litepaper).
  2. The 2025/422 sustainability indicators (kWh a year from hash rate and measured uJ, intensity per transaction, mix by region when known) published by the project each era, so an EU CASP can list without asking.
  3. No financial promotion: no price, no "buy", no "invest", no return language anywhere the UK can read it; the earnings page shows hash and IGN, not sterling.
  4. The pool reference implementation never holds a member's balance: payouts are direct from the coinbase split (the pool is a coordinator, not a custodian), which keeps pools out of CASP, safeguarding and VASP shapes in every regime above.
  5. The job market is peer-to-peer settled on-chain; no operator account, no dollar leg in the protocol; the dollar price is a quote, the settlement is IGN.
  6. The software is published under an open licence by a company that holds no coins by right (no dev fund, already decided) and runs no service the chain depends on.
  7. The region prompt and the banned-region list in Ember (Russia's ten regions and Moscow from 15 Aug 2026, China) with the sentence "mining may be restricted where you are; you are responsible for checking".
  8. A genesis "no privileged key" statement: no key can mint, pause or upgrade (true by design; say it because the regulators' tests turn on it).

9. Ten-year scenario table

Year Scenario A: GPUs stay general and cheap Scenario B: HBM-only high end, APU-only low end Scenario C: proving ASICs win
2028 48 GB flagship on GDDR7 (lane 7); used H100s at USD 7,500 to 10,500 flood rentals; hash at the rental equilibrium, home share falling with the rent curve. Igneum: the N ladder's first rung by signal; home miners inside 2.5x of the chip per joule; the chip pays at USD 100 M of cap. Alive. The parameter: N stepped by signal HBM4 only on datacentre parts; consumer GDDR7 unchanged; APUs (Medusa Halo, M6) take the laptop tier. Igneum: nothing changes for the hash; the Apple and APU tiers grow. Alive. The parameter: the cache rung, in case an APU's LLC reaches 256 MB First VPU or C1 boards ship at 10 to 50x per joule on MSM and NTT; the external job market goes to them within a year. Igneum: the 20 percent pool is sortition by weight, so home cards keep it; the dollar market is lost to ASICs. Alive, with the external income line (already "never the main income" by 2030, lane 7 3.11) gone sooner. The switch: the shard deadline as a function of the fleet median
2031 GDDR7-Next; 64 GB flagship; rental at 0.002 to 0.004 per MH/s-hour; hash 3 to 6x today's at the same price. The chip bar is USD 1.4 B at N5. Home miners on retail power are marginal unless the heat mode and the price prompt keep the UK and EU tiers on in winter. Alive; the home share is 10 to 20 percent (approximate). The parameter: the weight-aged reward rule (lane 7 3.1) A consumer HBM4-class halo card appears (unannounced today): 2 to 4x the read rate of a 2026 card; the home 2026 cards fall to 0.25 to 0.5x of a new card, the GPU cadence twice over. The chip's stack-level edge is the same as the card's, so the per-joule gap closes to about 1.5x. Alive; the installed base shifts to the new card. The parameter: the N bind-point re-measurement per generation Prover ASICs at 100x; the EF's L1 zkEVM runs on them; SP1-class software moves to ASIC backends. Igneum: proving in consensus verified, home cards prove shards inside the deadline because the deadline is the fleet's; the pool's 20 percent stays with weight. Alive. The risk: if a soundness bug in the ASIC's backend lands, the stop switch of 5.3 is the defence
2036 96 GB cards; rent at 0.0003 to 0.0013; hash 10 to 35x; the emission at the 1 percent tail (86 M IGN a year). Security is 1 percent of market cap a year (tail-emission.md). Igneum lives if the cap is over about USD 50 M (a 24-hour attack at 0.032 percent of cap is USD 16,000 against a rental market that can supply it); dies below that only in the sense every PoW chain does: cheap to attack, nothing to steal. The parameter: the tail HBM everywhere above the laptop; the f = 1 chip and the halo card are the same memory; the ASIC question is closed by parity and the ladder's N is the remaining difference. Alive. The parameter: N at the verifier's 10 ms ceiling Every proof is an ASIC proof; the pool pays miners for shards an ASIC proves 100x cheaper; miners buy ASIC prover boards the way they buy cards (USD 500 to 2,000 boards, approximate guess). Igneum: proving is a second card, not a second income. Alive. The parameter: the shard deadline rule

Where Igneum dies, honestly:

Death Scenario What would be needed to survive it
The stored-dataset chip at N5 is built at a USD 100 M cap in year 1 and holds over a third of hash by year 2 A, 2028, if the N ladder is not at genesis or the signal never steps it the N ladder at genesis with its first rung live; the weight-aged reward; the honest sentence that the chip's bar is USD 100 M, so the cap passes it fast or not at all
Idle datacentre hash sets the rent and the home tier leaves; weight concentrates in three hosting firms; a hosting failure is a 30-day finality pause A, 2031 the LEAVE item (shipped), the reward rule that pays aged keys, the heat mode and price prompt that keep winter home miners on, and a public "weight by hosting provider" chart so the concentration is seen
A CRQC in 2033 to 2036 (28 to 49 percent inside ten years, GRI) before the key succession has flipped any, 2033+ the genesis key succession and the aggregated-vote format shipped by 2030, the flip signalled the year NIST deprecates (2030), not the year a machine appears
A soundness bug in the one proof system while the pool pays on the proof alone C, any year two zkVMs on the active list and the 1/3-weight stop switch
The cap never passes USD 20 M: the tail's 1 percent is USD 200,000 a year of security, a day's rental all, any year nothing in the protocol; it is the market's verdict, and the design should say so rather than promise a floor

10. Verdict table

Axis Finding Source Per-tier consequence Recommendation
Memory roadmap tRC flat for about 20 years; HBM4 doubles channels a stack (2026); HBM4E adds pins and a custom base die (2027 to 2028); no consumer HBM, no GDDR8 before 2029 to 2031 JEDEC, SK hynix roadmap, Lee et al. a 2026 card keeps its read rate against a 2028 card; 8 and 12 GB leave the installed base by 2030, not the hash nothing at genesis; re-measure the N rungs per generation (text, a measurement duty)
The cache consumer LLC 96 to 128 MB; datacentre 256 MB today chip-model-v3, MI300X (approximate) an LLC at the cache size gives that card's owners a 2 to 3x shortcut a cache-size rung on the signal ladder (genesis parameter)
The dataset 2 GiB to 16 GiB over 28 years is under every card and never binds before year 12 algorithm.md 5.6 the prover footprint binds first nothing; fix the public card-lifetime sentence (text)
The chip's bar pays at USD 17 M of cap bare, USD 100 M with the N5 shadow core in years 1 to 2, 2.3x higher every two years with the glide; never needs N2 or A16 model, siliconanalysts mask costs the 9070 XT tier is displaced first, Apple never per joule, the 5090 inside 2.5x the N ladder at genesis (parameter); the p* sentence in the threat model (text)
C-HBM4E the memory controller moves under the stack from 2027 to 2028 TSMC/GUC via Tom's the chip's dollars per MH/s fall under USD 2.8 by 2028 (approximate) disagreement with lane 7 row 2.3 recorded; no new parameter, N covers it
Used-GPU flood 15 to 20 M datacentre GPUs installed by end-2026; residual 25 to 35 percent at 60 months; an H100 mines at 0.65x to 3x a 5090, unmeasured TechInsights, Omdia, mercatus 1 percent of a retired fleet is 2 to 35x the equilibrium hash measure an H100 and an A100 this month (measurement); the reward rule (parameter, lane 7 3.1)
Rental curve 0.0117 to 0.0003 to 0.0013 per MH/s-hour by 2036 model on cloudzero's H100 history attack dollars unchanged at equilibrium; home share falls 10 to 35x re-price 51-percent.md from a measured rental yearly (text); the cost-vs-rent line on the earnings page (software)
Energy UK 26.3 p, EU 29 c average, US 17.7 c; heat credit is a third off in the heating season; ERCOT pays industrial curtailment, not homes Ofgem, Eurostat, EIA, ABC13 retail-power tiers (UK, DE) are 5 to 9x the industrial price even with heat credited heat mode, curtailment input, per-region price prompt (software); 2025/422 indicators published (text)
Mining rules Norway (new sites), Sweden (tax), Russia (regions and Moscow), China (illegal), Kazakhstan and Paraguay (licensed, tariffed); US SEC: mining is not a securities offer the table in 4.4 a home miner's risk is regional the region prompt and banned list in Ember (software)
Proving cost USD 1.69 to about 0.005 a block in 20 months; four teams at sub-10 s; EF's 2026 metric is kWh a proof; no prover ASIC has shipped a block benchmark ethproofs, EF, Cysic a 12 GB card proves alone under 10 s at every rate; beside its miner only at 3x a year the shard deadline as a function of the measured fleet median (parameter); two zkVMs and the 1/3 stop switch (switch)
Proof-system risk three soundness-class events in 18 months across SP1 and RISC Zero LambdaClass, HackenProof, EF none for miners today; a chain failure only if a proof ever replaces execution for full nodes never let it (text in spec 10.1); the active-list and stop switch (switch)
Light clients 25 KB per two days (Ethereum), 22 KB (Mina), a 1 MB STARK in under 100 ms on a phone (StarkWare); a verifying Igneum wallet costs under 0.2 percent of a day's battery a16z, Mina, cryptotimes, model nothing a miner notices; every miner serves proofs serve the segment proof from every node by default (software); the Plonk wrapper for phones, the STARK for nodes (parameter)
Post-quantum ML-DSA-44 2,420 B signatures; 57 GB a day of naive votes at 8,192 voters, 223 MB with 250x aggregation; CRQC 28 to 49 percent within ten years; NIST deprecates 2030 FIPS 204, Cloudflare, GRI, NIST IR 8547, model a miner signs one succession item at migration; nothing hardware-specific the sig_scheme byte, the key succession, the PQ precompile, the VDF fallback at genesis (parameters); the aggregated-vote format by 2030 (switch)
Regulation of a no-issuer coin MiCA 4(3)(b) exempts block-reward assets; UK regulates platforms, dealing, custody, staking from 25 Oct 2027, not mining; SEC mining statement; CLARITY stalled; FATF by 2030 the texts in 8.1 a pool must not custody; a miner selling is a seller; the publisher is nobody's issuer the eight-item list of 8.3 (text); the non-custodial pool reference (software)
Scenarios alive in every 2028 and 2031 cell with one parameter each; dies on an unstepped N ladder, on concentration after the home tier leaves, on a late PQ flip, on a single proof system, or on a cap under about USD 20 M section 9 the home tiers are the ones each death removes first the five "needed to survive" rows of section 9

11. Three headline findings for the coordinator

  1. A stored-dataset chip with the class v4 shadow core pays for itself at about USD 100 M of market cap in Igneum's first two years and about USD 700 M in years 5 to 6 (30 percent share, USD 30 M project, model), and it never needs a node below 28 nm plus N5, so the N ladder at genesis is the only lever and C-HBM4E (2027 to 2028) moves the chip's controller under the memory and cuts its dollar cost further.
  2. Ten years of rental deflation (USD 0.0117 to 0.0003 to 0.0013 per MH/s-hour by 2036) leaves every 51-percent dollar figure unchanged at the equilibrium and cuts the home card's share of the subsidy 10 to 35x; the day an idle datacentre card on 5 c power sets the rent, a UK home miner at 26.3 p loses 7x on electricity, and the heat mode buys back a third, not the gap.
  3. A post-quantum vote at 8,192 voters costs 57 GB a day in naive ML-DSA-44 (19.4 MB a checkpoint) and about 223 MB a day with 250x SNARK aggregation, against a 28 to 49 percent chance of a cryptographically relevant quantum computer inside ten years (GRI 2025) and NIST deprecation after 2030, so the key-succession item, the scheme byte and the aggregated-vote format belong at genesis and the flip belongs in 2030.

File: /Users/joshm/Projects/igneum-wt-mission/docs/analysis/mission/future.md.

Sources (accessed 7 October 2026 unless dated otherwise)

Lane and repository inputs: docs/analysis/horizon/frontier.md (lane 7), docs/analysis/horizon/algorithm.md (lane 2), docs/analysis/chip-model-v3.md, docs/analysis/51-percent.md, docs/bench-log.md ("Rental cost of hash, 6 October 2026"), horizon-2026-10.md, finality-in-proof.md, tail-emission.md, vote-or-burn.md, prover-tiers-real-cards.md.

Memory and GPUs:

Chip costs:

GPU supply and rental:

Energy, heat, mining rules:

Proving:

Light clients:

Post-quantum:

Regulation: