Solana Validator Economics

The short version. A validator's revenue is three lines, not one: inflation commission (~30% of it), block rewards (~64%) and MEV tips (~2%). Two breakeven figures circulate and both are wrong in opposite directions — the high one counts inflation commission alone, the low one was computed before the Validator Admission Ticket was specified. Computed against real costs and checked against two operator settlement records — fifteen epochs against the revenue model (§3), and sixteen more spanning both slot-time transitions (§8) — the number is 76,528 SOL of delegated stake today. Three things move it, in order of how soon they arrive: the SIMD-0525 slot-time ramp (§8), which is the largest and the nearest; commission policy (§4), which is the largest thing an operator actually controls; and the SIMD-0550 disinflation vote (§10) — which passed on 28 August, is not yet active, and does less to breakeven than the schedule already in force.

Every figure here is computed rather than quoted, and the model is one page of arithmetic available on request. Known weak points and open questions are listed at the end; corrections are credited by name.

1. Where a validator's money comes from (per SOL of delegated stake)

At the SFDP commission cap of 5% inflation / 10% MEV — the reference case throughout this brief, with the three alternative commission settings priced in §4 — the revenue stack per delegated SOL per year is:

Shares are computed against the full five-line stack of 0.931%, so they describe an operator earning both BAM and Edge. For an operator running neither, the core-three shares are inflation 31.2%, block rewards 64.5% and MEV 1.6%. The MEV row is stated at the 10% cap: observed at ~0.010% on a ~150K SOL book, cap-equivalent 0.010–0.021% depending on the commission that operator runs. This brief uses 0.015% and flags it as the least stable line in the model.

Why the headline is the core three and not the full stack. Adding BAM and Edge takes the rate to 0.931% and drops breakeven from 76,528 to 73,406 — about 4%, real but conditional, and conditional in different ways. Edge is on its way to being universal: 434 of 462 connected validators already publish, and once Jito and Harmonic ship their integrations it becomes automatic. BAM is not — it requires a top-tier Jito Steward ranking most operators do not hold and cannot elect to hold, and it is explicitly named an early-adopter program. An operator already earning both should read the lower number as theirs; one planning a business should underwrite the core three.

Two breakeven numbers circulate, and both are wrong in opposite directions. The high one counts inflation commission alone — about a third of revenue — and implies roughly 246K at this brief's basis. The low one, nearer 50K, predates the Validator Admission Ticket rather than assuming anything naive about it — it was modelled when Alpenglow was expected to end vote costs outright and no ticket had been specified (§6). VAT at 1.6 SOL/epoch preserves about 80% of the vote line, and preserves more than that once epochs shorten (§8), so the figure is stale rather than wrong-headed. The near-elimination framing that does circulate — the $60K-to-$1K comparison — is protocol-side messaging about the size of the saving, not the basis of anyone's breakeven. The core stack against real costs puts it at ~77K today and ~65K post-Alpenglow. Chorus One independently computed ~56.6K minimum profitable outside stake at these commissions — between the low-VAT case (37,490) and the post-Alpenglow case (65,078). The assumption sets differ, so quote both and reconcile rather than claiming they agree.

BAM early-adopter claims. Running Jito's BAM with a top-tier Steward ranking earns a JitoSOL claim every epoch — observed across ten epochs at 0.398–0.769, averaging 0.538, about 98 SOL a year. No public validator calculator renders it. Three caveats: it is denominated in JitoSOL, which trades above SOL, so counting it 1:1 understates it; the program is named early adopter; and the Steward gate means a new or sub-scale operator cannot add it by deciding to.

DoubleZero Edge. Subscriptions are priced in USDC (~$30–100 per device per epoch) and split each epoch — 50% to network contributors, 32.5% to validators originating shreds, 17.5% to protocol client teams, with 10% burned. The validator share arrives as 2Z on roughly a ten-epoch lag: 27.7–42.3 per epoch, averaging 34.7, about 0.033 SOL per epoch or 6 SOL a year. Publishing is already table stakes at 434 of 462 connected validators, and Jito and Harmonic are integrating it into their clients — so being early is not a moat. Placement is: Edge's latency advantage is largest in Asia (>100ms under congestion against ~20ms in Europe), so an Asian publisher sits where the product is worth most.

Edge reached public beta in April 2026 with ~378 publishers (~43% of stake); by the Q2 close, 462 validators were connected, 434 publishing, and 447 distinct subscribers had used it, covering 59% of mainnet stake. Coinbase's validator connected in June.

2. Which stack is actually yours

The headline rate is a reference case, not a description of any particular operator. Two switches decide which row a reader belongs in — whether they charge MEV commission, and whether they earn the two conditional lines — and one input decides the column, which is what they pay for hosting. Both columns below use the same method; only the hosting figure differs (§6).

The spread across this table is 64,853 to 111,166 — a factor of 1.7 — and none of it is disagreement about the network. It is entirely commission policy, two conditional programs and one hosting contract. Any single headline number is therefore a claim about a profile, not about Solana.

3. The model against a real record (15 epochs, audited)

Rates above are calibrated to published network data. To test them, they were checked against one operator's actual settlement record — 15 consecutive epochs, 274,694 SOL of delegated stake, 5% inflation commission, 0% MEV commission, Singapore. The model's shape held. Its level did not.

Costs moved the other way and partly offset it. Voting came in at 398 SOL/yr, which the model now adopts. Infrastructure came in at $21,636 against the $30,000 assumed — a negotiated bare-metal contract covering the mainnet node and its failover, worth about 110 SOL a year. Applying the same method to that operator's own inputs: ~68,714 SOL at their actual costs, ~73,598 including program fees, ~79,938 at the $30,000 assumption.

The most useful finding in this section. The inflation row once ran 12% above measurement, read at the time as the model being optimistic. It was not. The model was using last year's inflation rate — 4.13% at ~65% staked, a 6.3% gross yield — when the network is at ~3.73% and ~67%, giving ~5.57%. This operator's own commission stream implies 5.57% independently, which is how it was caught. Correct the input and the row reconciles to within 1.2%. An earlier claim that model and measurement agreed within about 3% does not survive: it was a stale revenue input offsetting an understated cost line. On corrected inputs the model runs 7.6% above measurement, and the entire remaining gap sits in the block-reward line — the activity calibration flagged throughout as the widest error bar.

Two hosting inputs, and why both stay. The measured column uses the operator's real $21,636 contract; §6 and §10 use a $30,000 case. They answer different questions — what one operator spends, versus a headline that does not generalize one favorable private contract to the market. Hosting is published as a sensitivity in §6 so every figure can be read at either cost base.

A longer record on the same validator, and it moves the block-reward line the other way. The 15-epoch check above found block rewards 10% under model. A 115-epoch settlement export covering 1 January to 19 August 2026 on the same vote account puts the recent run-rate — the last nine epochs, at roughly 292,000 SOL of stake — at 0.68% of stake against the 0.600% modeled, about 13% over. Same line, opposite sign, seven times the sample. The honest reading is not that one measurement supersedes the other but that the fee line swings by roughly a quarter either side of the model within a single year, which is wider than the ±10% this brief had been treating as its error bar. The same export confirms the cost side tightly: the vote line held between 2.13 and 2.17 SOL per epoch across all 115 epochs, which is where the 2.154 figure used throughout comes from.

Two lines from that export are deliberately not quoted here. Its revenue split across the full period is distorted by a stretch of epochs carrying no commission at all, whose cause is unresolved, and a figure that may be a reporting artifact does not belong in a document arguing from measurement. It will be published once the record is reconciled.

4. Commission policy is the largest single lever

Under the SFDP caps an operator may charge up to 5% on inflation and 10% on MEV. Both get surrendered in practice, for the same reason — to qualify for delegation — but they are not remotely the same trade, and the difference is not widely priced.

Surrendering base commission is roughly 26 times more expensive than surrendering MEV commission — about 34,638 SOL of additional breakeven against 1,307. At mid-2025 MEV levels the ratio was nearer 5×; at current levels the MEV concession is close to free. The ratio holds at any hosting level — at $21,636 the base concession costs 30,602 and the MEV concession 1,154, still about 26× — because the asymmetry is a property of the revenue rates and the cost base cancels out of the comparison. Earlier versions published 31× on this row, which contradicted that reasoning; the ratio is the same at every hosting level and every SOL price. An operator weighing a delegation program's commission requirement should establish which knob is being asked for before agreeing to anything.

Check whether the trade is actually required before making it. The commission-for-delegation bargain is far less universal than it was: most programs now score on performance, decentralization, geography and community-goods contribution rather than commission rate, and operator experience puts the 0%-base-rewarding set at a named minority — BlazeStake, Marinade SAM, Edgevana (sunsetting), SOL Strategies and JPool among them. So 0% base is a choice with a ~34,600 SOL price tag, not an entry fee.

Who actually runs 0% base is unmeasured, and two incompatible stories circulate. One says it is sub-scale operators, pushed there by the belief that it is the entry fee. The other — argued by independent operators in coverage of the validator-set collapse — says it is large, well-capitalised validators running 0% as a customer-acquisition loss leader, which is what prices the small ones out. Both may be true at once, and nobody has published the shape. What is not in doubt is the consequence: the bar becomes roughly 111,000 SOL against the ~50K in circulation, and the operator has surrendered the one line the SIMD-0550 debate is about.

This is measurable. getVoteAccounts returns the inflation commission and activated stake for every validator on the network; cross-tabulating commission against stake size settles it in one query, and splitting out the 100%-commission private validators is the only real subtlety. Until someone publishes that distribution, treat any claim about what most validators charge — including the ones here — as anecdote. If you run it first, send it and it goes in with your name on it.

Partially answered by the proposal's own authors. SIMD-0550's analysis reports that 43.3% of validators outside the supermajority run 0% inflation commission (§11). That is the first published cut of this distribution and it is a large number — but it is one slice of the set, it is not cross-tabulated against stake size, which is the part that decides between the two hypotheses above, and it comes from a party with an interest in the result, since a 0%-base cohort is precisely the cohort SIMD-0550 does not touch.

5. Commission policy meets governance — the two proposals cut opposite ways

This is the claim here most worth arguing with. §4 prices the commission concession at today's yields. But an operator charging 0% base commission earns nothing from the inflation line — so SIMD-0550, which cuts only that line, costs them nothing. They are immune to the proposal everyone is worried about. Meanwhile SIMD-0123 strikes block rewards, which are 100% of a 0%-base operator's revenue.

Two readings, both true. Disinflation makes the base-commission concession look about 51% cheaper over three years — the price tag falls from 34,638 SOL to 17,118 — because it erodes the very line the 0%-base operator already gave away. SGP-0002 passed on 28 August, so that erosion is now scheduled rather than hypothetical and the argument in §4 weakens on its own terms — from the epoch the gate ships, not from today (§10).

But block-revenue sharing does the reverse and it is not close. At 50% retention a 5%-base operator needs about 115,221 SOL; a 0%-base operator needs about 217,039 — which is the inflation-only folk-math number, arrived at from the opposite direction. The commission concession is a leveraged bet that SIMD-0123 never ships, taken by the cohort least able to absorb it if it does — and SIMD-0185 already shipped the plumbing.

One caveat on the mechanism, and it cuts against the comfort in that reading. Block rewards cannot be shared with a delegator by protocol — that is what SIMD-0123 would change. But nothing prevents a validator from agreeing to rebate a share of them by contract, in a side letter or an SLA, and whether that happens at the institutional end of the market is not publicly known. If it does, the ~64% line is already reachable without any vote, the commission rate stops being a reliable read of what an operator actually nets, and the hedge described here is weaker than the arithmetic implies. This is flagged as an open question rather than a finding; anyone with visibility into how large staking mandates are actually papered is invited to correct it.

The planning conclusion. The two proposals are not additive risks pointing the same way; they are a hedge and a cliff. Base commission is protection against SIMD-0123 and a wasting asset under SIMD-0550. Both halves of that were written before the vote, and one of them has now resolved. Disinflation had failed twice — SIMD-0228 voted down in March 2025, SIMD-0411 closed for inactivity in January 2026 — and this brief leaned on that record in judging the odds. SGP-0002 carried at the third attempt on 28 August 2026, by about a third of a percentage point (§10), so that reading of the probability was wrong. The conclusion survives, but it has to be restated on narrower ground rather than repeated. It no longer rests on disinflation being unlikely; it rests on the two risks differing in shape. SIMD-0550's effect is gradual, not yet active, and about a third of a three-year move the existing schedule was already making on its own. SIMD-0123 is a step change against the ~64% line, with the plumbing already Accepted and only the policy unset. A wasting asset that wastes slowly is still worth holding against a cliff nobody has priced — but that is a weaker claim than the one this brief made before the vote, and it is stated as such. Stacking both at the terminal path with 0123 at 50% retention needs ~160,133 SOL.

6. What it costs (annual, SOL-equivalent at $105/SOL)

Alpenglow cuts the voting line by ~26%, not 98% — and total cost by ~15%. The famous “$60K → $1K” claim compares old vote fees to a VAT later set at 1.6 SOL/epoch (≈296 SOL/yr) — about three-quarters of the measured vote cost, not a near-zero one. Hosting doesn't move at all, which is why the total falls from 684 to 582 rather than collapsing. That 582 describes a 400ms epoch count, which mainnet left at epoch 1020; because VAT is charged per epoch and epochs shorten, the same ticket costs 591 SOL a year at 200ms and the total returns to 877 (§8). The “450 SOL minimum” circulating is a May-2025 modeling artifact from before VAT existed; the real Alpenglow floors are the VAT eligibility check (the vote account must cover 1.6 SOL/epoch + rent) and a 2,000-seat cap by stake.

The VAT concept was accepted within the SIMD-0326 Alpenglow governance vote; the implementation (SIMD-0357, Anza) activated at epoch 1006 — read directly off a mainnet node's feature list, alongside SIMD-0387 BLS pubkey management at epoch 999 and SIMD-0337 fast-leader-handover markers at epoch 1005. What has not activated is SIMD-0326, the consensus algorithm itself. An earlier version of this brief said VAT was not yet live; that was wrong, and the correct statement is narrower and more useful: Alpenglow's scaffolding is deployed and gated while the consensus change waits. A vote account already carries its BLS public key, but the network still votes on-chain and still pays per-slot vote fees — the settlement record in §8 shows an unbroken TowerBFT vote line and no VAT charge through epoch 1024. Treat VAT as provisioned but not yet charging. One operational consequence worth acting on before activation rather than after: VAT draws on the vote account, not the identity, and the eligibility check requires it to cover 1.6 SOL per epoch plus rent — a vote account run near empty by routine commission withdrawal is an admission problem waiting to happen. Where the level settles is still live: at 1.6/epoch voting remains a material cost line; at the 0.135/day level some researchers advocate, it stops being the dominant cost and hosting becomes the binding constraint instead.

Hosting is a sensitivity, not an assumption. It is roughly half of total cost today and the majority after Alpenglow (286 of 684 SOL, and 286 of 582), it is the only cost line an operator controls — votes and VAT are protocol-set — and it is the one input every reader already knows. Varied across the range a real operator can contract at:

Every $1,000/yr of hosting is about 1,066 SOL of breakeven at $105 SOL and the core rate. The spread here — 60,538 to 87,188 — is comparable to the activity error bar (76,528 to ~99K), which makes hosting the largest single swing in the model. It is not an error bar, though: it is an input the reader can replace with their own invoice.

Why the headline stays at $30,000. The $21,636 figure is one negotiated contract on one operator's two nodes — n=1, and a good one. Promoting a favorable private cost to the market case would tilt the headline toward the ~50K folk figure this brief exists to argue against, and would invite the objection that the author's own costs are doing the work. The conservative case stays on the front of the model, the measured contract sits beside it, and a reader who knows their own hosting bill should use the row that describes them.

7. What the hosting line actually buys — and a public floor price to measure it against

§6 treats hosting as a single input and varies it across a range. It is worth opening one line item, because a public estimate now exists to check it against, and because the composition of the line is not what the estimate assumes.

On 30 August 2026 Anatoly Yakovenko posted a back-of-envelope figure — described as such — for the storage and indexing cost that should set a floor under Solana’s price per byte of on-chain state: under $1e-7 per KB — a ten-millionth of a dollar per kilobyte. The post gives no time unit; it is read here as per year, because a storage cost without a time dimension is not a cost, and because on that reading it lands at about $100 per terabyte-year — a fair amortized figure for enterprise NVMe. As a disk number it is defensible. That reading is this brief's, not his, and every multiple below depends on it. Measured against this operator’s own invoices it is low by two orders of magnitude, for a reason the disk framing hides.

The RAM row is the cleanest measurement here because it is one line on one invoice. In August 2026 this operator’s RPC node went from 256 GB to 512 GB at +$186.70/month, bought for a single purpose: holding Solana’s account indexes in memory. That is $2,240/yr against 268 million KB, or $8.3e-6 per KB-year. The floor prices disk; the binding constraint on a node that answers indexed reads is RAM, which runs roughly two orders of magnitude more per byte. Two of three secondary indexes fit in the result, and only just: the process holds 450 GiB resident against a 470 GiB ceiling — 95.7% of cap, with peak usage at the same level. The third index exceeded that ceiling during generation and was dropped. That is what a binding constraint looks like from the invoice side.

The four-node row attributes the entire lease of four mainnet machines — about $29,400/yr — against the 474 GB accounts store each of them carries independently. That is a different machine set from §6's hosting line, which covers only the consensus node and its failover; the four here add the RPC and governance nodes that a validator cost model deliberately excludes, which is why the two totals overlap without matching. It overstates in one direction, since that lease buys cores, bandwidth and ledger rather than state alone, and understates in another, since four full copies of the same bytes is what a network of a thousand nodes does in miniature.

These are attributions, not line items: the provider bills machines, not bytes. And the RAM row describes an RPC node running secondary indexes, which most validators do not run at all. A vote-only validator sits far closer to the floor than to the top row.

The comparison still runs the way the floor argument intends. A 165-byte token account costs this fleet on the order of one-thousandth of a cent per year to carry. The user who created it paid a rent-exempt deposit of 0.00204 SOL — about $0.21 at this brief’s $105 basis — once, and never again. Even measured against an RPC operator’s inflated real cost, that deposit exceeds a decade of carrying cost by roughly three orders of magnitude. Storage is not what makes a validator expensive. Votes and cores are, which is what §6 and §8 are about.

The protocol appears to agree, and the mechanism is already written. Five feature gates sit inactive on mainnet today under SIMD-0437, stepping the rent rate from 6,960 lamports per byte-year down through 6,333, 5,080, 2,575 and 1,322 to 696 — a tenfold reduction, staged the same way SIMD-0525 stages slot times. None has activated and no schedule is published. If they do activate, the deposit on that 165-byte token account falls from about $0.21 to about $0.021 at this brief's $105 basis, and state gets an order of magnitude cheaper to occupy while costing operators exactly what it costs them now.

What the protocol does not price is the second word in the estimate. Rent compensates storage, in a form the operator never receives — the deposit sits in the user’s own account. Indexing is compensated by no protocol line at all, and on the numbers above it is the expensive half: the $186.70/month bought index capacity, not disk. That cost falls entirely on operators who serve reads, and it sits outside the breakeven arithmetic in §10 by construction, because a non-voting RPC node earns nothing from consensus while carrying the same state and more of it in memory.

8. The slot-time ramp is a vote-cost ramp

SIMD-0525 reduces Solana’s target slot time from 400ms to 200ms in four feature-gated steps — 350, 300, 250, 200 — each requiring supermajority activation and each taking effect an epoch after it lands. The first testnet stage activated on 5 August 2026 and mainnet activations began in the week of 17 August. As at 30 August 2026 mainnet is two stages in. The feature gates, read off a mainnet node, put stage 1 (350ms) active from epoch 1019 at slot 440,208,000 and stage 2 (300ms) from epoch 1023 at slot 441,936,000; because each stage takes effect an epoch after it lands, the first shortened epochs are 1020 and 1024. Stages 3 and 4 are present and inactive. This section was written as a forecast; the tables below are now partly a record. The design detail that decides the economics is easy to skim past: epochs remain fixed at 432,000 slots, so every stage shortens the epoch in wall-clock terms — roughly 42 hours at 350ms, and about 24 hours at 200ms.

On-chain votes are per-slot, and nothing scales them. A TowerBFT validator’s vote cost per epoch does not change at all; what doubles is the number of epochs in a year. At the measured 2.154 SOL per epoch, holding hosting at $30,000 and the core revenue rate of 0.893%:

Read that column as a run-rate, not a bill. It is what breakeven looks like while the condition holds, and the condition is partly temporary: on-chain voting ends when Alpenglow activates, but the epoch count does not fall back, so a fixed Validator Admission Ticket carries part of this increase forward (see the correction below). The Solana Foundation's Agave 4.2 notes state that Alpenglow does not activate in 4.2 and is expected in Agave 4.3, targeted around 21 October 2026. So the ordering is currently known — the ramp lands first and the relief follows, which means operators do pay this, but briefly.

Costed as cash rather than as a rate, it is a small number. The extra burn runs about 0.16 SOL/day at 350ms, 0.36 at 300ms, 0.65 at 250ms and 1.09 at 200ms. Over a ramp beginning in August and ending with an October activation that totals roughly 35–55 SOL — material to an operator already under water, immaterial to one comfortably above the line. Both numbers belong in the model because they answer different questions: the run-rate answers what breakeven looks like this month, the 35–55 SOL answers what it actually costs.

Three reasons it still belongs. The mechanism is not widely understood, and an operator watching their vote account drain faster in September should know why rather than assume a misconfiguration. The exposure is entirely a function of the gap, and Alpenglow's date has moved repeatedly — early-to-mid 2026, then Q3, then late Q3 or early Q4, now 4.3 in October; a slip to December takes the total to ~85 SOL and to January ~135. And most operators will pay the intermediate stages rather than the headline one, since reaching 200ms before Alpenglow requires all four gates to clear inside the window.

What the ramp actually did to one validator. The mechanism above was a prediction: the per-slot vote line should not move, the epoch count should rise, and time-denominated revenue should fall per epoch while holding steady per year. A sixteen-epoch settlement record spanning both transitions — epochs 1009 to 1024, one mainnet validator, 31 July to 30 August 2026 — lets it be checked rather than asserted.

^* The commission baseline is the eight consecutive epochs 1012–1019, which ran between 4.394 and 4.423 — a 0.6% band. That stability is what makes the subsequent break readable as duration rather than noise. The sample sizes differ and the second one is thin: the 350ms column is four epochs, the 300ms column is one. Read the −26.4% as a first observation rather than a mean, and re‑read it once epochs 1025 onward settle.

Both halves of the prediction hold. Vote fees are charged per slot and the slot count per epoch is fixed, so the per-epoch bill does not move — and it did not. Inflation accrues per unit time, so a shorter epoch pays proportionally less of it — and it did, by −13.6% where a 48-to-42-hour epoch predicts −12.5%, and by −26.4% where 48-to-36 predicts −25.0%. Both fall about a point further than the proportional prediction; the 15%/yr disinflation taper across the elapsed window accounts for roughly half of that excess and the remainder is not explained here. The annual totals move in opposite directions: commission per year is roughly preserved because fewer SOL per epoch meets more epochs per year, while vote spend per year rises because the per-epoch figure is constant. That is the whole of this section, measured on one vote account.

The feature gates corroborate the settlement record independently. The commission break in the table above falls at epochs 1020 and 1024 — exactly the epochs the activation slots imply, and derived from an inflation-rate ratio in a billing PDF rather than from the chain. Two unrelated methods, one answer. That also retires the obvious objection: a change in delegated stake would produce a similar signature in the commission line, but it would not land on the two epoch boundaries the gates independently name. One limit remains — the commission ratio measures epoch duration relative to the pre-ramp baseline, so it is robust as a ratio while its absolute anchor depends on what the 400ms-era epochs actually ran at, which is why the flatness of the 1012–1019 baseline matters. And leader rewards, the largest revenue line, are far too volatile over sixteen epochs to read anything from — they ranged from 4.49 to 19.05 SOL per epoch.

Achieved slot times run above target, and that makes the ramp slightly cheaper than the tables say. Measured across the first 43% of epoch 1025 — 184,533 slots in 16h 15m — the pace is 317 ms/slot against the 300ms target, and a second validator on another continent read 319 ms/slot in the same window. Thirty-second samples range 319–334, so the epoch average is the durable figure and the instantaneous reading is noisy. Every epochs-per-year figure in the tables above is target-scaled, so at this stage the real count is about 229–230 rather than 246.3, annual vote spend about 493–496 SOL rather than 531, and breakeven about 87,200–87,550 rather than 91,415 — roughly 4% below the modelled row. The published per-day burn figures survive better than the annual ones: 0.16 SOL/day at 350ms computes to 0.154 at the measured 2.154 SOL per epoch. The direction of the error is worth naming, because SIMD-0550's own text notes the inflation schedule is running 276 days behind for exactly this reason: target slot times are not achieved slot times, and every model built on the target overstates the epoch count.

Alpenglow reduces this exposure. It does not remove it — and an earlier version of this brief said otherwise. That version read the SIMD-0525 text as scaling the Validator Admission Ticket down with each stage (1.6 SOL/epoch at 400ms, then 1.4, 1.2, 1.0 and 0.8 at 200ms) to hold the wall-clock admission cost at roughly 0.8 SOL per day, and concluded that every post-Alpenglow figure survived the ramp untouched. That conclusion is withdrawn. An operator running the Alpenglow community cluster reports that VAT reduces to a fixed 1.6 SOL per epoch, and that the consequence of a rising epoch count is under active investigation at Anza. The scaling ladder is not settled implementation, and this brief should not have presented it as one.

The correction matters because VAT is charged per epoch while epochs shorten. A fixed ticket therefore inherits the same doubling the vote line has, and the relief the ramp was supposed to hand back does not arrive in full:

So the post-Alpenglow figure published throughout this brief — 65,078 — is a 400ms-slot figure, and mainnet left 400ms at epoch 1020. It is a reference point rather than a forecast. VAT does not begin charging until Alpenglow activates, expected with Agave 4.3 around 21 October 2026, by which time the network will be at 300ms or faster — so the ticket will never be charged against a 400ms epoch count, and the nearer row is 300ms at 76,042. What is no longer defensible is treating 65,078 as the number that survives the ramp. Read as a band, the post-Alpenglow answer is ~65,000 at 400ms and ~98,000 at 200ms if VAT stays fixed at 1.6, and where inside that band it lands is a live implementation question rather than an error bar. At the measured $21,636 hosting contract the same band runs 56,167 to 89,233.

Two consequences follow, and both are milder than the alarming reading. Alpenglow's cost advantage does not grow across the ramp — it is ~15% at 400ms (684 against 582) and about 19% at 200ms (1,082 against 877), because both sides of the comparison inflate with the epoch count. And the ordering matters less than this brief previously claimed: reaching 200ms before Alpenglow costs more, but having the order reverse no longer avoids the increase, it only softens it. The epoch-count effect is a property of the schedule, not of the consensus mechanism, and it outlives the consensus change.

This is the widest open question in the document and it is deliberately published as a band rather than resolved to a point. If Anza settles the VAT treatment, the band collapses to one column and this section becomes a table rather than an argument. Corrections from anyone closer to the implementation are welcome and credited by name.

Every rate here is annualized at 184.75 epochs per year, and every per-epoch measurement in §1 was taken at 400ms epochs. Once the ramp proceeds that stops being a constant — activity-linked lines should roughly halve per epoch as the count doubles, while the vote line will not, because it is per-slot and fixed. The annual figures are the durable ones. No table has been re-based, because the stage the network will sit at is unknown.

9. Sensitivities — activity, price, and the disinflation indifference point

Activity sensitivity (the honest error bar): the fee and MEV lines scale with network usage. At mid-2025 activity, core-stack breakeven is ~77K SOL today / ~65K post-Alpenglow at 400ms slots (at $105/SOL and $30K hosting); if activity has cooled a third, ~99K / ~84K. Both are still a fraction of the ~246K the inflation-only folk math gives at this price — the 2–3× correction is robust to price, hosting and activity; the third digit is not.

SOL price moves breakeven materially in today's vote regime — hosting is USD-priced, so a lower SOL price means more SOL to cover it — and much less post-VAT, when the dominant cost becomes SOL-denominated. Recomputed from the cost model above at the core rate, holding hosting at $30,000 and the vote line at 398 SOL: 89,320 SOL at $75, ~82,760 at $87.87, 76,528 at the $105 basis, ~66,930 at $150 and ~61,340 at $200. Two of those are measurements rather than scenarios: $87.87 is the mean SOL price across the sixteen settlement epochs in §8, and $75 was this brief's basis until 30 August 2026 — retained here so figures quoted from earlier versions can still be located. The direction is worth stating plainly: hosting is billed in dollars, so a higher SOL price buys it with fewer SOL and breakeven falls. Re-basing from $75 to $105 moved every published figure in this document down by about 14%. Note that the hosting sensitivity and the price sensitivity are the same mechanism seen twice: both move the USD-denominated share of the cost base.

The disinflation indifference point, for anyone weighing the vote. Because hosting is USD-priced, a higher SOL price offsets a lower yield. Running the SIMD-0550 terminal path against the cost model, the price at which breakeven returns to today's level is about $189 pre-Alpenglow and about $169 post — roughly 60–80% appreciation. That threshold rose when the basis rose: measured from a higher starting price, a larger move is needed to offset the same yield cut. Stated neutrally: if disinflation delivers the price effect its advocates argue for, an operator comfortably above the line comes out ahead in USD terms, because their revenue is SOL-denominated. An operator at 60K SOL does not, because the breakeven increase reaches them at today's price and the price effect does not.

10. Breakeven — delegated stake required (core stack, 5% / 10% commissions)

Two adjustments, stated separately. Programs: subtract about 4% from every cell if the operator also earns BAM and Edge (rate 0.931%) — 73,406 today, 62,423 post-Alpenglow. Hosting: at $21,636 the current-yield row becomes 67,611 today, 56,160 post-Alpenglow, 28,573 low-VAT; the reduction grows to ~10,957 on the terminal disinflation row, because a lower revenue rate magnifies any change in cost.

Read the SIMD-0550 rows carefully: the existing schedule does most of this on its own. Doing nothing takes post-Alpenglow breakeven from 65,078 to 73,972 over three years — +13.7%, with no vote and no proposal. SIMD-0550 doubles the decay rate, adding 73,972 → 79,977, or +8.1%. The total three-year move is +22.9%, but only about a third of it is the proposal. Disinflation cuts the inflation line only, and that line is ~30% of revenue; costs do not move at all, since hosting is USD and VAT is SOL.

Status, 30 August 2026 — it passed, and it is not in effect. SGP-0002, the on-chain referendum wrapping SIMD-0550 “Double Disinflation”, closed on 28 August with 176.29M SOL for, 66.19M against and 20.63M abstaining — 67.00% of participating stake against a 66.667% bar, a margin of about a third of a percentage point. Participation reached 60.7% of the 433.49M eligible, across 1,326 validators, the highest in Solana's governance history. Two companion results matter as much: SGP-0001, the Solana Constitution, passed at 85.97%, and SGP-0003 (SIMD-0553, resource and inclusion fees) failed at 53.90% — so the inflation line was cut and the fee line was left alone (§11). The predecessors are worth keeping in view: SIMD-0411 was closed for inactivity in January 2026 and SIMD-0228 was voted down in March 2025, which makes this the third attempt and the first to carry. The electorate mechanic this brief flagged is what decided it — stakers can override their validator at solanagov.com, and JitoSOL stakers did, while Kraken's validator moved from support to opposition and back inside the final hours.

Passage is a mandate, not an activation, and the distinction is doing real work here. SGP-0002 instructs; SIMD-0550 implements. The change requires a feature gate named double_disinflation_rate, and that gate does not appear at all — active or inactive — in the feature list of a mainnet node running the current Agave 4.2.1, which means no shipping client can activate it today. Three things have to happen in order: the implementation lands in a client release, stake adopts that release past the threshold at which activations begin, and the gate then fires on an epoch boundary. None has a published date, and the proposal was still missing a required conformance section at review. A mechanical detail most coverage omits: nothing drops on the day it activates. The design records the rate the existing 15% schedule yields at the activation point, sets the taper to 30%, and recomputes the curve to pass through that same point — so the schedule steepens from activation forward with no discontinuity, and every figure in the table above describes the path after that, not a step on the day. Until the gate ships, the row that governs is the existing 15% schedule.

The proposal that would actually hurt: SIMD-0123. If block revenue becomes shareable and a validator retains half, breakeven jumps to ~115,221 SOL at current yield — a bigger single hit than SIMD-0550's whole path — because it strikes the ~64% line, not the ~30% line. Stack both and it is ~160,133. The vote-account state supporting block-revenue distribution is already Accepted as SIMD-0185: the mechanism exists, only the policy is unset. This is the governance item to watch and to price into any commission strategy.

11. The disinflation proposal's own model — and why its headline is three times this one

SIMD-0550 publishes a breakeven table of its own. It assumes $18,000/yr of cost, an average commission of 2.75%, SOL at $80, and 201 SOL of annual vote cost (Alpenglow's VAT × 182.5 epochs), and reports 274,000 SOL to break even today — rising to 363,000 in year one, 519,000 in year two, and 698,000 at the terminal floor. This brief says 76,528 today. The gap is not a disagreement about disinflation. It is a disagreement about what a validator earns.

Back-solve it. $18,000 at $80 is 225 SOL; plus 201 SOL of vote cost, that is 426 SOL of annual cost. 426 ÷ 274,000 implies a revenue rate of 0.1555% of stake. At a 2.75% commission on the 5.84% nominal yield the proposal cites, the inflation line alone is 0.1606% — agreement within 3%, which indicates the published breakeven table counts inflation commission only, with no block rewards and no MEV. The inflation-only figure here, reached from the opposite direction in §4, is 245,609 SOL. Once the revenue basis matches, the two models converge.

That matters because block rewards are ~64% of revenue in the stack measured here and the inflation line is ~30%. A breakeven built on the 30% line will always show a validator set closer to the edge than one built on the whole stack, and the 698,000 figure is the one that will be quoted.

The proposal's second table does use the whole stack — and reaches a far milder conclusion. Its validator-set analysis takes real epoch-976 rewards from the Trillium API and totals Jito MEV tip commissions, block rewards and inflation commissions. On that basis it finds 2 validators out of 738 moving to unprofitable in year one, 13 in year two and 30 in year three, after which the terminal rate is reached. Two tables, one document, two revenue bases — and the more complete one supports the milder reading. Worth stating in both directions: the alarming number understates validator revenue, and the reassuring number is the better-founded one.

SIMD-0553 was rejected, and this brief's treatment of it now describes a road not taken. SGP-0003 failed on 28 August at 53.90% of participating stake — roughly thirteen points short of the two-thirds bar, on participation of 61.14%. What it would have done is still worth understanding, both because the fee line remains the largest revenue line in this model and because the proposal will probably return. SIMD-0553 struck a different line of the same stack. It would have replaced the leader's existing 2,500-lamports-per-signature component with a flat 2,500-lamport inclusion payment plus a compute-based resource fee that is burned outright. Validator-community analysis across 20 epochs found non-vote transactions averaged 1.1065 signatures each, so the flat payment would have cut that category of validator revenue by about 9.57% — on the order of 35.6 SOL a day network-wide. Two caveats: it would have struck the base-fee component, not the priority fees that make up most of the 0.600% block-reward line, and the 20-epoch measurement has not been independently reproduced here. The structural point survives the vote, and the outcome sharpened it: 0550 cuts the inflation line and 0553 would have trimmed the fee line, so an operator at 0% base commission was immune to the one that passed and fully exposed to the one that failed. The combined result is materially better for operators than either proposal considered alone would suggest — the ~30% line takes a scheduled cut while the ~64% line is untouched. A first-hand measurement gives a sense of what was left unpriced. Sampling twenty mainnet blocks every three hours from 27 to 31 August 2026 — 33 captures, 660 blocks, 861,905 transactions — 19.0% of transactions failed and those failures consumed 26.7% of block compute, each paying the same flat 5,000-lamport base fee as a successful transaction. Individual captures range from 9.5% to 28.7% of transactions and 15.8% to 40.3% of compute, and the failure rate tracks network activity closely, so any single twenty-block sample is close to meaningless on its own and the pooled figure is the one to read. The quantity SIMD-0553 existed to price is not small.

Three smaller basis notes, none of which change a conclusion. The proposal pairs a post-Alpenglow vote cost (201 SOL) with today's inflation rate, where these tables keep them in separate columns. Its inflation path reaches 2.86% / 1.99% / 1.5% at 68% staking, slightly behind the 2.61% / 1.83% / 1.50% modeled in §10, because it starts from 3.82% at epoch 980 and assumes a 4.5-month activation grace period. And it notes the schedule is running roughly 276 days behind its intended timeline because historical slot times exceeded 400ms.

A disclosure, since this section argues with a proposal. SIMD-0550 is authored by engineers at Helius, and roughly two thirds of the signaled support came from Helius's own validator. That is not an accusation — the economic argument is made openly and Anza's CEO gave concept support independently — but concentration of that degree in a support phase is a fact a reader weighing passage odds should have.

12. Where delegation actually comes from — and what's fragile

Segmenting the stake supply (Chorus One and Foundation figures, directional and not independently verified here): market-native delegation ~78%, LST pools ~13%, Solana Foundation (SFDP) now ~6% and falling from 44% at launch — with the SFDP match ratio stepping down toward 0.5:1 and vote-cost coverage sunsetting. The mechanics that made pool delegation extra valuable are the exact mechanics being wound down.

The pinch is where it has always been: operators above ~500–600K SOL of diversified delegation clear breakeven with margin in every scenario above; the cohort below depends on the fragile rows — and every fragile row is administratively granted, not market-won.

The fragile row, evidenced first-hand. One operator's DoubleZero delegation ran 34,737 SOL at the end of March, peaked at 46,898 in April, and stands at 1,973 in mid-August — a ~96% reduction inside two quarters. The cause was a criteria change: region was dropped as a qualifying factor, and the Solana Foundation stake behind the program was redirected to other initiatives. Nothing about the operator changed. This is the cleanest available proof that program stake is not revenue — it is a grant with a review cycle.

Read the same program's other half before concluding the sky is falling. DoubleZero simultaneously runs Edge, where validators earn a share of actual subscriber fees, distributed programmatically each epoch. Subsidy down, market revenue up — capital that was granted being replaced by capital that is earned, and the operators who benefit are those positioned to deliver something a paying counterparty wants.

Community pools share SFDP's sunset risk through three channels: the match that amplified their delegations is stepping down; programmatic and ecosystem SOL routed through them as plumbing leaves when the program does; and their own economics (2–4% of rewards on small TVL) never funded professional management. Retail was never the base.

For any program-dependent validator: treat program delegation as depreciating, and verify concentration on-chain before relying on it. Budget as though the grant expires at the next review, because for a growing number of programs it does.

Known weak points and open questions

The weak points that remain, in order of how much they would move the answer. First, the MEV normalization: observed at ~0.010% of stake, cap-equivalent 0.010–0.021% depending on the source operator's commission, which is not known; this brief uses 0.015%. Second, the activity calibration on the fee line — the difference between a ~77K and a ~99K headline, and the sole remaining gap in §3. Third, the commission distribution, partially answered by SIMD-0550's own analysis but still not cross-tabulated against stake size. Fourth, epochs-per-year is no longer a constant — 184.75 describes 400ms slots, and the SIMD-0525 ramp makes it stage-dependent.

Ranked honestly, the epoch-count question now sits first rather than fourth. Whether the Validator Admission Ticket stays fixed at 1.6 SOL per epoch or scales down with slot time is worth roughly 33,000 SOL of post-Alpenglow breakeven (§8) — more than the MEV normalization, more than the activity calibration, and more than the entire three-year disinflation path. It is also the one item on this list that a single implementation decision resolves outright.

Hosting is deliberately not on that list. It is the largest single swing in the model, but it is not an uncertainty about the world — it is an input the reader already knows and should substitute. What is unknown is the distribution of hosting costs across the validator set, which nobody has surveyed. Corrections are welcome and are credited by name.

Assumptions & honest limits

$105/SOL — a stated modelling basis, close to but not identical with spot. On 30 August 2026 SOL traded between $104.20 and $106.25 across Coinbase, Kraken and the settlement provider; $105 is a round figure inside that spread, chosen so the basis reads as a modelling input rather than a tick that is stale by morning. This replaces the $75 basis used until 30 August 2026, which had drifted roughly 40% below the market and was quietly overstating every breakeven in the document by about 14%. Re-basing was the right call and it was made late; the earlier figure is kept in the §9 sensitivity so anything quoted from a prior version can still be located. Only the hosting conversion moves with price — every SOL-denominated cost and revenue line is price-independent, which is why the vote and VAT rows are unchanged. $30,000/yr hosting including backup/testnet node — an assumption, not a measurement, and now published as a range in §6 (measured on one operator's contract at $21,636, which lowers breakeven to 67,611); 5% inflation / 10% MEV commission (SFDP caps); gross staking yield ~5.57% at ~3.73% inflation and ~67% staked — cross-checked against public tokenomics data and against one operator's measured commission stream, which implies 5.57% independently. 184.75 epochs per year throughout, for both rates and SOL-denominated costs — a 400ms-slot figure, and no longer a constant once the SIMD-0525 ramp proceeds (§8).

The headline rate is the core three lines only (0.893%). BAM (~0.036%) and Edge (~0.002%) are measured and real but stated separately, and they are not equally available: Edge is becoming standard and BAM is gated on a top Jito Steward ranking and named an early-adopter program. Including both gives 0.931% and drops breakeven about 4%. §2 gives the rate for each profile.

Fee and MEV lines are calibrated to network activity and scale with usage, so the honest read is a RANGE: ~77K today at mid-2025 activity, ~99K if activity has cooled a third. Quote the range, not a point — and quote the hosting row you actually pay.

The MEV line is measured on a ~150K SOL book over ten epochs (observed ~0.010% of stake) and normalized to the 10% commission cap at 0.015%. That normalization is uncertain — the cap-equivalent is 0.010-0.021% depending on the commission that operator actually runs, which is not known. It is the least stable line in the brief.

Excludes labour, and excludes other non-consensus revenue (RPC, swQoS, order flow) — which for many operators is the actual business, with the validator as its credential. Note the asymmetry this creates with §7, which costs the storage and indexing side of an RPC node: that section brings a non-consensus cost into a model that excludes the matching non-consensus revenue. It is stated there and restated here because the two must not be netted — §7's figures describe what state costs to carry, not what an RPC business earns, and neither enters the breakeven arithmetic.

Every figure here is computed, not quoted; the model is one page of arithmetic and available on request.