GnosisDAO Treasury — Independent Risk Teardown

Everything below is built from publicly disclosed positions (covering ~84%, about $137.7M, of the treasury by value; data basis and limitations are spelled out at the end). It’s descriptive, not advice. If you spot an error in the data basis, say so and it will be corrected. And if the treasury WG would rather see this with their own assumptions (off-chain positions, custom scenarios), I would be happy to re-run it. Hope it’s useful input for the community.

TL;DR (five model estimates, each conditional on the assumptions and data coverage spelled out at the end; none of them is a verified fact, a prediction, or advice):

1. Economically (and obviously), the Gnosis treasury is one large ETH position. ~66% of covered value is ETH or an ETH wrapper, and the wrappers moved at correlation ≈ 1.0 with ETH over the measurement window. The sector split (37% liquid staking vs 26% L1/L2) doesn’t capture that overlap.
2. That position (again, obviously) depends on one protocol. Lido holds 52% of covered value. In the modeled incident (temporary freeze plus a mild depeg), the estimated loss is roughly 19% of covered value.
3. It moves slowly. Under the liquidity model’s crisis assumptions (5% of daily volume, order books only), roughly $18.5M is exitable to cash within a week. If those assumptions hold, that figure, not the $137.7M headline, is the scale GIP-151 redemptions would draw on.
4. The volatility follows from the composition. A 20%+ intra-year drawdown occurs in 97–99.8% of simulated paths, and the simulated 1-in-20 bad year lands between −41% and −64% depending on model choice.
5. The scenarios agree on magnitude. Run independently, the engines put modeled stress losses in a $20–90M band, an order of magnitude above the ~$5M one-day VaR.

The full analysis follows.

Snapshot date: 2026-07-10/11 (prices as of 2026-07-10 daily close; analysis run 2026-07-11) Data basis: publicly disclosed treasury positions (token holdings, protocol allocations, and vesting as published on the community treasury dashboard), tracked-asset subset Covered value: $137.7M across 17 assets, approximately 84% of the on-chain treasury by the July-10 snapshot. Not covered: bridged/regional stablecoins (USDC.e, EURe, GBPe, ZCHF, BRZ, BRLA), the Bakkerland RWA tokens (bCSPX/bIB01/bIBTA), the Centrifuge RWA pool, the Frankencoin ZCHF debt leg, HOPR/PNK/OLAS/GIV/small caps, and the off-chain TheDAO loan.

Backdrop

A risk check seems warranted given three recent events:

1. GIP-151 approved pro-rata treasury redemptions, turning “how liquid is the treasury, really?” from a theoretical question into a targeted practice.
2. The April hack wave (Drift $295M, KelpDAO $293M) demonstrated cascade risk: Aave’s TVL fell ~46% without any direct exposure. Protocol dependencies transmit stress even when code doesn’t fail.
3. The Gnosis ecosystem’s own incidents (Safe SquidRouterModule, May 25; Gnosis Pay delay module, June 1) were a reminder that infrastructure risk includes module and patch-velocity risk at home.

Five scenarios were run against the treasury’s public book. Each section reports one; the last section is what they say together. Throughout, headline figures are model outputs under stated parameter choices; where a different defensible choice would move a number materially, that sensitivity is noted inline.

1 · Sector concentration

Concentration math: the sector HHI comes out at 0.258, above the 0.25 “concentrated” threshold, with an effective sector count of just 3.9. The single largest sector (liquid staking) is 37% of the book, and once you note that LSTs are ETH exposure, the economic picture is starker than the sector table: ETH-complex assets (ETH + WETH + stETH + wstETH) are ~66% of covered value.

Stress applied, “stablecoin crisis” preset (stablecoins −15%, DeFi −40%, LSTs −20%, cross-sector propagation on): Modeled loss: −$22.1M (−16.0%), of which LSTs contribute −$10.3M, DeFi −$5.0M, stablecoins −$3.9M. On top of that, −$2.4M propagates into the L1/L2 sleeve despite zero direct shock. The shock sizes are preset choices, not forecasts; the propagation term is the model’s way of expressing that correlated neighbors don’t leave “unshocked” sectors untouched.

* Classification per CoinGecko category data; WETH lands in DeFi rather than L1/L2, a taxonomy quirk that slightly understates the true ETH-complex weight shown above.

2 · Protocol concentration

With holdings attributed to the protocols that actually hold them (Lido, sDAI, Spark, Sky, Aave, Safe locked allocation, Balancer, Uniswap, Rocket Pool, CoW vesting):

• Lido holds $71.3M, 51.8% of the covered treasury (69.5% of the protocol-attributed sleeve).
• Protocol HHI: 0.502, an effective protocol count of 2.0. The engine flags Lido as a single point of failure (>50% of sleeve).
• Next largest dependencies: sDAI $11.8M, Spark $5.4M, Safe locked $4.5M, Aave $2.9M, Sky $2.7M.

Stress applied, Lido incident (temporary freeze with 80% eventual recovery + a mild 20% stETH depeg during the event): Estimated loss: $25.7M, or 18.7% of covered value, from a single protocol event whose severity assumptions are deliberately moderate (compare: stETH traded ~7% off peg in June 2022 without any protocol failure). Different recovery and depeg assumptions scale this number up or down roughly proportionally.

If the diversification debate wants a number to anchor on, this scenario offers one. The point isn’t whether Lido is safe (historically it has been); it’s that in the modeled event a single protocol moves about a fifth of the covered treasury while the second-largest dependency moves ~3%. Whether that concentration is acceptable is a judgment for GNO holders; the scenario only puts a number on it.

3 · Exit liquidity in two regimes

Modeled as disciplined liquidation (fixed participation in daily volume; slippage from venue depth), including per-holding lock status (vested COW, locked SAFE) and DeFi unwind delays:

Position-level structure (crisis-model estimates): the model puts the fast sleeve at USDC (~half a day), WETH (~2 days) and USDT; ETH at ~6 days for its liquid portion and ~27 days for a full exit; and the ecosystem positions as effectively frozen at disciplined participation: SAFE ~93 days, USDS ~330 days, COW ~6.7 years. These day-counts scale roughly inversely with the participation assumption: allow 10% of daily volume instead of 5% and they halve; accept more market impact and they shrink further.

The GIP-151 lens: the redemption mechanism gives GNO holders a claim on “liquid treasury assets.” Under this model’s crisis assumptions, that phrase corresponds to ~$18.5M inside a week without heavy market impact; beyond that, the model has redemptions either waiting weeks-to-months (the ETH/LST complex) or paying a slippage bill measured in tens of millions. Both figures are conditional on the participation and volume-data assumptions above; the point is the order of magnitude, not the third digit.

Important fairness note: these figures measure market-exit liquidity (order books only). They do not model redemption paths: Lido’s withdrawal queue converts stETH→ETH 1:1 in days, and sDAI redeems near-instantly. Read the stETH/wstETH/sDAI lines as “time to exit to cash through markets,” which is the relevant measure in a redemption or depeg scenario where everyone uses those same redemption queues.

4 · Correlation under a crisis regime

Baseline (365-day correlations): 1-day 95% VaR $4.9M (3.5%), expected shortfall $6.1M, annualized portfolio volatility 41.1%, diversification ratio 1.47.

Applying crisis-regime correlations: VaR rises to $5.0M, a change of just +2.2%.

That non-result is the finding. Crisis correlation stress barely moves this portfolio because there is almost no diversification left to destroy: ETH, WETH, stETH and wstETH already sit at pairwise correlation ≈ 1.00 in the baseline. Economically, the covered book behaves like one large ETH position (~66%), a stablecoin sleeve (~19%), and ecosystem tokens that were themselves 0.65–0.71 correlated to ETH over the measurement window. The usual crisis mechanism, where assets that looked independent suddenly start moving together, has nothing to grab here, because the book never looked independent.

5 · Monte Carlo, two models

Two deliberately different simulation models, same portfolio, same seed. The gap between them is itself a finding:

• Model-based (GBM): lognormal paths on the full 365-day covariance matrix, zero drift. Pure volatility structure, no market view.
• Empirical bootstrap: resamples blocks of the portfolio’s actual daily returns from the last 365 days, which inherits the real fat tails and the realized drift of what was a bear-market year for the ETH complex.

How to read the two columns: GBM answers “what does 41% volatility do to a portfolio with this correlation structure, with no view on direction?” Its +13% median is a lognormal artifact, not a forecast. The bootstrap answers “what if next year statistically resembles the last one?” Its −35% median simply replays the fact that the trailing year was deeply negative for two-thirds of this book. Neither column is a prediction; reality’s distribution sits somewhere in the band between them, and which column a treasury plans against is a risk-appetite choice for the DAO. The width of the band is itself the finding:

• A 20%+ intra-year drawdown occurs in 97–99.8% of simulated paths under either model. That counts as near-certain only within the models, though: both extrapolate from the trailing year’s volatility regime, and a calmer regime would lower the frequency.
• The simulated 1-in-20 bad year sits between −41% and −64% depending on model choice. The spread is model risk, and a concrete illustration of why any single-model VaR figure for a book like this deserves skepticism.

The combined picture

The five engines describe the same treasury from five angles, and the findings point the same way. As throughout, these are model estimates conditional on the stated assumptions and on the public data being complete and correctly attributed:

1. It behaves as one position. ~66% of covered value is the ETH complex, and the measured correlations treat the wrappers as the same asset (ρ ≈ 1.0). The sector split (37% LST vs 26% L1/L2) doesn’t capture that overlap.
2. That one position lives mostly in one protocol. Lido holds 52% of covered value; effective protocol count 2.0. The modeled Lido incident (freeze + mild depeg) prices out at ~19% of covered value.
3. The position is slow to exit through markets. In the model, even in normal markets 76% of the book can’t exit within two weeks at disciplined participation; under the crisis parameterization, seven-eighths can’t exit within a week. On those assumptions, GIP-151’s “liquid treasury assets” corresponds to ~$18.5M of crisis-liquid value.
4. The volatility follows from the composition. At 41% annualized vol, both simulations make a 20%+ intra-year drawdown the dominant outcome (97–99.8% of paths), with the simulated 1-in-20 bad year between −41% (model-based) and −64% (empirical bootstrap of the actual trailing year). On these numbers, the ~19% stablecoin sleeve is the buffer between the DAO’s operating budget and that distribution.
5. The stress scenarios agree on magnitude. Stablecoin-crisis: −$22M. Lido incident: −$26M. 1-year 95% VaR: −$57M to −$89M depending on model. Taken together, and only as far as their assumptions hold, the engines describe stress losses in the $20–90M band, against a one-day VaR of ~$5M.

None of this says the composition is wrong. A DAO whose mission is Ethereum infrastructure holding staked ETH is a coherent strategy, and the hard-lock share is small (4.6%). Nor is any of it a recommendation. What the numbers describe is a risk profile that is deliberate, concentrated, and slow-moving. If the community finds these estimates useful, three quantities lend themselves to being tracked on a standing basis instead of argued once:

• Crisis-liquid coverage: liquid-within-7-days value ÷ plausible redemption/operating demand (today, under this model’s assumptions: ~$18.5M against whatever GIP-151 may summon)
• Single-protocol exposure: Lido share of the covered treasury (today: 52%; in the scenario’s arithmetic, each 10pp held at a second staking venue would reduce the modeled single-event loss by ~$5M)
• Stablecoin runway: stablecoin sleeve ÷ monthly operating spend, read against how common a 20%-drawdown year is in the simulations (97%+ of paths)

Method & limitations

• Every headline figure in this document is a model output, conditional on (a) the public data being complete and correctly attributed and (b) the stated parameter choices (participation rates, shock sizes, recovery/depeg assumptions, correlation windows). Different but defensible parameter choices move the results materially; treat the numbers as orders of magnitude, not point facts.
• Public on-chain positions only; ~84% coverage of the July-10 snapshot by value. Bridged/regional stables, RWA positions, the ZCHF debt leg, and the off-chain loan are excluded. The missing sleeve is mostly stability-adding, so concentration findings here are, if anything, slightly overstated in the treasury’s favor on liquidity and slightly understated on diversification.
• LP positions decomposed 50/50 into constituent legs (estimate). Protocol attribution follows the public dashboard’s protocol section, scaled to on-chain totals.
• Liquidity model: fixed participation-of-ADV with venue-depth slippage; exchange volume only. Protocol redemption queues (Lido withdrawals, sDAI) are not modeled, and they materially improve real-world stETH/sDAI exit paths (see §3 note). Volume data for several assets uses aggregated market data rather than order-flow (flagged internally; treat thin-asset day-counts as order-of-magnitude).
• Correlations from 365 days of daily returns; crisis regime estimated from a 5-year window. Monte Carlo: two models, 25,000 paths each, seeded and reproducible: GBM on the full covariance matrix (zero drift) and a block bootstrap of the actual trailing-365-day return series (which inherits that year’s realized drift). Distributional metrics, not forecasts; the inter-model spread is reported as model risk.
• One classification quirk (a $20k position tagged to the wrong sector) was left as-is; immaterial at 0.01% of value.
• All figures at 2026-07-10 prices. Snapshot-based; no protocol-code audit; nothing here is investment advice.