Solana came within 4.51 percentage points of losing the ability to finalize transactions on Wednesday morning, after a single misconfigured internet route at one hosting company took 28.83% of staked SOL offline. The network stops finalizing at 33.34%.

The more uncomfortable number predates the outage: the hosting network that went dark already held 27.34% of all stake, above the 25% ceiling set by the Solana Foundation's own delegation program.

The safety cap that was supposed to prevent this was already broken

Staking platform Marinade Finance, which operates one of Solana's largest liquid staking protocols and has granular visibility into validator behavior, reconstructed the incident from validator data after the fact. A broken route originated at Teraswitch's Miami site, then spread through an internal relay in Amsterdam and cut off twelve facilities from London to Tokyo. North America was unaffected. Teraswitch identified the bug in approximately ten minutes, though full recovery took 33 minutes. CoinDesk's report of Marinade's reconstruction provides the independent write-up of those timestamps.

Finality, in plain terms, is the point at which a transaction can no longer be reversed by the network. Solana requires a two-thirds supermajority of stake to be online and voting to reach finality — a threshold rooted in Byzantine fault tolerance, the consensus framework shared by most major Proof-of-Stake networks. Once one-third becomes delinquent, the votes are insufficient. Blocks may still be proposed, but they do not become final. Payments, liquidations, and oracle updates all wait. That is the freeze Marinade measured against — not a website outage.

That distinction matters for anyone who treats Solana as a payments rail rather than a ticker. What Solana is, and what SOL is used for on the network, is the base layer this incident actually hit: fees, staking votes, and app settlement, not a price chart.

The concentration underlying Wednesday's route failure was not a surprise. Autonomous system AS20326, Teraswitch's network, carries 118,890,767 SOL — 27.34% of everything staked on Solana. An autonomous system (ASN) is the unit of routing control on the internet, effectively one operator's network footprint; the Solana Foundation measures delegation concentration by ASN precisely because it reflects how much stake shares a single operator's infrastructure. During the fault, 94% of it went offline in the same minutes. The Solana Foundation Delegation Program caps any single autonomous system at 25% for exactly this failure mode. That cap was already being exceeded. As of 22 July 2026, public hosting tallies placed Teraswitch first at approximately 27.1%, followed by UAB Cherry Servers at 12.7% and Latitude.sh at 11%. The concentration map was visible a month before the route broke.

How the 4.51-point margin was calculated

The 4.51 figure is not Marinade's branding. It is simply 33.34 minus 28.83. Dividing 4.51 by 33.34 shows the network was 13.5% of a halt-threshold away — equivalent to saying it traveled 86% of the distance to a freeze. The approximately 20 million SOL of headroom represents 4.51% of total stake, using Marinade's implied stake base from the 118,890,767 SOL on AS20326 at 27.34%. If that ASN share or the delinquent peak is revised, the headroom figure adjusts accordingly. The 86% headline is not being treated as a separate measurement.

Marinade's own thread put the near-miss in public before most market desks had noticed it.

1/ Solana got 86% of the way to a halt this morning and it barely registered anywhere. 28.83% of staked SOL went delinquent. Finality stops at 33.34%. We added up the rewards lost across all 90 affected validators. 333 SOL. pic.twitter.com/EEC2gYwQgz — Marinade 🛡️ (@MarinadeFinance) August 12, 2026

1/ Solana got 86% of the way to a halt this morning and it barely registered anywhere. 28.83% of staked SOL went delinquent. Finality stops at 33.34%. We added up the rewards lost across all 90 affected validators. 333 SOL. pic.twitter.com/EEC2gYwQgz

Most operators waited for the route to heal. Three did not

Around 90 validators went dark. Marinade was able to measure 74 of the operators. Fifty-nine of them, holding 80.2 million SOL, came back within the same narrow window in Amsterdam, Frankfurt, and Tokyo because they waited for routing to reconverge rather than failing over to alternative infrastructure. Helius, the second-largest validator on the network, was down for the full 33 minutes. Three validators came back clean: Laine and Cogent Crypto, both run by Sol Strategies, and Lion3d. Decrypt published the operator-level breakdown.

The 333 SOL in missed rewards, approximately $25,600 at Wednesday's price, will be covered by validator bonds at the end of the epoch. That is the number the market can price. It is also the wrong number. Had delinquency crossed one-third, nothing would have finalized for any SOL holder anywhere, and no bond covers that scenario. The last time Solana halted outright, in February 2024, a coordinated restart took close to five hours. A different host, Hetzner, cut Solana nodes in November 2022. The failure mode is not new. The share sitting on one ASN is.

Active validator counts have been shrinking over the same period. Public tallies in January placed the set near 800, the lowest since 2021, down from more than 2,500 at the 2023 peak. Fewer operators, more stake per host, and backup systems that do not switch on together explain how a 33-minute routing bug came within 20 million SOL of a network-wide freeze. Infrastructure concentration is not unique to Solana; researchers have flagged similar exposure on Ethereum, where a large share of validators run on a handful of cloud providers. But the margin between Solana's hosting map and its halt threshold is what turned a routine routing failure into a near-miss measured in single digits.

A halt would not resemble a website going down. Open liquidations on perpetual desks would sit unfinalized. Oracle updates would stall. Bridge messages requiring a finalized Solana slot would wait. The February 2024 restart took hours of human coordination among operators who still had to agree on which slot to resume from. Wednesday never reached that meeting. It got close enough that the meeting agenda is no longer theoretical. Marinade stated as much: if delinquency goes past one-third, nothing finalizes for anyone holding SOL, and there is no bond for that.

The market treated it as a non-event. The math does not

SOL was trading near $76, up approximately 0.6% on the day. No user funds were at risk. The chain kept producing blocks. For anyone watching the price alone, Wednesday was uneventful — and that reading is fair as far as it goes. A halt did not occur. Rewards will be made whole. Calling it a crisis would oversell what the tape showed. Calling it harmless would ignore how little additional delinquency it would have taken.

The structural assessment is less reassuring. Counting stake by hosting provider already understates the correlation, because additional stake on other providers went down in the same window. Marinade said it will review limits across autonomous systems, data centers, and backup paths. The timing is awkward: validators are preparing the Alpenglow finality upgrade, due by October, which promises confirmations on the order of 150 milliseconds. The Bit Journal's earlier Alpenglow briefing set out that speed claim. Faster finality does not help if one provider's routing table can stall the votes that produce it.

What happens next

Three things are worth watching, and none of them is the 333 SOL. The first is whether AS20326's share of stake moves back under 25%, or whether the Foundation's cap remains a rule that the largest host already exceeds. The second is whether operators who sat through the full 33 minutes publish a failover test rather than a statement. The third is Alpenglow itself: if the upgrade lands on the same hosting map, the network will be faster at the same single point of failure.

Delegators seeking a practical test can ask their validator one question: what happens to you if Teraswitch's Amsterdam relay fails again? If the answer is that they wait for it to come back, they are describing Wednesday. A second question is cheaper to verify from the outside: which autonomous system is this validator on, and what share of network stake already sits there? If the answer is AS20326, the 25% rule is not protecting the delegator — it is describing the risk already taken. How Solana staking actually works, including the two-day cooldown, is the other half of that decision. Unstaking does not dodge a halt that has already started; it only changes who the user is delegated to after the next cooldown. Large delegations through a handful of custodians create the same correlated-offline risk. Who wins when Solana ETF staking concentrates validator choice is the institutional version of Wednesday's hosting map.

Note: This article reports on network operations and published validator data. SOL is volatile. Missed rewards in this incident are expected to be covered by validator bonds, but a future halt would not be. Nothing here constitutes investment, staking, or technical advice. Figures are as reported on 12 August 2026.