DoubleZero: We Audited the 2Z Network's Own Telemetry (2026)

DoubleZero is a dedicated-fiber network that Solana validators connect to instead of routing consensus traffic over the public internet, and as of today it carries 58.98% of Solana's active stake. That number is not from a press release. DoubleZero publishes its entire operational telemetry at data.doublezero.xyz with no API key, no account and no rate-limit headache, and almost nobody reads it.

So we did. On 27 October 2026 we pulled the validator census, 24 hours of metro-pair latency buckets, every link and contributor record, the shred revenue ledger, and the per-validator reward table — then checked the parts we could against Solana mainnet RPC and the 2Z token's own mint account. Everything below is a number you can re-fetch.

Three things came out of it. The latency improvement is real but smaller and more route-dependent than the headline figure suggests. The jitter improvement is the opposite — uniform, large, and the one that actually matters for a blockchain. And the control plane that decides which links and validators exist on this network does not live on Solana at all; it lives on a separate chain run by four validators.

DoubleZero is a protocol for building permissionless networks out of privately contributed fiber. Contributors commit underutilised wavelength links between data centres, terminate them at both ends with DoubleZero Devices running the protocol's agent software, and cross-connect into the mesh at DoubleZero Exchanges. A validator installs a daemon, gets a tunnel, and its traffic rides dedicated fiber when a route exists and falls back to the public internet when it does not.

The pitch is narrow and technically honest. Validator hardware stopped being Solana's bottleneck a while ago; the remaining limits are bandwidth and the variance in how long packets take to cross the internet. The whitepaper makes two claims about how fiber fixes that: inbound transactions can be filtered for spam and duplicates at the edge by shared FPGA hardware instead of each validator provisioning for the whole firehose, and outbound messages can be explicitly routed and prioritised to cut jitter out of consensus. If you have read our breakdown of Firedancer, this is the layer below it — Firedancer makes a validator faster, DoubleZero makes the wire between validators faster.

Mainnet-beta launched on 2 October 2025. The protocol's token, 2Z, is an SPL token on Solana used to pay for network access, and the network has since grown a second product line called Edge that sells real-time market data over IP multicast — Solana shreds first, then Kalshi prediction-market feeds, Hyperliquid perps and Binance futures.

The code is public: github.com/malbeclabs/doublezero, Apache-2.0, mostly Rust, created 18 February 2025, and last pushed at 02:01 UTC on the morning we ran this audit. 233 issues are open.

423 of Solana's 664 active validators are connected, and they hold 257,073,445.67 SOL of the network's 435,850,937.37 SOL in active stake. That is 63.70% of validators by count and 58.98% by stake, measured at 15:02:56 UTC on 27 October 2026.

Before using a vendor's census we check it against the chain. DoubleZero's validator list contains exactly 664 vote accounts; a getVoteAccounts call to api.mainnet-beta.solana.com returns exactly 664 current vote accounts, and the two sets have a symmetric difference of zero — not one pubkey in one list and missing from the other. The RPC reports 436,873,593.07 SOL including 21 delinquent validators; DoubleZero reports 435,850,937.37, and the difference is 1,022,655.70 SOL, which is precisely the delinquent stake. The list is the current set, exactly, with delinquents dropped. It reconciles to the SOL.

That matters because the figure in circulation is wrong by now. The number repeated across coverage is 62.6%, measured 22 August 2026. DoubleZero's own stake endpoint carries its history, and the connected share has been drifting down:

Measured	Connected stake (SOL)	Share of active stake
7 days ago	269,135,322.49	61.37%
24 hours ago	253,289,383.91	57.94%
27 Oct 2026, 15:02 UTC	257,073,445.67	58.98%

Over seven days the network lost 12,061,876.82 SOL of connected stake — about $1.21 billion at $100.37 per SOL — and 2.39 points of share. Over the last 24 hours it gained 3,784,061.75 SOL back. Both things are true, which is why a single quoted percentage is a bad way to describe this network. It is a connection, and connections churn.

One more thing the census shows that no summary mentions: the connected set is not the big-validator set. The average connected validator holds 607,739 SOL; the average unconnected one holds 741,815 SOL. DoubleZero has more than six in ten validators but under six in ten of the stake, because some of the largest operators have not joined.

DoubleZero's Q2 2026 token holder update says that across measured corridors the network delivers "roughly 21.6% lower round-trip latency than the public internet." The data site publishes the raw material behind that kind of claim: for each metro pair it reports ten-minute buckets with DoubleZero and public-internet round-trip time side by side, full percentile spreads for both, jitter for both, and its own computed improvement percentage.

We took the last 24 hours: 57 metro pairs × 145 buckets = 8,265 rows, every single one carrying both sides. Then we recomputed the headline two ways.

The vendor's own method — improvement between the two average RTTs — gives a median bucket improvement of 17.18% (mean 20.12%). Recomputing on the p50 RTTs instead gives 15.45% per bucket, and taking the median of each pair's median improvement gives 12.77%.

The gap between those is a sampling artifact, and it is worth naming because it decides the headline. A typical bucket contains 120 DoubleZero samples and 3 internet samples. Across the whole window that is 1,518,343 measurements of one side against 21,412 of the other, a 71-fold difference. With three probes in ten minutes, one internet spike moves the average a long way and the median not at all — so an average-based comparison systematically flatters the fiber. It does not invalidate the claim. It means the honest version of "21.6% faster" is closer to 13–17%, depending on which statistic you pick, and that you should pick the median.

The bigger finding is that the improvement is extremely route-dependent. Of the 56 pairs with usable readings on both sides:

Median p50 improvement	Metro pairs	Example
40% or better	7	Seattle–Salt Lake 16.42 ms vs 40.68 ms
20% to 40%	14	Frankfurt–Tokyo 133.88 ms vs 233.66 ms
5% to 20%	23	London–New York 66.76 ms vs 70.70 ms
0% to 5%	8	Singapore–Tokyo 65.76 ms vs 67.64 ms
Slower than the internet	4	Hong Kong–Singapore 34.40 ms vs 30.11 ms

The four pairs where DoubleZero's median round trip is worse than the public internet are Hong Kong–Singapore (−14.2%), Chicago–Washington (−8.4%), Dallas–New York (−4.4%) and Frankfurt–Washington (−0.2%). A dedicated path is only shorter if it is physically shorter; where a carrier already has a good direct route, a mesh that hops through an exchange can lose. The fifty-seventh pair, Los Angeles–Singapore, returned no DoubleZero reading at all in any of its 145 buckets while reporting a 176.83 ms internet path, so we excluded it rather than scoring it as a tie.

Here the result is not ambiguous at all. All 57 of 57 metro pairs show lower jitter on DoubleZero, and the margin is not close: median p50 jitter is 21.5 µs on the fiber against 71 µs on the internet, a 3.3× tightening. DoubleZero's own 30-day dashboard puts the jitter improvement at 83.15%; our 24-hour recompute on medians puts it at 71.3%, which for once is the same story told two ways.

Look at the per-pair spread and the point gets sharper. DoubleZero's jitter sits at roughly 20 µs on almost every route regardless of distance — 20 µs on Frankfurt–Strasbourg at 2.81 ms, 34.96 µs on Frankfurt–Tokyo at 133.88 ms. The internet's jitter scales with how ugly the route is: 25 µs on Chicago–Pittsburgh, 108 µs Amsterdam–Dublin, 188.5 µs Singapore–Tokyo, 243.25 µs Marseille–São Paulo.

That is the real product. A blockchain does not care very much whether a packet takes 60 ms or 66 ms, because its timing budget is built around a worst case. It cares enormously whether the same packet sometimes takes 60 ms and sometimes 61 ms, because the variance is what forces every timeout, every repair request and every vote deadline to be padded. Solana has spent two years cutting its slot time — it is at 212.8 ms today, and Alpenglow moved finality onto certificates — and every one of those cuts shrinks the budget that jitter has to fit inside. A network that delivers 20 µs of jitter from Frankfurt to Tokyo is doing something the public internet cannot, and it is a far better claim than a mid-teens latency percentage.

The network has 154 links across 31 metros and 90 devices: 152 activated, 2 soft-drained, 101 long-haul WAN links and 53 short DZX cross-connects. Total committed bandwidth is 8.745 Tbps, of which 8.725 Tbps is on activated links.

Aggregate traffic across all of them is 71.71 Gbps inbound and 71.65 Gbps outbound. That is 0.82% of activated capacity. The median link runs at 0.138% utilisation, the busiest single link peaks at 28.1%, and 47 of 154 links carry no traffic at all in either direction. ISIS reports 300 adjacencies with zero devices overloaded and zero unreachable.

Read that generously and it is enormous headroom, which is the point of building ahead of demand. Read it strictly and it says the hard part was never the fiber. Note also that the "over 10 Tbps" figure in circulation does not match the network's own API, which says 8.745 Tbps today.

The fiber itself is more concentrated than "permissionless contributions" implies. Fifteen contributors are registered, twelve actually supply links, and three supply none:

Contributor	Links	Bandwidth	Share
tsw	24	2.400 Tbps	27.4%
dgt	19	1.900 Tbps	21.7%
glxy	13	1.220 Tbps	14.0%
cherry	8	0.800 Tbps	9.1%
jump_	68	0.700 Tbps	8.0%
7 others	22	1.725 Tbps	19.7%

The top three contributors supply 63.1% of the network's bandwidth. Jump holds the most links by a wide margin — 68 of 154 — but only 8.0% of the capacity, because most of its contributions are short DZX cross-connects rather than long-haul wavelengths. Link count and link capacity are different leaderboards, and only one of them is load-bearing.

This is the finding we did not expect. Every device, link, contributor and user in DoubleZero's API carries a Solana-style pubkey, and the whitepaper describes a "permissionless controller that lives on a public blockchain." So we looked the accounts up on Solana mainnet-beta, and they do not exist.

Pulling the program IDs out of the repo explains it. The serviceability program — the one that holds the network's actual configuration — declares ser2VaTMAcYTaauMrTSfSrxBaUDq7BLNs2xfUugTAGv, and getAccountInfo on Solana mainnet returns null for it. Same for the record program, dzrecxigtaZQ3gPmt2X5mDkYigaruFR1rHCqztFTvx7. They run on a separate chain, the DoubleZero Ledger, which the data site also publishes.

The DoubleZero Ledger is at epoch 242 with 104,768,826 blocks and 701,307,862 transactions behind it, running node version 2.3.6 at 0.367-second slots and about 17.9 transactions per second, with zero inflation and zero delinquent validators. It has 4 active validators and 399,999.99 tokens of total stake.

Put plainly: the network carrying 59% of Solana's stake is configured by a four-validator chain. That is not a scandal — a control plane is not a settlement layer, it needs liveness and auditability far more than it needs 664 validators, and four honest operators with a public ledger is strictly more transparent than a vendor database. But "onchain control plane" invites people to imagine Solana's validator set is the thing securing it, and it is not.

Two DoubleZero programs are on Solana mainnet, and they are the ones that touch money:

• Passport (dzpt2dM8g9qsLxpdddnVvKfjkCLVXd82jrrQVJigCPV), 177,333 bytes, last upgraded 31 July 2026 at 20:12:16 UTC.
• Revenue distribution (dzrevZC94tBLwuHw1dyynZxaXTWyp7yocsinyEVPtt4), 388,189 bytes, last upgraded 23 October 2026 at 18:19:48 UTC — four days before we measured.

Both sit under the BPF Loader Upgradeable, both have a live upgrade authority, and neither authority is a program. Each is a system-owned account holding exactly 0.001 SOL with zero bytes of data. That shape does not tell you whether it is a single keypair or a multisig vault — a Squads vault looks identical from account inspection — so the honest statement is that the revenue-distribution program is upgradeable by whoever controls one signature slot, and the data does not say how many humans that is. Neither program has a verification record in the Solana verified-builds registry at verify.osec.io, so the Apache-2.0 source on GitHub cannot be cryptographically matched to the bytecode that is actually deployed. Both are the same questions worth asking of any Solana protocol holding user value, and DoubleZero answers them better than most by publishing the code at all.

DoubleZero publishes every connected validator's earnings. Not a summary — the table, per validator, per epoch, with the payout asset. Nobody appears to have added it up.

563 validators have earned something, and the lifetime total across all of them is 1,963,974 2Z. At $0.0279647 per 2Z that is about $54,900 paid out in total, ever. The distribution is steep:

Measure	Lifetime 2Z	At today's price
Top earner (Figment)	129,628	$3,625
Mean validator	3,488	$98
Median validator	993	$28
Unclaimed across all validators	582,399	$16,287

The top ten validators took 33.6% of everything ever paid, the top decile took 69.0%, and the bottom half of the list shared 6.13% between them. 23 validators have earned exactly zero, and 40 have earned under 10 2Z — under 28 cents. Rewards track stake and leader slots, so this is the stake distribution reflected in a payout table rather than anything unfair, but it answers the operator's question bluntly: for a mid-sized validator, DoubleZero is currently a performance decision, not a revenue decision.

Two details in the field definitions matter. Earnings are reported in a field named total_earned_2z while a per-epoch field records which asset each payout was actually taken in — across the last ten epochs, 3,547 validator-epoch cells were paid in 2Z, 392 in USDC and 117 in wrapped SOL. The protocol denominates everything in 2Z and converts fiat payments into it, so the 2Z figure is the accrual and the payout asset is the validator's choice. And the latest finalised reward epoch is 1029 while the current Solana epoch is 1069 — rewards settle about 40 epochs, roughly 79 days, behind live.

DoubleZero Edge sells the Solana shred feed and a growing list of other venues' market data. Its ledger is public down to the metro. Recognised revenue from April through October 2026 is 787,897 USDC, with 52,126 already booked for November; run-rate MRR is 119,507 USDC and ARR 1.43M across 1,043 seats, 136 live subscriptions, 71 paying entities and 16 feeds.

Month	Metered seats	Invoiced subscriptions	Total (USDC)
May 2026	136,693	0	136,693
Aug 2026	73,989	33,643	107,632
Sep 2026	3,348	123,070	126,418
Oct 2026 (open)	0	117,953	117,953

The interesting thing is not the level, it is the handover. Metered per-epoch seat revenue went from 136,693 USDC in May to zero in October while invoiced subscriptions went from zero to 117,953. Edge quietly moved from selling seats by the epoch to selling subscriptions by the month, and total revenue held roughly flat through the switch. Frankfurt alone accounts for 257,090 USDC — 33% of everything — and 27 of 31 priced metros have produced any revenue at all.

Against that, the token. 2Z trades at $0.0279647, a $146.4M market cap and $279.6M fully diluted, ranked #217. Its all-time high was $0.893679 on 2 October 2025 — mainnet-beta launch day — and it is down 96.87% from there, 88.01% over the year, and 57.89% in the last 30 days. Its all-time low was set at 10:15:30 UTC on the morning of this audit. Across 15 DEX pairs there is $2.01M of on-chain liquidity against $205,384 of 24-hour DEX volume; CoinGecko reports $3.69M of total volume, so roughly 94% of 2Z trading happens on centralised venues.

The mint is J6pQQ3FAcJQeWPPGppWRb4nM8jU3wLyYbRrLh7feMfvd, eight decimals, supply 9,998,069,243.89 against a 10 billion cap — meaning 1,930,756 2Z have been burned in total, 0.0193% of supply. Freeze authority is revoked. Mint authority is not: it is still set to a live system-owned account, so the supply cap is a policy rather than a property of the token. That is worth knowing before treating burn mechanics as a deflation thesis, in the same way Solana's own fee burn is too small to matter against issuance.

So the network is at its operational peak while the asset that denominates it sits at its all-time low, on the same day. Both facts are measurable and neither explains the other, which is the kind of thing worth saying out loud rather than resolving.

Almost certainly not, and this is the part retail coverage gets wrong. A Solana validator or RPC operator connects with a simple install and no network reconfiguration. Everyone else does not. DoubleZero's own onboarding documentation says plainly: "User onboarding is currently permissioned for non-validators and RPCs." You fill in a form, watch your Telegram for a reply, and the docs add that "there may be fees associated with Permissioned User usage in the future."

That is the gap between "permissionless network" and what is actually open. Permissionless applies to validators joining and to contributors committing fiber. For a trader, a bot operator or an app developer, DoubleZero is a gated enterprise product — and Edge, the market-data business, prices seats in the hundreds to low thousands of USDC per metro per month.

Which means DoubleZero does not make your trades faster. It makes the consensus layer faster and steadier, and you get that for free whether or not you have ever heard of it: tighter jitter between validators means more predictable block propagation, which means the latency you experience as a trader has a slightly smaller tail. It does nothing for the hop between your machine and your RPC provider — that is still a question of which RPC endpoint you use and how close it is to you, and the same reason a dedicated node beats a public one.

For execution, the variable that still dominates is where the decision is made. A trader on a laptop reacting to a chart is thousands of milliseconds behind the chain no matter whose fiber the validators use. That is the entire argument for copy trading on Solana: uwuu watches a verified on-chain leaderboard and mirrors a trader's position in under 400 ms from infrastructure that is already next to the chain, so the human reaction time comes out of the loop. Our guide to Solana trading bots works through how that compares to running your own setup, and how MEV changes what you actually receive covers the other half of execution quality that has nothing to do with raw distance.

Publishing this much telemetry means publishing your rough edges, and three are worth flagging — not as gotchas, but because anyone else reading this API will hit them.

The validator version flag is a naming artifact. The publisher table carries validator_version_ok, defined as whether a validator's version meets the minimum for its client, and it reports 384 of 405 publishers as failing. Group by client and the reason appears: Firedancer 26.10.4 passes, while HarmonicFiredancer 26.10.4 — identical version string — fails all 33 times. Agave 4.3.2 passes; AgaveBam 4.3.2 fails 235 times, JitoLabs 4.3.2 fails 58 times, HarmonicAgave 4.3.2 fails 39 times. The minimum-version table is keyed on the base client names, so every fork falls through to false. The validators are fine; the flag is not.

A −99.57% throughput delta that is arithmetically impossible. The 30-day network-health headline reports a peak of 111.95 Gbps and a change of −99.57%. Back out the previous window and you get a peak of about 26 Tbps — three times the network's entire 8.745 Tbps of committed capacity. One of those two figures is wrong, and it is not the current one.

An advertised MCP server that is not there. DoubleZero's llms.txt points agents at a Model Context Protocol endpoint at data.doublezero.xyz/api/mcp "instead of scraping pages." It returns 404 to both GET and POST. The REST API works beautifully, so this is a dangling reference rather than a missing capability.

One field we deliberately did not build on: slots_needing_repair sums to 84.75% of observed slots with a median per-publisher rate of 98.44%, which sounds alarming. The documentation defines it only as "number of slots that required repair" — not repaired by whom, against what baseline, or whether a single late shred counts a whole slot. With 102 publishers at exactly 100% and two at exactly zero, that looks like measurement coverage rather than delivery failure. We are reporting the number and declining to draw a conclusion from it.

What is DoubleZero?

DoubleZero is a dedicated-fiber network for blockchains and other real-time distributed systems. Contributors commit private fiber links, validators connect with a daemon, and traffic routes over the fiber when a path exists and falls back to the public internet when it does not. Its token, 2Z, is an SPL token on Solana used to pay for network access.

How much of Solana is on DoubleZero?

423 of 664 active validators, holding 257,073,445.67 SOL — 58.98% of active stake — measured 27 October 2026. The 62.6% figure widely quoted was measured on 22 August 2026; the connected share was 61.37% a week ago and has been drifting down.

Is DoubleZero actually faster than the internet?

On median round-trip time, yes but modestly: 12.77% lower across metro pairs on its own 24-hour telemetry, and four of 56 pairs are slower on the fiber. On jitter it is decisively faster — all 57 of 57 pairs improve, with median jitter of 21.5 µs against 71 µs, a 3.3× tightening. Jitter is the number that matters for consensus.

Can I use DoubleZero as a trader?

Not today. Validators and RPCs connect permissionlessly, but DoubleZero's documentation states that onboarding is "currently permissioned for non-validators and RPCs" — a request form and a Telegram follow-up, with fees flagged as possible in future. Traders benefit indirectly through steadier block propagation, not through a connection of their own.

What do validators earn from DoubleZero?

Across 563 validators that have ever earned, the lifetime total is 1,963,974 2Z — around $54,900 at today's price. The median validator has earned 993 2Z, about $28. The top ten took 33.6% of all rewards. For most operators it is a performance decision rather than a revenue one.

Where does DoubleZero's on-chain control plane run?

Not on Solana. The serviceability and record programs do not exist on Solana mainnet-beta; they run on a separate DoubleZero Ledger with four active validators and about 400,000 tokens of stake. Two programs that handle money — passport and revenue distribution — are deployed on Solana mainnet and both remain upgradeable.

DoubleZero is one of the few crypto infrastructure projects where the claims and the audit trail are in the same place, and that is worth more than any single metric in this article. The gap we would keep an eye on is the one between 8.745 Tbps of committed fiber at 0.82% utilisation and $1.43M of ARR: the engineering is clearly ahead of the demand, and the token's price is pricing that in. Three things would change the picture — connected stake reversing its drift back above 62%, Edge revenue growing beyond a single Frankfurt-heavy feed business, and the control plane widening past four validators.

Meanwhile, all of this happens below the layer you trade on. If what you want is faster execution rather than faster consensus, the lever is where your decisions are made, not whose fiber carries the votes — which is what choosing a Solana trading platform actually comes down to.