Hydra and Mithril: What Cardano's Scaling Actually Delivers

Three articles into this series, you've seen Cardano's foundations, its governance overhaul, and its confidentiality bet. This last piece is about speed — the gap between what Cardano's scaling stack does and what it's been advertised as doing. Two technologies, Hydra and Mithril, shipped and work. A third, Leios, is still in testnet. And one number — a million transactions per second — has followed Cardano since 2020 despite being disowned by the people who calculated it. Separating the three is the point of what follows.

The Reality Check: What Cardano's L1 Actually Does Today

Start with the baseline, before any scaling claims enter the picture. Real-world throughput, per Chainspect, sits around 0.29–0.34 TPS averaged over the last hour, with a measured max of 11.62 TPS across the last 100 blocks and an average block time of about 20.01 seconds. That's the honest floor. Not a criticism — a measurement.

It isn't stagnant by neglect. Two governance actions in 2025–2026 raised the Plutus maxTxExecutionUnits memory parameter from 14,000,000 to 17,500,000, roughly a 25% increase in the computational headroom a single smart-contract transaction gets. That helps complex DeFi and NFT contracts fit inside a block without hitting execution-unit ceilings. It does not move the order of magnitude. Cardano's L1 in mid-2026 processes well under one transaction per second in practice, same as it has for years. Everything below — Hydra, Mithril, Leios — has to be read against that number, not instead of it.

Hydra Heads: Isomorphic State Channels, Not a General-Purpose L2

A Hydra Head is a multi-party state channel. A fixed set of participants lock funds on L1, run a mini-ledger among themselves off-chain, and settle only the net result back to the main chain when they're finished. If you've read our Lightning Network piece, the shape is familiar — Hydra is Cardano's version of a channel, not a rollup with an open validator set.

The distinguishing property is what the protocol calls isomorphism: transaction validation and Plutus script execution inside a Head follow exactly the same rules as on-chain. Nothing is reimplemented in a separate VM, and nothing needs a separate trust assumption — a contract that works on L1 works inside a Head unmodified.

That correctness guarantee has a structural cost. A Head runs on BFT-style consensus where every participant must agree on every transaction before it's committed. That caps how many participants a single Head can practically support, and it requires a static, pre-configured peer topology known in advance — you can't have strangers wander into a Head the way they can open a channel on a public payment network. Hydra's own known-issues documentation adds more edges: a Head can't finalize while it's holding minted-but-unburned native tokens, it doesn't support Byron-era addresses, and tolerance for offline peers is bounded. None of this is a defect. It's what a consensus channel for a known group of counterparties looks like — currently shipping as protocol version 2.2.0.

The Million-TPS Myth

The number that won't go away traces to a single 2020 IOG blog post. It reasoned that a single Hydra Head could achieve "up to roughly 1,000 TPS," and that running "1,000 heads in parallel — for example, one for each stake pool of the Shelley release — we should achieve a million TPS." Read carefully, that's a theoretical aggregate: 1,000 isolated groups each running their own channel, summed on paper. It was never presented as an L1 throughput claim, and it was never a promise about what one network could process as a whole.

By 2023 that distinction had been lost in circulation, and Cardano's own technical leadership stepped in to correct it. Matthias Benkort, then Cardano Foundation Technical Director, said publicly: "Please stop spreading the '1 million tps' narrative... it is misleading." That's about as credible a source as a debunk gets — not an outside critic, but the technical lead of the organization stewarding the protocol.

The number resurfaced anyway. At the Hydra Doom Tournament — qualifiers on 3 December 2024 in Buenos Aires, finals 20–25 January 2025 in Las Vegas, a $100,000 prize pool — every frame of a multiplayer Doom match was streamed as a Hydra transaction, and Charles Hoskinson said the network had "crossed the 1 million TPS mark globally." Analysts who looked closer noted those transactions never settled on Cardano mainnet: it was a testnet demo simulation, an impressive engineering stunt, not evidence of mainnet capacity.

The Hydra FAQ has since been rewritten to hedge directly. It now states the 1M figure "has been previously referenced as a theoretical maximum, but the reality is more nuanced," and that summing TPS across independent heads "isn't accurate." That's the honest version: aggregate throughput across many isolated, closed channels running in parallel says nothing about what the network as a whole, or any single open and permissionless path through it, can process. Each Head's ceiling is bounded by the hardware running it, not by cryptography — a thousand closed rooms running fast doesn't make the building's one front door any wider.

What's Actually Running on Hydra Today

Hydra entered what IOG calls its "adoption phase" in February 2026. The framing this time is more careful: IOG describes it as "a production-ready platform already delivering transaction throughput orders of magnitude higher than current layer 1 capacity" — a claim about relative improvement over Cardano's own sub-1-TPS baseline, not an absolute number.

The deployments are thin but real. DeltaDeFi runs a perpetuals DEX on Hydra, currently with restricted deposits — a deliberate early-stage security measure while the team builds confidence in the setup, not a technical ceiling. Masumi uses Hydra for agent-to-agent commerce: buyer and seller agents each run a hydra-node and co-sign snapshots as they transact. In April 2026, Hydra 2.0 shipped as an alpha, removing the old commit-phase/collectCom/abort flow in favor of directly opening a Head with gradual deposits — production hardening, not a new throughput headline.

The honest read: Hydra has moved from research demo to a handful of live, deliberately restricted production apps. That's genuine progress — not yet evidence of broad adoption, and none of it changes the L1 number from the first section.

Mithril: The Boring Infrastructure That Actually Works

Mithril gets a fraction of the attention Hydra does, which is backwards, because it's the piece that helps every node operator and light client, not just participants in a closed group.

The mechanism: stake pool operators each sign a snapshot, and their individual signatures are combined, weighted by stake, into one compact certificate. A node or wallet can verify that single certificate instead of replaying the entire history of the chain to confirm a snapshot is genuine — a stake-based threshold multi-signature standing in for a full replay.

The payoff is concrete. Before Mithril, bootstrapping a full Cardano node took days, by IOG's own description. With a Mithril client pointed at an aggregator, a full mainnet node now restores in under 20 minutes. A mainnet upgrade on 23 March 2026 finished decentralizing light-client verification, replacing centralized intermediaries with SPO-run signer networks, and cut average light-client bandwidth by about 40%. The current release, Mithril v2617.0 (23 April 2026), is still beta-grade for some artifact types — this is ongoing infrastructure work, not a finished product.

None of this raises throughput by a single transaction per second. What it does is make syncing and light verification fast for every node and wallet on the network. Less glamorous than a million TPS, and more broadly useful than anything Hydra ships today.

Leios: The Real Answer to "When Does L1 Get Faster"

If the L1 number from the first section is ever going to move, Leios is the mechanism meant to move it. It works by splitting block production into two streams. Endorser Blocks (EBs), produced roughly every 5 seconds by committee members, reference and bundle the actual transaction data. Ranking Blocks (RBs), produced roughly every 20 seconds using the existing Praos consensus mechanics, anchor and finalize the ordering of those EBs. Separating "what transactions happened" from "in what order are they final" is what lets the network process more data per unit of time without changing the security model underneath it.

The projected numbers are projections, not results: up to 1,500+ TPS, a 30–65x improvement over today's baseline, with IOG's treasury-proposal framing describing a wider "10x to 65x through a phased rollout." Those figures come with real qualifications attached — the codebase already exceeds 705,000 lines, and IOG has requested a ₳27.7M treasury proposal to take Leios from testnet prototype to a mainnet-ready release candidate.

A public testnet, "Musashi Dojo," launched 23 June 2026 across five stages — Earth, Water, Fire, Wind, Void. Mainnet is targeted for "end of 2026" per an IOG product manager, contingent on testing, audits, and, as Part 2 of this series covered, on-chain governance approval. That last part matters: since Chang, protocol upgrades don't ship on a team's say-so, they go through DReps and SPOs. Leios could still slip. It is not locked in.

Beyond Hydra: Midgard

Hydra isn't Cardano's only L2 bet. Midgard, built by Anastasia Labs, is an optimistic rollup: transactions are assumed valid on submission and can be challenged via fraud proofs during a window, with no central sequencer running the show — the same trust model that underpins Ethereum's rollup ecosystem. Like Hydra, it's isomorphic, so existing Cardano dApps can redeploy without rewriting their contracts.

As of September 2025, Anastasia Labs reported a "functional node" with live L1 settlement pathways. A public testnet is targeted for Q3/Q4 2026. As of this writing, that testnet hasn't launched. Midgard belongs in the same category as Leios: a credible roadmap item, not a current capability.

The Honest Verdict

Put the numbers side by side and Cardano's position is unambiguous.

Cardano's L1, at well under 1 TPS in practice, trails both Solana's real-world ~1,349 TPS and Ethereum's ~25.6–25.8 TPS by a wide margin, and trails Ethereum's combined L2 throughput of ~226.92 TPS too. Hydra can match or beat either network for a narrow class of problems — closed groups who know each other, like a DEX's matching engine or two commerce agents settling with each other — but it is not a substitute for open, permissionless scaling, and its ceiling is set by the hardware running each Head, not by any figure in a blog post. Mithril doesn't move any of these numbers; it makes every operator's and wallet's experience faster without touching throughput at all. Leios is the one credible, funded, testnet-stage path to closing the L1 gap, and it isn't live yet.

That's the distinction worth carrying away: the engineering behind Hydra and Mithril is sound, and Hydra is narrowly but genuinely useful today. The marketing built on top of it — a single number repeated past the point its own authors stood behind it — is a separate thing. It's the marketing, not the cryptography, that earned the scrutiny in this article.

Four parts, one throughline. Part 1 covered the base layer: Ouroboros consensus and the eUTXO model everything else is built on. Part 2 covered the governance overhaul that now decides whether upgrades like Leios actually ship. Part 3 covered Midnight's confidentiality bet, a genuinely different direction from either Hydra or Leios. This part covered scaling, and the honest summary applies to the whole series: Cardano's 2026 story is real and improving, but it is still mostly ahead of itself, not behind — the gap between a roadmap and a result is exactly where the network sits right now.