Hyperscale - Linear Scalability Initiative

500,000 TPS. Sustained.
Scalability without compromise.

On January 31, 2026, over 590 community nodes sustained more than half a million real DeFi transactions per second, while preserving atomic composability across 128 shards. Not a simulation: a public, reproducible test on commodity hardware.

SHARD 0x1ASHARD 0x2ESHARD 0x71SHARD 0x9BATOMIC COMMIT250K+ TPS

Adaptive Sharding and Linear Scalability: The network dynamically activates new shards on demand. The more shards operating in parallel, the higher the total throughput (TPS), always preserving atomic composability in the central consensus.

500K+
Sustained TPS in the January 2026 public test
Doubling the shards doubled throughput: linear scaling confirmed
590+
Community nodes, from datacenters to laptops
128
Shards with atomic composability preserved
Public proof of scale

Half a million TPS,
verified in public

The final Hyperscale public test - January 31, 2026 - closed the interim phase with results anyone can reproduce. Real DeFi swap transactions, community hardware, and atomicity preserved on every single transaction.

Final public test - January 31, 2026 · Results published February 20, 2026
500,000+
Sustained TPS
Steady-state throughput across the full network
~3,900
TPS per shard
Held constant while scaling from 64 to 128 shards
590+
Community nodes
Plus 384 bootstrap and 40 validator nodes on AWS
0
Atomicity violations
Every transaction fully succeeded or fully reverted

Linear scaling, confirmed

Doubling the shard count doubled the throughput - the defining prediction of the architecture. Per-shard performance stayed constant as the network grew.

64 shards~250,000 TPS
128 shards~500,000 TPS

Real transactions, not transfers

The workload used cross-shard atomic DeFi swaps - the hardest transaction class - rather than simplified token transfers that inflate benchmark numbers.

Commodity hardware

Validators ran on 4-core, 16 GB machines. Community participants joined from datacenter servers and ordinary desktop laptops alike - no specialized hardware required.

Consensus was not the bottleneck

The limiting factor was generating transactions fast enough to feed the network. The consensus layer itself still had headroom.

Open and reproducible

Code, tooling, logs and documentation were published in February 2026 so results can be independently verified - closing the Foundation's interim phase.

Architecture

From Cerberus research
to Xi'an engineering

A decade of consensus research, validated in public at scale, now being rebuilt as production software by the community. These are the pieces and how they fit together.

One goal, two generations of design

Cerberus (peer-reviewed in 2020) and the 2026 Hyperscale test validated the goal: linear scalability with cross-shard atomicity. Xi'an is a clean-room Rust rewrite - hyperscale-rs - that pursues the same guarantees with a new consensus design: HotStuff-2 shard chains coordinated by the POLARIS leaderless beacon chain. Metrics from earlier prototypes should not be attributed to Xi'an, and vice versa.

Explore the technology in depth at hyperscale.rs
Research foundation

Cerberus consensus

Braided BFT consensus across a 2²⁵⁶ shard space: only the shards touched by a transaction take part in its consensus, so unrelated transactions never queue behind each other.

Beacon chain

POLARIS - leaderless BFT

A prefix-consensus protocol with formally proven censorship resistance (arXiv:2602.02892). No leader means no single point of censorship - committees are drawn by BLS-backed verifiable randomness.

Live resharding

Adaptive sharding

Shards split under load and merge when demand drops - without halting the network. In-flight work is handed to child shards at a certified boundary. First cut completed in June 2026.

State model

Unified global Merkle trie

All state lives in one global Jellyfish Merkle Tree; each shard owns a prefix subtree. Splits and merges are logical operations - no data migration, no reindexing, no downtime.

Execution

Radix Engine + Scrypto

Asset-oriented execution where tokens are protocol-native objects that cannot be duplicated or lost. The engine and the Scrypto developer experience carry over intact to the sharded network.

Decentralization

Home-validator economics

Every node has equal consensus weight; stake buys nodes, not votes. A 50 Mbps home connection is enough, penalties are jailing rather than slashing, and post-quantum cryptography is designed in from the start.

How a cross-shard transaction commits

Shards do not trust each other - they independently execute and compare results. A malicious shard cannot corrupt state without a supermajority of every other involved shard.

1

Declare

The transaction declares upfront which substates it reads and writes, so the shards it touches are known deterministically - no routing table, no coordinator.

2

Lock

Each involved shard commits the transaction and locks the relevant substates, preventing conflicting access until the outcome is decided.

3

Execute everywhere

Shards share the locked substates with each other, then each one independently executes the full transaction with identical inputs.

4

Exchange certificates

Each shard votes locally on the result and produces an execution certificate. Certificates are exchanged between all involved shards.

5

Commit - or fail cleanly

If every shard reached the same outcome, a transaction certificate commits the results atomically. Any disagreement means no certificate, no commit: the transaction fails cleanly, with no partial state.

For financial institutions

Scale that solves
real settlement problems

Tokenized deposits, RWAs and CBDCs only work if the underlying ledger can absorb the volume of real payment systems - without ever splitting a trade into a paid leg and an undelivered one.

Atomic DvP at market scale

Payment and delivery settle as one indivisible event, even when cash and asset live on different shards. At 500,000 TPS the guarantee is identical to a single transaction: full settlement or full revert. Counterparty risk is not reduced - it is eliminated.

Headroom beyond card networks

Global card networks peak at roughly 65,000 transactions per second. The 2026 public test sustained over seven times that with real swaps - and capacity grows linearly by adding shards, so tokenized deposit and CBDC volumes never compete for space.

Deterministic execution for compliance

Transaction manifests declare their outcome before execution: readable by risk teams, auditable by regulators, immune to blind signing. Under load there are no fee auctions that price payments out or reorder them unpredictably.

Resilience measured in decades

Thousands of independent validators on commodity hardware, graceful degradation if a shard loses liveness, and a post-quantum cryptography path designed into the architecture - infrastructure lifecycles that match banking horizons.

T+0
Settlement finality in seconds, not days
24/7
No batch windows, no cut-off times
Demonstrated throughput vs. card-network peaks
0
Partial executions - atomicity is protocol-level
For the machine economy

Built for transactions
between machines

When AI agents pay per query, per dataset and per API call, transaction volume stops looking like crypto trading and starts looking like HTTP traffic. That is the scale Hyperscale is designed for.

x402: HTTP-native micropayments

The 'Payment Required' status code, finally implemented: AI pays content and data sources automatically at the moment of access, a fraction of a cent at a time. The standard is driven by Coinbase and the Linux Foundation, with Cloudflare backing and an implementation already running on Radix.

Agents with their own wallets

AI Ventures runs live on the Radix testnet: AI agents acting as employees of on-chain companies, with their own wallets, spending limits and on-chain accountability. The application layer is assembling before the scaling layer ships.

A ledger agents can read

Through the Model Context Protocol, AI agents query the Radix ledger directly - balances, transactions, on-chain state - without a human intermediary or a custom integration per model.

Intent that machines can trust

Transaction manifests declare outcomes upfront and execute atomically: no blind signing, no sandwich attacks, no ambiguity. Autonomous agents need deterministic results - exactly what asset-oriented atomic transactions provide.

If you want to build a layer 1 blockchain that can support 100 million transactions per second - call us.

Matthew Prince, CEO of Cloudflare - Bankless, May 2026

Cloudflare handles around 500 million requests per second. No fixed-capacity chain can meet that bar - only an architecture that adds throughput by adding shards can grow toward it.

500M
Requests per second on Cloudflare's network today
100M
TPS named by Cloudflare as the bar for web-scale settlement
2027
Estimated year AI agent traffic overtakes human browsing
Roadmap

The road to Xi'an,
milestone by milestone

Hyperscale validated the science in public; Xi'an turns it into a production network. Development is community-led, open source and funded per delivered milestone.

Already achieved

January 31, 2026

Final public test

500,000+ sustained TPS across 590+ community nodes and 128 shards - with real cross-shard swaps and linear scaling confirmed.

February 2026

Interim phase concluded

Code, tooling, logs and documentation published for independent reproduction, formally closing the Foundation-led phase.

April 2026

Community RFC for Xi'an

Proposal to deliver the production network: six milestones, ~$300k in XRD plus a 50M XRD mainnet bonus. The Foundation shifts to a community-led model.

May – June 2026

Milestone 1 underway, ahead of schedule

Community vote approved funding. POLARIS beacon chain integrated, and the first cut of live shard splitting and merging completed in June 2026.

Xi'an delivery milestones

M1

Adaptive sharding

In progress

Live resharding, validator shuffling, snap-sync and the POLARIS leaderless beacon chain.

Est. August 2026
M2

Radix Engine integration

Planned

Full Scrypto execution on the sharded network: transaction manifests, WASM execution and cross-shard fee collection.

Est. January 2027
M3

Gateway & Wallet

Planned

Rust gateway rewrite and Radix Wallet connectivity - the first normal-user interaction with Xi'an.

Est. April 2027
M4

Public testnet + home validators

Planned

Anyone can run a node: public Stokenet, explorer and a home-validator GUI. The network becomes touchable.

Est. July 2027
M5

Mainnet ready

Planned

Migration tooling, post-quantum primitives, dry-run migrations and an audit-ready codebase.

Est. November 2027

Dates are community estimates from the Xi'an proposal; funding is released per delivered milestone, so the schedule can shift with scope.

Take part

Run a node,
join the next public test

Hyperscale's public tests are carried by the community: anyone with a commodity machine can spin up a node, contribute real throughput and watch it on the live scoreboard. That is how 590+ nodes reached half a million TPS.

Public test status

  • Last public testJanuary 31, 2026
  • Throughput reached500,000+ sustained TPS
  • Community nodes590+ on commodity hardware
  • Next public testTo be announced on radixlabs.net

Minimum node requirements

  • 4-core CPU
  • 16 GB RAM
  • SATA SSD
  • 10 Mbps down / 5 Mbps up, wired

How to take part

1

Install OpenJDK 21

Java 21 is the only software requirement. Linux: sudo apt install openjdk-21-jdk. macOS: brew install openjdk@21. Windows: the official Oracle JDK 21 installer.

2

Download the node files

Get hyperscale.jar and default.config from the Radix Hyperscale Telegram channel and place them in a folder named Hyperscale.

3

Open the ports

Forward port 8080 TCP and port 30000 TCP/UDP on your router to your machine and allow them through the firewall. Disable any VPN during the test.

4

Start the node

From the Hyperscale folder run the command below. In the console you get ledger, network and network -stats to monitor it.

5

Join and watch

Track your node at http://localhost:8080/dashboard/index.html and register at radixlabs.net to join the test on the announced date.

Linux (Ubuntu/Debian)

$ sudo apt install openjdk-21-jdk

macOS

$ brew install openjdk@21

Run Hyperscale

$ java -Xms12G -jar hyperscale.jar -console
Step-by-step guides per platform:

The permanent public testnet, with a home-validator GUI, arrives with Xi'an Milestone 4 (est. 2027). Until then, tests are coordinated per event through radixlabs.net and the Radix Hyperscale Telegram channel.

Scale is no longer
the bottleneck.

Whether you settle tokenized assets for a bank or build agents that pay per query - the architecture that absorbs both is being built in the open, today.