TowerBFT just lost the testnet.

On September 24 Anza confirmed the Alpenglow handoff on Solana’s public testnet: at slot 444625255 the Alpenglow genesis block formed, TowerBFT was retired, and Votor began finalizing. Validator votes moved off-chain. Roger Wattenhofer, a co-author of the May 2025 proposal, called it a successful “Alpenswitch.” The feature gate (SIMD-0326) had been counting down about 5,000 slots. Only operators on Agave 4.3 took part. Firedancer and Frankendancer do not speak Votor yet. For this cluster, Solana is a one-client test.

What changes — and what does not

Today’s TowerBFT writes votes on-chain and needs 32 slots, about 12.8 seconds, before a block is rooted. Alpenglow replaces that voting and finality logic with Votor: validators send votes to each other, aggregate them into certificates, and finalize in one round if 80% of stake notarizes or two rounds at a 60% threshold. Target finality under normal conditions: about 150 milliseconds (simulations have printed as low as ~100 ms). “Confirmed” and “finalized” collapse into one. Slots stay near 400 ms. The SVM, transaction format, fees and user-facing send path stay put. Users do not upgrade a wallet.

Rotor is not this release. Block propagation remains Turbine. Rotor, the planned single-relay replacement, is a later, separately numbered proposal. Proof of History leaves the consensus path. Resilience math shifts: Alpenglow is specified to tolerate 20% Byzantine stake plus 20% crash, a different mix than Tower’s one-third. Vote cost moves from on-chain vote fees toward a validator admission ticket burned per epoch.

How they got here

Governance approved SIMD-0326 in September 2025 with 98% of votes in favor. A community cluster ran Alpenglow from May 2026. Agave 4.3 went general-adoption on September 21 after a staged 10% / 25% / all-mainnet-validator call. Public testnet is the first time the full test cluster, indexers and RPC shops see Votor commitment behavior. Triton One had already warned data consumers to update. Nodes that sit outside the admitted voter set learn finality from certificates in later block footers — or they lag.

Mainnet is not next Monday

September 28 appears on Anza’s Agave 4.3 calendar as a possible window to resume mainnet feature activation. That is not a confirmed Alpenglow mainnet date. Anza’s sequence is testnet, then devnet, then an observation period, then mainnet-beta. Firedancer still has to ship Votor. Indexers, bridges and exchanges that key off “32 slots to final” have to change code. A single-client testnet that works is necessary. It is not sufficient.

Why 150 milliseconds is the product

Exchanges credit deposits on finality. Bridges release on finality. DEX takers and payment apps care about the gap between “sent” and “done.” Sub-second irreversible settlement is how Solana argues it can sit next to Web2 checkout rather than next to a 12-second root. Pump.fun-class mint storms and perp venues are the noisy beneficiaries. The quiet one is any stack that today waits a dozen seconds because that is what Tower required.

Alpenglow on public testnet is the first time that claim is running on a cluster people can break in public. Mainnet still needs a second client, a clean observation window, and Rotor later if propagation is the next bottleneck. Until then, treat 150 ms as the design target that just left the lab — not as the slot time on the chain that holds $117 SOL.