Ethereum Scaling with Cerebras WSE-3, Groq LPU & Low-Latency Inference

Scaling Ethereum: Layer-2 Solutions and the Future of DeFi

Ethereum’s growth has brought a problem: the main chain can only process about 15 transactions per second. During peak demand, users pay hundreds of dollars in gas fees for a single swap. Layer-2 (L2) networks solve this by moving transaction execution off the main chain while inheriting its security guarantees. This post breaks down how the two dominant L2 approaches work, where they fall short, and what they mean for DeFi users today.

How Layer-2 Networks Work

Layer-2 networks are separate blockchains that run on top of Ethereum. They batch hundreds of transactions together, execute them off-chain, and then submit a compressed proof back to the main chain. This reduces the load on Ethereum’s base layer while allowing users to transact at a fraction of the cost. The key difference between L2 solutions lies in how they prove the validity of those batched transactions to the main chain.

The two major categories are Optimistic Rollups and zk-Rollups (zero-knowledge rollups). Both batch transactions, but they use fundamentally different trust models and verification mechanisms.

Optimistic Rollups

Optimistic rollups assume transactions are valid by default. They post transaction data to Ethereum but do not compute anything on-chain unless someone challenges a result. If a validator suspects fraud, they submit a fraud proof, which triggers a re-execution of the disputed transaction on the main chain. If the challenge succeeds, the fraudulent operator is penalized (their staked ETH is slashed).

This model has a critical property: anyone can challenge any transaction during a dispute window (typically 7 days). This “optimistic” design keeps on-chain computation minimal during normal operation, but introduces a delay for withdrawals. When you move funds from an Optimistic rollup back to Ethereum mainnet, you must wait for the dispute window to expire before the funds are released. This means users cannot bridge assets out quickly, which creates friction for arbitrageurs and frequent traders.

To illustrate: suppose you swap 10 ETH for 3000 USDC on Arbitrum (an Optimistic rollup). The swap executes instantly on the L2 at near-zero gas cost. But if you want to move that USDC back to Ethereum mainnet to use it in a lending protocol, you must wait 7 days. During that week, the USDC price could move against you, or you might miss a trading opportunity. Some bridges offer “fast exits” by lending you the funds upfront for a fee, but that adds cost and counterparty risk.

zk-Rollups

zk-Rollups use cryptographic proofs called zero-knowledge proofs to verify that a batch of transactions is valid. Instead of waiting for a challenge period, the rollup operator computes a succinct proof off-chain and submits it to Ethereum. The main chain verifies this proof in milliseconds, and if it passes, the batch is accepted as final. There is no dispute window, and withdrawals are immediate.

The trade-off is computational cost. Generating a zero-knowledge proof for a large batch of transactions requires significant hardware and electricity. This makes zk-Rollups more expensive to operate per batch than Optimistic rollups. However, recent advances in proof systems (like PLONK and STARKs) are steadily reducing this cost.

Consider the same 10 ETH to 3000 USDC swap on zkSync (a zk-Rollup). The swap executes on the L2 with similar gas savings. When you want to move the USDC back to Ethereum mainnet, the withdrawal is processed as soon as the next batch proof is verified, typically within minutes. No 7-day wait. For a trader who needs to rebalance positions across multiple protocols, this speed difference is decisive.

Comparison Table: Optimistic vs zk-Rollups

DeFi on Layer-2

L2 networks are not just scaling infrastructure; they are becoming the primary venue for DeFi activity. As of early 2025, over $15 billion in total value locked (TVL) sits across L2 networks, with Arbitrum leading at roughly $6 billion and Optimism, Base, and zkSync each holding between $2-4 billion. This shift is driven by two factors: lower fees and higher throughput.

A swap on Ethereum mainnet might cost $50-200 in gas during congestion. The same swap on Arbitrum or zkSync costs pennies. This makes previously uneconomical strategies viable, such as frequent small trades, yield farming across multiple protocols, and arbitrage between decentralized exchanges (DEXes).

Consider a yield farmer who wants to move liquidity between Aave and Compound based on the best lending rate. On mainnet, the gas cost of each deposit and withdrawal could eat up a significant portion of the yield. On an L2, the cost is negligible, so the farmer can rebalance positions daily or even hourly without worrying about fees.

Liquidity Fragmentation

The rise of multiple L2s has created a new problem: liquidity fragmentation. A DEX on Arbitrum has a separate pool from the same DEX on Optimism or zkSync. If a user wants to trade a large amount of an asset, they might find shallow liquidity on any single L2. The total liquidity across all L2s might be sufficient, but it is split across isolated environments.

Several solutions are emerging. Cross-chain bridges like Stargate and Hop Protocol use liquidity pools on each chain to transfer assets between L2s, but they add fees and latency. Intent-based architectures (popularized by projects like Uniswap X and Across) allow users to specify what they want to trade and let solvers compete to fill the order across chains. And shared sequencers (like Espresso Systems) aim to give L2s a common ordering layer, so transactions from different rollups can be sequenced together, making atomic cross-chain swaps possible.

Each approach has trade-offs. Cross-chain bridges introduce trust assumptions about the bridge operators. Intent-based systems rely on solvers being willing to front capital. Shared sequencers are still experimental. For now, most users stick to one or two L2s and accept the fragmentation.

Security Risks of Layer-2 Networks

L2 networks inherit Ethereum’s security, but they also introduce new risks. The most significant is sequencer centralization. In most L2s today, a single entity (the sequencer) is responsible for ordering transactions and submitting batches to Ethereum. If the sequencer goes offline, the L2 stops processing transactions. If the sequencer is malicious, it could censor transactions or reorder them for profit (MEV extraction).

Optimistic rollups mitigate this with a forced inclusion mechanism: users can submit transactions directly to Ethereum if the sequencer is unresponsive. But this costs mainnet gas fees and defeats the purpose of using the L2. zk-Rollups have a similar mechanism, but it is less tested in practice.

Another risk is smart contract bugs. L2 networks run their own virtual machines (often EVM-compatible, like Arbitrum’s Nitro or zkSync’s zkEVM). These implementations are new and may contain bugs that differ from Ethereum’s EVM. A bug in the L2’s bridge contract could allow an attacker to drain funds from the L2 to mainnet. The Wormhole bridge exploit ($320 million) and the Ronin bridge hack ($620 million) are both examples of bridge-level vulnerabilities, though those were cross-chain bridges, not L2-native bridges.

Finally, fraud proof systems in Optimistic rollups are not fully decentralized yet. In practice, only a small set of whitelisted validators can submit fraud proofs on Arbitrum and Optimism. If those validators collude or are compromised, fraudulent transactions could go unchallenged. Both projects plan to open validation to anyone, but that transition introduces its own risks.

The Road Ahead

The L2 landscape is still evolving. zk-Rollups are maturing quickly and will likely become the dominant approach once proof generation costs fall far enough to make them competitive with Optimistic rollups on a per-transaction basis. Optimistic rollups, meanwhile, are working to reduce the 7-day withdrawal delay through “fast exit” markets and partial finality mechanisms.

For developers building on Ethereum, the choice between L2s depends on the use case. If your application needs fast withdrawals and can tolerate slightly higher operator costs, zk-Rollups are the better fit today. If you want the most battle-tested infrastructure and can design around the 7-day withdrawal window, Optimistic rollups are safer. Either way, the future of DeFi will run on L2 networks, not on Ethereum mainnet.

For users, the practical advice is simple: if you are paying more than $5 in gas for a transaction, you should be using an L2. The ecosystem is fragmented, but the savings are large enough to justify the extra complexity of bridging and managing multiple wallets.

Sources and References

This article was researched using a combination of primary and supplementary sources:

Supplementary References

These sources provide additional context, definitions, and background information to help clarify concepts mentioned in the primary source.

• AWS will bring Cerebras’ wafer-size WSE-3 chip to its cloud platform
• Product – Chip – Cerebras
• Cerebras
• Groq LPU (Language Processing Unit) performance tested – capable of 500 tokens per second
• Inside Groq’s LPU and its impact on generative AI
• Can Groq Really Take On Nvidia?
• Nvidia's $20 Billion Groq Acquisition Just Paid Off. This New Chip Could Change the AI Inference Game in 2026.
• Nvidia Does $20 Billion Deal With Groq
• LPU | Groq is fast, low cost inference.
• Groq is fast, low cost inference.
• Cerebras WSE-3 (Liquid-Cooled) | Wafer-Scale Engine | Bulk Procurement
• A Comparison of the Cerebras Wafer-Scale Integration Technology …
• Cerebras stock slides after near-70% surge in biggest IPO of 2026
• Cerebras WSE3 Versus Nvidia B200