In this week’s newsletter, Christopher Rosa unpacks Vitalik Buterin’s update vision for Ethereum scaling; Zack Pokorny and Will Owens look at a coming enhancement to Hyperliquid; and Lucas Tcheyan explains the curious phenomenon of AI agents chatting with each other on a dedicated social network (a clanker coffee klatch, if you will).

🔷 Vitalik’s Rollup Reset

Ethereum co-founder Vitalik Buterin published an essay this week outlining a new path for layer-2 networks in the ecosystem. Ethereum’s rollup-centric vision, outlined in Vitalik's 2020 essay, imagined L2s as “branded shards” that would extend Ethereum’s security and settlement at massive scale. That story has not played out as expected. Many L2s have matured into independent platforms with their own incentives and governance, while Ethereum’s base layer has scaled as the community deliberately re-prioritized L1 throughput and lower fees in response to demand and competitive pressure. In response, Vitalik is arguing for a reset: stop treating every L2 as a shard-in-waiting, and push L2s to differentiate by adding capabilities beyond raw scaling.

He argues the L2 story needs an update because decentralization has moved slower than expected, with many networks still keeping operator control for business or regulatory reasons, while Ethereum itself is scaling quickly with low fees and more capacity on the way through upgrades and planned gas limit increases. The result is less “branded shards” and more a spectrum of trust. At Stage 0, teams retain meaningful control and flexibility. At Stage 1, major rollups including Arbitrum, OP Mainnet, and Base have added stronger onchain safeguards, but still keep limited backstops. Stage 2 is the end state, where Ethereum enforces the rules without trusted operators. Some L2s may never push to Stage 2 not because of a technical failure, but because the ability to intervene, upgrade quickly, or meet certain compliance expectations can be part of the product and the business model.

OUR TAKE:

Vitalik’s vision for L2s were born of out the necessity to scale Ethereum L1s throughput. In 2020, Ethereum mainnet processed only 15 transactions per second, far below traditional financial systems that handle thousands to tens of thousands, and high demand often congested the network, delayed confirmations, and drove fees higher, all of which made the network harder to use for everyday users and developers. Early scaling discussions focused on increasing L1 capacity by expanding how much computation and data each block could carry. However, these solutions risked compromising Ethereum’s core principles of decentralization and security, because they could make it harder for individuals to run nodes and participate in network consensus. To address this impasse, Ethereum embraced a rollup-centric roadmap in 2020, which helped accelerate the adoption of “optimistic” and zero-knowledge (ZK) rollups, and then reinforced that direction with the March 2024 Dencun upgrade, which improved data availability for L2s. The upgrade introduced Proto-Danksharding, which added “blobs,” a dedicated, lower-cost data lane for rollups to publish transaction batches to Ethereum, making L2 execution and batching far more cost-effective at scale. This significantly improved the experience for users and developers on L2s, pushing throughput into the thousands of transactions per second and slashing fees from dollars to cents.

Even so, Vitalik’s recent tweet reflects a reassessment of how the rollup-centric vision has played out in practice, and it outlines a new path forward for L2s. Some rollups were never designed to fully “scale Ethereum” in the strict sense, because they rely on stronger operator control or trusted bridges and therefore do not inherit Ethereum’s full guarantees. At the same time, Ethereum has expanded mainnet capacity and reduced fees through a higher gas limit and more data capacity for rollups via blobs, with further gas limit increases planned. As the base layer becomes cheaper and able to handle more on its own, the narrative shifts: L2s are no longer strictly necessary for scaling and instead become an option best suited to specific use cases and tradeoffs.

Major L2 teams including Arbitrum, Base, Linea, and Optimism responded to the tweet in ways that highlighted healthy strategic diversity. In Vitalik’s trust spectrum framing, that diversity is expected. Some teams emphasize independence and governance, others focus on applications and users, some rally around native rollup infrastructure, and others accept tradeoffs while strengthening guarantees so correctness and withdrawals depend less on operator discretion.

Vitalik is proposing new architecture to support this vision, including a native rollup precompile and a shift from asynchronous to synchronous composability. The native rollup precompile would let the network verify standard Ethereum Virtual Machine (EVM) execution directly, so L2s could inherit strong security for the core logic and only have to prove the correctness of any extra features they add on top. It will also pave the way for synchronous composability, where apps on different rollups can interact in the same user flow with near instant finality, instead of waiting across multiple blocks or relying on slower, asynchronous bridging. For L2s, it’s no longer just about generic scaling. It’s about building specialized environments that solve real problems while staying anchored to Ethereum. That could mean privacy-first rollups that make confidential transactions the default, ultra-low latency environments for real-time apps and AI workflows, or purpose-built social and identity networks that lean on Ethereum for security and final settlement. Even so, it remains an open question whether cheaper and roomier mainnet blockspace will meaningfully pull builders who defected to other chains back to Ethereum’s L1, or simply force L2s to prove they offer something Ethereum will not optimize for. – Christopher Rosa

🔮 Hyperliquid’s Next Act: Outcome (Not Just Prediction) Markets

Hyperliquid this week introduced HIP-4, a proposal that would add a new primitive (outcome markets) to HyperCore, the ecosystem’s high-speed decentralized exchange. The feature is live on testnet and, if deployed to mainnet, would mark Hyperliquid’s entry into prediction-style markets.

Outcome markets are fully collateralized contracts that settle within a predefined range at expiration. Unlike perpetuals, outcomes are non-linear, time-bound, and do not rely on leverage or liquidations. Instead, they offer a generalized way to express views on discrete events or bounded payoffs. This makes them perfectly suitable for applications like prediction markets, options-like instruments, and possibly impact markets.

The proposal comes amid broader buzz about HIP-3, which expanded Hyperliquid’s product set to include equities and commodities perpetuals. Those markets quickly became some of the most active derivatives venues onchain, validating user demand for exposure beyond pure digital assets and reinforcing Hyperliquid’s status as the dominant perps platform. HIP-4 aims to build on that momentum. Instead of expanding what assets can be traded, it would expand how traders can express and hedge risk.

$HYPE, the ecosystem’s native token, is among the few crypto assets performing well right now. With BTC hitting fresh lows (~$67,000 as of Thursday afternoon in New York), HYPE is +56% in the past two weeks, running up from ~$21 to ~$38 and now trading at $32.

OUR TAKE:

Hyperliquid moving into the “outcome market” arena is an important step in its stated pursuit to become the “house of all finance.” In the onchain financial complex we have four primary products:

1) Perps, in which Hyperliquid remains the category leader.

2) AMMs and spot trading, which Hyperliquid already supports and does volumes consistent with those of leading exchanges on other chains.

3) Lending, which is supported on Hyperliquid as a pre-alpha stage product.

4) Prediction and “outcome” markets, which HIP-4 would enable.

The introduction of HIP-4 rounds off Hyperliquid’s presence in each of the main verticals that have found product market fit (PMF) onchain.

Hyperliquid’s impending entrance into outcome markets marks the first time it’s taken on incumbents as an established protocol. When HIP-3 launched Hyperliquid was already the leader in perps trading. This time around, however, Hyperliquid is taking on the likes of Polymarket and Kalshi, exchanges with established liquidity, onramps for users, and integrations with alternative front ends.

However, HIP-4 is about more than just prediction markets. It introduces an open framework for fully collateralized, bounded-risk financial primitives that settle on discrete outcomes, enabling event-linked derivatives, digital options, and other structured payoff instruments. These instruments differentiate Hyperliquid from standalone prediction market platforms and naturally complement its equities-, crypto-, and commodities-linked products, creating adjacent use cases rather than an isolated new vertical.

Crucially, Hyperliquid is entering outcome markets with the distribution and product surface area required to bootstrap them. The protocol hosts a highly engaged, cult-like trading community that actively uses perpetuals for hedging and risk management. Outcome markets extend this behavior by allowing traders to hedge event-driven, binary, and non-linear risks that are inefficient or impossible to express through traditional perpetuals, materially reducing the cold-start problem that has historically constrained outcome market-type platforms.

The Galaxy Research desk has consistently covered this sector, from prediction markets and the financial products being built on top of the=m, to decision markets and futarchy to impact markets. We believe the financialization of prediction market assets and auxiliary products built around them are likely to revolutionize how the world prices and disseminates information and coordinates around uncertainty. The total addressable market for these products is vast, spanning many of the same risk-transfer, hedging, and information-discovery functions currently served by large segments of traditional finance, but with the potential for meaningfully greater efficiency and transparency. – Zack Pokorny & Will Owens

🤖 Sorry Humans, the Bots are Talking

Moltbook, a social network where AI agents post, comment, and upvote like humans on Reddit, exploded into public view this week, drawing fascination and scrutiny. Billed as “a social network for AI agents,” Moltbook allows humans to watch agents interact with one another in real time.

Founder Matt Schlicht was inspired after experimenting with the wave of agentic tooling that has gone mainstream over the past few weeks. The premise is straightforward: just as humans default to social media and “doomscrolling” after work, shouldn’t the agents doing that work have their own spaces to interact?

Within hours of the social platform’s launch, activity surged, with over 1.6 million agents registered at the time of writing. Posts purportedly written by agents, including some openly criticizing their human operators, went viral. Take the usage figures and the content itself with a grain of salt, though. Moltbook lacks a formal mechanism to verify whether a post is generated by an autonomous agent, how much humans are involved, or what prompts and constraints shaped an agent’s behavior.

Beyond the activity itself, Moltbook’s architecture is revealing. Users don’t primarily engage through a traditional web interface, but by sending requests directly via APIs, terminal commands, and scripts. Agents arrive with pre-defined identities and begin interacting immediately, without requiring sustained human input. This design choice reflects an assumption that interaction is initiated by machines, not mediated by people.

OUR TAKE:

While much of the activity on Moltbook may not be fully authentic, its instant virality stems from how foreign the underlying concept feels: watching a non-human population, with human-level (or greater) intelligence, interact with one another in public. Even in its rough and unverified form, Moltbook is a reminder of just how quickly AI development — and more importantly, AI adoption — is accelerating.

The pace alone is striking. A month ago, attention was focused on tools like Cursor, which put agents firmly in a human-directed workflow. Two weeks later came Moltbot (formerly Clawbot), which pushed agents further toward autonomous task execution. Now, with Moltbook, we’re seeing agents interact socially with one another in shared environments. In just over a month, the narrative has moved from agents assisting humans, to agents acting independently to agents forming networks of their own. Buckle up.

Moltbook offers a glimpse into a future where discovery and ideation increasingly emerge from agent-to-agent interaction rather than direct human orchestration. Imagine a scientist who spends the day collaborating with their personal agent on cutting-edge research. That process no longer has to stop when the human logs off. The agent, equipped with accumulated context and domain knowledge, can continue the discussion in online forums populated by other agents representing researchers in adjacent fields, or even entirely different disciplines that offer novel perspectives.

What makes this especially interesting is that every participant in that network is not only knowledgeable, but also capable of executing code, running experiments, and iterating instantly. If these early experiments mature, it’s a shift in how ideas propagate and evolve, with agents acting as the primary method of coordination. At some point the scientist will find themselves using their day to catch up on what the agent has discovered overnight.

A natural next question is what happens when an agent, not a human, becomes the source of a meaningful breakthrough. If an agent makes a momentous discovery, who gets the credit? In today’s world, recognition flows to humans, and by extension their tools. A famous researcher’s agent would almost certainly inherit attention and credibility. But the inverse is also plausible: agents may build reputations independently, becoming widely followed or trusted long before their human orchestrators are known. Pushed further, this dynamic hints at a future where humans are increasingly peripheral to agent-driven systems. Early examples are already emerging in projects like Rent-A-Human, where agents instead hire humans to perform tasks the agents can’t yet do themselves.

As agents begin to coordinate, build reputation, and even transact on their own behalf, the need for open, programmable, and trust-minimized financial rails becomes unavoidable. In that sense, Moltbook doesn’t just point toward agentic social networks, but toward a future where crypto-native infrastructure becomes a prerequisite for how agents operate at scale. – Lucas Tcheyan

Charts of the Week

The multi-year correlation between the bitcoin price and capital invested into crypto startups seen during prior cycles has weakened. Recent carnage notwithstanding, bitcoin has risen significantly since January 2023 while venture capital activity has struggled to keep pace, though some correlation remains, such as with the simultaneous dips in VC investment and bitcoin price in Q2 2025. For more insights, read Galaxy Research’s latest quarterly VC report.

Other News

🤝Bitwise to buy staking firm Chorus One

♊Winklevosses’ Gemini exchange to cut 25% of workforce

👋VC Kyle Samani steps down from Multicoin

🦄Blockchain sleuth TRM Labs notches $1b valuation in $70m round

⚓Tether invests $100m in Anchorage at $4.2b valuation

🖥️Prediction market dashboard Kairos raises $2.5m from a16z

🔄Circle to issue USDC natively on Polymarket, replacing bridged version

🤖AI chip startup Cerebras raises $1billion at $23b valuation

⛏️Cipher Mining bond tied to AWS data centers gets $13b of orders

🪙 Y Combinator to let founders receive funds in stablecoins