Dune Just Released 35 MCP-Native Data Engineers. Surf Bid in 4 Hours.
Dune's May 14 layoffs aren't a crypto bust. They're the first public release of agent-facing data engineers, and the bidding window is days, not months.
On May 14, 2026, Dune CEO Fredrik Haga announced on X that the company had cut 25% of its team and offered, in the same post, to personally recommend them to hiring managers. Within hours, Surf cofounder Ryan Li replied on the same thread pitching his agent-native crypto data stack. If you read that exchange as a crypto story, you're going to miss the trade. The ~35 people Dune just released are the scarcest archetype in AI infrastructure right now, and the half-life of the pool is measured in days.
Why these aren't "crypto hires"
The reflex when a data company in Web3 lays off is to file the resumes under "crypto talent" and move on. That label is wrong, and it's going to cost recruiters who use it.
Dune spent the last twelve months reorganizing around one product: Dune MCP, an open-standard server launched in March 2026 that exposes the platform's multi-chain data warehouse to AI agents through twelve tools covering table discovery, query execution, and visualization across more than 100 chains. The engineers who shipped it didn't just write SQL against Ethereum. They designed the tool surface that an LLM calls without hallucinating, modeled decoded onchain state so an agent could reason over it, and shipped the whole thing to production against a warehouse that millions of users hit.
That skill set, "agent-facing data modeling," transfers directly to any company building MCP servers in 2026. Vercel. Supabase. Snowflake's Cortex. Composio. The entire Anthropic ecosystem. If your sourcing filter is ("Web3" OR "blockchain") AND "data engineer", you will literally not see these people as candidates for your AI-infra role. They will see themselves that way, though, the moment a recruiter frames the conversation correctly.
The bar they already cleared
The non-obvious part of this RIF: the people being released are pre-filtered for the exact attribute every AI-infra founder is screening for in 2026.
Two months before the layoff, Haga posted that Dune had run 300 interviews in March for engineering and data roles with "AI fluency" as a non-negotiable bar. His words: "If you are not hands on AImaxxxing and exploring how AI changes your domain we won't hire you." That hiring filter ran through the company for months. Then the company laid off 25% for reasons Haga publicly framed as a pivot toward AI and institutional products, not performance and not cash. Dune has raised roughly $79 million total, including a $69.4 million Series B, and Haga said the company remains well capitalized.
Read those two facts together. This is the inverse of a normal layoff, which skews toward laggards and underperformers. The Dune list is pre-filtered for AI-fluent data engineers who shipped against a production MCP server. You will not find a cleaner cohort on LinkedIn this quarter.
The Dune list is pre-filtered for AI-fluent data engineers who shipped a production MCP server. There is not a cleaner cohort on LinkedIn this quarter.
What the role actually looked like
If you want to know what these people can do, read Dune's own analytics engineer spec, which is still live on third-party job boards. The work: "Create new data models when we onboard additional blockchains. Reverse-engineer smart contracts and protocol mechanics to determine what to extract, how to decode it, and how to model it. Work hands-on with transformation frameworks (dbt or SQLMesh). Partner with engineers to ensure pipelines are reliable, documented, and cost-aware from raw to curated layers."
Translate that into a sourcing rubric:
- dbt or SQLMesh at production scale, with cost-awareness baked in
- Decoding bytecode and IDLs (EVM plus Solana), which is just a harder version of "parse this messy upstream source"
- Raw to curated layering for a warehouse that millions of users query
- MCP tool design against that warehouse
That last bullet is the one nobody else has shipped at scale yet. The whole MCP server economy is roughly nine months old. The number of engineers in the world who have done this for real, in production, with real users, is small enough to fit in a conference room. Roughly 35 of them just hit the market at once.
This is the kind of search where Boolean strings on LinkedIn break down, because the title is whatever the company chose to call it ("analytics engineer," "data engineer," "protocol engineer," "MCP engineer," none of which is standard yet). Which is exactly the kind of query Refolk was built for: describe the person in plain English ("data engineer who built or contributed to an MCP server against a production warehouse, dbt or SQLMesh, comfortable decoding onchain data") and get a ranked shortlist instead of 4,000 false positives.
The bidding war is already public
Within hours of Haga's post, Ryan Li replied:
"Crypto research has evolved and operating in the AI era demands infrastructure built for agents, not humans clicking through dashboards. We need fast query engines, reliable SQL, structured outputs. At Surf, we've spent Q1 building an end-to-end crypto data stack purpose-built for AI agents."
Li is not a tourist. He was previously CTO of BitTorrent and cofounded CyberConnect and Cyber L2. He has also publicly said Surf is "aggressively hiring engineers in person in the bay area." Surf raised $15 million in December 2025 from Pantera Capital, Coinbase Ventures, and DCG. Since launching in July, the company says it has reached millions in ARR, reports 50% month-over-month growth, and claims 80% of top exchanges and leading research firms as users. Surf ships its own MCP server, @surf-ai/surf-mcp, which dynamically generates tools from its OpenAPI spec: twelve grouped tools across 86 endpoints.
Surf's headcount is under 30 people. Three to five senior hires from Dune meaningfully reshape that company. They know it. They moved on hour one.
When a direct competitor pitches your laid-off staff on the same thread as the layoff announcement, "we'll get to it next quarter" is not a sourcing plan. It's a concession.
The CEO endorsement is the actual arbitrage
Most recruiters will read "ping me if you're hiring top crypto talent" as boilerplate severance kindness and scroll past. Haga's track record says otherwise. He posted publicly under his real name, attached his reputation, and used the phrase "exceptional people I can wholeheartedly recommend." The handful of hiring managers who DM him in the next two weeks get founder-graded referrals, which is a tighter signal than any LinkedIn filter you can build.
The play is to send Haga (@hagaetc) a short note with three specifics: the exact role you're hiring for, the MCP or agent-data context, and a one-line ask ("two or three names you'd bet on for this"). You will get a better top of funnel from that one DM than from a month of inbound on a job post.
In parallel, mine the public artifacts. Dune's spellbook GitHub repo is the canonical source of the company's data models, and the contributor list is a who's who of both staff and the broader Dune Wizards community. Cross-reference that against the npm and GitHub MCP server registry contributors, and you have a sourcing list that no keyword search on LinkedIn will produce. This is also a case where Refolk pays for itself in an afternoon: "spellbook contributors who don't currently work at Dune, weighted toward people with AI/agent work in the last six months" is a one-line ask, not a two-week project.
The second buyer class nobody is sourcing against
Here's the part most AI-infra founders haven't priced in: they're not just competing with Surf, Nansen, The Graph, and Glassnode. Haga explicitly framed Dune's pivot around institutions moving onchain "as currencies, stocks, bonds, and commodities migrate to blockchain rails." Translation: BlackRock, Fidelity, Citadel, and Jane Street data teams are now plausible competing bidders for this exact profile.
Traditional finance pays cash, not equity, and pays it on a faster cycle. If you're a Series A AI-infra startup pitching equity against a Citadel data role, you need to be in the conversation in week one, not week six. Dune's customer list, which the company says includes more than 1 million users and more than 20,000 companies (Base, OP Labs, 1inch, Blockworks among them), means every ex-Dune engineer also walks in with warm intros to the buyer side. That's a hiring multiplier most recruiters forget to score for.
The adjacent pool, while you're at it
The Dune cohort is the highest-quality slice, but it isn't the only one. Gemini and Messari have both cited AI integration as a reason for reducing headcount in early 2026. Blockworks and DL News shut their entire newsrooms this year to focus on data products. There is a small, growing cohort of AI-fluent data and research engineers leaving similar firms over the next two quarters. The same sourcing rubric (MCP work, dbt/SQLMesh, agent-facing schema design, onchain decoding) catches all of them. Build the query once and rerun it monthly.
The window on the Dune-specific list is shorter. Surf bid in four hours. Haga is taking inbound now. If you're hiring for MCP data infrastructure and you haven't sent the DM, sent the cold note, or built the spellbook contributor list by next Friday, you are sourcing against the leftovers.
FAQ
Are these engineers actually willing to leave crypto?
Most of them weren't really "in crypto" the way the label suggests. They were building agent-facing data infrastructure that happened to point at onchain sources. Framing your outreach around MCP, agent tooling, schema design, and warehouse cost, instead of "we're a Web3 company too," will land better with this cohort than with almost any other crypto-adjacent pool.
How do I find ex-Dune engineers who aren't yet on "Open to Work"?
Three signals beat the LinkedIn flag: contributions to the public spellbook repo, authorship on Dune MCP-related issues and PRs, and recent commits to any MCP server in the npm registry. Combine those with a short window on "updated profile in the last 14 days" and you'll surface most of the cohort before they self-identify. Refolk does this combination in a single plain-English query across GitHub, LinkedIn, and the open web.
Is sourcing through Fredrik Haga directly actually realistic?
Yes, if you do it once and do it well. He publicly invited inbound and is on the record vouching for the cohort. Send one specific role, one specific ask, and don't follow up more than twice. The arbitrage is that most recruiters won't bother, so the few who do get founder-graded referrals at zero acquisition cost.
What's the deadline on this pool?
Functionally, three to six weeks for the top of the list. Surf has already pitched publicly. Institutional buyers in traditional finance will move on a slower cycle but pay cash that AI-infra startups can't match on equity alone. If you're competing for the top five names, your offer needs to be live in week two, not week six.