nextbig.dev
Vancouver, B.C. · Intelligence on AI and the machines that run it
nextbig.dev
The Briefing · Saturday, June 20, 2026

Builder.io ships an MIT-licensed framework that makes the agent a first-class user of your app

Builder.io ships an MIT-licensed agent-native framework; ClickHouse hits $250M ARR; SemiAnalysis says 99% of custom AI ASICs fail before silicon.

11 min read
The Rundown No. 120 · Audio Edition · 10 min All episodesRSSMP3
0:00 / 10:00
VTT
The Big Story
Builder.io ships an MIT-licensed framework that makes the agent a first-class user of your app

Builder.io released agent-native, an open framework where the agent and the UI are equal citizens of the same system. It is MIT licensed, sitting at roughly 3,800 stars, runs on any SQL database, and deploys anywhere including serverless. The pitch is blunt: you do not choose between a rich interface and an autonomous agent. Every app is both.

The mechanism is a single action surface backed by one shared SQL state. Every action works both ways, by click or by prompt, and a change from either side shows up on the other instantly. The agent joins as a peer editor in a CRDT multiplayer document rather than as a chat box bolted onto the margin. Protocols ship inside the framework instead of arriving as per-feature integrations: A2A, MCP, MCP Apps, remote MCP OAuth, the OpenAI and Claude Agent SDKs, AG-UI, and the Vercel AI SDK all hang off the same surface. That is the move. The hard part of adding agents to existing software was never the model. It was wiring one consistent set of actions that both a human and a model can call with the same permissions and the same state.

If you run a SaaS on React and Postgres, this lowers the floor for real agent support from a quarter of integration work to a weekend. You define actions once. The UI calls them on click, the agent calls them on request, and you stop maintaining two divergent code paths that drift apart every release. The license matters here. MIT means you can fork it into a regulated codebase without a procurement fight. Pair it with the day's other signal, Palmier Pro exposing a local MCP server at port 19789 so Claude or Cursor can drive a video timeline, and the pattern is clear: the editor itself becomes the API.

Watch the compute bill, because the shared-state design routes agent actions straight through your primary database. Each agent turn is now a read-write transaction on the same store your users hit, not a sandboxed side channel. Capacity-plan for agents as a second class of write-heavy users, and put the action surface behind the same rate limits and auth you give humans. The remote MCP OAuth support is the part to study first.

Over the next six to twelve months, the chat sidebar stops being a feature and starts being an architecture mistake. Products that treated conversation as a wrapper over a separate backend will rebuild around a unified action layer, and the ones that started there will ship agent capabilities faster than the incumbents can retrofit them. Claude Code landing in the Jira assignee dropdown is the same story from the work-management side. The agent is becoming a user with a seat, not a button.

@github Read source 1,050 engagement
Compute & Infrastructure

SemiAnalysis says 99% of custom AI ASICs die between the slide deck and the wafer

The claim is a cold shower for the non-Nvidia accelerator pitch: simulated performance rarely survives real silicon, and almost all custom ASIC programs fail to ship competitive parts. If you are weighing a startup's bespoke chip for inference, treat the benchmark slides as fiction until there is hardware in a rack. Nvidia's lock-in holds for another cycle.

The EUV chokepoint is back in focus as ASML denies shipping a tool to China

The Zeiss-to-ASML optics chain is the single point that gates advanced-node supply, so any China breakthrough would reshape chip availability worldwide. For now the story resolves to a denial: ASML calls a US government report of an EUV shipment inaccurate and damaging. No confirmed tool, no change to the supply map yet. Watch this one, do not trade on it.

RAM is now the most expensive part in a phone, and it just killed a product line

Nothing's CMF Phone line is dead because memory prices spiked past the point where the BOM closed. The same squeeze that flows into a budget handset flows into every server you rent. Memory, not logic, is becoming the binding cost, and anyone planning a hardware refresh or a self-hosted cluster this year should lock memory pricing early.

Intel puts advanced packaging on the same tier as wafers

Patrick Moorhead flags packaging as a growth lever under Intel's new leadership, and he is right about where the bottleneck moved. Accelerator output is now gated by how fast you can stitch dies together, not just how fast you can print them. If you depend on GPU supply, packaging capacity is the number that decides your 2027 allocation.

Amazon investigates engineers who testified against AI data centers

Three Amazon engineers who spoke against new AI datacenter buildouts say they were monitored and threatened with termination, and a civil-rights complaint has followed. The capacity story is no longer just power and land. Local and labor friction is becoming a real drag on siting, and that friction shows up as delayed megawatts. Factor permitting and political risk into any region you are counting on for compute.

AI & Models

John Jumper leaves DeepMind for Anthropic as senior talent walks out the door

The AlphaFold lead and Nobel winner joining Anthropic signals a push into scientific reasoning, and it lands days after Noam Shazeer's reported move to OpenAI. Two senior departures in a week is a pattern, not a coincidence. DeepMind's research bench is being raided by the labs with the most aggressive product timelines, and that reshuffle decides which company ships the next reasoning breakthrough.

Norway bans generative AI for kids 6 to 13, supervised only for teens

This is an age-tiered education policy, not a product rule: full ban for grades one through seven, supervised use for ages 14 to 16, starting the late-August school year. The stated reason is protecting reading, writing, and math while test scores slide. If your edtech roadmap assumes classroom AI in Europe, the regulatory floor just moved, and other ministries will watch Norway's results.

Open weights at Opus level are running on a Mac Studio at home

A 26B Gemma-4 model handling private documents on consumer hardware, plus Hugging Face's cofounder touting frontier-class open models for self-hosting, point at the same shift. When data cannot leave the machine, local inference is now a real option, not a compromise. For privacy-bound workloads the question is no longer whether you can self-host, but whether the orchestration around it is worth building.

Developer Tools

Palmier Pro turns a Mac video editor into an MCP server agents can drive

This open-source Swift editor exposes a local MCP server at port 19789 when open, so Claude, Codex, or Cursor can generate and edit clips directly in the timeline. The editor, server, and agent chat are open source; only generation is paid, with Pro at $29/mo and credits buying roughly 333 images or several minutes of video. The catch is macOS 26 Tahoe only. The real lesson is the pattern: ship a local MCP endpoint and your app becomes agent-drivable for free.

Project Valhalla lands a preview in JDK 28, disabled by default

After a decade, JEP 401 value classes are merging into OpenJDK, a pull request adding over 197,000 lines across 1,816 files. Value classes drop object identity so the JVM can flatten them in memory and skip heap allocation, which means == compares by components and synchronizing on them fails. It is a preview in a non-LTS release dated March 2027, so most teams meet a stable version only at JDK 29. Audit any code that leaned on identity equality before you celebrate.

ClickHouse turns 10, and the headline is $250M ARR

The anniversary post reads as nostalgia, but the load-bearing number is business scale: over 4,000 customers and $250M in annual run-rate revenue, more than triple year over year, with 1,000 customers added in a single quarter. Engineering kept pace too, with the join-heavy TPC-H SF100 workload now 26x faster than version 22.4. If you are picking an analytical database, this is a vendor that is compounding, not coasting.

Zero-touch OAuth lands for MCP

Enterprise-managed auth for the Model Context Protocol means agents can authenticate against your IdP without a per-app consent dance. This is the plumbing that makes agent-native frameworks safe to deploy inside a company. If you are exposing an MCP server to internal agents, read this before you ship, because it is the difference between a demo and something security will sign off on.

DuckDB internals, part one on why it is fast

A clear walk through vectorized execution, the in-process model, and the design choices that make DuckDB quick on analytical queries without a server. Worth a read if you are deciding between an embedded engine and a warehouse, or if you just want to understand where the speed actually comes from before you cite it in a design doc.

Claude Code becomes a Jira assignee

Coding agents now route work straight from the assignee dropdown or a comment, so a ticket can be handed to Claude like any teammate. Small change, large signal. The agent is getting a seat in your work-management tool, and the metric that matters next is how many of those tickets close without a human reopening them.

Security

Let's Encrypt renewals threw errors today

A batch of certificate renewals failed during an incident, which for anyone on short-lived certs is a production risk, not a footnote. Check your renewal logs and confirm your automation retried cleanly. If your cert expiry monitoring is downstream of the same provider, this is your reminder to add an independent check.

Google Workspace is threatening to block Firefox access

A report that Workspace may restrict Firefox raises the browser-lock question for any team standardized on Google's suite. If your org runs Firefox by policy, test your critical Workspace flows now and have a fallback, because a quiet access change can break logins for a whole department in a morning.

Startups & Capital

Reliance pushes AI into calls, apps, and homes for 500M users

Ambani's plan puts AI features in front of more than 500 million Indian subscribers through telecom distribution. Distribution at that scale is the moat, not the model. For builders eyeing India, the question shifts from whether AI reaches users to whether you can ride a carrier's rails or get squeezed out by one.

Fusion startups have raised $7.1B, concentrated in a few firms

Most of the capital sits with a handful of companies, which tells you where the long-horizon power bets for datacenters are landing. Fusion is still years from a meter, but the funding concentration signals which firms hyperscalers will partner with first if any of this works. A power story to track, not a 2026 supply line.

Quick Hits
The Takeaway

The agent is becoming a user of your software, not a wrapper around it. Between Builder.io's shared action surface, Palmier exposing a local MCP server, Claude Code joining the Jira assignee list, and zero-touch OAuth for MCP, the pieces for treating an agent as an authenticated peer are all shipping this week. If you run a SaaS, define your actions once behind a single permissioned surface and put an MCP endpoint in front of it now, before you maintain two diverging code paths for humans and models.

The Call C-20260620

Memory, not GPUs, becomes the named constraint on inference economics in 2026. At least one top-five model provider will publicly attribute a price increase or a capacity limit to HBM or DRAM supply before the end of Q3.

The case

A RAM price spike just killed Nothing's CMF phone line because memory is now the most expensive component, and SemiAnalysis data shows semiconductor output climbing while broad manufacturing stays flat, meaning demand is concentrating in memory and advanced packaging. The consensus still frames the squeeze as a GPU and compute story and underweights memory as the binding input.

What proves us wrong

By September 30, 2026, no provider among OpenAI, Anthropic, Google, AWS, or Azure has publicly cited memory or HBM supply as a reason for a pricing change or capacity limit.

Settles by September 30, 2026
The Tape T-20260620
▼ Short MRVL Marvell Technology medium conviction

Marvell joins the S&P 500 on Monday at roughly 70x forward earnings with the index bid already front-run, up 7% Friday, while its CFO liquidated about half his stake days earlier. The passive flow is the liquidity insiders are selling into, and the custom-silicon story carrying the multiple just lost its most credible cheerleader.

SemiAnalysis [28] reminding the market that simulated ASIC slides rarely survive real silicon lands hardest on Marvell, whose two flagship programs, Microsoft Maia and AWS Trainium, are the documented stumbles, with Trainium3 reportedly lost to Alchip. The crowd reads June 22 inclusion as validation; a CFO dumping ~48% of his holdings on a 99x trailing multiple reads the other way. Marvell runs ~59% gross margin against Broadcom's ~78%, the sign of a price-taker in the programs it does win.

Wrong if MRVL closes above its June 22 inclusion-day high through its Q2 FY27 print in early September, or it announces a new hyperscaler ASIC design win. Settles Settles by September 2026 (Q2 FY27 print).
▲ Long AVGO Broadcom low conviction

The same note that tempers custom silicon quietly endorses Broadcom, the one design house whose flagship program, Google's TPU, SemiAnalysis rates at ~90% utilization and 44% lower TCO than a GB200. A 99% failure rate across the field is a moat for the program that actually ships.

The market prices AVGO and MRVL as a single custom-ASIC basket; [28]'s reminder that most programs die in silicon should split them. Broadcom designs the TPU and OpenAI's chip, runs ~78% gross margins, and carries a backlog around $73B, while the cited failures cluster at Marvell's Maia and Trainium. Demand concentrating into proven programs accrues to Broadcom, not the field.

Wrong if A public hyperscaler shifts a lead program away from Broadcom (Google or OpenAI to a rival design house), or AVGO guides AI revenue below its stated $60-90B FY27 serviceable range. Settles Settles by September 2026 (next AVGO print, early September).
◆ Watch Private Anthropic low conviction

Hiring AlphaFold's John Jumper six weeks before a targeted October listing puts a Nobel-grade science story into the IPO book and lifts the comp from frontier LLM vendor toward AI for discovery, the one adjacency where safety framing is a selling point rather than a cost. At a $965B mark, the narrative that reprices the multiple matters as much as the run rate.

[29] Jumper leaving DeepMind for Anthropic is the most senior science hire in the field this year. Consensus reads it as a DeepMind talent-drain headline; the positioning read is Anthropic broadening its story into biology and chemistry ahead of the raise. DeepMind keeps the AlphaFold IP and franchise, so the desk's standing GOOGL long is unmoved by one departure.

Wrong if Anthropic's listing slips past 2026, or no science/biology research initiative or defined Jumper role is disclosed before the IPO window closes. Settles Settles by October 2026 (targeted listing window).
Desk signals from the day's verified wire — falsifiable, dated, settled in public. Analysis, not individualized investment advice.

Get this briefing in your inbox

What changed in AI and compute, what it costs, and what to build. One email per week. No spam, unsubscribe anytime.