Thursday, June 18, 2026

TLDR

SpaceX is acquiring AI coding startup Cursor for ~$60B in stock (Jun 16) — possibly the largest-ever acquisition of a venture-backed startup, folding Cursor into the xAI/SpaceX AI division Trending
Google retires the Gemini CLI today and forces migration to the closed-source Antigravity CLI (Jun 18) — free/Pro/Ultra tiers stop being served; enterprise license holders are exempt Pricing
Z.ai's GLM-5.2 open weightsA model whose trained weights are published for anyone to download and run.NTIA ↗ are live (Jun 17) — MIT-licensed, ~1M-token context, tops the open-model intelligence index and beats GPT-5.5 on coding at roughly one-sixth the cost Open Source
The AI data-center buildout is stalling on physical infrastructure (Jun) — ~half of planned 2026 US capacity is delayed/canceled; transformers on 3–5yr lead times, switchgear sold out through 2028 Infra
OpenCode passes ~172K GitHub stars to become the most-starred open-source coding agent (Jun) — model-agnostic across 75+ providers, ahead of Gemini CLI and Codex Open Source

Things your engineering team should look at or do this week.

Migrate off Gemini CLI / Gemini Code Assist today. As of June 18 the old CLI and IDE extensions stop serving requests for free, Pro, and Ultra users (enterprise Standard/Enterprise licenses are exempt). Move active workflows/CI to the Antigravity CLI — note it's a closed-source Go rewrite without full feature parity — or confirm you're on an exempt enterprise license. (Google Developers Blog)
Benchmark GLM-5.2 open weights for coding/agent workloads. MIT license, ~1M context, and strong SWE scores at ~1/6 the per-token cost make it worth a self-hosted eval before renewing closed-API spend — but route the hosted Z.ai API carefully given China data-residency concerns. (Computerworld)

Z.ai GLM-5.2 — open weights now live (Jun 13 launch / open weights ~Jun 17) Open Source — Zhipu/Z.ai's 753B-parameter flagship ships under MIT with a usable ~1M-token context and two thinking modes, tops the open-model tier of the Artificial Analysis Intelligence Index (~51), and reportedly outperforms GPT-5.5 on long-horizon coding at roughly one-sixth the cost per tokenA small unit of data, such as part of a word, that AI models read and generate. Many AI services price by tokens in and out.NVIDIA ↗. (Computerworld | Crypto Briefing)
OpenCode becomes the most-starred open-source coding agent (Jun) Open Source Trending — The model-agnostic CLI passed ~172K GitHub stars (ahead of Gemini CLI and Codex) with reported ~7.5M monthly developers, drawing users with no subscription lock-in, 75+ provider support, air-gapped/local deployment, and LSP-aware (compiler-aware) generation. (Abhishek Gautam writeup | LogRocket rankings)

Google: Gemini CLI & Code Assist retired June 18; Antigravity CLI is the replacement (Jun 18) Pricing — Free, Pro, and Ultra users lose Gemini CLI/IDE-extension access today; Gemini Code Assist for GitHub blocks new installs and winds down over the following weeks. Enterprise (Standard/Enterprise) licenses are unaffected. The replacement is closed-source and lacks 1:1 feature parity at launch — a migration cost for anyone with CLI-based pipelines. (Google Developers Blog | The Register)

SpaceX to acquire Cursor for ~$60B in stock (Jun 16) Trending — Days after its blockbuster IPO (SpaceX market cap ~$2.5T, now the #4 US company), SpaceX struck what may be the largest-ever acquisition of a venture-backed startup, aiming to accelerate its merged xAI/SpaceX AI division. Expected to close in Q3. A major shift in the coding-tools landscape — watch for roadmap and pricing implications for Cursor users. (TechCrunch | CNBC)
Nebius completes Eigen AI acquisition (Jun) Infra — Nebius Group closed its purchase of Eigen AI (announced May), adding model-optimization tech to its AI inferenceRunning a trained model on new input to get an output, as opposed to training it.NVIDIA ↗ platform and a new Bay Area engineering base — part of the ongoing race to drive down inference cost. (Simply Wall St)

AI data-center capacity is bottlenecked on physical infrastructure, not chips or capital (Jun) Infra — Roughly half of planned 2026 US data-center capacity is expected to slip or be canceled; of ~12 GW expected, only about a third is under active construction. The binding constraints are transformers (3–5yr lead times), switchgear (sold out through 2028), and grid interconnections (up to ~4yr) — even as hyperscalers plan $650B+ in 2026 capex. Plan capacity and vendor commitments assuming tight, lumpy supply. (Bain & Company | Datacenters.com)

EU finalizes Code of Practice on marking/labelling AI-generated content; transparency rules hit August 2026 (Jun 10) Policy — The Commission published guidance on marking AI-generated content ahead of the AI Act's transparency obligations taking effect in August 2026. If you ship generative features to EU users, start scoping content provenance/labeling now. (EU AI Act portal)

SWE-benchA benchmark that tests whether a model can resolve real GitHub issues by editing code in popular Python repositories.Jimenez et al. 2023 ↗ headline scores are mostly vendor self-reports — mind the scaffolding gap (Jun) Trending — Claude Fable 5 tops SWE-bench Verified at ~95% and SWE-bench Pro at ~80.3%, but 99 of 100 Verified entries are self-reported, and the same model family scores 10–30 points lower on a standardized harness (e.g. Scale's). Treat leaderboard numbers as upper bounds and run your own harness before trusting them for model selection. (Morph SWE-bench Pro | Scale leaderboard)

ScaleSWE — small open model, competitive coding agent (Jun) — Mines ~6M real GitHub PRs across ~5,200 repos (distilled to ~71,500 trajectories) to fine-tuneTraining an already trained model further on extra data so it fits a specific task, domain or style.Anthropic ↗ Qwen-30B to ~64% on SWE-bench Verified (near-3x over base). A useful blueprint for teams wanting a cheaper, self-hosted coding agent instead of frontier-API spend. (arXiv listing)
AgileThinker — reactive + planner heads run concurrently (Jun) — Argues single-paradigm agents can't stay both correct and on-time under rising time pressure; running a cheap reactive head alongside a deeper planner holds up best. Relevant for latency-sensitive agent loops (ops monitoring, real-time support, robotics). (arXiv listing)