Cursor becomes the week's stress test
Cursor dominated the fortnight because the ownership, supplier and product stories all collided. On 2 Sep, multiple articles reported that SpaceX had acquired Cursor for $60 billion, with co-founder Aman Sanger joining SpaceX as part of the deal. The same day, reports said OpenAI had ended its partnership with Cursor and would cut off access to its models. On 4 Sep, further coverage said the contract termination was expected to take effect in November 2026.
The product response was not cosmetic. On 2 Sep, Cursor added 16 documentation pages around cloud agents and Grok Bot, introduced 53 product pages and removed 67. On 3 Sep, its changelog added support for self-hosted machines, letting users keep codebases, build outputs and secrets on internal infrastructure, alongside My Machines for connecting a personal laptop or VM. On 4 Sep, Cursor released SDK bridge v1.0.31 with standalone executables and a stable sdk.v1 proto contract for @cursor/sdk and cursor-sdk.
The same week also showed the cost of being central infrastructure. On 2 Sep, Cursor reported three minor service degradation incidents and was named in multiple recently published CVEs, including CVE-2026-73217 and CVE-2026-73218 for macOS Auto-Run Sandbox mode issues. On 3 Sep, it reported a major incident affecting Automations, Cloud Agents and Review Agents for around 80 minutes. On 4 Sep, it reported five incidents, including a critical service degradation resolved after about 206 minutes.
GitHub pushes Copilot deeper into enterprise workflow
GitHub's Copilot story was one of expansion under operational pressure. On 2 Sep, GitHub said Copilot could now approve pull requests, updated model access for Team plans and made Claude Fable 5.1 generally available in Copilot. On 5 and 6 Sep, GitHub announced the general availability of GPT-6 Astra in GitHub Copilot, a privacy-safe star history API endpoint and updates including the ability to run several agents at once.
The enterprise motion was visible in customer material rather than a single launch. On 2 Sep, GitHub added Accenture and AMD logos, with case studies saying Accenture used GitHub Copilot with 12,000 developers and AMD reduced development cycles from weeks to days. On 3 Sep, it added InfoTrack and a case study on providing its global developer team with a common language. On 4 Sep, Jibo, Arduino and Trustpilot appeared as new logos. On 5 Sep, Ahold Delhaize was added.
There was management and reliability noise around that push. On 1 Sep, GitHub's CEO resigned and Microsoft began shifting the platform's leadership. The same day, GitHub resolved a 42 minute incident involving elevated errors for OpenAI models provided by Copilot. On 3 Sep, it resolved a 174 minute minor incident affecting the Grok Copilot AI Model Provider. On 5 Sep, a major disruption to Copilot Code Review lasted about 108 minutes.
Vercel positions itself as the agent substrate
Vercel spent the week making a specific infrastructure bet: agents need sandboxes, gateways and cheaper compute options. On 4 Sep, its changelog added support for Cursor Cloud Agents running in Vercel Sandbox, using scale-to-zero workers and isolated microVMs. The same update introduced Basic build machines with 2 vCPU and 8 GB for Pro and Enterprise teams as a lower-cost build option.
The AI Gateway kept widening. On 2 Sep, Vercel added Qwen 3.8 Max 0902 and Claude Fable 5.1 through the AI Gateway via a single API key, with spend tracking and traces on every request. On 3 Sep, Meta's Muse Spark 1.3 became available through the same gateway. On 5 Sep, Vercel added Ling 3.0 Flash Sante from Inclusion AI, free through 4 Oct, and GPT 6 Astra from OpenAI for coding and research use cases.
The public surface shifted alongside the platform surface. On 4 Sep, Vercel launched a new contact page and removed the Svelte TurboRepo starter template page. On 5 Sep, it added REST API documentation for listing sandboxes and removed a Python workflows documentation page. On 6 Sep, it added product pages for G2, npm, retail, SWR and a coding agent template, while removing five pages that included Next.js templates. The security backdrop was material: on 1 Sep, three articles reported a Vercel security incident involving a third-party AI tool and a $2 million ransom demand, with the CEO confirming an ongoing threat beyond the initial compromise.
Lovable and Replit were quieter, but not still
Lovable's week mixed capital momentum with product reliability pressure. On 2 Sep, multiple sources reported that Lovable had completed a $400 million funding round and reached a $13 billion valuation. On 6 Sep, further reporting again cited $400 million in funding, without specifying the round, stage or investors. On 7 Sep, Lovable reported a major editor incident lasting about 27 minutes and a separate minor incident involving Supabase project creation failures lasting about 231 minutes.
Its market surface moved towards education and localisation. On 2 Sep, Lovable added five blog pages focused on Fable 5.1 in multiple languages and five product pages for a motion tutorial signup, while removing ten blog pages, mostly in Spanish. On 3 Sep, it added motion tutorial pages in English, Spanish, French and Portuguese. On 5 Sep, it added a Tech Procurement Specialist role and removed a Global Mobility and Immigration Lead role.
Replit's changes were smaller but pointed at packaging discipline. On 2 Sep, it added the ability to run one background task at a time to the Starter tier and removed 2 parallel agents from the Core tier. The same day, it posted roles for Learning Experiences Creator, AI Agent Security Architect and Legal Operations Manager in Foster City, and released river v0.221.0 with typed custom handshake rejection codes and session cleanup improvements. On 7 Sep, it released nixpkgs-replit 1.0.1 with CI tweaks and gzip compression for tarballs.
The week's hidden pattern was page churn
The loud story was Cursor, SpaceX and OpenAI. The quieter signal was how quickly vendors rewrote the public machinery around agents. Cursor's sitemap changed heavily twice in five days: on 2 Sep it added 53 product pages and removed 67, then on 6 Sep it added 112 product pages and removed 111, including commands, skills and marketplace items. In between, on 3 Sep, it added 10 documentation pages for self-hosted cloud agents and Grok Bot security while removing 6 self-hosted guide pages.
Vercel showed the same rhythm from a different angle. From 1 to 6 Sep, it added pages for Academy, Changelog, Marketplace, customers, contact, networking, sandboxes, G2, npm, retail, SWR and a coding agent template, while removing integration pages, templates, blog pages, a Python workflows document and Next.js template pages. Lovable pushed multilingual tutorial material across 2 and 3 Sep. Replit removed the marchmadness product page on 2 Sep.
That matters because positioning did not wait for quarterly launches. By 4 Sep, Cursor had moved Grok Bot access, network and audit controls into Enterprise while removing priority support and account management from that tier. By 5 Sep, Vercel had sandbox API documentation live. By 6 Sep, the category's public shelves had been rearranged around agents, sandboxes, gateways, tutorials and enterprise control.
| Company | Change | Date |
|---|---|---|
| Cursor | Access, network and audit controls for Grok Bot were added to the Enterprise tier. Priority support and account management were removed from the Enterprise tier. | 2026-09-04 |
| Replit | Run one background task at a time was added to the Starter tier. 2 parallel agents were removed from the Core tier. | 2026-09-02 |
| Vercel | Basic build machines with 2 vCPU and 8 GB became available for Pro and Enterprise teams as a lower-cost build option. | 2026-09-04 |
| Vercel | The Pro offer added free .app and .dev domains, increasing eligible TLDs to eight. | 2026-09-03 |
The market is moving from AI coding as an editor feature to AI coding as managed infrastructure. Cursor's reported SpaceX acquisition and OpenAI cut-off made model supply a board-level dependency. GitHub pushed Copilot further into enterprise workflow with PR approvals, GPT-6 Astra and multiple agents. Vercel built around sandboxes, gateways and cheaper build capacity. Replit and Cursor both changed plan packaging around agent capacity and controls. The operators who win the next phase will not just ship better agents. They will prove control over models, execution environments, secrets, audit trails, reliability and pricing boundaries.
Each week this page takes a position and grades it in public once the horizon passes. Misses stay up. The full record.
This is the public read. OpsControl customers see this market live: every signal, graded and evidenced, the day it happens.
Track your own market
OpsControl