OpenAI Buys Statsig, Ends Cursor: AI News Sep 2, 2026

1. Today's Headlines

• OpenAI acquires Statsig for $1.1B and reshuffles its product leadership. OpenAI agreed to buy experimentation platform Statsig in an all-stock deal valued at roughly $1.1 billion, and named its founder and CEO Vijaye Raji (a decade-long Meta engineering veteran) as the new CTO of Applications, reporting to Applications CEO Fidji Simo. The move coincides with a broader leadership reset: Kevin Weil shifts from Chief Product Officer to VP of AI for Science, and Srinivas Narayanan becomes CTO of B2B Applications. The strategy is explicit — internalize the A/B-testing and real-time-decisioning machinery needed to run ChatGPT and Codex at hundreds-of-millions-of-users scale. (Sources: Reuters, TechCrunch, The Verge; Sep 1–2; subject to vendor confirmation)
• OpenAI ends the Cursor partnership, effective Nov 12. Following SpaceX's $60B acquisition of Cursor parent Anysphere (closed Aug 14), OpenAI invoked a change-of-control clause to wind down Cursor's direct access to its models, with a proposed shutoff date of Nov 12, 2026. OpenAI says it cannot be confident SpaceX will comply with its terms of service, citing past conduct by Musk-owned companies (the post-acquisition Twitter API cutoff and sworn testimony about xAI's ToS violations). Future models — including the unreleased Astra — will not reach Cursor under the partnership. Developers can still use their own OpenAI API keys or OpenAI's IDE extensions for Cursor. Within hours, Anthropic said it would boost Claude compute inside the editor. (Sources: OpenAI statement, Reuters, Times of India; Sep 1; subject to vendor confirmation)

2. Model Releases

• GLM-5.3 Flash tops OpenRouter as Chinese open models hit ~68% of global token volume. GLM-5.3 Flash (Z.ai, MIT, 320B/18B active) led OpenRouter's weekly usage chart, part of a broader wave where Chinese open-weight models processed an estimated 68.1 trillion tokens in a single week — about 68% of global weekly model calls and the 19th consecutive week at #1, with the gap to US usage widening to ~4.8x. (Sources: OpenRouter data, Chinese tech media; Aug 29; subject to vendor confirmation)
• DeepSeek V4-Flash-Vision-Exp keeps spreading. DeepSeek's MIT-licensed 305B multimodal build (Aug 31) — matching Claude Opus 4.8 on vision-agent benchmarks at a fraction of the cost — is drawing adoption among self-hosting teams, extending the open-weight agentic-vision push. (Source: DeepSeek / Hugging Face; Aug 31; subject to vendor confirmation)
• Google's Gemini Omni 1.1 Flash expands video editing. Released Aug 27, the model adds 4K video understanding, scene extension to ~40s, and first/last-frame interpolation, leaning on 360p drafts plus 4K upscaling for fast iteration. (Sources: Google, HeadsUpAI; Aug 27–29; subject to vendor confirmation)
• A wave of smaller open releases. Per daily roundups, the last 48 hours also surfaced Apple's FastVLM, Nous Research's Hermes 4 family, Tencent's R-4B VLM, MiniCPM-V4.5, Microsoft's VibeVoice TTS, and Meituan's LongCat-Flash — most framed as efficiency or modality-specific drops rather than frontier jumps. (Sources: aggregated AI newsletters; subject to vendor confirmation)

3. Industry & Capital

• OpenAI's ads business hits a $1B annualized run rate. Roughly 200 days after launch, ChatGPT ads span 40+ countries and now reach free and Go-tier users among the app's ~1 billion weekly actives — a fast-emerging revenue stream as OpenAI diversifies ahead of any IPO. (Source: Forward Future; Sep 1; subject to vendor confirmation)
• Nvidia deepens its chip-ecosystem grip. On Aug 31 Nvidia invested $3.5B in MediaTek and opened its NVLink Fusion platform so MediaTek can design custom XPUs that plug into Nvidia rack-scale "AI factories" — extending Nvidia's fabric beyond its own GPUs. Separately, Nvidia guaranteed up to $105B in 20-year leases for an 8-gigawatt OpenAI data center in Ohio and mobilized a $500B third-party financing platform, shifting its moat from silicon to capital. (Sources: Nvidia/MediaTek joint statement, aibriefing.dev; Aug 31–Sep 1; subject to vendor confirmation)
• Anthropic cements its capital position. Its May 28 Series H — $65B at a $965B post-money valuation, with run-rate revenue crossing $47B — remains the backdrop as rivals race the IPO window; the company also signed up to 5GW of new AWS capacity. (Sources: Anthropic official, Silicon Report; subject to vendor confirmation)
• G20 AI summit puts frontier labs in front of ministers. At the Sep 1–2 G20 Innovation Ministerial in Chapel Hill, US Commerce Secretary Howard Lutnick hosted fireside chats with OpenAI's Sam Altman and Nvidia's Jensen Huang, putting model and chip strategy directly before trade and technology ministers. (Source: The Neuron; Sep 1; subject to vendor confirmation)
• Apple cuts 200+ Siri and Vision Pro roles to refocus on foundational AI. The reallocation shifts talent toward core AI development and infrastructure, while iOS 27 beta 7 tests a major Siri upgrade with Visual Intelligence. (Source: aibriefing.dev; subject to vendor confirmation)
• Meta shows its hardware hand. Meta unveiled MTIA 300, its first in-house training chip with built-in NIC chiplets plus a MetaRoCE RDMA transport, and released Muse Glimmer, a 30B open-weight model under Apache 2.0 built to run multi-step agentic tasks on a single consumer GPU. (Source: aibriefing.dev; subject to vendor confirmation)

4. China Power

• Open-weight momentum is now a volume story. Beyond GLM-5.3 Flash's OpenRouter lead, DeepSeek V4-Flash-Vision-Exp's multimodal open release and Tencent's Aug 28 open-sourcing of Hy4-preview (770B/49B active, Apache 2.0 — the largest permissively licensed model so far) extend a run where Chinese open models dominate global token share. (Sources: IThome, ModelScope; Aug 28–31; subject to vendor confirmation)
• Zhipu's GLM-5.3 gains institutional weight. After open-weighting GLM-5.3 (Aug 28/29), Zhipu was added to the MSCI China index (effective Aug 31 close), and its "Ox-Alpha" anonymous model was confirmed as GLM-5.3. (Source: Yicai, IThome; Aug 29–30; subject to vendor confirmation)
• National token volume crosses 500 trillion/day. China's data regulator said daily model token calls exceeded 500 trillion, up over 5,000x in two years, as a new "compute optimizer" profession emerges and enterprise AI adoption lowers the barrier from "building a power plant" to "plugging in a card." (Source: CCTV News; Aug 30; subject to vendor confirmation)

5. Today's Observation

The frontier race is quietly changing shape. OpenAI's $1.1B Statsig deal is not about analytics — it's about acquiring the experimentation engine to run a 1B-user AI product like a consumer-internet company, where every model-routing decision and onboarding nudge is measured. That same day, OpenAI used a contract clause to cut Cursor off, and Anthropic immediately moved to capture the displaced users. Model access is hardening into strategic leverage, not a commodity. Meanwhile Chinese open-weight models keep compounding their share (~68% of global token volume), so the industry is bifurcating into tightly-controlled closed frontiers and an open ecosystem that is winning on cost and reach.

Editor's Take

The Statsig deal is the more interesting story than the headline suggests. OpenAI isn't buying analytics — it's buying the machinery to run ChatGPT like a consumer-internet company at 1B-user scale, where every model-routing choice and onboarding nudge is A/B tested. Combined with the Cursor cutoff, the signal is clear: frontier labs now treat model access as leverage, not a commodity. I'd watch whether "Astra" — the unreleased frontier model — ever reaches Cursor at all; OpenAI just wrote the playbook for keeping the next model behind a wall.