AI Startups

AI Startups to Watch - August 16, 2026

AI startups to watch August 16: SpaceX closes its $60B Cursor deal, an open Chinese model beats frontier rivals, Unitree's $9B IPO lands. Five disruptive moves.

Daniyal

· 4 min read
Rocket launching upward carrying a glowing code editor window, startup disruption theme, dark background with orange and blue neon accents

SpaceX closed the largest AI acquisition in history this weekend: $60 billion in stock for Cursor, the AI code editor. Meanwhile, an open-weight model from Beijing beat every closed frontier model on a cyber benchmark, and China's first humanoid-robot IPO is about to land. Five moves that reshape who owns the AI stack.

The $60B close: SpaceX now owns Cursor

SpaceX officially closed its all-stock acquisition of Cursor on Aug 15, confirmed on Cursor's own blog. The deal process started in April with a $60B purchase option and accelerated in June, days after SpaceX's blockbuster IPO. Cursor, the leading agentic coding platform with its Composer agents and "AI teammates", now gets access to what SpaceX calls the largest fleet of GPUs in the world, the compute business that already rents capacity to Anthropic and Google. Remember that SpaceX absorbed xAI earlier this year, so the stack is now models (Grok 4.6), compute, and the top coding tool in one company.

Who it hurts: OpenAI, whose models were Cursor's largest distribution channel, plus Microsoft and GitHub Copilot, and every independent AI coding vendor from Windsurf to Cognition and Codeium. A neutral multi-model IDE just became a Grok-first distribution point.

An open Chinese model that beats the frontier at cyber

Z.ai released GLM-5.3 on Aug 14, a post-training upgrade of GLM-5.2, and it posted record open-source scores: the top open-source result on Terminal Bench 3.0, plus CyberGym vulnerability discovery at 84.5%, ahead of Anthropic's Claude Mythos 5 at 83.8% and OpenAI's GPT-5.6 Sol at 83.6%. ExploitBench is at 54.4%, roughly double GLM-5.2. The lab says the model has surfaced 2,436 vulnerabilities across 269 open-source projects, 1,097 of them critical or high severity, some undetected since 1981. That list reportedly includes a serious vulnerability in Cursor, discovered the same week Cursor joined SpaceX. Weights are delayed about two weeks for a safety review, with the most sensitive cyber functions gated behind a "trusted access" program.

Who it hurts: the "closed equals safer" narrative at Anthropic and OpenAI, and their enterprise pricing power. Axios ยท CryptoBriefing

Unitree: China's first humanoid-robot IPO

Unitree priced its Shanghai STAR Market IPO at 150.8 yuan, about $22.34 per share, valuing the Hangzhou robot maker at roughly 61B yuan, about $9.04B. It is selling 40.45M new shares to raise around 6.1B yuan, about $905M, with DeepSeek taking a $20.8M strategic cornerstone stake. Unitree's 2025 revenue was 1.7B yuan, about $252M, with net profit of 278.2M yuan. The trading debut is expected Aug 17-21, and pre-IPO perpetual contracts on Hyperliquid are already trading around $92-94, implying a roughly $38B market valuation before the first trade. Tech in Asia coverage

Who it hurts: every humanoid startup that hoped to go public without a comp, and Tesla's Optimus narrative on cost and scale. This IPO is the template for Figure, Agility, and the rest, and DeepSeek's stake ties frontier LLM capability directly to physical robots.

Blacksmith: the QA gate for AI-generated code

Blacksmith raised a $45M Series B led by Peak XV Partners, with Y Combinator and GV participating, at a $550M valuation, about 10x the $60M set on its Series A less than a year ago. The San Francisco company runs continuous validation and CI workloads on purpose-built hardware, effectively the pick-and-shovel QA gate for the flood of AI-generated code hitting build and test pipelines. The 10x mark-up in under a year is the clearest leading indicator that capital is concentrating in AI coding infrastructure, not just the coding tools on top. PR Newswire

Writer: cutting enterprise agent costs with open weights

Writer released Palmyra X6 on Aug 13, claiming an average 52% lower cost per agent task, 48% faster execution, and 10% higher quality than its prior generation, with cost per task dropping from about $0.25 to $0.12. The telling detail: Palmyra X6 is a post-training variation of Z.ai's open-source GLM-5.2, not a frontier lab model. An enterprise AI platform using an open Chinese base model to undercut OpenAI and Anthropic enterprise deals is the clearest proof yet that open weights plus vertical harness economics, not model size, are now the cost lever in enterprise AI. Writer's announcement

The pattern

Three of this week's five moves share one theme: the open-weight stack is eating the frontier from below, and the fastest money is following the plumbing, not the models. For more context, see yesterday's AI Startups to Watch and the Startups archive. The open question for the week ahead is Unitree's debut: if the implied $38B valuation holds on day one, the robot-IPO window is open, and every humanoid lab with a revenue line will be lining up.

If this story matters to you, these earlier editions cover the same ground in more detail:

More on AI Startups

Browse every AI Startups story →