8/14: CursorSpaceXAI

SpaceXAI acquires Cursor, Z.ai releases GLM-5.3, Thrive releases an investor letter, Anthropic publishes its frontier risk report

It’s Friday. Between DeepSeek-V4 Pro, Gemini 3.7 Flash, Grok 4.6, Josh Kushner buying the Lakers, Grok Bot, River AI raising $1.1B, Cognition and Lovable raising, Bernie demanding an AI pause, Zuckerberg explaining Meta’s vision for superintelligence, Claude making progress on the Riemann hypothesis, and OpenAI previewing GPT-5.6 Cyber, it’s been a huge week. Be sure to monitor with us live on X and YouTube, and follow us on Instagram.

SpaceXAI Acquires Cursor

Today, SpaceXAI completed its $60 billion acquisition of AI coding company Cursor, one of the largest tech deals in history. Both xAI and Cursor were founded less than five years ago.

Cursor began in 2022 as Anysphere, a project led by four MIT students — Michael Truell, Sualeh Asif, Aman Sanger, and Arvid Lunnemark — building tools around nascent AI models. They started out building tools for mechanical and CAD engineering, which failed due to their lack of domain expertise, before pivoting to software engineering. In March 2023, the same month GPT-4 was released, Anysphere published the first version of Cursor, then just a fork of Microsoft Visual Studio Code optimized for large language models. Very few people knew at the time how big agentic coding would be, or how quickly it would take over.

Cursor raised an $8M seed round in October 2023 led by the OpenAI Startup Fund, then a $60M Series A in August 2024 led by a16z, Thrive, OpenAI, and multiple top angels. It grew rapidly, taking just over a year (Oct 2023 to Jan 2025) to scale from $1M to $100M ARR, and nine months to scale to $1B ARR (Oct/Nov 2025). The company forecasts over $6B ARR by the end of 2026.

All of this scaling was heavily reliant on frontier labs like OpenAI and Anthropic, who provided the models used in Cursor’s products. This was especially problematic given that both companies started developing products — Claude Code and Codex — that directly competed with Cursor on agentic coding. In October 2025, Cursor released Cursor 2.0, pivoting to more fully autonomous coding, and Composer, their first model. Composer was slightly less intelligent than top frontier models like GPT-5 and Claude Sonnet 4.5, but much faster and more efficient, helping Cursor retain its independence.

xAI’s origins go much further back. In 2015, fearful of potential Google dominance over superintelligence, Elon Musk and Sam Altman co-founded OpenAI as an open-source AI research nonprofit. Musk provided most of OpenAI’s early funding. In 2018, frustrated by its perceived lack of progress, Musk attempted to take direct control of the company. The other co-founders rejected the proposal, and Musk resigned. Under Sam Altman, it converted to a for-profit company, took a major investment from Microsoft, and began developing paid products: first the GPT-3 API in 2020, and then the unexpectedly successful “research preview” of ChatGPT in 2022.

Not wanting to allow Altman to win the race to superintelligence, Musk founded xAI in March 2023. In November 2023, the first version of Grok was released, fully integrated with the X app. It was later open-sourced. In May 2024, xAI raised a $6 billion Series B. That summer, they moved at breakneck speed to build Colossus 1, then the largest data center in the world, in just 122 days. Grok 3, trained on Colossus, was released the following February. The next month, xAI acquired X. Four months later, they released Grok 4, the largest model ever trained.

But despite the enormous amounts of capital and compute and the hardcore work culture, not all was well. xAI’s models lagged substantially behind OpenAI and Anthropic, and almost the entire founding team had left by the beginning of 2026. Musk conceded that the company was not built right the first time around, and needed to be rebuilt from the foundations. He needed a team with a proven track record of building great models. So he turned to Cursor.

On April 21, xAI and Cursor announced a partnership on model training, plus a deal: either SpaceX (xAI’s parent company) would acquire Cursor for $60 billion, or walk away for a $10 billion breakup fee. Cursor Composer 2.5 was released in May, performing very close to the frontier at a fraction of the cost. Grok 4.5, released in July, was the best Grok model ever by far. And Grok 4.6, released earlier this week, is the best model in the world other than Opus 5, Fable, and Sol. The SpaceXAI-Cursor acquisition ended up saving both companies: Cursor is no longer directly threatened by coding model competitors, and SpaceXAI now has a world-class model team once again. Elon Musk has said that SpaceXAI is currently training a 10T parameter Grok model. When it comes out, we’ll be monitoring.

More Situations

• Z.ai releases GLM-5.3. GLM-5.3 is the exact same base model as GLM-5.2, released June 13, but with post-training improvements. It performs similarly to Kimi K3 on most benchmarks, but is substantially improved on TerminalBench and on cyber capabilities. Z.ai found that the model performs substantially better than GLM-5.2 and Kimi K3 on cybersecurity benchmarks (though still behind Mythos and Sol), and was able to chain together multiple exploits. In response, the company built the Z.ai Security Disclosure Ledger as a public record of vulnerabilities that are found and fixed, and will release the weights in two weeks after safety evaluations and hardening are complete.

• Thrive releases its first-ever investor letter and invests $215M in Amazon. Thrive Capital, founded by Joshua Kushner in 2009, has over $65 billion under management and stakes in OpenAI, SpaceX, Stripe, Databricks, Anduril, and more. In his first letter, Kushner laid out the firm’s opportunity set: emerging tech, infrastructure and applications, traditional industries to be transformed by tech, and scarce assets that will become more valuable with tech. Since 2009, Thrive has compounded at 41% gross and 33% net IRR.

• Anthropic publishes its August 2026 Risk Report focusing on three threat models: misalignment in high-stakes settings, risks from automated R&D, and chemical and biological weapons production. They find that risk is still low, though no longer very low, and could increase if models learn how to scheme more coherently. Anthropic evaluated an unreleased “Model 2” that is slightly more capable than Claude Mythos 5, that they have no plans to release.

• France’s highest court blocks ban on social media for under-15s. The French Constitutional Council ruled that the ban was too broad and age verification would violate the privacy of adults. The French Senate and Assembly passed the bill last month with overwhelming margins (243-2 and 279-81, respectively).

• Jane Street posted a $15B loss in July due to declines in the value of Situational Awareness LP’s holdings. Jane Street was a major backer of SALP, an unusual move for a firm which usually trades only its own capital.

• Nvidia disclosed a $21B stake in SpaceX and a $30B stake in Intel. Also, Google disclosed a $94B stake in SpaceX, up ~100x from its $900M investment in 2015.

• Apple is working with Alibaba to train its own model in China. It will be used as part of Apple Intelligence for Apple devices sold in the country.

• Alibaba releases the weights of Qwen3.8-27B, the smaller version of its recent model. Weights of the full 2.4T version are also available on Hugging Face.

Today’s Experts

• Damon Falck and Samer Sabri (Members of Technical Staff, Inherent Labs)

• Brannin McBee (Co-founder and CDO, CoreWeave)

• Jarosław Jaworski (CEO, Edge Aerospace)

• Mehul Nariyawala (Co-founder and President, Matic)

• Rev Lebaredian (VP of Physical AI Simulation, Nvidia)

• Akash Ramdas (Co-founder, Discovered Materials)

• Christian Keil (Partner, a16z American Dynanism)

• John Gedmark (CEO, Astranis)