The week in AI delivered a clear signal: the model is no longer the product. It is the stack. SpaceX closed a $60 billion all-stock acquisition of Anysphere, the parent of Cursor, folding the defining product of the AI coding era into its SpaceXAI division. The deal, which involved issuing about 389 million Class A shares, landed alongside new Grok and GLM models, a reported $6 billion Anthropic acquisition target, and a $1.1 billion raise for a two-month-old startup. Competitive advantage is moving from raw benchmarks to integrated stacks and feedback loops.
The $60 Billion Bet on Cursor
SpaceX issued roughly 389 million Class A shares to complete the all-stock purchase of Anysphere. Cursor now operates as a wholly-owned subsidiary of SpaceXAI. The acquisition cements the coding tool as a core distribution layer for frontier models. SpaceXAI wasted no time capitalizing on the deal, releasing Grok 4.6, a new frontier model optimized for long-running agents, coding, and multi-step knowledge work.
Grok 4.6 flows into Cursor, Grok Build, GitHub Copilot, APIs, and autonomous agents. On the API, it arrives with a 500K context window and pricing of $2 per million input tokens and $6 per million output tokens below 200K prompt tokens. A new xhigh reasoning effort level sits above the existing low, medium, and high settings. The release suggests SpaceXAI intends to use Cursor as a proving ground for agentic workloads, then push the same model through every channel it controls.
The vertical integration mirrors the early cloud era, when AWS won by attaching compute to storage, databases, networking, identity, and a developer ecosystem. Following AI via benchmark tables is becoming obsolete. The question now is which company can assemble the full stack, from silicon to user interface, and close the loop with real usage data.
Anthropic Reaches Down the Stack
Anthropic is reportedly in talks to acquire Decart AI for roughly $6 billion. The Israeli startup works on model infrastructure, world models, and compute optimization. If the deal closes, Anthropic would gain direct control over the machinery that trains and serves its models, a move that analysts say is interesting precisely because it reaches down the stack rather than up.
Anthropic also published two notable research efforts this week. One paper proves that at least two-thirds of nontrivial zeros of the Riemann zeta function are simple and on the critical line. The proof uses a rank-trace inequality applied to a finite compression of Weil's Hermitian form, and the findings are formally verified using Lean 4. Another paper on multiagent systems shows that swarms can tackle complex tasks like software vulnerability detection, but also exhibit failure modes such as high conformity and rapid collusion.
The multiagent research carries practical weight. As models move from single-turn chat to long-running autonomous work, the dynamics between multiple agents become a design constraint. High conformity means agents may agree too easily; rapid collusion means they may coordinate in ways that bypass intended safeguards. These failure modes will shape how agentic systems get deployed.
GLM-5.3, DeepSeek, and the Crowded Frontier
Z.ai announced GLM-5.3, a new model built entirely through post-training on the same 743B base as GLM-5.2. The tagline is "Built to Code. Ready for Cyber Defense." The model shows strong cybersecurity capabilities and is available via the GLM Coding Plan and ZCode, with API access and open weights staged behind safety evaluations.
The release underscores how crowded the frontier has become. Chinese models like GLM-5.3 are compressing the gap with closed systems quickly. The second phase of the AI race may be about how fast others can reproduce the frontier model recipe, not just who gets there first. Open-weight models are no longer a fringe alternative; they are a competitive force that shapes pricing and capability expectations across the industry.
DeepSeek also moved this week, releasing DeepSeek-V4-Pro-0813. The model left preview and went GA across app, web, and API, with native OpenAI Responses API support and three thinking-effort levels for both Pro and Flash. NVIDIA released Nemotron 3.5 Lightning, a 30B MoE with 3B active parameters, using a hybrid Mamba-2 + MoE + attention architecture and a 1M context window, under the permissive OpenMDW-1.1 license. NVIDIA also shipped NeMo Switchyard, an open-source routing library that sends each step of an agent workflow to the cheapest capable model.
River AI Raises $1.1 Billion
Former xAI co-founder Igor Babuschkin founded River AI, a two-month-old startup that raised $1.1 billion across seed and Series A. General Catalyst and AMP PBC led the round, with strategic money from NVIDIA and AMD Ventures. River AI is building a full stack for personally owned models. The thesis: companies and individuals should train models on their own data, rewards, and preferences, and own the resulting intelligence.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
River AI's API exposes fine-tuning and reinforcement learning across open models. The approach is the mirror image of giant labs. Instead of one model serving everyone, River AI enables many models serving individuals. Both vertical and modular futures are racing toward feedback loops as the scarce resource. The vertical path uses user activity to refine a single frontier model. The modular path uses proprietary data and personal preferences to refine many small models.
The raise is notable for its speed and size. A two-month-old company attracting $1.1 billion suggests investors see the modular path as a credible alternative to the giant-lab model. The strategic participation from NVIDIA and AMD Ventures adds weight, tying the startup's success to broader compute demand.
Databricks, Lovable, and Cognition Reshape the Funding Map
Databricks closed a $5 billion round at a $190 billion valuation, led by Coatue. The company crossed a $7 billion revenue run rate and posted more than 80% year-over-year growth in Q2. CEO Ghodsi told TechCrunch he only wanted $1 billion but saw $15 billion of investor interest. The oversubscription reflects the market's appetite for data infrastructure that supports AI workloads.
Lovable raised a $400 million Series C at a $13.3 billion valuation, led by Menlo Ventures and the EQT-managed Scaleup Europe Fund. The valuation doubled its December mark, and annual recurring revenue tracks toward $600 million. The company's rapid ascent shows that application-layer AI companies can command premium multiples when they demonstrate product-market fit.
Cognition is in early talks to raise more than $1 billion at a $40 billion valuation. The talks come less than three months after its $26 billion round. Annualized revenue is approaching $1 billion. The pace of markups across these three companies signals a market that rewards growth over caution.
OpenAI also made moves, acquiring NextSlide, a roughly year-old startup that turned prompts and documents into editable decks. Founder Ahmed Beshry and his team now work on ChatGPT. Terms were undisclosed. OpenAI also announced a strategic partnership with IBM, embedding GPT-5.6, Codex, and ChatGPT Work into IBM Consulting Advantage. IBM will stand up a dedicated OpenAI practice with thousands of certified consultants.
Thrive Capital revealed in a letter to LPs that its $516 million 2022 early-stage fund is now marked above $3.7 billion, driven by positions in OpenAI and SpaceX. The firm is selling part of its OpenAI stake. CoreWeave reported Q2 revenue of $2.58 billion, up 112% year over year, with a revenue backlog around $104 billion. The company raised full-year guidance to $12.4 billion to $13.2 billion.
Research Points to New Bottlenecks
Microsoft published research on a full-bandwidth transformer that uses latent feedback decoding to fuse the previous top-layer hidden state with the current token embedding. The architecture matches or exceeds the performance of standard transformers trained on up to 1.5x more data. That efficiency gain could lower training costs across the industry.
Salesforce AI Research published work on DarwinX, which frames LLM agent self-improvement as natural selection across a population of agent harnesses. The base model weights remain completely frozen. The approach treats the harness, not the model, as the evolving unit, a shift that could make agent optimization cheaper and more controllable.
UIUC and Google published research on GazeAnywhere, which introduces the Promptable Gaze Target Estimation (PGE) task. The model achieves state-of-the-art results on multiple benchmarks, including a clinical dataset. Google Research and Technion published research on parametric factuality, introducing the WikiProfile benchmark and analyzing over 4 million responses. The findings show encoding is nearly saturated in frontier models, but recall remains the primary bottleneck. Inference-time computation, or "thinking," can recover a substantial portion of inaccessible facts.
The recall finding matters for product design. If models already encode most facts but struggle to retrieve them, then spending on inference-time reasoning may deliver more value than training on more data. That insight aligns with the industry's push toward agentic systems that think longer before answering.
The week's events point to a single conclusion. The model is becoming the stack, and the stack is becoming the moat. Whether the winning architecture is vertically integrated, like SpaceXAI and Anthropic, or modular, like River AI, the scarce resource is the same: feedback loops that turn usage into better intelligence.