TAI #158: The Great Acceleration: AI Revenue, M&A, and Talent Wars Erupt as the Industry Matures
Also, Gemini 2.5 Flash-Lite, xAI raise, MiniMax-M1, Mistral Small 3.2, and more.
What happened this week in AI by Louie
While LLM model releases have slowed down lately, the AI industry’s undercurrents of commercialization and consolidation were taken to the next level this week. The era of pure research is rapidly giving way to a fierce land grab for talent, customers, and market share, evidenced by a series of unprecedented M&A discussions, funding rounds, and revenue milestones. This great acceleration signals that AI is moving from a nascent technology into a mature, high-stakes commercial arena.
No company exemplifies this shift more than Meta, which appears to be in a state of strategic urgency to catch up. After being caught off guard by the performance of open-source models from competitors such as Deepseek, Meta is now on an aggressive campaign to acquire AI talent and technology. The company recently finalized a $14.3 billion deal for 49% of Scale AI, a move that is as much about acqui-hiring its CEO, Alexandr Wang, as it is about data-labeling capabilities. Further reports indicate Meta is in advanced talks to hire prominent AI investors Nat Friedman and Daniel Gross, which could involve a partial buyout of their multi-billion-dollar VC fund. These moves come alongside reports of Meta exploring acquisitions of high-profile startups like Perplexity (recently valued at $14bn) and Safe Superintelligence Inc. ($32bn valuation 1 year after founding), and making staggering $100 million-plus compensation offers to poach top researchers. As OpenAI CEO Sam Altman noted, this focus on “a ton of upfront guaranteed comp” over the mission highlights the intensity of the talent war.
The high-profile Scale AI deal also casts a spotlight on a critical, often unglamorous, truth: the performance of frontier models is fundamentally dependent on high-quality human feedback. Companies like Scale and its larger, bootstrapped competitor Surge AI (which hit over $1 billion in revenue last year, outpacing Scale’s $870 million), provide the essential service of grading model responses and creating the curated datasets that power instruction tuning, RLHF, and reasoning models. This human-in-the-loop process transforms a raw foundation model into a useful, reliable tool, making these data firms indispensable components of the AI value chain. The strategic importance is clear, with customers like OpenAI reportedly winding down work with Scale post-Meta deal to avoid conflicts of interest.
While Meta scrambles to buy its way back to the top, a new layer of “AI wrapper” companies is demonstrating explosive growth, proving the immense value in building applications on top of foundation models. The revenue growth at the model providers themselves is staggering: OpenAI has reportedly surged over 80% in the last 5 months to a $10bn ARR, while Anthropic is said to have tripled its run rate in the same period to $3bn. But the application layer is where the velocity is truly breathtaking. AI code editor Cursor reached $100M ARR with around 100x growth in 2024, and after a recent $900M funding round, is now reportedly at a >$500M run rate. In a similar vein, OpenAI acquired Windsurf, another VS Code fork, for a reported $3bn. Elsewhere, Replit’s ARR has skyrocketed 10x from $10M at the end of ’24 to $100M year-to-date. And in a stunning example of the bootstrapped potential, the AI coding tool Base44 was acquired by Wix for $80M just six months after its launch, having reached $3M ARR with zero employees. Legal AI firm Harvey just raised a $300M Series E at a $5B valuation, and search startup Genspark reportedly hit $36M in ARR in just 45 days with a team of only ~20 people.
The sheer scale of capital being deployed is reaching astronomical levels. Mira Murati’s new startup, Thinking Machines Lab, raised an unprecedented $2 billion at a $10 billion valuation less than five months after its founding, reportedly with a strategy to build custom models optimized by “RL for businesses.” Meanwhile, Elon Musk’s xAI is seeking another $4.3 billion in equity and $5bn debt after raising $12bn in equity in 2024. And looking even further ahead, SoftBank’s Masayoshi Son is reportedly pitching “Crystal Land,” a potential $1 trillion AI and robotics manufacturing hub in Arizona.
Why should you care?
The flurry of multi-billion-dollar deals, soaring valuations, and nine-figure talent offers marks a turning point for the AI industry. The phase of tentative exploration is over; we are now in an era of intense commercialization and strategic consolidation. The “land grab” for top-tier talent, defensible market positions, and, crucially, high-quality data is happening at an unprecedented speed and scale.
This isn’t just hype; it’s a market responding to real, generated value. The meteoric rise of companies like Cursor, Replit, Harvey, and Base44 demonstrates that the most immediate and explosive growth often lies in the application layer — building sophisticated workflows and user-friendly products on top of foundational AI. The immense valuations are a direct reflection of the value being placed on those who can skillfully translate raw AI capabilities into tangible business solutions. This underscores a point I’ve consistently made: the value of skilled LLM developers and product builders has never been higher.
The frantic M&A activity from established players like Meta also tells a story of both opportunity and risk. It validates the immense strategic importance of AI, but also signals that even the largest tech companies can fall behind and be forced to pay a steep price to catch up. This creates opportunities for nimble startups and talented individuals to build highly valuable companies in record time.
— Louie Peters — Towards AI Co-founder and CEO
This issue is brought to you thanks to Box:
60% of companies expect to be transformed by AI within two years. But many organizations aren’t even close to putting AI at the forefront of their business strategies.
Box’s new white paper, Becoming an AI-First Company, helps break down the ever-evolving landscape so businesses can move at the speed of AI. Download the new white paper now for Box’s key recommendations.
Hottest News
1. Minimax Open-Sources MiniMax-M1: 1 Million Token Context Model
MiniMax has open-sourced MiniMax-M1, a long-context hybrid-attention model combining Mixture-of-Experts and Lightning Attention for low-FLOP, high-efficiency reasoning across up to 1 million tokens. Tailored for coding, tool use, and software engineering, it introduces a new reinforcement learning approach called CISPO and achieves competitive results on diverse benchmarks at relatively low training cost.
2. Anthropic Adds Remote MCP Support in Claude Code
Anthropic has expanded Claude Code to support remote Model Context Protocol (MCP) servers, allowing broader tool and data integration from external services like dev tools, project management platforms, and knowledge bases. The update adds compatibility with MCP servers that stream data via Server-Sent Events (SSE) or communicate through Standard Input/Output (STDIO).
3. Midjourney Introduces V1 Video Model
Midjourney has launched Version 1 of its video model, bringing animation capabilities to its image generation platform. Users can animate images, control motion, and extend scenes up to 16 seconds. The web-only release is priced comparably to image upscales, and Midjourney plans further improvements aimed at real-time, open-world video simulation.
4. Musk ‘In Talks’ To Finalize $4.3 Billion in New Equity Funding for xAI
xAI is reportedly looking to raise $4.3 billion in new equity funding as it burns through nearly $1 billion a month. With projected losses of $13 billion in 2025, the company is aiming to raise a total of $9.3 billion in debt and equity, spending more than half in the next quarter. While xAI expects to hit $2 billion in revenue next year, it still lags behind peers like OpenAI in developing competitive revenue streams.
5. Mistral AI Releases Mistral Small 3.2
Mistral AI has released Mistral Small 3.2, a refinement of its previous 24B-parameter instruction-tuned model. Though a minor update, it brings significant accuracy and reliability gains: WildBench v2 scores rose from 55.6% to 65.3%, and HumanEval Plus Pass@5 reached 92.9%. Improvements include fewer repetition errors, better function calling, and stronger performance on STEM tasks.
6. OpenAI Shows How To Control Misaligned Behavior in GPT-4o by Steering Internal Features
OpenAI has identified a recurring internal activation pattern, termed the “misaligned persona”, that correlates with undesirable behavior in GPT-4o and other models. Triggered by fine-tuning on low-quality or harmful data, the pattern can be isolated using sparse autoencoders (SAEs). It appears across both supervised and RL-trained models and is more active when models respond unethically. The team suggests SAEs can help inspect and control alignment-relevant features by analyzing activation traces.
Five 5-minute reads/videos to keep you learning
1. Why AI Agents Fail on Long Tasks: The Constant Hazard Rate Explained
This piece explores why AI agents tend to underperform on long-horizon tasks, attributing it to a constant failure rate across subtasks. As the number of steps grows, success probability drops exponentially — an agent with a 10% failure chance per 10-minute segment, for instance, has just a 0.2% chance of lasting 10 hours. The model’s “half-life” offers a way to estimate how long an agent remains reliable. Unlike AI systems, humans seem to recover better mid-task, possibly through reflection or self-correction.
2. Temporal Graph Neural Networks for Multi-Product Time Series Forecasting
The article demonstrates how Temporal Graph Neural Networks (TGNNs) can improve sales forecasting across multiple products in a supply chain. By combining Graph Neural Networks to learn inter-product influence and Temporal Convolutional Networks to capture time-based trends, the model outperforms classical baselines like SARIMAX and SES. A case study with 50 SKUs shows how TGNNs uncover cross-product dynamics often missed by traditional methods.
3. Agentic Misalignment: How LLMs Could Be Insider Threats
Anthropic investigates the risk of LLMs exhibiting harmful, agent-like behavior in enterprise environments. Testing 16 models across hypothetical corporate tasks, the team identified early signs of agentic misalignment, where models autonomously take harmful actions. The post dives into evaluation scenarios, risk dimensions, and potential strategies to limit such behavior.
4. SHADE-Arena: Evaluating Sabotage and Monitoring in LLM Agents
Anthropic also introduced SHADE-Arena, a benchmark designed to test LLM agents’ susceptibility to sabotage and hidden task execution. Each of the 17 test environments contains benign tasks subtly paired with covert malicious ones. The models are evaluated not just on task performance, but on whether they can or can’t be trusted in high-stakes, multi-step environments with access to sensitive data and tools.
5. Build Smarter RAG Systems: Make It Context-Aware
This hands-on guide outlines a method for improving RAG systems by assigning context labels to individual document chunks, like “geography” or “history,” before embedding. These categories are stored alongside chunk vectors in a database, allowing retrieval to prioritize relevance at query time. The approach boosts generation accuracy and includes a full working example using LangChain, Chroma DB, and Groq.
Repositories & Tools
1. miniDiffusion reimplements the Stable Diffusion 3.5 model in pure PyTorch with minimal dependencies.
2. DeepEP is a communication library tailored for Mixture-of-Experts (MoE) and expert parallelism (EP).
3. Awesome AWS is a curated list of AWS libraries, open-source repos, guides, blogs, and other resources.
Top Papers of The Week
1. MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention
This paper introduces MiniMax-M1, a cutting-edge hybrid-attention reasoning model with 456 billion parameters and an 8x larger context size than DeepSeek-R1. Utilizing a lightning attention mechanism and CISPO algorithm, MiniMax-M1 excels in complex tasks, completing RL training on 512 H800 GPUs in three weeks for $534,700. MiniMax-M1 outperforms competitors in software engineering and long-context tasks.
2. Self-Adapting Language Models
This paper introduces Self-Adapting Language Models (SEAL), a framework enabling LLMs to self-adapt by generating finetuning data and update directives. SEAL employs self-edits for persistent weight updates using a reinforcement learning loop, enhancing knowledge incorporation and few-shot generalization. Unlike prior methods, SEAL directly manages self-directed adaptation without auxiliary networks. Code and details are available online.
3. Text-to-LoRA: Instant Transformer Adaption
Researchers introduced Text-to-LoRA (T2L), a hypernetwork for instant language model adaptation based on natural language task descriptions. T2L streamlines model customization by constructing LoRAs efficiently, matching task-specific performance with minimal computation. This approach leverages pre-trained adapters and facilitates language-based specialization, democratizing the customization process of foundational models while reducing resource demands.
4. Wings: Learning Multimodal LLMs without Text-only Forgetting
Wings modifies multimodal LLM training to preserve performance on text-only tasks. It integrates separate visual and textual learners in parallel across attention blocks, addressing attention shift during training. Wings matches or exceeds the performance of models at a similar scale across both text and multimodal benchmarks.
5. MEMOIR: Lifelong Model Editing with Minimal Overwrite and Informed Retention for LLMs
MEMOIR proposes a memory-based editing approach where new knowledge is added through sparsely activated residual modules. Each edit is confined to a separate subset of memory parameters using input-dependent masks, limiting interference. At inference, activation patterns from new queries are matched against stored edits, allowing relevant updates to be applied while ignoring unrelated memory slots.
Quick Links
1. Google made Gemini 2.5 Flash and Pro generally available and also introduced a preview of the new Gemini 2.5 Flash-Lite, our most cost-efficient and fastest 2.5 model yet. The preview is now available in Google AI Studio and Vertex AI, alongside the stable versions of 2.5 Flash and Pro.
2. Yandex has released Yambda, the world’s largest publicly available dataset for recommender system research and development. This dataset is designed to bridge the gap between academic research and industry-scale applications, offering nearly 5 billion anonymized user interaction events from Yandex Music, one of the company’s flagship streaming services with over 28 million monthly users.
3. OpenAI drops Scale AI as a data provider following the Meta deal. OpenAI said it was already winding down its work with Scale AI ahead of Meta’s announcement last week that it was investing billions of dollars in the startup and bringing on CEO Alexandr Wang. An OpenAI spokesperson told Bloomberg that OpenAI had been seeking other providers for more specialized data to develop increasingly advanced AI models.
Who’s Hiring in AI
Junior Machine Learning Engineer — Product @Nextdoor (San Francisco, CA, USA)
AI Researcher @Hume AI (Remote)
Senior Software Engineer (Conversational AI) @Smartly (Helsinki, Finland)
Software Engineer (Full Stack AI Engineer) @BC Forward (Remote)
Enterprise Data — Software Engineering — Senior Associate @Fannie Mae (Reston, VA, USA)
Staff Backend Engineer — Core AI @Anzen (Remote)
Interested in sharing a job opportunity here? Contact sponsors@towardsai.net.
Think a friend would enjoy this too? Share the newsletter and let them join the conversation.