AI News August 29, 2026: Nvidia buys Hugging Face for $12.9B, DeepSeek eyes $74B valuation, Apple launches Core AI
Alexandre
··
Reading time: 9 min
Click to enlarge
Credit: SiliconANGLE
Nvidia is set to acquire Hugging Face for $12.9 billion. The Information broke the story on August 26, followed by confirmations from TechCrunch, Fortune, and CNBC. Neither company has signed publicly, but the deal values the open-source platform at roughly 86 times its annual revenue.
DeepSeek, the Chinese challenger, is closing a $7.4 billion raise at a $74 billion valuation, with an IPO planned in Shanghai for 2027. Apple ships the Mac Studio M5 Ultra with 512GB of unified memory and a new Core AI framework for running frontier models locally. Anthropic rolls out Claude in Chrome to all paid subscribers, adds a built-in browser to Cowork, and launches Claudeforce with Salesforce.
Here's what happened.
Anthropic everywhere: Claude in Chrome, built-in browser, Claudeforce with Salesforce
Click to enlarge
Credit: The Decoder / Anthropic
Anthropic made Claude in Chrome generally available to all paid subscribers on August 26, 2026, graduating from beta. The extension lets Claude act autonomously in the browser, with a safety classifier verifying each action against the user's instructions. Alongside the extension, Anthropic added a built-in Chromium browser to the Claude Cowork desktop app (Gigazine, The New Stack).
A browser Claude owns
The built-in browser in Cowork opens automatically in a side panel when Claude determines a task requires web access. It can navigate between pages, click buttons, fill forms, all without accessing the user's personal browser. The Chrome extension remains useful when users want Claude to act directly on an already-open page (updating a CRM, editing an online document), but the native browser becomes the default for research and web-based tasks.
The browser is available to Pro, Max, Team, and Enterprise subscribers. The key detail: Anthropic is building its own web access layer instead of depending on Google Chrome, reducing technical dependency on Google's ecosystem (The Decoder).
Claudeforce: Claude enters Salesforce
On the enterprise front, Salesforce announced "Claudeforce" during its Q2 earnings on August 27, an integration that embeds Claude directly into the CRM platform with 37 pre-built sales skills. The product is already accessible to pilot customers, with an open beta planned for September 2026. Salesforce, which holds an equity stake in Anthropic, beat forecasts with adjusted earnings of $2.53 per share and raised its annual revenue guidance. The stock jumped 14% in after-hours trading (The Star).
For context, Anthropic's annualized revenue run rate (the annual revenue projected from the current monthly pace) reached $65 billion by end of July 2026, roughly $25 billion ahead of OpenAI (Yellow.com). The Claude ecosystem keeps expanding, from the browser to the CRM to developer tools (full Claude Code overview).
While Anthropic extends its software footprint, Apple is making a different bet: bringing AI compute back to local hardware.
Click to enlarge
Credit: Apple
Apple Mac Studio M5 Ultra: 512GB of unified memory to run LLMs locally
Apple unveiled the Mac Studio with the M5 Ultra chip on August 25, 2026, featuring up to 512GB of unified memory, an 80-core GPU, and 1.2 TB/s memory bandwidth. The chip delivers 4.3x the peak AI compute of the M3 Ultra and 9.8x the M1 Ultra, enabling developers to run large language models locally without sending data to the cloud. Pre-orders opened August 27, with general availability September 22. Starting price: $5,499 (Apple Newsroom).
Core AI: a new framework for AI on Apple Silicon
The hardware launch comes with Core AI, a unified framework (a single API for building, training, and deploying AI models across all Apple Silicon components: CPU, GPU, and Neural Engine). Core AI works alongside MLX, Apple's open-source library for model inference and fine-tuning on Mac.
Early independent benchmarks show a 70-billion-parameter model at 4-bit quantization running at 42-52 tokens per second via MLX on the M5 Ultra (TechTimes). That's comparable to cloud API performance like GPT-4o, except everything runs locally with no data leaving the machine.
Thunderbolt 5 clustering: pooling RAM across machines
The Mac Studio M5 Ultra supports clustering via Thunderbolt 5 and RDMA (Remote Direct Memory Access, a protocol that lets two machines share memory without going through the CPU). By connecting multiple Mac Studio units, developers can create a shared memory pool that accelerates distributed inference by a factor of 3 according to Apple. For researchers and AI developers working on models too large for a single machine, it's a concrete alternative to cloud GPU clusters.
512GB of unified memory in a Mac Studio. A 70-billion-parameter model running locally at over 40 tokens per second. Apple isn't just selling hardware anymore; they're building the infrastructure for AI to run on the desk.
On the funding side, the money isn't flowing only into hardware. In China, a challenger is making its move.
DeepSeek eyes $74 billion valuation: the Chinese challenger raises $7.4 billion ahead of IPO
DeepSeek, the Hangzhou-based AI startup, is finalizing a funding round of 50 billion yuan ($7.4 billion) that would value the company at 500 billion yuan, approximately $74 billion. Revenue for the first seven months of 2026 reached 475 million yuan ($70.7 million), ten times the full-year 2025 total. The story was reported by the Wall Street Journal and confirmed by Fortune and Breaking The News (WSJ, Fortune).
Heavyweight Chinese investors
Existing investors returning include Monolith Management, Shixiang Capital, and CATL, the Chinese battery giant. New backers include CPE, Legend Capital, Stony Creek Capital, and several funds tied to local government financing platforms. The investor diversity shows support beyond the typical tech VC circle.
DeepSeek's annualized recurring revenue sits around $500 million (NewsBytesApp). For comparison, Anthropic posted a $65 billion run rate in July. The orders of magnitude aren't the same, but the growth trajectory (10x in seven months) is remarkable.
Shanghai IPO in sight
DeepSeek has hired banks and advisors to prepare an IPO on the Star Market in Shanghai, the segment of the Chinese stock exchange dedicated to tech and innovation companies. Filing could happen as early as late 2026, with a public debut targeted for 2027 (BigGo Finance). If the raise goes through, DeepSeek would become China's most valuable AI startup.
DeepSeek multiplying revenue by 10 in seven months and preparing a $74 billion IPO. The AI race doesn't run only through Silicon Valley.
And while DeepSeek raises billions, Nvidia is playing a different game entirely.
Nvidia buys Hugging Face for $12.9 billion: open-source AI changes hands
Nvidia has agreed in principle to acquire Hugging Face for $12.9 billion, according to The Information (August 26), confirmed by TechCrunch, Fortune, Ars Technica, and SiliconANGLE. The acquisition values the open-source model platform at over $13 billion after adjustments, roughly 86 times its estimated $150 million annual revenue. Neither company has publicly confirmed the deal, which could still fall through (TechCrunch, Fortune).
Why 86x revenue?
Hugging Face is the go-to platform for sharing and deploying open-source AI models. Over 800,000 models are hosted there, along with datasets, demo spaces, and inference tools used by millions of developers worldwide. Annual revenue went from roughly $100 million to $150 million, but it's not the revenue that justifies the price: it's the strategic positioning.
For Nvidia, controlling Hugging Face means locking down a key link in the AI value chain. Nvidia already makes the GPUs, provides the CUDA framework, and sells DGX servers. Adding the model distribution platform to that stack means controlling the full path from model to silicon (Ars Technica, SiliconANGLE).
Open questions
The open-source community is watching closely. Hugging Face has always championed openness and interoperability: models accessible to everyone, compatible with any hardware. If Nvidia starts favoring its own GPUs in inference recommendations or restricts access to certain features, alternatives like ModelScope or Replicate could benefit.
$12.9 billion for a platform doing $150 million in revenue. 86x annual revenue. Nvidia isn't paying for today's revenue but for control of open-source AI distribution tomorrow.
The architecture of AI code agents (full analysis) relies heavily on models hosted on Hugging Face. This acquisition could reshape the entire ecosystem's dependencies.
Beyond deals and valuations, the real question remains: do AI agents deliver on their promises at industrial scale? Uber just provided a data-driven answer.
Uber runs 3,600 AI agent skills and posts 25,000 automated code reviews per week
Uber presented its "Software Factory" numbers at the AI Engineer 2026 conference on August 27. 3,600 AI agent skills built by engineers, over 30,000 executions per day, 7x growth in weekly active users between February and August 2026, and agent requests up 9.4x. All while total AI spend remained flat since April. Presented by Uday Kiran Medisetty, Distinguished Engineer at Uber (Uber Engineering Blog, StartupHub.ai).
uReview: the multi-agent code review engine
The core of the system is called uReview, a multi-agent code review engine built in-house. By August 2026, uReview posts roughly 25,000 AI-generated review comments each week and handles over 70% of pull requests across Uber's six language-specific monorepos (one monorepo per language). Review costs dropped 60%, and comment accuracy improved 70% compared to an earlier implementation.
The original problem: first-time-to-review latency had tripled from 3 hours in 2024 to 9 hours in 2026 as the codebase and engineering headcount grew. uReview agents catch common issues and only escalate to humans when deeper judgment is required.
AI Context Graph: 24 million nodes, 80 million edges
The enabling technology is the AI Context Graph, a knowledge graph (a data structure that connects entities through typed relationships) of 24 million nodes and 80 million edges. The graph aggregates data from over 30 internal systems: code, pull requests, design docs, incident logs, datasets. Agents query the graph in natural language, cutting information retrieval time from over 20 minutes to under one minute.
Cost per 1,000 requests dropped 34% and cost per session fell 52% over the same period. Uber demonstrates that scaling AI agents in production without blowing the budget is possible. It's orchestration optimization, not raw model power, that makes the difference at this scale. Token economics is a central challenge for any developer using AI tools daily (10 ways to cut your token bill by 5x).
Quick hits
McKinsey publishes its State of AI 2026: the global survey (1,719 professionals, 97 countries) reveals that 40% of large enterprises (over $1 billion in revenue) are scaling AI agents in at least one function, up from 27% the previous year. 32% of respondents are now building software internally with agentic coding tools rather than buying packaged solutions (McKinsey, The Register).
OWASP updates its LLM Top 10: prompt injection (an attack where malicious text hijacks the AI's behavior) remains the number one threat for the third consecutive year. Three major incidents in 2026: the "PerplexedBrowser" exploit against agentic browsers in March, the "BioShocking" technique in June that stole SSH keys through poisoned web pages, and a Microsoft Copilot web interface flaw in August. AI agent security (August 17 news) shows no signs of slowing down (SC Media, VentureBeat).
OpenAI's ChatGPT Work adoption gap: launched in July at $20/month, ChatGPT Work lets white-collar workers deploy AI agents connected to their tools (email, Slack, calendar, Notion). But OpenAI's own internal study reveals a massive gap: 98% of OpenAI employees use Codex agents, versus only 17% of organizational subscribers and under 1% of individual users (TechCrunch).
Space startup funding hits a record: $20.3 billion in 2026, driven by Anduril ($5 billion), Yuanxin Satellite ($1 billion), and K2 Space ($500 million). SpaceX, recently listed, trades around $140 per share, below its $225 peak (TechTimes, Crunchbase).
Conclusion: the week AI infrastructure changed hands
Three major shifts emerged this week. The first is vertical: Nvidia, which already controlled silicon and frameworks, now takes distribution with Hugging Face. The second is geographic: DeepSeek proves that Chinese AI is no longer following but raising, growing, and preparing an IPO. The third is architectural: AI agents graduate from prototype to production system, as Uber demonstrates with 25,000 automated reviews per week and a 24-million-node knowledge graph.
The common thread across these announcements is consolidation. Players controlling multiple links in the chain (hardware, models, distribution, integrations) are pulling ahead. Apple builds the hardware and the framework. Nvidia buys the model platform. Anthropic embeds Claude in the browser, the desktop, and the CRM. The question for independent developers and smaller teams is which spaces of freedom will survive in this ecosystem.
What about you: does the Hugging Face acquisition by Nvidia feel like a risk for open source or a chance to industrialize it? Drop me a message on Twitter/X or in the comments.
Alex
Key takeaways
Nvidia acquires Hugging Face for $12.9 billion, 86x its annual revenue
DeepSeek raises $7.4B at $74B valuation, revenue up 10x in seven months
Apple ships Mac Studio M5 Ultra with 512GB unified memory and Core AI framework
Anthropic launches Claude in Chrome GA, Cowork browser, and Claudeforce with Salesforce
Uber runs 3,600 AI agent skills with 25,000 automated code reviews per week
