AI Weekly: 06/01/26
Anthropic raises $65B near $1T, SoftBank pledges $87B for French AI, Cognition lands $1B for Devin
Good morning and welcome to this week’s edition of AI Weekly! In this week’s news, Anthropic raised a $65 billion Series H at a $965 billion valuation, potentially the company’s last private round before an IPO, as annualized recurring revenue crossed $47 billion. SoftBank committed EUR 75 billion (roughly $87 billion) to building data centers across France with up to five gigawatts of compute capacity, one of the largest single country AI infrastructure investments ever announced.
Cognition raised $1 billion at a $25 billion valuation for its Devin coding agent, now generating $492 million in ARR, while Apple previewed a sweeping Siri overhaul for iOS 27 with a standalone app, Dynamic Island integration, and Google’s Gemini powering the intelligence layer. Robinhood opened its brokerage platform to AI agents through MCP, letting autonomous systems trade stocks and manage finances on behalf of users.
Also this week, Anthropic released Opus 4.8 with a new Dynamic Workflows feature for parallel subagent coordination, Meta announced an AI pendant and four new smart glasses models, ElevenLabs launched Music v2 with mid-track genre switching, and ByteDance revealed it is developing its own inference chips with roughly 1,000 chip designers on the project.
More on this past week’s top AI headlines below!
- ZG
Here are the most important stories of the week:
AGENTS
Apple previewed a sweeping overhaul of Siri for iOS 27, including a standalone Siri app with a full screen chat interface, Dynamic Island integration, and Google’s Gemini model powering the intelligence layer through an on-device distillation deal. Link.
The new standalone app (codenamed “Campos”) replaces the current overlay with a full screen experience that includes a conversational chat interface, the ability to control apps, summarize content, and compose messages.
Apple struck a deal with Google to distill Gemini’s capabilities into smaller models that run directly on device, preserving Apple’s privacy first approach while closing the gap with ChatGPT and other cloud based competitors.
The company is also building its own foundation models internally, but the Gemini partnership provides a near term bridge; the iOS 27 update is expected to debut at WWDC in June.
Robinhood announced that AI agents can now trade stocks and manage finances on its platform through MCP (Model Context Protocol) integration, and introduced an “agentic” credit card designed for autonomous AI spending. Link.
The MCP integration allows AI agents from any provider to connect to Robinhood, execute trades, monitor portfolios, and make financial decisions on a user’s behalf with granular permission controls.
Robinhood also unveiled a credit card that AI agents can use to make purchases autonomously, with built in spending limits and approval workflows to prevent runaway transactions.
The move makes Robinhood one of the first major financial platforms to embrace fully autonomous AI trading, raising questions about regulatory oversight of agent driven financial activity.
Meta announced an AI pendant and four new smart glasses models as part of a push to sell 10 million wearable AI devices in the second half of 2026, expanding well beyond the Ray-Ban collaboration. Link.
The AI pendant is a standalone wearable that listens, sees, and responds without requiring glasses, targeting users who want an always available AI companion in a minimal form factor.
The four new smart glasses models span different price points and use cases, from fashion forward frames to sport specific designs with enhanced camera and audio capabilities.
Meta’s 10 million unit sales target for the second half of 2026 would represent a massive acceleration, signaling that the company views wearable AI as a major product category beyond VR headsets.
Meta launched an “Enterprise Solutions” unit with forward deployed engineers who embed directly with corporate customers to build custom AI applications on Meta’s models and infrastructure. Link.
The new unit mirrors the forward deployed engineering model pioneered by Palantir, where Meta’s engineers work on site with enterprise clients to build bespoke AI solutions tailored to their needs.
The initiative targets large corporations that want to adopt Meta’s open source Llama models but need hands on implementation support for complex, mission critical deployments.
The move marks a significant expansion of Meta’s enterprise ambitions, adding a high touch services layer on top of its open source model strategy to compete more directly with Anthropic and OpenAI in the enterprise market.
SPEECH/AUDIO
Sesame, the conversational AI startup founded by Oculus creator Brendan Iribe, launched its iOS app with four distinct AI voice agents, all available for free, with plans for a smart glasses product by 2027. Link.
The four agents each have unique personalities and voice characteristics, designed to feel like distinct conversational companions rather than a single assistant with different settings.
The app is free at launch, with Sesame betting on user acquisition before introducing premium tiers; the company has raised significant funding backed by Iribe’s track record building and selling Oculus to Meta for $2 billion.
Sesame’s smart glasses roadmap positions it as a future competitor to Meta’s Ray-Ban partnership, with voice first AI interaction as the core differentiator.
ElevenLabs launched Music v2 with a genre switching capability that allows AI generated tracks to shift between musical styles mid-song, a first for any major AI music platform. Link.
The genre switching feature lets users specify transitions within a single track (for example, moving from jazz to electronic to orchestral), creating dynamic compositions that evolve over the course of the song.
Music v2 also includes improved vocal synthesis, longer generation lengths, and more granular control over instrumentation and arrangement.
The update intensifies competition with Suno and Udio, which have faced ongoing copyright lawsuits from major record labels over their training data practices.
CODING/DEVTOOLS
Cognition raised $1 billion at a $25 billion pre-money valuation for Devin, its AI coding agent that now generates $492 million in annualized recurring revenue. Link.
The round values Cognition at roughly 50x its ARR, reflecting investor confidence in the AI coding agent market’s growth trajectory; Devin was first introduced in March 2024 as the “first AI software engineer.”
Cognition has rapidly scaled from its initial demo to nearly half a billion in revenue, driven by enterprise adoption of autonomous coding agents that can handle complex, multi-step software engineering tasks.
The raise intensifies the AI coding wars, with Anthropic’s Claude Code, OpenAI’s Codex, and Cursor all competing for developer adoption in a market that could reshape how software gets built.
Anthropic released Opus 4.8 with a new “Dynamic Workflows” feature that enables parallel subagent coordination, allowing the model to dispatch multiple specialized agents simultaneously for codebase-scale migrations and complex engineering tasks. Link.
Dynamic Workflows lets Opus 4.8 break down large tasks into parallel subtasks, each handled by a specialized subagent, then synthesize the results into a coherent output, dramatically speeding up work like migrating entire codebases or refactoring large projects.
The feature represents a shift from single threaded AI reasoning to orchestrated multi-agent execution within a single model interaction, letting developers hand off entire project level tasks rather than individual prompts.
Opus 4.8 is positioned as Anthropic’s most capable model for professional software engineering, competing directly with OpenAI’s Codex and Google’s Jules for high stakes development workflows.
INFRASTRUCTURE
SoftBank committed EUR 75 billion (roughly $87 billion) to building data centers across France with up to five gigawatts of compute capacity, one of the largest single country AI infrastructure investments ever announced. Link.
The investment will fund multiple hyperscale data center campuses across France, leveraging the country’s relatively abundant nuclear power supply and favorable energy costs compared to other European nations.
SoftBank CEO Masayoshi Son personally announced the deal alongside French President Macron, who has made attracting AI investment a central policy priority for France.
The commitment follows SoftBank’s $100 billion Stargate pledge in the United States and positions France as a major hub for AI compute in Europe.
Groq is raising $650 million after Nvidia’s $20 billion “not quite acquisition” deal fell apart, pivoting its business model from chip design toward operating as an inference neocloud provider. Link.
The pivot means Groq is positioning itself more as a cloud service that sells inference compute (powered by its custom LPU chips) rather than as a chip company that sells hardware, similar to how Cerebras has evolved.
Nvidia’s attempted $20 billion deal, which would have given it significant control over Groq without a full acquisition, did not proceed; Groq is now raising independently to fund its infrastructure buildout.
Groq’s LPU (Language Processing Unit) architecture is designed specifically for inference speed, offering significantly lower latency than GPU based alternatives for serving large language models.
South Korean startup Xcena raised $135 million at a $570 million valuation, developing memory chips purpose built for AI workloads that aim to solve the memory bandwidth bottleneck limiting inference performance. Link.
The company’s chips are designed to address the growing gap between compute speed (which has improved dramatically with GPUs and custom accelerators) and memory speed (which has not kept pace), a bottleneck that limits how fast AI models can run.
Xcena offers a specialized alternative to the high bandwidth memory (HBM) produced by Samsung and SK Hynix, targeting a different slice of the AI memory architecture.
The raise positions Xcena in a rapidly growing market for AI specific silicon, alongside companies like Cerebras, Groq, and Etched that are building alternatives to Nvidia’s GPU centric approach.
Snowflake signed a $6 billion deal with AWS to use Amazon’s custom Graviton CPU chips, marking a major shift for the data cloud company toward Arm based compute infrastructure. Link.
The multi-year agreement represents one of the largest cloud compute commitments by a single company, reflecting Snowflake’s massive data processing needs as it expands its AI and machine learning offerings.
Graviton chips, designed by Amazon’s in-house Annapurna Labs team, offer better price performance than traditional x86 processors for data intensive workloads.
The deal underscores the growing enterprise appetite for Arm based cloud infrastructure, a trend reshaping the server chip market as companies optimize for cost efficiency at scale.
ByteDance is developing its own inference chips in a dual track effort using both Arm and RISC-V architectures, with roughly 1,000 chip designers working on the project. Link.
The company aims to reduce its dependence on Nvidia GPUs for running inference on its AI products, which span TikTok’s recommendation engine, the Doubao assistant, and other AI services serving hundreds of millions of users.
ByteDance is pursuing both Arm (for near term deployment) and RISC-V (for longer term cost savings and licensing independence) architectures simultaneously, hedging its bets on which will prove more efficient.
The effort follows a similar path taken by Google (with TPUs), Amazon (with Inferentia and Trainium), and Meta (with MTIA), as large tech companies invest in custom silicon to reduce inference costs at scale.
Baseten, the AI inference platform startup, is raising $1 billion at an $11 billion valuation after reaching $600 million in annualized recurring revenue. Link.
Baseten provides infrastructure that helps companies deploy and serve AI models at scale, competing with platforms like Modal, Replicate, and major cloud providers’ AI serving solutions.
The $600 million ARR represents rapid growth for the company, driven by surging enterprise demand for reliable, low latency AI inference infrastructure as more businesses deploy models in production.
The raise would value Baseten at roughly 18x its ARR, reflecting investor enthusiasm for the “picks and shovels” layer of the AI stack where revenue is recurring and margins improve with scale.
OpenRouter more than doubled its valuation to $1.3 billion as the AI model routing platform now serves 100 trillion tokens per month across 8 million users. Link.
OpenRouter acts as a unified API layer that lets developers and consumers access models from OpenAI, Anthropic, Google, Meta, and dozens of other providers through a single interface, routing requests to the best model for each task.
The 100 trillion tokens per month figure represents staggering scale, driven by both developer API usage and OpenRouter’s consumer chat interface that lets users compare model outputs side by side.
The platform’s growth reflects a broader trend toward model agnostic infrastructure, as customers increasingly want flexibility to switch between AI providers rather than being locked into a single vendor.
OTHER
Anthropic raised a $65 billion Series H at a $965 billion valuation, potentially the company’s final private funding round before an IPO, as annualized recurring revenue crossed $47 billion. Link.
The round makes Anthropic one of the most valuable private companies in history, trailing only SpaceX at its pre-IPO valuation and approaching the trillion dollar mark.
Anthropic’s ARR growth from $30 billion at the end of Q1 to $47 billion represents an extraordinary acceleration, driven by demand for Claude Code and enterprise AI products across legal, finance, and healthcare verticals.
The “potentially last round before IPO” framing suggests Anthropic could go public in late 2026 or 2027, joining Cerebras and potentially OpenAI in the wave of major AI IPOs reshaping public markets.
OpenAI launched a self-serve advertising manager, going downmarket to let small businesses buy ads that appear within ChatGPT conversations and expanding beyond the company’s initial enterprise and subscription focus. Link.
The self-serve platform lets small businesses create and manage ad campaigns without needing a dedicated sales relationship with OpenAI, similar to how Google and Meta democratized digital advertising for local businesses.
OpenAI is positioning ads as a major new revenue stream alongside subscriptions and API access, targeting the massive small business advertising market that has powered Google and Meta’s growth for two decades.
The move represents a strategic shift for OpenAI, which had previously focused on premium subscriptions and enterprise deals; the timing follows Anthropic’s own push into small business with Claude for Small Business earlier this month.