17 Sep 2025

AI News Digest

🤖 AI-curated 8 stories

Today's Summary

Groq just raised a cool $750 million, pushing its valuation to $6.9 billion, as investors are all in on the AI inference chip game. Meanwhile, Workday is making moves by snapping up Swedish AI startup Sana for $1.1 billion, aiming to blend AI agents into their HR and finance platforms. And over at Replit, a $250 million funding round is helping launch Agent 3, an autonomous AI that’s set to make coding smoother for everyone from hobbyists to enterprise developers.

Stories

Groq raises $750M, valuation surges to $6.9B as investors double down on AI inference chips

Groq, a U.S. AI‑chip startup focused on high‑speed inference hardware, announced a $750 million funding round that lifts its post‑money valuation to about $6.9 billion. The round — led by Disruptive with participation from BlackRock, Neuberger Berman, Deutsche Telekom Capital Partners, Samsung, Cisco and others — underscores continued investor appetite for hardware that accelerates real‑time AI workloads. The new capital will help Groq scale production and compete in the fast‑moving inference market, where customers are seeking lower‑latency, cost‑efficient alternatives to incumbent GPU providers.

Workday to acquire enterprise AI startup Sana for about $1.1B to add agents and AI learning

Workday signed a definitive agreement to buy Swedish AI company Sana for roughly $1.1 billion, bringing Sana's AI search, no‑code agents and Sana Learn platform into Workday’s suite. Workday says the deal will let it embed proactive, personalized AI agents and learning tools across its HR and finance platform—potentially accelerating adoption of AI‑driven productivity and upskilling for Workday’s large enterprise customer base. Sana will continue developing its products inside Workday, which plans to use its data/context to scale Sana’s capabilities to millions of users.

Read more → Workday (press release / PR Newswire)

ParaThinker — training LLMs to ‘think in parallel’ to beat tunnel‑vision and boost reasoning

Researchers propose ParaThinker, a new paradigm that trains LLMs to generate multiple diverse reasoning trajectories in parallel and then synthesize them into a final answer. The approach targets the “tunnel‑vision” problem of long sequential chain‑of‑thought reasoning (where early mistakes lock models into bad trajectories). According to the paper and secondary coverage, ParaThinker yields substantial accuracy gains on hard math/reasoning benchmarks (reported ~12.3% for 1.5B and ~7.5% for 7B models with 8 parallel paths) while adding only modest latency (single‑digit % overhead) — enabling smaller models to match or exceed much larger sequential baselines. Why it matters: this reframes test‑time compute scaling from “deeper” chains to “wider” parallel reasoning, offering a practical path to better reasoning under tight latency/compute constraints and influencing how future models and inference systems will be designed.

REFRAG — rethink RAG decoding to compress context, extend context length and cut inference latency

A team (listed on the arXiv preprint) introduces REFRAG, a decoding framework for retrieval‑augmented generation (RAG) that compresses retrieved passages into chunk embeddings, selectively expands only the chunks that matter via a lightweight policy, and thereby dramatically reduces decoder input length and KV‑cache cost. Reported results show big speedups in time‑to‑first‑token (examples up to ~30× in the paper’s experiments) and the ability to handle much longer contexts (authors report up to ~16× context extension) without loss in perplexity or task accuracy. Why it matters: REFRAG addresses a core production bottleneck for long‑context RAG systems (latency and memory), making retrieval‑heavy applications — multi‑document QA, long‑form summarization, agent memory — far more practical at scale.

AI chip startup Groq raises $750M, valuation soars to $6.9B

Groq, an AI inference chip maker, announced a $750 million financing that pushes its post-money valuation to about $6.9 billion. The round—led by Disruptive with participation from BlackRock, Neuberger Berman, Samsung, Cisco and others—boosts Groq’s war chest as customers and cloud providers race to diversify hardware for inference beyond dominant players. Why it matters: the deal underlines continued investor appetite for specialized AI hardware and signals more capital flowing into inference-focused processors, which could intensify competition with incumbents and influence cloud procurement and data‑center strategy.

Workday to buy AI knowledge-and-agent startup Sana for about $1.1B

Workday signed a definitive agreement to acquire Sana—an AI company building enterprise knowledge search, learning and agent tools—for roughly $1.1 billion. Workday says integrating Sana’s search, no-code agents and personalized learning capabilities will let it deliver proactive, AI-driven employee experiences across HR and finance workflows. Why it matters: the deal is a strategic move to embed AI agents and knowledge tooling into core enterprise systems, accelerating the trend of legacy SaaS vendors snapping up AI-first startups to keep customers tied into integrated platforms and to compete on agent-enabled automation.

Read more → Workday (PR Newswire / investor release)

Replit raises $250M and launches Agent 3 — an autonomous AI that tests, fixes code and builds workflows

Replit announced a $250 million funding round (valuing the company at about $3B) and introduced Agent 3 — an autonomous coding agent that can test and fix code, build custom agents and automate developer workflows. The raise (led by Prysm Capital with participation from Google’s AI Futures Fund and others) underlines investor appetite for code‑generation and ‘vibe‑coding’ platforms that let non‑engineers and engineers ship software faster. Agent 3 strengthens Replit’s position as a hands‑on coding platform (useful for learning, prototyping, and enterprise automation) and signals continued momentum in tools that embed AI deeply into developer workflows.

Microsoft adds role-based Sales, Service and Finance Copilots to Microsoft 365 (moving them to the Frontier)

Microsoft announced new role‑based Copilot solutions for sales, service and finance — integrating CRM/ERP data into Microsoft 365 Copilot so teams can automate tasks like deal prep, CRM updates, variance analysis and report generation. The company is positioning these as frontline, job‑focused tools (and is adding management tooling for agents) to help businesses adopt AI in day‑to‑day workflows. For learners and builders this means more off‑the‑shelf agent templates and enterprise connectors to learn from or adapt; for organizations it lowers the barrier to deploying practical AI assistants that interact with business systems.

Read more → Microsoft 365 Blog (official)