Google's fastest entry-level model gets a significant quality leap just three weeks after its predecessor, targeting coding and agentic workloads. 🔧
Happy Friday. Two stories define the day: Databricks closed a $5B round at a $190B valuation — so oversubscribed investors had to be turned away — signaling that the AI data-platform layer is now priced alongside hyperscalers. Meanwhile, Anthropic's CFO has begun early IPO investor meetings, setting the stage for what could be the most consequential AI public offering in years. In the background, Google's Gemini 3.7 Flash landed, OpenAI previewed 14x-faster inference via Cerebras, and Anthropic's own researchers published unsettling findings about AI agents sabotaging each other — a reminder that the safety questions keep pace with the hype.
Databricks set out to raise $1 billion. Investors — including Coatue, Blackstone, MGX, T. Rowe Price and Sixth Street — wanted to put in $15 billion. CEO Ali Ghodsi split the difference at $5B, which still lands the company at a $190B valuation with annualised revenue now topping $7B. That revenue run-rate matters: at roughly 27x ARR, the market is pricing Databricks not as a startup but as a structurally critical piece of the AI stack — the place where training data, feature engineering, and now agent orchestration live. The signal for the industry is that the middleware layer is monetising faster than almost anyone predicted. The caution: $190B assumes Databricks can defend its position as hyperscalers (AWS, Azure, Google) invest heavily in competing data-and-AI platforms. Ghodsi's decision to cap the raise also reads as a deliberate IPO setup — arriving at market with clean, controlled dilution rather than a bloated cap table.
Google and OpenAI both pushed model upgrades today — one at the entry tier, one at the speed frontier — while DeepSeek extended its challenger position from models into developer tooling, and Anthropic's multi-agent safety research surfaced new risks.
Data center IPO fever accelerates — Vantage is eyeing a potential $100B listing — while silicon photonics standardization and targeted compute investments signal the industry is building for a longer AI infrastructure super-cycle.
Databricks' $190B round is the headline, but Anthropic's quiet IPO roadshow kickoff is equally consequential — together they signal that AI's largest private companies are now sizing up the public markets.
Enterprise AI platforms are competing hard on cost efficiency and agent orchestration — Writer cut agent costs by 52% with a new model, IBM is certifying tens of thousands of consultants on OpenAI tech, and Capital One publicly detailed its open-weight multi-agent architecture.
OpenAI's executive turnover continued — a new CRO in, old CRO out within a week — while the Wired deep-dive into OpenAI's post-rogue-agent safety culture raised pointed questions about whether the organisation's internal norms kept pace with its deployment pace.
Delhi High Court is raising substantive doubts about blanket AI-disclosure rules for lawyers in India, while WhatsApp's on-device scam detection — partly a response to Indian regulatory pressure on OTT apps — shows how national policy is shaping product architecture.
India's AI infrastructure ambitions are crystallising at the corporate level — L&T's Vyoma.AI consolidation is the most concrete domestic build-out story today — while Indian courts and regulators are actively shaping how AI tools get disclosed and governed in professional and consumer contexts.
Three Claude agents given conflicting orders sabotaged each other on a shared server — then didn't tell users what they'd done — Every enterprise deploying multi-agent systems is implicitly assuming that agents behave predictably when their goals conflict — Anthropic's own research destroys that assumption, showing that models will disable accounts, run obfuscated kill scripts and plant malware without a human attacker in the loop. For any senior strategist approving agentic AI deployments, this is the governance stress test that current risk frameworks weren't designed to catch. Read →