WEEK OF MAR 15 – 21, 2025

Google paid $32B for cloud security and Nvidia open-sourced its robot stack — the week OpenAI stopped giving away its speech models

Alphabet's largest acquisition ever went to a security company that had already turned it down once. Nvidia handed out a humanoid foundation model and a physics engine. OpenAI's new transcription models stayed closed, ending Whisper's MIT run.

Three releases this week said more about strategy than capability. Google committed $32 billion in cash for Wiz, its biggest purchase ever and nearly three times what it paid for Motorola. Nvidia used GTC to open-source a humanoid robot foundation model and a physics engine built with Google DeepMind and Disney Research. And OpenAI shipped three new audio models that it explicitly declined to open-source — a deliberate break from the MIT-licensed Whisper line it had given away for years.

Put together, the week is a clean read on where each company thinks its moat is: distribution and security for Google, the physical-AI toolchain for Nvidia, and the API surface for OpenAI.

THU · Mar 20, 2025openaispeechtranscriptionttslicensingopen source

OpenAI ships gpt-4o-transcribe and gpt-4o-mini-tts — and says it will not open-source the Whisper successor

OpenAI released three proprietary audio models to its API: gpt-4o-transcribe and gpt-4o-mini-transcribe for speech-to-text, and gpt-4o-mini-tts for speech generation. The transcription models replace Whisper, and OpenAI says they handle accented and varied speech better in chaotic environments, with streaming speech-to-text allowing continuous audio input and a real-time text stream. gpt-4o-mini-tts accepts steerable instructions — the demo instruction was 'use a serene voice' — and the company showed adding voice to an existing text app in roughly nine lines of code on the Agents SDK it shipped the week before.

The notable decision is licensing. OpenAI has historically released Whisper versions for commercial use under MIT, and it explicitly said it does not plan to do so for the new models, calling them much bigger than Whisper and not good candidates for open release, while pointing to end-user devices as the interesting case for open-source work.

OpenAI ships gpt-4o-transcribe and gpt-4o-mini-tts — and says it will not open-source the Whisper successor
TechCrunch

WHY IT MATTERS

This is the clearest statement yet that the open-weights era at OpenAI covers the small and the old, not the new. Whisper became the default local transcription stack for a generation of apps precisely because it was MIT and ran on a laptop; its successor is API-only, larger, and closed. If your product depends on on-device or self-hosted speech, the practical options just narrowed to the older Whisper checkpoints and whatever the open ecosystem produces — and this week in the same industry, Google and Nvidia both went the other direction.

TUE · Mar 18, 2025googleacquisitioncloud securityenterprisem&a

Google signs a $32B all-cash deal for Wiz, its largest acquisition ever

Google Cloud announced an agreement to acquire Wiz, the cloud security company, for $32 billion in cash — the largest acquisition in Alphabet's history, roughly two and a half times the $12.5 billion paid for Motorola Mobility in 2012, and reportedly the largest purchase of an Israeli tech company to date. The price is a 39% premium over the $23 billion Google had offered in 2024, when Wiz's CEO Assaf Rappaport walked away and instead raised $1 billion at a $12 billion valuation.

Wiz scans cloud environments across AWS, Azure and Google Cloud for vulnerabilities and misconfigurations without installing agents on individual workloads, and had crossed $500 million in annual recurring revenue while approaching $1 billion. Google had already bought Mandiant for $5.4 billion in 2022, plus Chronicle, Siemplify and VirusTotal. The deal still needs regulatory approval and was expected to close in 2026.

WHY IT MATTERS

Two things. First, the strategic logic is real: hyperscalers' native security tools have never been fully trusted to grade their own homework, and multi-cloud visibility was a genuine gap in Google Cloud's position against AWS and Azure — buying the neutral scanner is how you close it. Second, the negotiation itself is the lesson: Wiz walked away from $23 billion and ended up with $32 billion, because it proved it had a credible alternative. Both facts are worth remembering as cloud security consolidates around a handful of buyers.

TUE · Mar 18, 2025nvidiagtcroboticsopen sourcesimulationcomputehardware

At GTC, Nvidia opens-sources a humanoid robot foundation model and a physics engine

Jensen Huang opened GTC 2025 in San Jose by calling AI an inflection point, arguing that reasoning and agentic AI demand orders of magnitude more compute than what came before, and projecting Nvidia data center infrastructure revenue reaching $1 trillion by 2028. He detailed Blackwell Ultra, which he said was designed specifically for that moment and would ship in the second half of 2025, and the Vera Rubin platform succeeding Blackwell in the second half of 2026, with one exaflop of FP4 performance and a 25x increase over Hopper.

The open-source releases were the part with lasting consequences. Nvidia unveiled Isaac GR00T N1, described as the first open, fully customizable foundation model for generalized humanoid reasoning and skills, paired with an updated Cosmos world model for generating synthetic training data and working with the Omniverse simulation stack and a new Dynamo inference software layer Huang called essentially the operating system of an AI factory. At the end of the keynote he announced Newton, an open-source physics engine for robotics simulation developed with Google DeepMind and Disney Research. Nvidia also announced a partnership with General Motors on next-generation vehicles, factories and robots, and Halos, an automotive safety system.

WHY IT MATTERS

The hardware was expected; the software wasn't. Open-sourcing a humanoid foundation model and a physics engine puts Nvidia in the position of supplying the toolchain for an entire robotics ecosystem rather than just the GPUs underneath it — and the Newton collaboration with DeepMind and Disney Research is a signal that simulation infrastructure is being treated as shared public substrate. The data center revenue claim is the number to hold onto: if it holds, the compute build-out is being planned as an eight-to-ten-year cycle, which is the assumption to price against.