Welcome back, revolutionary…
Jensen Huang just bet a trillion dollars on a future where AI agents, not humans, decide what to buy and who to buy it from. Seven new chips, a rack-scale supercomputer, and an agentic platform that Adobe, Salesforce, and SAP are already building on. The chatbot era just got its expiration date.
In this week's edition:
NVIDIA's seven new chips and the platform built for a world where agents spend the money
The White House plan that could make 50 state AI laws disappear overnight
New Ahrefs data on what AI Overviews are actually citing (hint: not your top-ranking pages)
Who was behind that mystery model, and why it matters
$38 billion in infrastructure deals, an open-source copyright reversal, and a march through San Francisco
NVIDIA's $1 Trillion Bet: Seven Chips and a Platform Built for Agents
The chatbot era got a formal expiration date this week. At GTC 2026 in San Jose, Jensen Huang unveiled seven new chips, a $1 trillion order pipeline through 2027, and a platform designed for a world where AI agents, not humans, do most of the computing. The message was plain: the future runs on inference, and NVIDIA plans to own the rack.
Need to know:
The centrepiece is Vera Rubin, a rack-scale supercomputer pairing NVIDIA's new Vera CPU with Rubin GPUs. The system packs 1.3 million components, delivers 3.6 exaflops of compute, and claims 10x better performance per watt than Grace Blackwell. It ships later this year.
Huang also revealed the Groq 3 LPU, the first chip from NVIDIA's $20 billion acquisition of Groq last December. The LP30 chip carries 512 MB of on-chip SRAM per die with 150 TB/s memory bandwidth. Paired with Vera Rubin, it boosts tokens-per-watt by 35x. That's purpose-built silicon for inference at scale.
The software side got NemoClaw, an enterprise stack for building autonomous agents on top of the open-source OpenClaw platform. Adobe, Salesforce, ServiceNow, SAP, and 13 other enterprise players are already building on it. NVIDIA also released Nemotron 3 Nano (4B parameters) and Super (120B) as open models tuned for agentic workloads.
Purchase orders between Blackwell and Vera Rubin are projected to reach $1 trillion through 2027, double last year's $500 billion estimate. Huang previewed Kyber, the next rack architecture after Rubin, integrating 144 GPUs in vertical compute trays.
Every new agent built on this infrastructure is another machine making decisions about which businesses to recommend, which products to surface, and which sources to trust. As inference costs drop and agent deployments scale, the question of whether your business is visible to AI systems gets more urgent by the quarter.
Trust & Power
The White House Wants One Set of AI Rules. States Have Other Ideas.
The Trump administration released its National AI Legislative Framework on March 20, handing Congress a blueprint that amounts to one core argument: let us handle this, and keep the states out. The framework covers seven areas, from child safety to data centre permitting, and aims to become law this year.
Need to know:
The biggest ask is broad federal preemption of state AI laws. The administration wants to prevent a patchwork of 50 different regulatory regimes, preserving state authority only for generally applicable laws, zoning, and procurement.
Child safety is the one area with genuine bipartisan energy. The framework calls for age-assurance tools, platform features to reduce exploitation and self-harm risks, and extending existing child privacy rules to AI systems.
On copyright, the framework punts. Rather than settling whether training on copyrighted material is fair use, it defers to the courts. The creative industries will not be satisfied.
No new federal AI agency. Oversight routes through existing regulators with relevant expertise. The framework also proposes streamlined permitting for data centres and protections so residential ratepayers don't absorb AI's energy bill.
For businesses building AI strategies, the framework matters because it signals which rules will stick. If federal preemption holds, companies get one compliance target instead of 50. If it doesn't, the current patchwork gets worse.
Breakthrough
Xiaomi's Mystery Model Was Hiding in Plain Sight
A trillion-parameter AI model called Hunter Alpha appeared on OpenRouter on March 11 with no press release, no branding, and no attribution. The AI community spent a week convinced it was DeepSeek's unreleased V4. It wasn't. Xiaomi's AI division confirmed on March 18 that Hunter Alpha was an early build of MiMo-V2-Pro, the flagship in a new family of three models.
Need to know:
MiMo-V2-Pro runs 1 trillion total parameters with 42 billion active, supports a 1-million-token context window, and benchmarks near GPT-5.2 and Claude Opus 4.6 on coding and agent tasks. Its API pricing is one-fifth of Opus 4.6.
The MiMo division is led by Luo Fuli, a former core contributor to DeepSeek's R1 and V-series models. Her move to Xiaomi in late 2025 brought significant architectural know-how from one of China's most capable open-source labs.
Xiaomi simultaneously launched MiMo-V2-Omni (unified multimodal) and MiMo-V2-TTS (expressive text-to-speech), signalling ambitions well beyond a single flagship model. The company is committing $8.7 billion to AI.
A phone manufacturer just shipped a frontier-class model at commodity pricing. For businesses tracking which AI systems their customers interact with, the competitive surface just got wider, and less predictable.
Round Up: Building the Machine
Meta signs $27B infrastructure deal with Nebius: Meta locked in a five-year agreement for AI infrastructure built on NVIDIA's Vera Rubin platform. Nebius delivers $12 billion in dedicated capacity starting early 2027, with Meta committing to buy an additional $15 billion in available compute. Meta's AI capital expenditure is expected to reach $115-135 billion in 2026 alone.
IBM closes $11B Confluent acquisition: The deal brings Apache Kafka's real-time data streaming to IBM's watsonx platform, positioning IBM for the agentic AI era where autonomous systems need live data feeds. Over 6,500 enterprises, including 40% of the Fortune 500, already run on Confluent. IBM paid a 34% premium at $31 per share.
Mistral ships Small 4 under Apache 2.0: One model now does the work of three. Mistral Small 4 unifies the capabilities of Magistral (reasoning), Pixtral (multimodal), and Devstral (agentic coding) into a single 119B-parameter mixture-of-experts model with only 6B active parameters per token. It's 40% faster and triples throughput compared to Small 3.
Linux Foundation gets $12.5M for open-source AI security: Google, Microsoft, OpenAI, Anthropic, and AWS pooled funding to help open-source maintainers cope with the flood of AI-generated vulnerability reports. The grants flow through Alpha-Omega and OpenSSF to build maintainer-centric AI security tooling for hundreds of thousands of projects.
UK scraps copyright exception for AI training: After 11,500 consultation responses and vocal opposition from artists including Elton John and Dua Lipa, the UK government dropped its proposed text-and-data-mining exception. No replacement policy exists yet. The government is now watching the licensing market develop while the House of Lords pushes for mandatory licensing before training.
Stop the AI Race marches across San Francisco: On March 21, protesters walked from Anthropic's headquarters to OpenAI to xAI, calling on Dario Amodei, Sam Altman, and Elon Musk to commit to a conditional pause on frontier AI development. The march follows Anthropic dropping its Responsible Scaling Policy and OpenAI weakening safety commitments during its for-profit restructure.
Tool Shed
Klariqo: Build a no-code AI phone and chat receptionist in 3 minutes that books appointments, qualifies leads, and integrates with 50+ services.
Tracium: Track AI agent costs, errors, and traces across tenants with a single line of code and auto-updating pricing tables.
Friendware: Press Tab anywhere on macOS and get context-aware AI autocomplete that reads your screen and predicts your next sentence.
Adject: Transform a single product photo into professional e-commerce visuals with realistic models and lifestyle backgrounds in seconds.
Livedocs: Connect CSVs and databases to an AI copilot that generates charts, metrics, and plain-English insights without setup or code.
Solvea: Deploy an AI resolution engine that handles phone, email, SMS, and live chat support around the clock with built-in CRM and analytics.
Raydian: Build full-stack web apps through chat with a visual editor, integrated backend, database, authentication, and one-click deploy.
FindNiche: Analyse TikTok Shop data, track ads across four platforms, and surface winning products for e-commerce sellers using AI-driven research.
Quick Bytes
Google AI Overviews now cite top-10 ranking pages only 38% of the time, down from 76% seven months ago, with YouTube emerging as the single most-cited domain for sources outside the top 100.
Gartner predicts 50% of enterprise cybersecurity incident response will involve AI-driven applications by 2028, as custom AI tools become both the target and the first line of defence.
Accenture and Databricks launched a joint business group on March 17 to accelerate enterprise AI adoption, betting that the data platform will be the chokepoint for agent deployment at scale.
Anthropic hired Google DeepMind's Matt Botvinick and OpenAI's Zoe Hitzig for its new Anthropic Institute, a research arm studying how advanced AI reshapes jobs, economies, and security, led by co-founder Jack Clark.
Only 15% of web pages retrieved by ChatGPT end up cited in its responses, and citation performance starts declining within 5 days without content updates, making freshness the new currency of AI visibility.
