TLDR AI 2026-03-17

Every AI company eventually hits the same wall: billing. Token-based pricing, usage metering, per-customer rate cards... infra work that eats months of engineering time.

Metronome is the platform that companies like Anthropic and Databricks rely on to handle it all: real-time metering, pricing changes in minutes, and billing that can scale from seed stage to revenue in the billions.

🆕 And with their new self-serve option, it's even easier for AI startups to get up and running:

>> Sign up for a free sandbox

>> Model your pricing

>> Test billing flows before committing anything

✅ Startup-friendly pricing kicks in when you go live.

Create a billing experience that can grow with you

OpenAI plans to refocus its efforts around coding and business users. Its leaders are actively looking for areas to deprioritize. The company's 'do everything at once' strategy helped it gain a reputation as the pioneer of the AI era. However, it is under growing pressure from rivals, so the company requires a clearer strategic direction.

NVIDIA outlined a broad GTC 2026 slate spanning open foundation model partnerships, agent tooling, new reasoning and safety models, robotics systems, and healthcare AI for drug discovery and simulation.

Sachin Katti joined OpenAI in November to serve as its head of industrial compute. Before OpenAI, Katti spent more than 15 years as a Stanford professor and four years working at Intel. Katti now works on finding additional data center capacity and lining up ways to get more components like AI chips and memory. This has proved challenging, as data center operators are contending with power grid constraints, memory chip shortages, and growing pushback from local communities.

Alibaba is setting up a business unit to bring its AI services and development endeavors under a single umbrella. The new Alibaba Token Hub will comprise the research team that develops the company's flagship Qwen models, its consumer-facing app division, and other major AI-related products. It will also oversee Alibaba's Slack-like DingTalk app and devices under the Quark brand. The revamp will help quicken interaction between the various teams within Alibaba's broader AI effort.

This developer asked their AI agent how it wants to remember things. The agent redesigned its own memory system, ran a self-eval, diagnosed its blind spots, and improved recall from 60% to 93%, all for just $2. The experiment shows what happens when you treat AI as a participant in its own cognitive architecture.

OpenAI explained that Codex Security was designed to analyze repositories directly instead of triaging static analysis reports, focusing on system architecture, trust boundaries, and validating findings before surfacing them to humans. The approach targets semantic security flaws where defenses appear present but fail to actually enforce the intended protection.

This article draws parallels between nuclear weapons development and AI, arguing that AI poses similar global stakes. Anthropic, a key player in AI, hesitates to grant US government access, reflecting private control dilemmas akin to the hypothetical McBombalds Corp scenario. This raises the debate on whether private entities should control technologies with such immense global impact or if government oversight is more appropriate.

Agentic AI and GenAI are powerful but unpredictable. It's not just hallucination - they regularly take entirely new paths through established workflows.

This Dynatrace report lays out six pragmatic observability recommendations for practitioners managing agentic AI and GenAI workloads. Learn to look beyond monitoring, spot escalating costs, and catch critical issues early. Read the report

Want to check it out firsthand? Experiment with AI observability tools in the Dynatrace Playground - where you can explore sample data without installing any software.

OpenShell is a safe, private runtime for autonomous AI agents that provides sandboxed execution environments that protect data, credentials, and infrastructure. It is governed by declarative YAML policies that prevent unauthorized file access, data exfiltration, and uncontrolled network activity. The OpenShell project ships with agent skills for everything from cluster debugging to policy generation. It will eventually build toward multi-tenant enterprise deployments.

Leanstral is an open-source coding agent designed for Lean 4, a proof assistant capable of expressing complex mathematical objects. It is designed to be highly efficient and is trained for operating in realistic formal repositories. Leanstral's weights were released under an Apache 2.0 license. It can be accessed in an agent mode within Mistral vibe and also through a free API endpoint.

The subagents pattern is now widely supported in coding agents. Subagents are now generally available in OpenAI Codex. There are default subagents called 'explorer', 'worker', and 'default', but it is unclear what differentiates them. Users can define custom agents with custom instructions and specific models.

Mistral Small 4 integrates the capabilities of Magistral, Pixtral, and Devstral models, offering unified multimodal, reasoning-optimized AI with configurable reasoning effort. It employs a Mixture of Experts architecture with 119B parameters, supporting both text and image inputs, and features efficient scaling. Mistral Small 4 achieves competitive performance with reduced output length, is open-source, and is available on platforms like vLLM, llama.cpp, and Transformers.

Apple will invest $14 billion into AI this year, a tiny amount compared to the $700 billion Amazon, Alphabet, Meta, and Microsoft are investing. The company appears to believe that the AI infrastructure build-out will produce inadequate returns. It is spending less due to a conviction that AI models will commoditize and shrink, that existing product lines will absorb the workloads the cloud was built to serve, and that the durable franchise belongs to whoever owns the customer. Apple is betting on its AI-capable devices rather than centralized infrastructure.

Nvidia's focus this year at its GTC event shifted to inference, the type of computing required to run models and allow them to respond to user queries. The AI industry is now less concerned with training AI models, which is what GPUs are best at, and more preoccupied with running them and generating profits from end-users. Inference requires different hardware than chips optimized for training. How far ahead the company remains in the AI-infrastructure race will depend largely on how effectively it is able to pivot its product road map from training to inference.

We benchmarked five MCP architectures against real enterprise queries. Most were accurate 60–75% of the time. CData Connect AI: 98.5%. The data and testing approach are public.

The proposed deal could give OpenAI a faster route into corporate adoption while providing PE firms with a potential lifeline for companies in their portfolios that are exposed to AI disruption.

Nvidia Dynamo 1.0 accelerates generative AI and reasoning models in large-scale distributed environments by delivering low-latency, high-throughput distributed inference.

Open-H-Embodiment is a community‑driven healthcare robotics dataset created for the training and evaluation of AI autonomy and world foundation models for healthcare applications.

Drew Bent is the education lead at Anthropic.