Table of Contents
- 1. Big Tech will spend $700 billion on AI in 2026
- 2. Battle of the models: GPT-5.3, Claude 4.6 Opus, Gemini 3.1 and Grok 4.20
- 3. Subagents in coding: how AI solves the large-context problem
- 4. Local AI is gaining momentum: RAG systems without the internet
- 5. Decentralized agent networks — the future beyond orchestrators
The world of artificial intelligence does not stand still for a single day. March 2026 brought a series of explosive events that are changing the rules of the game: massive investments, a race among language models, a revolution in approaches to AI development, and debates about the technology’s economic return. Here we look at the main hot AI news of the week.
1. Big Tech will spend $700 billion on AI in 2026 — but Goldman Sachs says the effect is “zero”
The world’s largest technology companies are ready to invest a record amount in artificial intelligence $700 billion in 2026 — almost twice as much as the $400 billion spent in 2025. The amount is comparable to the GDP of a small European country.
However, Goldman Sachs chief economist Jan Hatzius cooled the optimism: according to him, AI spending in 2025 made a “practically zero” contribution to U.S. GDP growth. In an interview with the Atlantic Council, the economist said bluntly: “We do not actually view AI investment as highly growth-stimulating.”
The paradox is obvious: Deutsche Bank analysts calculated that AI investment accounted for almost all of the growth in the American economy, while Harvard professor Jason Furman pointed to AI contributing 92% of U.S. GDP growth in the first half of 2025. Time will tell who is right, but the debate has already erupted across the entire financial world.
Bottom line: Despite the debate about the economic impact, investment in AI continues to grow. Technology companies are betting on long-term returns, even if the short-term results are not yet obvious.
2. Battle of language models: GPT-5.3, Claude 4.6 Opus, Gemini 3.1 and Grok 4.20
The large language model market has turned into a real boiling pot. Over just the past month, several major releases have happened at once:
- OpenAI released GPT-5.3 — the latest update to its flagship model
- Anthropic responded with the powerful Claude 4.6 Opus, strengthening its position in the reasoning-model segment
- Google updated its flagship to Gemini 3.1 Pro following the release of Gemini 3
- Elon Musk and the xAI team released the Grok 4.20 beta
Particular attention deserves Gemini 3.1 Pro: despite Google’s aggressive marketing, many real users are skeptical. “I’ve been running the model for the second day and I honestly can’t see any difference from the regular version 3,” is a typical comment in professional communities. Critics believe the update is a “renamed badge” meant to maintain visibility in the news cycle.
For users this means a rich choice of tools and real competition, which pushes all companies to improve their products.
3. Subagents in AI coding: a revolution in the approach to large projects
Developers of large projects know the pain well: you ask an AI assistant to add OAuth authentication with three providers, and by the middle of the task it starts mixing up middleware names, duplicating helpers, and “forgetting” decisions made three messages earlier.
This is not a bug in a specific model — it is a fundamental architectural problem with a single context. And now Cursor and Claude Code have independently arrived at the same solution: subagents.
The idea is simple: instead of one agent trying to keep everything in its head, several specialized agents are launched, each responsible for its own part of the task. One handles architecture, another tests, and a third documentation.
At the same time, Cursor and Claude Code implemented the concept differently:
- Cursor emphasizes parallel task execution with manual control
- Claude Code provides a more autonomous model with hierarchical subagents
This topic is being actively discussed in the Russian-speaking developer community: an article on Habr collected more than 4,000 views in just a few days.
4. Local AI without the internet: RAG systems are gaining popularity
Alongside cloud giants, interest in local AI systems is growing. Developers are building solutions that work entirely offline — without sending data to OpenAI or Google servers.
One illustrative example is a RAG system (Retrieval-Augmented Generation) with a graphical interface, built on the combination of GitHub Copilot and Claude. The system indexes textbooks and corporate documents, then generates reports, presentations, and summaries from them — all locally, without the internet.
Another trend is smart speakers on Raspberry Pi with an open stack: Ollama + Gemma3 + Moondream + Whisper.cpp + Silero TTS. Such a device understands speech, sees through a camera, and responds by voice — fully autonomously, without the cloud.
Why this matters: Local solutions give businesses control over data, eliminate dependence on external providers, and work without an internet connection. In an era of sanctions and unstable networks, this is more relevant than ever.
5. Decentralized networks of AI agents: the next frontier
The most philosophical, but also the most far-sighted, discussion of March 2026 is about decentralizing multi-agent systems.
The essence of the problem: even when modern agent systems look like “swarms” of autonomous agents, in practice they remain centralized — there is one orchestrator that manages everything. This creates a single point of failure and limits scalability.
The alternative is the protocol HyperCortex Mesh Protocol (HMP v5.0), an open specification for interaction between autonomous AI agents without a central coordinator. In spirit, it is close to IPFS and ActivityPub, but it solves the task of coordinating cognitive agents.
For now, this is more of an architectural concept than a finished product. But it is precisely such fundamental changes that often precede a new wave of technological opportunities.
Summary: what these news items mean for you
The hot AI news of March 2026 paints a clear picture:
- Investment is growing — $700 billion shows that big business believes in a long-term bet on AI
- Model competition — this is good for users: quality is improving, prices are falling
- Architecture is changing — subagents and decentralization are not a trend, but a response to real technical limitations
- Locality matters — control over data is becoming a priority for businesses
Artificial intelligence in 2026 is no longer an experiment. It is a working tool, infrastructure, and the object of multi-trillion-dollar investments. Keeping track of its development is important for everyone who wants to stay up to date with technology.
Frequently Asked Questions (FAQ)
How much will Big Tech spend on AI in 2026?
According to forecasts, the largest technology companies will spend about $700 billion on artificial intelligence in 2026. That is almost twice the $400 billion invested in 2025. The main players are Microsoft, Google, Amazon, Meta, and Apple.
What are subagents in AI coding?
Subagents are specialized AI agents that work in parallel under the control of the main agent. Each subagent is responsible for its part of the task (architecture, tests, documentation), which makes it possible to overcome the limitations of a single context and handle large projects.
How is Gemini 3.1 different from Gemini 3?
According to Google, Gemini 3.1 Pro is an improved version of the flagship model with enhanced reasoning and generation capabilities. However, many users still do not notice a significant difference in real-world tasks, calling the update a "minor patch."
What is a local RAG system?
RAG (Retrieval-Augmented Generation) is an architecture in which a language model is supplemented with search over a knowledge base. A local RAG system runs on your computer without the internet: you upload documents, the model indexes them, and answers questions based on your data.
Which language model is the best in 2026?
At the beginning of 2026, several top models are competing at once: GPT-5.3 (OpenAI), Claude 4.6 Opus (Anthropic), Gemini 3.1 Pro (Google), and Grok 4.20 (xAI). The choice depends on the task: Claude and GPT lead in reasoning, while Gemini is strong for multimodal tasks.