Latest briefing
July 13, 2026 · 5 stories (site) · 5 stories (base)
On July 13, 2026, the AI agent industry is entering a phase of industrial consolidation: NVIDIA + LangChain cut the cost of an enterprise agent by ten with NemoClaw, UST makes Claude the reasoning layer of its chip, healthcare, telecom and banking platforms, Prime Intellect and Lyzr bring together $230 million to sell enterprises their own agentic stack. Three converging movements — price, reliability, infrastructure — sign the industry's coming of age.
🔥 Top story
01
NVIDIA and LangChain publish a recipe for enterprise agents 10 times cheaper — and turnkey
Running an AI agent in production today costs as much as a Paris–New York flight for every case processed. On July 8, NVIDIA and LangChain released NemoClaw: a complete kit combining their open Nemotron 3 Ultra model, their Deep Agents orchestration tool and a secure environment so the agent does not break anything. Announced result: the same work as before for about $4.50 instead of $43 — ten times cheaper. For an SME or a large enterprise that hesitated to start because of the entry ticket, the door is now open: the tooling is ready, the price is comparable to a standard cloud bill, and the first customers (Abridge in healthcare, Amdocs in telecoms, EY for audit) are already validating the model.
02
Claude becomes the brain of agents that act in the physical world — chips, hospitals, banks, telecoms
Until now, language models mostly served to answer on a screen. On July 9, UST — a global digital transformation group present in 30 countries and serving Fortune 500 clients — announced that it is making Claude the reasoning layer of its platforms for validating chips, managing health records, orchestrating telecoms networks and processing banking operations. UST promises to cut chip-validation time by half to nearly three quarters, and to train 20,000 of its engineers on the tool. For an IT director, it signals that frontier models are now judged reliable enough by a major partner to be plugged into systems that touch the physical world — not just to reply to a chat.
03
Prime Intellect raises $130 million to sell enterprises their own AI agent stack
Today, training a custom enterprise agent means going through OpenAI or Anthropic — and handing your data to an outside player. On July 8, Prime Intellect, a startup founded in 2024, closed a $130 million Series A at a billion-dollar valuation. It sells a complete stack — compute, isolated environment, trial-and-error training, evaluation — that companies can install on their own premises. First customers include Ramp, Zapier and the startup Flapping Airplanes; annual revenue already reaches $100 million. For a CIO or CISO, this is the most credible option so far to take back control of agents without depending on an American model provider.
04
A startup lets its own AI agent answer investors — and raises $100 million
Raising funds usually looks like an exhausting round of meetings in San Francisco. On July 9, Lyzr, a startup that helps companies build agents, did something new: it let its SivaClaw agent answer 130 investors, draft the memos and even measure which pitch-deck slides held attention. The founder barely left his office. Result: $100 million raised at roughly a $500 million valuation, on $400 million of expressed interest. For an entrepreneur or investor, it is the first real-world proof that an agent can run a complex negotiation — and a signal that nine-figure rounds no longer necessarily require a VC roadshow.
05
LangChain gives agents a notebook they fill in themselves — Gmail, Notion, Git, X included
Today, when you entrust a task to an AI assistant, it forgets everything as soon as the conversation ends. Tomorrow, it could keep a living memory of what it has learned — and update it without being asked. On July 10, LangChain (the company that invented the reference language for AI agents) introduced OpenWiki Brains, a free and open-source extension that turns any local wiki into an automatic notebook. The agent writes into it on the fly what it reads in your emails, your Notion documents, your code, your tweets and the sites you visit — like a very diligent intern who takes notes through the night. For a team testing several agents, it ends the obligation to re-explain everything at every session.
📡 To watch
Meituan LongCat-2.0: a Chinese model with 1.6 trillion parameters trained without any American GPU
Meituan, the Chinese food-delivery and local-commerce giant, unveiled in late June LongCat-2.0: a model with 1.6 trillion total parameters (33 to 56 billion active) trained end to end on a cluster of 50,000 Chinese accelerator cards — likely Huawei Ascend chips. One-million-token context, designed to code as well as the best. It is the first quantified confirmation that China can train a trillion-scale model on its own. If Meituan releases the weights, it is a new option in the diversification stack — to be confirmed in the coming weeks.
The agentic inference price war has started — NemoClaw cuts the bill by ten
With NemoClaw at $4.50 per run versus $43 for the runner-up, the question "which model should I choose?" becomes "which software harness around an open model?". For those building agents, the entry ticket has just dropped by an order of magnitude in a single announcement. To watch over the next seven days: a pricing response from Anthropic or OpenAI on their paid agentic offering, and adoption by the first non-early-adopter customers.
UST joins Anthropic's "top tier" partner list — physical AI goes into production
With 20,000 engineers set to be trained on Claude and chip-validation cycles cut by two to seven times, UST becomes one of the world's largest Claude deployers on systems that touch the physical world. For the agentic ecosystem, it is enterprise validation that frontier models are judged reliable enough for regulated use (health, banking) and for operations with a high cost of error (semiconductors).
Two rounds in one week confirm agent infrastructure as a standalone investment category
On July 8, Prime Intellect raised $130 million (Series A); on July 9, Lyzr closed $100 million (Series B). Two deals totalling $230 million in 48 hours, both on enterprise agent infrastructure and tooling. For an investor or vendor, it signals that the agent wave is no longer just marketing promise — money is now following the platforms that build the plumbing.
📊 Trend
On July 13, 2026, the AI agent industry is entering a phase of industrial consolidation. Three movements are converging in one week: NVIDIA + LangChain cut the cost of an enterprise agent by ten with NemoClaw; UST makes Claude the reasoning layer of its chip, healthcare, telecom and banking platforms; Prime Intellect and Lyzr bring together $230 million to sell enterprises their own agentic stack. For those building with AI, three lessons emerge: (1) the price of an agent in production has just dropped by an order of magnitude, opening the door to uses that were too expensive before, (2) frontier models are now judged reliable enough by major partners to be plugged into systems that act in the physical world, (3) agent infrastructure (compute, RL, memory, evaluation) is becoming a standalone investment category. The agentic stack of July 2026 looks less and less like a laboratory experiment and more and more like a classic industry — with its costs, its leaders, its reference customers.