Modelos, agentes, investigación y negocio de la IA.
Although AI models are increasingly efficient, the rise of agentic workloads is causing an explosive increase in token consumption. A single task can require millions of tokens, driving monthly costs to thousands of dollars per user.
NVIDIA has announced that its new Rubin architecture will use coolant at 45°C, hotter than a jacuzzi, to dissipate heat from AI chips. This counterintuitive strategy reduces the need for mechanical cooling and water consumption, marking a milestone in data center efficiency.
British startup CuspAI has launched the AI Materials Foundry, a coalition that includes giants like Nvidia, Meta and Samsung. The goal is to create a 'search engine' for materials that don't yet exist, combining computer simulations, machine learning and robotic experimentation. The project promises to revolutionize industries such as batteries, semiconductors and construction.
Researchers at Writer published a paper demonstrating reductions of up to 61% in cost per successful task by optimizing the AI harness, without touching the base model. The key lies in combating 'tokenmaxxing' and redesigning agentic loops.
The Model Context Protocol (MCP), developed by Anthropic, is becoming the de facto standard for AI assistants to securely access external data and services. A new update reduces implementation friction, bringing code-free interoperability closer to businesses and developers.
A conversation with an AI agent can score perfectly in an individual evaluation yet signal a broken product. Leaders from LangChain, Conviva, and CoreWeave explained at VB Transform 2026 that the industry is moving from isolated traces to contrastive analysis across user cohorts, along with cheaper, more specific judge models.
Hugging Face confirmed a cyberattack where an autonomous AI agent executed thousands of actions, exploited vulnerabilities, and stole credentials. This incident marks a milestone in the evolution of cyber threats, proving that AI-driven attacks are no longer theory but reality.
Capital One has launched VulnHunter, an open-source AI tool that scans source code for exploitable vulnerabilities, traces attack paths, and proposes fixes. The move seeks to redeem the bank's security reputation after the 2019 breach that affected 106 million people.
Moonshot AI's Kimi K3 has outperformed top US models in coding tests at significantly lower cost, with plans to be open-weight. This threatens US tech dominance and pressures sector prices and valuations downward.
Last Saturday, opposition to data centers in the United States escalated nationwide with coordinated protests in 42 states. Protesters denounce massive energy and water consumption, environmental impact, and lack of regulation. The movement threatens to slow the expansion of artificial intelligence infrastructure.
At VB Transform 2026, infrastructure leaders from LinkedIn, Walmart, and Zendesk concluded that legacy infrastructure, not AI models, is the main obstacle to scaling AI agents. Each company faced distinct bottlenecks, but all pointed to the same root: systems built for human pace, not agent speed.
Anthropic has donated the Model Context Protocol (MCP) to the Linux Foundation, joining Google and OpenAI in standardizing AI protocols. This move, following the 'commoditize the complement' strategy, aims to position Anthropic at the center of the AI agent ecosystem.