Tech • IA • Crypto
Chips, OCR, CAPTCHA, espionage, what's new: AI news decoded
9.3/10TL;DR
OpenAI, Mistral, and major tech partners are accelerating a shift toward specialized AI infrastructure, smarter document processing, and new human–AI collaboration models.
KEY POINTS
OpenAI unveils in-house inference chip “Jalapeño”
OpenAI has designed its first proprietary AI chip, developed with Broadcom, focused exclusively on inference rather than training. This application-specific integrated circuit aims to bypass GPU limitations, particularly data transfer bottlenecks that slow response times. Early tests are already underway, marking a strategic move toward full control of AI infrastructure.
A bid for massive computing power
The company targets up to 10 GW of inference capacity, far exceeding current large-scale AI data centers such as Elon Musk’s Colossus clusters, which total around 2 GW. This signals a shift toward vertically integrated AI ecosystems, where providers control chips, servers, and cloud delivery.
Inference becomes the new battleground
Demand for inference computing is now estimated at 15 to 20 times that of training workloads. This shift is driving competition among players to optimize speed, cost, and scalability for real-time AI applications, including ambitions like live code generation.
Groq re-enters the race with new funding
Chipmaker Groq, previously linked to Nvidia through a licensing deal, has raised $650 million to expand into inference cloud services. Rather than focusing solely on hardware, Groq is positioning itself as a full-stack competitor to OpenAI in AI infrastructure delivery.
Mistral launches OCR4, a document AI breakthrough
French firm Mistral introduced OCR4, a powerful system that converts scanned documents, PDFs, and images into structured, machine-readable data. It supports 170 languages and delivers outputs in structured formats like Markdown, enabling seamless integration into workflows.
Structured data unlocks enterprise value
OCR4 can identify document elements such as titles, tables, and signatures, producing organized datasets usable for automation, analytics, and AI training. This is particularly valuable for industries with large volumes of non-digitized records, including administration and finance.
Performance and scalability gains
The system can process up to 2,000 pages per minute on high-end GPUs and shows strong preference scores in blind evaluations. Its ability to follow custom extraction schemas makes it highly adaptable for enterprise use cases.
New protocol aims to eliminate CAPTCHAs
A collaboration between Cloudflare, Google, Mozilla, Microsoft, and Shopify introduces Private Access Control Tokens (PACT). This system verifies whether a user is human without requiring visible tests, reducing friction in online experiences while preserving privacy.
AI espionage tensions surface
Evidence suggests a coordinated attempt, attributed to Alibaba’s Qwen team, to extract insights from advanced AI models. The operation reportedly involved large-scale probing of systems, raising geopolitical concerns and possibly influencing restrictions on certain AI releases.
Claude integrates as a workplace collaborator
Anthropic has introduced “Claude Tag,” embedding AI agents directly into Slack channels. Each channel gains a dedicated AI that can access shared resources, perform asynchronous tasks, and proactively assist teams, moving beyond chatbot interactions toward persistent digital coworkers.
CONCLUSION
The AI industry is rapidly evolving toward specialized hardware, structured data pipelines, and embedded collaboration tools, intensifying both competition and geopolitical stakes in the race for technological leadership.