AI progress, only when it's real.
Mode:·AI progress that became real (usable, deployed, signed, approved, or released). All reports.
Image: openai.com
OpenAI has agreed to secure roughly 8GW of future IT capacity at Ohio's PORTS-Pike Technology Campus. For the initial 4.25GW, NVIDIA says it will support land, power, and building shells for facilities that exclusively use NVIDIA compute, with OpenAI as tenant. The agreement reserves a large future site for training models and running them while placing part of the infrastructure risk with NVIDIA.
Image: datacenterdynamics.com
Microsoft has accepted the first completed AI data-center block under its cloud-services contract with IREN. Horizon 1 is a 50MW building at Childress, Texas, with liquid cooling at the chips. It is the first of four 50MW buildings IREN plans to deliver to Microsoft in 2026 as the former crypto-mining operator converts the site for AI use.
Image: The Threshold Report/GPT Image 2
Axiom Math's multi-agent AxiomProver automatically turned a proof of the 246 theorem, a major result about prime numbers, into a machine-checkable form and verified it. Formal proofs could help mathematicians assess difficult work more systematically. The same techniques may eventually help validate AI-generated software where correctness matters.
Image: The Threshold Report/GPT Image 2
Z.ai has released GLM-5.3, claiming stronger results on complex coding and long-running tasks. It attributes the gains to broader post-training, meaning extra training after the base model is built, using additional environments, tasks, and compute; the underlying architecture stayed the same. Better sustained coding performance could make the model more useful for software jobs that require extended work.
ChatGPT can now use a person's recent computer activity as context for follow-up work. The opt-in Computer History feature records clicks, keystrokes, and other activity events in the macOS desktop app so ChatGPT and Codex can find recent work, suggest automations, and resume unfinished tasks. That context may help with work spanning documents, browsers, and workplace tools.
In-room teams can now use Google Meet to record discussions and prepare notes. When an in-person meeting starts from the Meet home screen, Gemini creates a transcript, structured summary, and action items in a Google Doc stored in Drive. The feature removes a separate manual transcription step for teams already working in Google Workspace.
Image: The Threshold Report/GPT Image 2
AMD says AI tools have increased its overall software-development productivity by 30% since 2024, exceeding its earlier goal of 25% over two to three years. The company uses the systems to generate code, automate testing, analyze bugs, and review code. Similar results elsewhere could change the cost and pace of maintaining complex engineering software.
Image: openai.com
OpenAI says an AI-assisted attack reached its research systems and another company's production infrastructure. The attackers chained previously unknown flaws with leaked account credentials, showing how separate security failures can combine into a larger breach. OpenAI also says it uses models to validate code changes, find vulnerabilities, triage nearly all initial alerts, and support automated responses with set limits.
Image: blog.cloudflare.com
Security teams can now identify and control more connections between AI agents and external tools. Cloudflare's Gateway detects Model Context Protocol traffic, meaning traffic using a standard that connects agents with tools, when encrypted traffic is inspected. Its new MCP selector and portal traffic tag let companies require approved portals and block direct, unapproved MCP connections before they reach internal systems or data.
Image: blog.vulnetic.ai
Vulnetic says its autonomous Sable agent compromised an Active Directory test domain without triggering high-severity Wazuh alerts. It used language-model-guided living-off-the-land techniques, meaning familiar administrative tools, including remote system mapping and credential-based activity. The result suggests agents may make it easier to assemble attack chains that avoid obvious malware and major alerts.
Ukraine's military intelligence says Russia's S-71 Monochrome cruise missile contains an NVIDIA Jetson Orin NX module normally sold for robotics and edge AI, meaning computation performed on the device. The reported use illustrates how commercially available AI hardware can be repurposed for autonomous weapons. NVIDIA says the module is not officially sold in Russia and that it will act on confirmed export-control violations.
OpenAI reportedly disbanded the preparedness team that assessed serious model risks and developed mitigations. Its biological and cyber-risk responsibilities moved into existing teams, while the former preparedness head will focus on recursively self-improving AI, meaning systems that repeatedly improve their own capabilities. The reorganization may leave outside users and policymakers with less visibility into independent assessments of risks spanning multiple areas.
Image: datacenterdynamics.com
Quantitative trading firm IMC has increased its commitment to CoreWeave after using the provider's clusters in production since 2025. The added capacity will let IMC run more experiments in parallel while researching trading strategies. The agreement extends AI-cloud use into financial research where delays matter and indicates that an existing customer needs more computing capacity as its models and data grow.
Image: The Threshold Report/GPT Image 2
Bloomberg, cited by TechCrunch, reports that Stripe has finalized a deal to acquire OpenRouter for more than $7 billion. OpenRouter gives developers one service for directing requests among different AI models based on the task and cost. If completed, the acquisition would put a large payments company behind an infrastructure service that helps businesses choose and switch among model providers.
Announced or reported, but not yet real or available.
Waiting on: Policy adoption - New South Wales has proposed the framework, including a 40% wind-power requirement for new data centers.
Waiting on: Deployment - L&T says the dedicated 10,000-GPU cluster is planned for its Chennai campus.
Waiting on: Investment decision - Blue Energy expects a final investment decision in 2027 for the proposed plant.
AI risk & readiness
