☕️ Nvidia backs OpenAI with $105B

Hi there, this is your daily ☕️ AIpresso.

In today's AIpresso:

💰 Nvidia backs OpenAI with $105B

📚 Amazon destroys rare books to train AI

🔓 LiteLLM hack exposed 2,500 organizations

🖥️ Qwen model runs coding agents locally

💻 Cursor launches GitHub rival Origin

Plus: 💡 4 strategies & tactics, 🎁 7 other news you might like, 🧰 6 tools, and 📚 5 papers.

💰 Nvidia backs OpenAI with $105B LINK

Nvidia will guarantee up to $105 billion to help OpenAI lease capacity at a massive Ohio data center in Pike County, deepening the partnership between the two companies.

OpenAI plans to lease up to 8 GW at the SoftBank-backed site, with the first 800 MW online by 2028; the complex will run exclusively on Nvidia GPUs, potentially yielding ~$600bn in revenue through 2030.

The backing phases in from 2028 to 2030 over a 20-year term, though the $105bn covers only the initial 4.25 GW portion and Nvidia is not guaranteeing the full site or all of OpenAI's lease obligations.

📚 Amazon destroys rare books to train AI LINK

Amazon is reportedly buying and scanning large quantities of books, including rare and antique volumes, for AI training data, sometimes destroying them in the process, according to a 404 Media investigation that tracked a planted book to a Las Vegas warehouse.

The report echoes Anthropic's Project Panama, which cut book spines for scanning and was ruled fair use by a judge, while warehouse workers told 404 Media their main task was preparing bulk book shipments for faster scanning.

Alongside the scanning push, Amazon has overhauled its AI strategy, winding down Nova Premier, Omni, and Reel and folding its AGI org under Peter DeSantis, with the Frontier Model Research group now led by Pieter Abbeel after Rohit Prasad's December 2025 departure.

🔓 LiteLLM hack exposed 2,500 organizations LINK

A March supply-chain attack on the LiteLLM AI gateway compromised its PyPI account and may have exposed credentials from more than 2,500 organizations and hundreds of thousands of CI/CD pipelines, security researchers report.

Attackers published malicious LiteLLM versions 1.82.7 and 1.82.8 on March 24 to steal cloud keys, repo tokens, SSH keys, and Kubernetes secrets, with Hudson Rock finding a 153GB archive of 433,909 files tied to 2,488 domains.

The stolen tokens create downstream access to cloud platforms, source repos, and Kubernetes environments, making credential rotation critical, though scale estimates vary widely, from a 40-minute exposure window to Mocchis' 195TB haul.

🖥️ Qwen model runs coding agents locally LINK

Alibaba released Qwen3.8-27B on Hugging Face under Apache 2.0, a 27B dense multimodal model with a 262K-token context that runs coding and agentic workflows locally, fitting in ~17GB at 4-bit quantization.

Third-party benchmarker Artificial Analysis scored it 52 on its Intelligence Index, matching GPT-5.6 Luna at max reasoning, and 51 on the Agentic Index, edging out Anthropic's Claude Opus 4.8, while Alibaba reported 61.7 on SWE-bench Pro.

The model passed 3M downloads in three days and serves via vLLM and SGLang, though it defaults to an xhigh reasoning setting that burned 160M output tokens across testing, roughly 30x slower and 4.5x costlier in one small agent trial.

💻 Cursor launches GitHub rival Origin LINK

Cursor has launched Origin, a Git-based cloud service for storing and managing code repositories, marking its first major product release since the company's $60 billion sale to SpaceX in June.

Accessible via a new tab in the Cursor desktop client or a dedicated CLI, Origin offers branching, a GitHub connector that two-way syncs projects, and both desktop and cloud-sandbox agents that handle long-horizon tasks with the workstation off.

Origin ships with pre-packaged connectors for Vercel, Depot, and Buildkite, with more integrations and unspecified "agent-native features" planned, though Cursor has not detailed those capabilities or a timeline.

💡 Strategies & Tactics

> Developing Nemotron 3.5 Lightning NVFP4 with QAD Using NVIDIA Model Optimizer: Compressing NVIDIA's Nemotron model with quantization-aware distillation shrinks it from 66 GB to 22 GB and quadruples throughput while keeping accuracy.

> DeepSeek V4 Pro 0813 vs GPT-5.6 Sol on DeepSWE: Cost, Coding, and Routing: Run cheap DeepSeek first and escalate to pricier GPT-5.6 Sol only when tests fail, hitting higher accuracy for under half Sol's cost.

> I ditched Ollama as my default runtime, and the replacement starts models in a fraction of the time: Switching to BaseRT, a Mac-only runtime for running AI models locally, loads models and processes prompts far faster than Ollama on Apple Silicon.

> The economics of a voice agent: what a minute of conversation actually costs: Judge a voice agent by its cost per successful outcome, not its per-minute rate, since failures, retries, and human handoffs quietly inflate the real price.

Other news you might like

• Stripe will reportedly acquire AI gateway startup OpenRouter for $7B+LINK

• Multi-Vector (Late Interaction) Embedding Models with Sentence TransformersLINK

• Same Cluster, 33 Points More Utilization: What Changed Was the OrderLINK

• U.S. CISA adds a Ray-Project Ray flaw to its Known Exploited Vulnerabilities catalogLINK

• Snowflake targets AI costs with model routingLINK

• Microsoft Copilot reveals secret input that allowed it to be hackedLINK

• Claude can now delete your production voice agent from a chat windowLINK

🧰 Trending tools

Zetik: a personal intelligence agent that continuously tracks news, podcasts, papers, and code, delivering concise briefings through feeds, push, newsletters, or RSS.LINK

Hoplite: moves your local coding agent setup, sessions, MCP servers, dependencies, and CLIs, to the cloud so agents run in parallel without laptop interruptions.LINK

Basedash Tasks: builds dashboards and analyzes customer data from plain-language descriptions, so you can skip writing SQL queries manually.LINK

Deskcommcrm: Open-source AI sales OS, self-hosted CRM with native AI agents + WhatsApp (WAHA). Open alternative to Kommo, Octadesk & Intercom for any business that sells by chat. MCP-ready, multi-tenant, LGPD.LINK

world-intel-mcp: a server providing 100+ tools for real-time global intelligence across markets, finance, conflict, cyber, and climate domains.LINK

LLMVault: an intentionally vulnerable training platform for practicing AI security, prompt injection, RAG security, agent security, and GenAI penetration testing.LINK

📚 Trending papers & reports

The ambiguity curse shows that language models struggle most to predict situations with many plausible next words, meaning you should trust their probability estimates least exactly when outcomes are genuinely uncertain and open-ended.LINK

Live-shopping avatar agents can be trained to follow rapidly changing product and compliance rules without retraining, letting a compact model outscore a top general system on real streaming Q&A, ~95 versus ~93.LINK

A model's sense of the current year comes from two disconnected internal clocks, one tied to verb tense and one to direct questions, so no fix updates both at once, undermining reliable time-based reasoning.LINK

Database automation code gets its first real exam, and tests of eight AI systems show they still stumble on writing working procedural database programs, a task ordinary coding and query benchmarks never checked.LINK

Hallucination pinpointing flags the exact words an AI made up and traces each back to the source text it should have relied on, showing not just whether an answer is wrong but where.LINK

See you tomorrow for a new dose of ☕️ AIpresso!