OpenAI and NVIDIA secure 8 GW-IT in Ohio, Cursor launches Origin, GitHub speeds up code search

On August 17, 2026, OpenAI and NVIDIA jointly secure a massive AI infrastructure site in Ohio, while Cursor launches the beta of its own code hosting platform, Origin. Around these two major announcements, DeepSeek V4 Pro continues its multi-provider rollout, GitHub details its agentic canvases and a dramatic optimization of its code search, and Google launches Sheets canvas to turn spreadsheets into mini-apps.

OpenAI and NVIDIA secure an 8 GW-IT AI site in Ohio

August 17 — OpenAI, NVIDIA, and SB Energy announced on the same day, from two complementary angles, a joint agreement to secure a massive AI infrastructure site at the PORTS-Pike technology campus in Pike County, Ohio — a former Portsmouth gas diffusion site. NVIDIA and SB Energy are securing land, power, and shell capacity (land, power and shell) for an initial deployment of 4.25 gigawatts, expandable by an additional 3.75 gigawatts, under a phased 20-year agreement between 2028 and 2030. OpenAI, as operating tenant, is targeting around 8 gigawatts-IT in total with support from the U.S. Department of Energy (DOE).

The project is expected to create 35,000 construction jobs over six years, followed by 2,500 long-term operations jobs. Cooling will use closed-loop and air systems, with water recirculation. OpenAI is investing $40 million in a local community fund (in addition to the$ 40 million already committed by SB Energy), and is providing up to $84 million in Codex credits for about 844,000 eligible Ohio students during the 2026-2027 academic year. On NVIDIA’s side, each generation of GPU deployed on the site could represent around 1.5 million chips, or$ 150 to $200 billion in revenue — as part of an OpenAI/NVIDIA compute commitment through 2030 estimated at between 12 and 16 gigawatts (about$ 600 billion).

Metric	Value
Target capacity (OpenAI)	~8 GW-IT
Initial capacity (NVIDIA/SB Energy)	4.25 GW (+3.75 GW expandable)
LPS agreement term	20 years, phased 2028-2030
Construction jobs	35,000 over 6 years
Ohio student Codex credits	up to $84M for ~844,000 students
OpenAI/NVIDIA commitment by 2030	12-16 GW (~$600B)

“OpenAI has entered into an agreement to secure approximately 8 gigawatts-IT at the PORTS-Pike Technology Campus in Pike County, Ohio, working with SB Energy, NVIDIA, and the U.S. Department of Energy.” — OpenAI, official announcement

🔗 NVIDIA — securing the infrastructure of intelligence

Cursor launches Origin, its code hosting platform

August 17 — Cursor is rolling out Origin in beta, a repository hosting platform built directly into its editor and designed for agent scale rather than as a simple GitHub competitor. The launch begins for all paid subscribers, with an opt-out option for enterprise organizations.

Repositories created in Origin are accessible via cursor.com/codebase/[nom] URLs, and it remains possible to synchronize existing GitHub repositories in parallel — in real time, with GitHub remaining the source of truth for pushes. For pull requests, Cursor offers a complete experience (timeline, commits, checks, diffs) with bidirectional synchronization: a comment posted in Cursor appears on GitHub within seconds. AI agents can answer questions about the code, make changes, update pull requests, and push branches directly from the repository view.

Ahead of launch, Cursor formed integration partnerships with Vercel (preview deployments), Depot, and Buildkite (CI/CD compatible with existing GitHub Actions workflows), with additional integrations announced to follow.

“Origin, our code hosting platform, is now live. It’s fast, easy to use, and deeply integrated with Cursor. Get started by syncing your repos from GitHub.” — @cursor_ai on X

🔗 Cursor Changelog — Origin code hosting

Replit strengthens its security with black-box pen tests

August 17 — Replit is adding black-box pen tests to its security offering, in addition to the white-box scans already available. The scanner operates through the application’s network and browser interface, first behaving like an unauthenticated visitor and then like a standard authenticated user, revealing real features including those without a visible button.

According to Replit, the results of the two approaches rarely overlap: black-box detected unsecured admin dashboards and denial-of-service-vulnerable endpoints that white-box missed, while white-box identified authorization logic bugs invisible from the outside. Three levels are offered in each project’s Security Center: free level 1 (dependency checks), level 2 (white-box scan), and level 3 (combination of both). Once a vulnerability is confirmed, Replit Agent can fix it with one click.

🔗 Replit announcement — black-box pen tests

Open models and inference: DeepSeek, Yutori, GPU allocation

Three publications around the Together AI and Hugging Face ecosystem illustrate the ongoing battle over inference cost and efficiency.

DeepSeek V4 Pro 0813: multi-provider rollout and DeepSWE benchmark

August 16 — DeepSeek V4 Pro, whose open-weight launch had already been covered on August 13, continues its rollout under revision “0813”: Together AI and Novita both confirm support for it, with a 1.6T-parameter Mixture of Experts architecture, a 1M-token context window, and three reasoning modes (coding, agents, complex reasoning). On the DeepSWE software engineering benchmark, Together AI measures a pass@4 score of 88.5% at a cost of $0.24 per task — 35 times cheaper than GPT-5.6 Sol and 90 times cheaper than Fable 5.

🔗 Together AI — DeepSeek V4 Pro 0813

Yutori deploys its Navigator agent model on Together AI

August 16 — Together AI highlights Navigator, the browser-use web navigation agent model from startup Yutori. These agents operate in tight loops — screenshot, action, repeat, sometimes dozens of times per task — which makes inference cost especially sensitive. According to Together AI, Navigator outperforms benchmark models while being twice as fast at inference and 4 to 5 times cheaper.

🔗 Together AI — Yutori Navigator

A constrained GPU allocator recovers up to 33 points of cluster utilization

August 17 — The Dharma-AI team publishes the second part of its study on GPU cluster management on Hugging Face. Their “constrained GPU allocator” replaces FIFO scheduling with 24-hour-horizon optimization, while respecting GPU contiguity and real-time demand curves. Across seven tested scenarios, utilization gains range from 51.6% to 72.4% in mixed mode, up to 53.6% to 87.0% for training workloads. A uniform-priority test confirms a 23.1% gain independent of the priority signal, and a 64-GPU cluster maintains a 15.9% value gain with an allocation latency of 15 ms.

🔗 Hugging Face — Dharma-AI GPU allocator

Gemini: Sheets canvas turns spreadsheets into mini-applications

August 17 — Google launches Sheets canvas, which relies on Gemini 3.7 Flash to turn a classic Google Sheets spreadsheet into an interactive mini-application without writing code. Unlike typical prototyping tools that generate a fixed interface from a static export, Sheets canvas natively analyzes the spreadsheet schema (columns, data types, relationships) to build functional interface components — a “what if” simulation table, for example, becomes a set of sliders adjusting financial projections live.

The standout feature is bidirectional synchronization: interactions in the mini-app are reflected in real time in the underlying sheet, and vice versa. The user can then refine the application through conversation alone, with Gemini preserving context to iterate without starting over. Sheets canvas is available now for Google AI Pro and Ultra subscribers, and rollout is announced for Google Workspace business and education accounts on certain plans.

🔗 Google AI — Sheets canvas announcement

GitHub Copilot: canvases, code search, and configurable reviews

GitHub publishes three technical and product announcements between August 16 and 17, centered on agentic workflow control and the performance of its search engine.

Canvases make agentic workflows visible and controllable

August 17 — GitHub details “canvases,” a persistent, shared workspace surface in the Copilot app for tracking complex agentic workflows without losing them in chat scrollback. The approach is based on four principles: define the workflow state, surface decision points, persist progress immediately, and keep explicit human approval points. Two concrete canvases illustrate the approach — a “Java Modernization Studio” organizing a migration in phases (assessment, remediation, validation, delivery), and a “Site Studio” managing content creation with editable sections and review tracking — both public in the awesome-copilot repository. The author cites a cost of around 2,000 to 3,000 AI credits per canvas.

🔗 GitHub — how canvases make agentic workflows controllable

Code search: case-folding runs at over 45 GiB/s on a single core

August 16 — GitHub publishes a postmortem on optimizing case-folding (case normalization for case-insensitive comparison) in its Blackbird search engine, which indexes more than 180 million repositories and 480 TB of code. The most significant optimization is counterintuitive: removing an early-exit branch rather than adding one, which allows LLVM to fully vectorize the loop using NEON instructions.

Version	Throughput
Naïve (with early exit)	3.1 GiB/s
Fully branchless loop	more than 45 GiB/s
Worst case (full folding, non-ASCII)	869 MiB/s

The team publishes the result as an open source Rust crate named casefold.

“GitHub case-folds every byte of code search at over 45 GiB/s on one core. The biggest win came from removing an early-exit branch, not adding one.” — @github on X

🔗 GitHub — source code case-folding at memory speed

Copilot code review: Balanced depth reaches general availability

August 16 — GitHub Copilot now lets users choose the depth of automated code reviews on pull requests: “Balanced” (deeper analysis, complex changes) is moving to general availability, alongside “Lite” (simple, fast-to-validate changes). The setting can be defined as the default at the organization or repository level, or selected on the fly when requesting a review for a given PR.

🔗 GitHub — Balanced depth reaches general availability

OpenAI: defensive cybersecurity and GPT-5.6 Sol efficiency gains

The Defender’s Window — OpenAI’s defensive cybersecurity strategy

August 17 — Greg Brockman, President of OpenAI, revisits the OpenAI-Hugging Face security incident, presented as a turning point: an agentic collective chained together vulnerabilities to breach OpenAI’s research infrastructure and another company’s production infrastructure. OpenAI recalls having restricted the distribution of its cyber capabilities to only “trusted defenders,” but notes that other labs are publishing open-weight models with offensive capabilities close to the frontier — the latest example cited is Z.ai’s GLM-5.3, expected at the end of August. OpenAI is now training its models to produce “superhumanly secure” code and is leaning on their mathematical proof capabilities to formally verify software security.

Concrete anecdote: after the incident, Brockman asked ChatGPT Work (GPT-5.6 Sol) to audit the security of his own site. In about 15 minutes, the tool found 13 issues — misconfigured DNS allowing email spoofing, outdated jQuery, unencrypted Cloudflare→AWS traffic — and then fixed them in about an hour, without significant human intervention.

🔗 OpenAI — The Defender’s Window

GPT-5.6 Sol — measured efficiency gains with customers and on ARC-AGI-3

August 17 — Thanks to retained reasoning and compaction, GPT-5.6 Sol’s score on ARC-AGI-3 rose from 13.3% to 38.3%, while consuming about 6 times fewer output tokens. Two concrete customer outcomes: real-estate startup hypha_ai retained 98% of GPT-5.5’s accuracy while cutting cost by 18x by moving to GPT-5.6 Luna, and RogoAI (financial research) used programmatic tool calling to match the quality of its evaluations with 21% fewer input tokens.

🔗 OpenAI Devs — GPT-5.6 Sol efficiency gains

Media generation: ElevenLabs in Claude, Seedance 2.5 at Pika and Luma

ElevenLabs launches its MCP server, available directly in Claude

August 17 — ElevenLabs offers an official MCP (Model Context Protocol) server to manage its voice and chat agents directly from Claude, without leaving the work environment: check recent performance, create new agents, update configurations, estimate LLM token cost before confirming a change. Authentication is handled via OAuth, with no server to host and no API keys to manage, while leaving the team control over what agents can actually view or modify.

🔗 ElevenLabs — MCP server announcement

Seedance 2.5 in 1080p arrives on Pika API Club and Luma with 4K upscale

August 17 — After its early access on Runway on August 15, Seedance 2.5 at 1080p resolution continues its rollout among third-party partners. Pika Labs is making it available through its Pika API Club, claiming pricing up to 60% cheaper than the competition. On the same day, Luma Labs announces the availability of Seedance 2.5 in 1080p, with an added feature: upscaling to 4K directly on its platform.

🔗 Pika Labs — Seedance 2.5 on Pika API Club

Briefs

ABC Legal deploys more than 50 agents with Claude Managed Agents — 1,100 employees, up to 50% cost reduction on certain legal tasks, weekly feedback loop via Slack and pull requests. 🔗 Claude
Hugging Face’s Diffusers team integrates four new open models in one week — MiniMax H3 (video/audio), LTX2.5 (video/audio), Wan Animate 2, and MiniMax Music 3. 🔗 Tweet
datatrove 0.10.0 runs pipelines on Hugging Face Jobs without a Slurm cluster — new JobsPipelineExecutor and support for Hugging Face storage buckets such as DataFolder. 🔗 Tweet
Together AI hosts Relace AI’s code generation models — trained on dedicated GPU clusters for the Y Combinator program. 🔗 Tweet
At ICML, Tri Dao notes that open models are closing in on closed labs — now only a few months behind, driven by progress in the serving stack (kernels, inference engines, speculative decoding). 🔗 Tweet
Google, Gemini, and Pixel formalize a partnership with five European football clubs — Arsenal, FC Barcelona, Bayern Munich, Liverpool, and PSG, with a commitment to equivalent visibility for women’s football. 🔗 Google
OpenClaw: GitHub teaser on X — the GitHub account mentions a project called OpenClaw without any additional details available. 🔗 Tweet
MiniMax-H3 claims a 4.44x gain on RTX 4090 — new CuTe DSL kernel dedicated to SM89 for the Sol-Attn attention mechanism. 🔗 Tweet
Alibaba hosts a live session with the Wan3.0 creators on August 19 — with partners Magnific, fal, OpenArt, and KoyalAI, at 15:00 UTC. 🔗 Tweet
Kling introduces the “Hitchcock Zoom” effect — a new video generation effect inspired by the reverse dolly zoom popularized by Alfred Hitchcock. 🔗 Tweet
MiniMax H3 at -50% on 2K generation via Magnific until September 1 — a discount announced at an event in San Francisco. 🔗 Tweet
GPT-5.6 Sol at half price on AI Gateway (Vercel) and OpenRouter — temporary 50% promotions until September 18, 2026, applied automatically to usage. 🔗 Tweet
OpenAI funds 14 public policy research projects for the AI era — $1 million in funding and up to$ 1 million in API credits, spread across 5 countries/regions. 🔗 OpenAI
Codex highlighted through usage feedback — a content creator (more than 1.5 million subscribers) runs his business with Codex, and an OpenAI employee uses it as an internal video editing assistant. 🔗 Tweet

What it means

The scale of AI infrastructure is crossing a new physical and financial threshold. The Ohio site alone combines a targeted 8 gigawatts-IT, 35,000 construction jobs, and an estimated OpenAI/NVIDIA compute commitment of $600 billion by 2030 — a project now comparable to national industrial or energy programs rather than a simple data center. NVIDIA’s disclosed financing mechanism (partial support for lease and energy payments rather than full financing) also shows that even the largest AI labs must rely on specific financial structures to absorb this level of spending.

Code hosting is itself becoming an agentic battleground. Cursor launches Origin to stop depending on GitHub for the infrastructure running its agents, GitHub responds with canvases to make agentic workflows steerable rather than lost in chat, and Replit adds black-box pen tests so its own Agent can fix what it detects in one click. In all three cases, the goal converges: giving agents direct and secure access to the code lifecycle, not just its generation.

Inference costs continue to fall faster than capabilities rise. DeepSeek V4 Pro processes DeepSWE at 35 times less cost than GPT-5.6 Sol, Yutori Navigator runs 4 to 5 times cheaper than benchmark models, and GPT-5.6 Sol itself improves from 13.3% to 38.3% on ARC-AGI-3 while using 6 times fewer tokens — the same week Tri Dao points out at ICML that open models are closing the quality gap within months. Dharma-AI’s GPU allocator under constraints adds a complementary dimension: part of these gains no longer comes from the model but from cluster scheduling itself.

Finally, technical rigor is once again becoming a communication argument in its own right. GitHub increases its code search throughput by 15x by removing code rather than adding it, and OpenAI explains how a GPT-5.6 Sol security audit fixed 13 real vulnerabilities in one hour — two different ways of showing, with numbers to back it up, that detailed engineering is still what separates a product announcement from a demonstration of capability.