A quiet day for model releases turned into a loud one for everything underneath them. The biggest number was NVIDIA agreeing to guarantee up to 105 billion dollars of OpenAI's new Ohio data centre, blurring the line between chip supplier and lender. Availability was the other story: GitHub fell over worldwide and took Copilot, Actions and coding agents down with it, while Google switched off its Imagen 4 endpoints on schedule. The only genuine product launches were connectors, Grok gaining a native Whop integration and ElevenLabs shipping an MCP server into Claude, both of which broke on the companies' own channels first. Add AWS rationing CPU as agents strain capacity and a clutch of applied-AI funding rounds, and the week's real contest looks less about which model is smartest and more about who controls access, routing and uptime.
NVIDIA guarantees up to 105 billion dollars of OpenAI's Ohio campus
An SEC filing reported on 17 August shows NVIDIA agreeing to guarantee up to 105 billion dollars in lease and power obligations at a 10-gigawatt data centre in Pike County, Ohio that OpenAI has leased for twenty years, with the first 800 megawatts targeted for 2028. NVIDIA is also putting 1.5 billion dollars into SoftBank subsidiary SB Energy, which will build and operate the campus on a decommissioned uranium-enrichment site. Total project cost including chips could exceed 500 billion dollars.
For users, the relevant read is on capacity and pricing: the compute behind the models you use is now being financed through structures that tie the chip supplier's balance sheet to the model provider's growth. That tends to keep capacity expanding and unit prices falling in the near term, and it concentrates risk in ways worth watching if you are betting a product roadmap on a single provider staying cheap.
GitHub went down worldwide and took Copilot and Actions with it
Microsoft confirmed at 09:40 EDT on 17 August that GitHub was impaired globally, with roughly 20 per cent error rates on the web experience and API, around 50 per cent on archive and repository downloads, and degraded SAML, OIDC and SCIM authentication. Actions, Copilot, Issues and Pull Requests were all affected, which broke CI pipelines and any coding-agent workflow that reaches into a repository. No cause was disclosed.
The practical lesson for anyone running agentic coding in production is that the agent is only as available as the platform underneath it: an agent that cannot read a repo or open a pull request is not degraded, it is stopped. If your build or release process now depends on an agent completing a GitHub action, yesterday was a reminder to keep a manual path that still works.
The Imagen 4 endpoints went dark, and the replacement is not a drop-in swap
If you have any image generation running through the Gemini API, check it this morning. Google's own deprecation table lists imagen-4.0-generate-001, imagen-4.0-fast-generate-001 and imagen-4.0-ultra-generate-001 with a shutdown date of 17 August 2026, and the recommended replacement is gemini-3.1-flash-image. The catch is that this is a migration rather than a model swap: the dedicated generate_images() path is gone, so anything calling the old Imagen surface needs rewriting rather than a string change in a config file.
Cost moves too, with the replacement priced higher per image than the Imagen 4 tiers it retires. This is the sort of change that quietly breaks a scheduled job or a client-facing feature days after the fact, so it is worth an explicit check rather than an assumption.
Grok picks up a native Whop connector
This is small on its own and telling in aggregate. Whop is now live as a native connector in Grok, announced by the official SpaceXAI account on the evening of 17 August and confirmed by Whop the same night. Grok's connector list has been growing steadily since the spring, and the direction of travel matters more than any single integration: the assistant is increasingly the place where you act on data rather than a place you paste data into.
For anyone selling through Whop, it means account questions and revenue queries can be handled conversationally without an export step. For everyone else, it is another data point that connector coverage, not raw benchmark scores, is becoming the thing that decides which assistant people actually keep open.
ElevenLabs ships an MCP server that manages voice agents inside Claude
The new ElevenLabs MCP lets a team review agent performance, create new agents, update configurations and estimate LLM costs before changes go live, all from inside Claude rather than the ElevenLabs dashboard. The cost-estimation piece is the interesting part, because it moves a decision that normally happens after the fact into the moment of change. Authentication is handled through OAuth with no server to run and no API keys to manage, which lowers the barrier for non-engineers to touch agent configuration.
If you already run voice or chat agents on ElevenLabs, this is worth ten minutes today. More broadly, it is another sign that MCP has become the default way vendors expose their control planes to assistants rather than building bespoke plugins for each one.
Agentic workloads have turned the CPU into the bottleneck
IEEE Spectrum reported on 17 August that AWS has told engineers to conserve CPU cycles at all costs after wait times for CPU capacity climbed sharply. The underlying point is that an agentic pipeline is mostly not inference: AMD testing cited in the piece puts seven of the eight stages of a realistic agent run entirely on CPU, covering orchestration, tool calls, parsing and retrieval. Intel has sold out of server CPUs through year end and AMD has doubled its forecast.
If you are building agents and have been sizing infrastructure around GPU spend, this reframes the budget. It also helps explain why per-token prices keep falling while the cost of actually running an agent in production does not fall at the same rate.
Industry themes
The interesting movement has shifted from models to plumbing. The two genuine product announcements in the window were both connectors, Whop into Grok and an MCP server from ElevenLabs into Claude, and both surfaced on the companies' own X accounts before any press coverage. Without the social scan, neither would appear in this digest at all.
Availability was the story rather than capability. GitHub and Copilot went down globally and Google's Imagen 4 endpoints were switched off on schedule, and the practical effect on a working day was larger than any benchmark gain would have been. The lesson is the same in both cases: an AI feature is only as reliable as the platform and the endpoints beneath it, so a fallback path is not optional.
The money, meanwhile, moved to the layer underneath everything, with NVIDIA guaranteeing 105 billion dollars of OpenAI's Ohio campus, Stripe reportedly buying the routing layer, and AWS rationing CPU because agents spend most of their time outside the GPU. Taken together, the competitive question this week is less about which model is smartest and more about who controls access, routing and uptime.