Nvidia co-signs OpenAI's $105B Ohio bet

The AI capital machine dominated the day: Nvidia agreed to backstop up to $105B for OpenAI's record Ohio data center, Anthropic's run rate leapt past $65B ahead of its IPO, and Groq raised at half its old valuation. Meanwhile an AirTag exposed Amazon shredding rare books for training data, and open weights kept graduating from benchmark toys to real build tools.

Nvidia backstops $105B for OpenAI's record Ohio data center

OpenAI signed a 20-year lease for the PORTS-Pike campus in Pike County, Ohio, built and owned by SoftBank's SB Energy on a decommissioned uranium-enrichment site. Nvidia becomes the exclusive chip supplier and guarantees up to $105B of the finished facilities' residual value on the first 4.25 IT-GW (of 8 IT-GW total, backed by ~10 GW of new gas generation), plus a $1.5B stake in SB Energy; first 800 MW is slated for 2028. The Wall Street Journal notes nine tech firms now carry roughly $3 trillion in mostly-AI commitments off their balance sheets.

Why it matters: Nvidia is now simultaneously OpenAI's supplier, investor and loan guarantor — Jensen Huang insists 'OpenAI will pay the lease,' but the structure is the clearest test yet of whether AI's circular financing holds up if demand doesn't fill the racks.

AirTag traces Amazon's bulk rare-book buys to a book-shredding AI scanning lab

404 Media convinced a bookseller to plant an Apple AirTag in a ~1,000-book bulk order; it ended up at Amazon's VGT3 team inside the LAS8 facility in Las Vegas, whose door logo is a T. rex devouring a book. Workers there cut spines off books to speed destructive scanning, and Amazon uses the pages to train its Nova models. Amazon's statement said only that it 'purchases books through commercial channels'; the practice mirrors Anthropic's court-revealed 'Project Panama.'

Why it matters: Pre-2022 printed text is now a scarce, contamination-free training asset worth destroying originals for — the data land grab has moved from scraping the web to physically shredding the archive.

Anthropic's run rate tops $65B, passing OpenAI ahead of IPO

Anthropic's annualized revenue run rate surpassed $65B at the end of July, up from $47B in May and just $9B at the end of 2025, per Bloomberg; Q2 preliminary revenue hit $11.5B, more than double Q1. Investors expect 2026 to close between $100B and $120B, with projections of $190-200B for 2028. Both Anthropic and OpenAI have filed confidential IPO paperwork, and Anthropic — last valued at $965B — is expected to list first, possibly this fall, seeking $2T+.

Why it matters: Anthropic's enterprise-and-coding bet has pulled it ahead of OpenAI's reported $40B run rate; token efficiency (cost per completed task, not per token) is shaping up as the next margin battleground.

Groq raises $350M at $3.5B, half its old valuation, and leans into Nvidia clouds

Groq raised $350M led by Disruptive, with planned Nvidia participation, at a $3.5B valuation — down from $6.9B last September, after Nvidia hired founder Jonathan Ross and top talent in a $20B licensing deal. The company insists it isn't a down round but a reset for the 'post-Nvidia-licensing-deal' Groq, which has pivoted from building its own LPU inference chips to operating Nvidia systems as a neocloud. It now runs 13 data centers and plans to scale from 54 MW to 200+ MW in 2027.

Why it matters: A one-time custom-silicon challenger now reselling Nvidia GPUs is a blunt signal about how hard it is to compete on inference chips — and neocloud economics (capex, debt, fast-depreciating hardware) remain unproven.

Independent 'AI Observatory' says labs' usage reports hide the messy half

A Stanford/MIT-led project aggregated 24,521 consented conversations (85,633 turns, 52 models, 2023-2025) to independently check how people actually use chatbots. Applying Anthropic's Economic Index methodology dropped 48% of conversations; those filtered-out chats were far more likely to involve health and relationships, adult or illicit topics, harassment (27.5% vs 5.7%), and sexual content. Usage also varied sharply by model — Grok for news and misinformation, Gemini for roleplay, Claude for coding, ChatGPT for homework.

Why it matters: Policymakers lean on vendor-curated usage reports with no external corroboration; this is a first attempt at an independent ground truth, and it suggests the sanctioned narratives skew heavily toward work.

Tencent open-sources UI-Mate-27B, an Apache-2.0 desktop GUI agent

UI-Mate-27B, built on Qwen3.6-27B, observes live screenshots and emits structured mouse/keyboard actions for native desktop control, in both general computer-use and demonstration-guided modes that re-plan from the live screen rather than replaying coordinates. It was trained with SFT then online RL in executable GUI environments, reports strong Ubuntu/Windows benchmarks, and ships pyautogui-compatible actions with OpenAI-compatible serving. Tencent also released EVIE-Preview-4.5B, a compact ColBERT-style visual-document retrieval model.

Why it matters: Computer-use agents have mostly been closed API demos; an Apache-2.0 27B with weights lets developers run and fine-tune desktop automation locally instead of renting it.

Qwen 3.8 27B graduates to real build tool, as Qwen damps 35B-A3B hopes

Days after release, local users are running Qwen 3.8 27B through full long-horizon jobs: one reported an 8-hour, 131M-token agentic project with zero generation failures, priced at $0 locally versus an estimated ~$677 on Claude Opus 4.6. Others published tuned llama.cpp configs fitting 73k context in 16GB VRAM via aggressive quantization and native MTP speculative decoding, while Empero distilled the flagship down to 9B/4B/2B checkpoints. A Qwen developer, meanwhile, told the community not to wait for a 35B-A3B MoE.

Why it matters: The story has shifted from 'good benchmarks' to 'cheap, reliable long-horizon coding on consumer hardware' — but the roadmap signal suggests the much-requested sparse MoE variant may not be coming.

AWS wires OpenClaw agents to pay HTTP 402 paywalls with x402 stablecoin rails

A joint AWS/OpenClaw walkthrough connects agents to Amazon Bedrock AgentCore payments via the aws-agents-pay plugin, letting them settle sub-cent USDC payments for paid APIs, content and MCP tools within human-approved limits. The design keeps wallet credentials and session-creation authority outside the model-facing runtime, assumes prompt injection is possible, and bounds spend by recipient, asset, network, per-payment ceiling, cumulative budget and expiry; it supports x402 and Machine Payments Protocol on Base and other EVM chains plus Solana.

Why it matters: Agentic micropayments are moving from spec to shipping product, and the security model — bound the runtime's authority, treat all paid content as untrusted — is the interesting part for anyone building autonomous agents that spend money.

Also worth a look

Browse previous days →