NVIDIA and OpenAI's $1.5B Ohio AI Compute Campus
NVIDIA is putting $1.5 billion into SB Energy to build the PORTS-Pike Technology Campus in Pike County, Ohio — a data center on the site of a decommissioned uranium-enrichment plant — and will exclusively supply it with AI compute, with OpenAI as the anchor customer on a 20-year lease. The first phase brings 4.25 gigawatts of IT capacity online, scaling toward 8 GW, backed by at least $4.2 billion in new grid infrastructure and a co-investment from SoftBank Group. It's one of the largest single AI-compute buildouts announced this year and a concrete data point in the compute-scarcity race driving frontier-lab spending; construction starts in 2028.
Sources & depth
Teaching Everyone to Fish for Tokens
Nathan Lambert argues NVIDIA's roughly $26 billion in open-model investment isn't charity — it's a bet that if enough companies can build and run their own models, NVIDIA sells more chips than it would if a handful of closed labs monopolized inference. The piece frames deals like the Ohio campus above as the compute side of the same "teach a person to fish" strategy: NVIDIA profits whether or not any single model wins, as long as token generation keeps growing. The risk it flags is that the strategy depends on open-model builders staying profitable enough to keep training — if the economics don't hold, the ecosystem could contract back toward narrow, efficiency-focused models instead of frontier ones.
Sources & depth
GPT-5.6 Sol Pricing Cut by 50%
OpenAI has cut GPT-5.6 Sol's list price in half across every provider — input tokens from $5 to $2.50 per million, output from $30 to $15 — continuing a month of aggressive repricing across the GPT-5.6 line (Luna already cut 80%, Terra 20%). Sol is OpenAI's flagship for agentic coding and long-horizon tasks, with a 1M-token context window, and the cut lands as open-weight models like Qwen3.8 (below) close the quality gap at a fraction of the cost. It reads as OpenAI pricing to defend usage share rather than margin.
Sources & depth
Qwen3.8 27B Scores 52 on Artificial Analysis, #1 in Its Class
Alibaba's dense Qwen3.8 27B variant scored 52 on the Artificial Analysis Intelligence Index — ranked #1 among the 135 models tracked in its size class, well above the ~9 median for comparable models — with a 256K context window, image input, and an Apache 2.0 license that makes it freely self-hostable. It's noticeably more verbose than peers at inference time (160M output tokens across the benchmark suite vs. a 43M median), which matters for anyone budgeting compute against it, but it slots alongside last month's 2.4T/95B-active Qwen3.8 Max as a genuinely competitive open weight-class leader.
Sources & depth
Amazon Is Destroying Rare Books to Train AI
404 Media planted a tracking device in a rare book and followed it to an Amazon warehouse in Las Vegas run by a team called VGT3, where staff cut the bindings off books to scan them faster — destroying the physical book in the process — as part of an operation the investigation ties to AI training-data sourcing. Ars Technica and TechCrunch independently picked up the same shipment, adding corroboration to a concrete, trackable data point in the broader fight over where AI training data actually comes from and what gets destroyed to get it. Amazon hasn't issued a public response to the specific claims.
Sources & depth