Frontier AI split along two capital paths. OpenAI raised at unprecedented scale to secure compute, while Nvidia and IBM described ways to extract more value from power and smaller models. ARC-AGI-3 supplied the counterweight: immense investment has not closed a basic gap in learning how an unfamiliar environment works through interaction.
1. OpenAI closes a reported $122B round at an $852B valuation
OpenAI completed a financing reported at $122 billion and an $852 billion valuation. TechCrunch reported that retail investors supplied $3 billion, an unusual channel for a private technology round of this size.
The round buys strategic time, not cheap inference. Capital can secure accelerators, power, and long-term capacity while masking the gap between product revenue and the cost of serving it. The retail allocation also spreads exposure to a private valuation whose operating margins remain undisclosed.
Sources: Bloomberg on the financing and valuation · TechCrunch on the retail allocation
2. Nvidia proposes scheduling AI work around grid conditions
Nvidia and energy-sector partners described AI facilities that can lower or shift electricity demand when grids are constrained. Training checkpoints and deferrable batch inference are the workloads most compatible with that model; latency-sensitive serving offers far less room to move.
No sustained operating record accompanied the announcement. The commercial test is whether payments for flexibility exceed the cost of idle hardware and delayed work. That equation links cluster scheduling directly to utility tariffs, service commitments, and the fraction of demand that can actually be deferred.
Sources: NVIDIA on power-flexible AI factories
3. IBM targets document extraction with a 3B vision model
IBM released Granite 4.0 3B Vision as an AI model for document understanding. At three billion parameters, it offers a plausible local or private deployment path for forms, tables, and scanned records that would be expensive or sensitive to send through a frontier API.
Its size changes the deployment boundary, but document mix sets the quality ceiling. A model that succeeds on clean forms can fail on handwriting, rotated scans, nested tables, or organization-specific layouts; those errors flow directly into downstream records and approvals.
Sources: IBM Granite 4.0 3B Vision
4. ARC-AGI-3 finds frontier agents below 1% on interactive tasks
The ARC-AGI-3 preprint introduced interactive environments in which agents explore, infer hidden rules, and act over time. The authors reported that human testers solved every environment, while the evaluated frontier systems earned less than 1% of the benchmark's available score.
The contrast isolates a capability that static question answering can conceal. An agent can generate convincing plans yet fail to update them when actions reveal new rules. Model comparisons on ARC-AGI-3 therefore turn on exploration policy and scaffold design as well as the underlying language model.
Sources: ARC-AGI-3 benchmark preprint