AI Sector Daily Digest: July 31, 2026 — MiniMax H3, OpenAI price cuts, and Europe’s compute plan

In brief

1. MiniMax releases H3 for multimodal video generation

• MiniMax released H3, a model that can take text, images, video, and audio as inputs and generate clips up to 15 seconds long in 2K resolution with native stereo sound. It can also edit existing content and transfer movement from one video to another. 1
• The company said it plans to release H3's model weights within days. At publication, that was a planned follow-up rather than an already available download; if delivered, it would let developers customize the underlying system. 1
• MiniMax said 2K generation would cost less than one-third of mainstream rivals and that H3 was designed to run on several Chinese-made chips. For video teams, the important variables are now availability, output quality, and inference cost together. 1

2. Nscale agrees to buy Anyscale and its Ray-based software platform

• Cloud infrastructure provider Nscale agreed to acquire Anyscale, which provides software for running distributed AI workloads across many computers. The companies did not disclose financial terms; Reuters reported that Bloomberg put the price at about $1.65 billion. 2
• Anyscale's platform covers data processing, training, inference, and reinforcement learning across thousands of GPUs. Its roughly 200-person team will join Nscale, while Anyscale will keep its brand and existing customer operations. 3
• The deal is subject to regulatory approvals and is expected to close in the second half of 2026. Nscale is adding a software layer to its own GPUs, data centers, and power supply, a sign that AI-cloud competition is moving beyond renting chips toward bundling the stack that runs workloads. 23

3. OpenAI cuts prices on smaller models

• OpenAI cut the price of its smaller GPT-5.6 Luna model by 80% and its mid-tier Terra model by 20%. The flagship Sol model's price was unchanged. 4
• Luna input fell from $1 to $0.20 per million tokens, and output from $6 to $1.20. Terra input fell from $2.50 to $2, and output from $15 to $12 per million tokens. 4
• The cuts make lower-cost models more competitive with Anthropic and Chinese providers, but they do not eliminate budget uncertainty: Reuters notes that usage-based pricing can still produce higher total bills when models take on more work per task. 4

4. EU opens a call for up to seven AI Gigafactories

• The European Commission opened a call for tenders to establish up to seven AI Gigafactories. The program is backed by up to €10 billion in EU and national funding and is expected to unlock at least €20 billion in private investment, according to the Commission. 5
• The planned sites are intended to give start-ups, scale-ups, small and medium-sized businesses, industry, academia, and public authorities access to infrastructure for training, inference, and fine-tuning advanced models. 5
• This is a tender, not a completed buildout: the Commission says the facilities would combine advanced processors, software, cloud stacks, connectivity, and energy-efficient data centers while operating under EU data-protection, safety, security, and ethics standards. 5

5. Echoverse reports a large training lift for computer-use agents

• A new arXiv preprint, Echoverse: Deep, Evolving Environments for Training Computer-Use Agents at Scale, builds stateful applications whose tasks are graded against each application's own database. A co-evolution loop then uses rollout failures to repair the environment, its tasks, and its verifier. 6
• In the authors' reported experiment, a 9B-parameter model trained on 12 environments improved from 36.5% to 67.1% across 14 evaluation splits, coming within 14 percentage points of the much larger frontier model used as its teacher. 6
• The result is a preprint, not a settled benchmark consensus. Its practical claim is narrower and useful: shallow synthetic environments can lower live-site accuracy, while deeper environments and grounded graders can improve transfer; the authors released four environments as a benchmark. 6

The read-through