The week's big AI-infrastructure money went to compute and connectivity, not models
No frontier model shipped over the weekend — the model tracker showed zero releases in the last 24 hours. The money, though, kept moving, and it moved to the layer underneath the models. Two companies came out of stealth this week with rounds aimed squarely at AI infrastructure: Volta, an AI-cloud-compute startup, raised a Series A at a $2.4 billion valuation led by Azora, Andreessen Horowitz, Altimeter and Nvidia, and Lumilens, a connectivity platform for AI infrastructure, emerged with more than $700 million, per Crunchbase's weekly tally. Volta pitched the round as building “the utility of compute” — capacity you rent by the unit rather than a model you call.
The pattern matters more than any single round. As models converge on quality and per-token prices keep falling, the scarce, capital-hungry part of the stack is the compute fabric and the wiring around it — and it is being built by a widening set of players outside the three big US clouds. For a European buyer, that reframes the sovereignty question: it is less “which model” and increasingly “whose compute, whose network, and whose grid is my AI actually running on.” Neutral compute layers are useful only if you can point them at infrastructure that sits under the right jurisdiction.
Enterprises are buying tools to watch and govern their AI agents
The other place money and product went this week was control. Microsoft put Project Perception, its agent-based cybersecurity platform, into public preview on 3 August — a system that runs “red,” “blue” and “green” AI agents to find, investigate and remediate risks, alongside a smaller in-house cyber model, MAI-Cyber-1-Flash. Days earlier, Okta signed a definitive agreement to buy Permiso Security for a reported ~$200 million, adding threat detection across human, machine and agentic identities. AI-agent security took roughly a quarter of all cybersecurity seed deals in Q2, the largest single category.
The through-line for regulated buyers is simple: before an agent can touch a core system, someone has to be able to say what it did, under whose identity, and stop it when it goes wrong. Governing what an agent can see and do is becoming its own product layer — and, like the compute layer, that governance is only as sovereign as the place the logs and controls actually live.
“Announced open” still isn't “downloadable”
The open-weight field remains the clearest route to running frontier-class AI on your own hardware — but this week was a reminder to check the artifact, not the press release. Alibaba's Qwen3.8-Max, a 2.4-trillion-parameter model billed as the first of its Max tier slated for open weights, was promised on Hugging Face and ModelScope the week of 10 August; as of this writing the Qwen organisation page shows no such checkpoint, and the model is still API-only. The counter-example is Ant Group's inclusionAI, whose MIT-licensed Ling 3.0 family keeps actually shipping — Flash's open weights landed last week, and the small Ling 3.0 Tiny reached OpenRouter on 6 August. For a team planning to self-host, the useful test is whether the weights are on Hugging Face today, not whether a launch post says they will be.
Quick Hits
- The power bill is getting funded too. — Alongside the compute rounds, two energy plays tied to AI-driven demand closed $1 billion each this week — Austin's Base Power ($1B Series D at a $13B valuation, home battery storage) and nuclear developer Valar Atomics ($1B Series B led by Sequoia), per Crunchbase. Compute is only half the constraint; electricity is the other.
- Autonomous pentesting keeps raising. — Cybersecurity firm Horizon3 — whose platform runs continuous automated penetration tests — closed a $250 million Series E led by NightDragon and NEA at a valuation above $2 billion, roughly triple last year's valuation, another sign that AI-era security is where enterprise budgets are moving.