Cheaper tokens, sovereign compute, and the human work behind agents

AI compute moves off-premises and into public plans

The infrastructure under AI is increasingly owned by someone other than the companies using it. Network World reports that most corporate IT now runs outside company-owned data centers, with AI workloads cited as a major driver of infrastructure strategy and uptime expectations. Hyperscalers are building for that demand. Google Cloud expanded storage and networking for large AI training and inference clusters, targeting throughput and latency for big GPU and TPU fleets. Europe wants an alternative to US and Chinese providers, and the EU plans seven AI gigafactories backed by $11.4B to build sovereign capacity for training and serving frontier models.

Vendors are adjusting their leadership as well. Scale AI hired former Google Cloud COO Francis deSouza as chief executive. Putting a cloud operations executive in charge of a data labeling, evaluation and infrastructure services company may point to a stronger push toward large enterprise and government buyers, though the hire itself is the only concrete change so far.

What running agents actually costs and requires

GPT-5.6's Luna tier cut prices 80% to $0.20 per million input tokens and $1.20 per million output, and the Terra tier fell 20%. The same analysis argues that aggregate enterprise spend keeps rising because lower prices make high-volume agentic workloads viable, so unit price is a poor predictor of the bill. Price is not the only way engineers now choose models. One developer explains picking models mostly on speed rather than intelligence, arguing that for everyday coding and agent tasks the capability gap between top models matters less than round-trip latency. Both pieces suggest that call volume, more than the price or score of a single call, may shape how teams evaluate models.

Other writers place the bottleneck outside the model entirely. One essay holds that platform engineering matters more as agents generate more code, because paved roads, guardrails and consistent deploy paths are what let a flood of generated changes land safely. QM takes a narrower approach to collaboration: it is a multiplayer agent harness that keeps shared team memory without trying to model the whole company. At the organizational level, another essay proposes forward deployed executives embedded with customers to clear political and organizational blockers to AI rollouts, framing this as a possible billion-dollar services category.

Watermarks, video models and who gets paid

Ars Technica's testing found Google's SynthID watermark hard to strip or spoof, but the piece argues that watermarking cannot solve disinformation on its own, since bad actors can use unwatermarked tools or strip metadata and most people never check. Generation tools are also becoming more controllable: ByteDance's Seed team introduced Seedance 2.5 with one-take creation and flexible referencing, aimed at keeping characters, scenes and style consistent across shots.

Sources of human-written content are questioning the terms of AI summarization. As Reddit's stock fell, its CEO questioned the value of Google's AI Overviews for content sources, saying the company is 'still looking for that win-win.'

Memory savings and devirtualized bytecode

Two engineering write-ups stand out for technique rather than news. The Curiosity project cut its heap size by 60%, and the post walks through the allocation patterns and data-structure choices behind the reduction, a useful case study for long-running services under memory pressure. Separately, back.engineering details static devirtualization of Tencent's VM-based obfuscation, recovering virtualized bytecode for analysis without executing it. It is a dense reference for reverse engineering, anti-tamper and binary analysis work.