Media Zone | 2026-06-06
Social and video spent the day on the money: compute scarcity to 2031, a government stake in OpenAI, and the coding-cost reckoning.
Today's signal
- Dominant story: AI infrastructure is now sovereign finance, government stake, $920M/month leases, memory shortage to 2031.
- Pattern: video and tweets converge on agentic velocity, 3,000 commits/day, doubled Cowork limits, Copilot bills.
- Counter-signal: SABER-style worry surfaces socially via Schneier's self-carrying AI worm.
- Quiet area: no curated reposts, no Reddit signal, no KV-cache practitioner chatter today.
- Video leans tooling (OpenClaw, Chrome DevTools MCP, Paxel); Google's CVPR booth carries the vision news.
Routing, KV cache, compression, GPU
Compute and memory scarcity is the binding constraint
- NVIDIA + SK hynix: HBM memory crisis runs until 2031.
- TSMC confirms it secured ASML High-NA EUV tools, R&D for now.
- Text diffusion pitched as 10x lower latency, compute-bound not memory-bound.
- Nemotron 3 Ultra (550B open weights) already topping Kilo's OpenRouter leaderboard.
LLMs, agents, safety
Agentic velocity and the coding-cost reckoning
- OpenClaw maintainers claim 3,000 commits/day via parallel agent "swim lanes."
- Engineers reframed as "factory managers"; the bottleneck is taste, not code.
- Anthropic doubled Claude Cowork limits for a month.
- Copilot's June-1 usage billing bit; Uber burned its 2026 budget by April.
Building for agents as a distinct user
- Chrome DevTools MCP: optimize for "tokens per successful outcome," not raw output.
- Semantic summaries and self-healing errors beat dumping raw trace files.
- Schneier flags a prototype AI worm that carries its own LLM.
- Test-time compute let a model disprove an 80-year-old Erdős conjecture.
Multimodal / vision / audio
Google's CVPR booth and the on-device push
- BlazeEdit: 195M-param on-device image editor, full pass in 290ms on Pixel 10.
- D4RT (CVPR Best Paper): one transformer for 4D depth + tracking + camera.
- TIPSv2 finding: distillation beats standard pretraining for patch-text alignment.
- Reve 2.0 pitched as the best 4K image model, layout-based editing.
Industry and business
The AI build-out becomes sovereign finance
- Trump: "not that far apart" from Bernie's 50% public-private AI proposal.
- OpenAI negotiating a direct US government equity stake.
- SpaceX leasing Google 110,000 Nvidia chips at $920M/month.
- Alibaba's Qwen3.7-Plus: proprietary multimodal agent, priced below Western frontier models.