EAIDaily — 2026-08-26

Focus: AI Coding · Embodied Intelligence · Frontier Models Date: 2026-08-26 (Wednesday) Coverage: Past 24 hours of high-signal AI news Editor: WoLoveAI @ AI News Daily

Today’s Headline

NVIDIA’s Vera Rubin first silicon test + Skild AI’s S1 + SSI’s potential TTT breakthrough converge on a single thesis: the Agent era has formally arrived, and the infrastructure, models, and robots built for it are now emerging together. Hardware is being co-designed around agent workloads, foundation models are shifting from pretraining to test-time adaptation, and robots are learning from a single video.

🔥 Top Stories (8 Items)

1. NVIDIA Vera Rubin NVL72 First Silicon: 30× Throughput, 35× Lower Token Cost

At Hot Chips 2026, NVIDIA unveiled the first on-chip benchmark for its next-gen Vera Rubin NVL72 rack, running DeepSeek-V4-Pro (1.6T params) on real agentic coding workloads. The result: up to 30× higher throughput per megawatt vs. GB300 NVL72, and up to 35× lower cost per million tokens. The gains come from rack-scale co-design — disaggregated prefill/decode, distributed KV cache, NV-aware routing, NVFP4 quantization, and 6th-gen NVLink (10× faster, 3× lower latency than Ethernet).

Why it matters: NVIDIA is formally declaring that traditional 8K/1K-length benchmarks are dead — an Agent’s token consumption is 15× a chat’s. This defines the unit economics for the next 12–24 months: a per-million-token cost cliff means 24/7 “digital employees” become economically viable. The Agent infrastructure layer is now officially a moat.

2. NVIDIA Ships First Agent-Specific CPU (Vera) + Groq 3 LPX Mass Production

Two chip launches in one day. Vera CPU (88 Olympus cores, 1.2TB/s LPDDR5X bandwidth) is positioned as the first CPU built for agent orchestration, and SpaceXAI is already deploying it at scale and planning to take a Vera Rubin NVL72 rack to orbit by 2028. Groq 3 LPX (the $20B Groq acquisition, now in full production) hits 3,400 tokens/s on Gemma 4 31B with a 100K-token context — coding agents that took hours now finish in minutes. Nebius has deployed it; a 5,000-token decode task drops from 50s → 1.5s (34× speedup).

Why it matters: NVIDIA is no longer selling GPUs — it’s selling “AI factories.” The Agent era requires new silicon at every layer (CPU, GPU, inference accelerator, interconnect). Whoever controls the rack-scale stack controls the next decade of agent deployment.

3. Skild AI Ships S1: One Video Demo → 10-Minute Robot Task, No Fine-Tuning

Skild AI released S1, a robotics foundation model that watches a single video demonstration and executes multi-step manipulation tasks up to 10 minutes long — potting plants, pour-over coffee, pancakes, mechanical kit assembly — with no weight updates. On Skild’s unseen-task benchmark, in-context S1 reached 66% per-step success at 100K pretraining hours vs. 9% for language-prompted VLA baselines. One video ≈ 380 post-training teleop demos in performance equivalence.

Why it matters: This represents the first serious shift of robot foundation models out of the “BERT era” (heavy pretraining + task-specific post-training). Robots can now be deployed in ~11 minutes from demo to execution, collapsing the bottleneck on long-tail industrial and home tasks. Skild has raised $1.7B (valuation >$14B, SoftBank-led) — this is a direct competitor to Physical Intelligence (π), Figure, and 1X.

4. Rumor Mill: Ilya Sutskever’s SSI Teases First Model — “Test-Time Training” Breakthrough Imminent

a16z partner Martin Casado posted that he got access to a “mind-blown” new model, calling it potentially “the most significant drop this year.” Within 12 hours, multiple sources (Andrew Curran, Dan McAteer, Milad Khademi Nori) pointed at Ilya Sutskever’s Safe Superintelligence (SSI) — a company with zero public products but a $32B valuation and a $5B NVIDIA strategic investment (announced July 27). Gavin Baker confirmed in early August that SSI planned to release in August. The rumored breakthrough: Test-Time Training (TTT) — instead of stuffing context into a frozen-weight model’s “cheat sheet,” the model actually performs gradient updates while reading, turning deployment into continual learning.

Why it matters: If TTT works, it rewrites the moat: pretraining compute, data-center scale, million-token context — all become secondary. A small model that learns on the job could beat frontier giants. SSI’s “15-year-old genius” thesis aligns with TTT. This is the most consequential rumored release of 2026.

5. Anthropic Unifies Memory Across Claude Chat + Claude Cowork (First Real Cross-Surface Continuation)

Claude Chat and the desktop agent Cowork now share a single, bi-directional memory store. Previously, each product had its own memory; now, what you discuss in chat flows into Cowork tasks, and what Cowork learns flows back. Memory updates in real time during conversations (not just at the end). Sensitive topics (health, beliefs, identity) are off by default, with explicit opt-in. Available today on Free/Pro/Max across web/desktop/mobile.

Why it matters: This is the first major lab to ship true cross-surface agent continuity. Heavy users no longer have to re-explain project context, manager preferences, or client history when switching from chat to Cowork. Claude Code’s memory remains separate (for now). It pushes the industry toward the “one assistant across surfaces” model — a direct counter to Perplexity’s local-first agent strategy on the same day.

6. Perplexity × NVIDIA “Portable Computer”: Fully Local AI Agent, Zero Token Cost

Perplexity launched Portable Computer, a local-first AI agent co-designed with NVIDIA for DGX Spark (and soon Windows RTX machines with ≥24GB VRAM). It runs the entire agent harness, planner, tool router, search index, and a post-trained Qwen 3.8 27B locally — with zero credit consumption for on-device work. When the local model hits its limit, it asks permission before escalating to cloud frontier models. Initial skills: research, data science, coding; connectors: GitHub, Gmail, Google Drive, Slack. NVIDIA Nemotron 3.5 Lightning (30B) coming soon.

Why it matters: NVIDIA publicly declared: “Local AI reached an inflection point.” This is the chipmaker betting its consumer hardware narrative against the trillion-dollar data-center narrative. For developers, it means predictable cost and data sovereignty. For NVIDIA, it’s an entirely new SKU category — desktop AI workstations for the post-cloud era.

7. Harvey Tenet: $11B Legal AI Unicorn Ditches Closed Frontier Models for Chinese Kimi K3

Harvey (the largest legal AI unicorn, $11B valuation, $350M ARR, 2400 firms/200K lawyers) released Tenet, its first proprietary model post-trained end-to-end on Kimi K3 (Moonshot AI) using 134 NVIDIA B300 GPUs over two months. Tenet operates at <1/4 the cost of leading foundation models while hitting state-of-the-art on LAB Contracts. AI policy researcher Simon Hedlin called it “a textbook example of open-weight value” — adding that “the US has fallen behind in frontier open-weight models.”

Why it matters: A Western legal-AI unicorn — previously all-OpenAI/all-Anthropic — is now built on Chinese open-weight infrastructure. Cursor, Devin, Cosine, and Thinking Machines have all adopted Kimi or DeepSeek. The Chinese open-weight stack is no longer just competitive — it has become the default substrate for US vertical-AI products that need full model control and low per-token cost.

8. China Humanoid Robots: 8.86s 100m Record + Games Grow 138% — But “Walk→Work” Gap Emerges

At the 2nd World Humanoid Robot Games in Beijing, Tiangong (Beijing Humanoid Robot Innovation Center) clocked 8.86s in the 100m (vs. Bolt’s 9.58s human record); 400m at 38.15s; 1500m at 2:21.64; high jump 2.88m; long jump 7.97m. The Games fielded 666 teams from 16 countries / 6 continents, 2056 robots (+138% YoY), 51 events (was 26). Yet the Industry Minister’s framework now calls for 100+ standards by 2028, and Morgan Stanley notes ~65% of shipments remain concentrated in entertainment/education/data capture — not real factory work. Unitree (王兴兴) just set the industry consensus threshold: when robots can complete 80% of daily tasks autonomously in 80% of unfamiliar scenes via voice — that’s the “ChatGPT moment.”

Why it matters: The athletic performance is real, but the “Walk vs. Work” gap is now the defining frontier. Investors are looking past locomotion demos toward物流/工业 场景 where scale actually happens. Skild’s S1 (above) and the雄安 trained-robot factory are early signals of where the next $10B will be deployed.

📊 Cross-Cutting Themes

🎯 Key Numbers to Remember

• NVIDIA Vera Rubin NVL72: 30×/MW throughput, 35×/token cost reduction vs GB300
• Groq 3 LPX: 3,400 tokens/s on Gemma 4 31B, 100K context
• Vera CPU: 88 Olympus cores, 1.2TB/s bandwidth, SpaceXAI deploying to orbit by 2028
• Skild AI S1: 66% step success on unseen tasks at 100K hours, single-video → 11-min deployment
• Skild funding: $1.7B raised, $14B valuation
• Harvey Tenet: <1/4 cost of frontier, 2 months training on 134 B300s, LAB Contracts SOTA
• Tiangong 100m: 8.86s (-0.72s vs Bolt)
• 2nd Humanoid Robot Games: 666 teams / 16 countries / 2056 robots (+138% YoY)
• China humanoid H1 2026: 97% of global ~19K shipments
• NVIDIA → SSI: $5B strategic investment (July 27)

Sources: SemiAnalysis AgentX benchmark, NVIDIA dev blog, Perplexity research blog, Anthropic official, Skild AI technical report, Harvey engineering blog, CCTV, IT之家, 新智元, 36氪, 21世纪经济报道

— @WoLoveAI | Tracking AI Coding × Embodied Intelligence · 2026-08-26