1. Nvidia in Talks to Back Perplexity at a $30B+ Valuation
Nvidia is reportedly negotiating to join a new Perplexity financing round that would value the search startup north of $30 billion — more than a 50% jump from the roughly $20 billion mark it hit about a year ago. Perplexity's annualized revenue has climbed past $750 million, up from under $250 million at the start of 2026, growth partly credited to Perplexity Computer, its cloud agent for automating desktop work. The talks reportedly also touched on a technology licensing deal.
2. OpenAI's First Custom Chip, Jalapeño, Undercuts Nvidia on Power
SemiAnalysis published a deep dive on Jalapeño, OpenAI's first in-house inference accelerator, built with Broadcom on TSMC's N3P process and taped out in just 16 months. The part is reported at 13.4 PFLOPs in a 700W envelope, against the 900–1,150W Nvidia's Rubin draws. It is the clearest signal yet that frontier labs intend to own their inference silicon rather than rent it.
3. Nvidia Raises AI System Prices 15%+ — and Finances a Quarter of Its Own Demand
Nvidia's server builders have told major hyperscalers that flagship Vera Rubin and Grace Blackwell configurations get a 15%-plus price increase from early 2027, driven by DRAM costs the company can no longer absorb. Separately, reporting this week estimates that roughly a quarter of Nvidia's business next year will come from AI labs it is itself financing — sharpening the circular-revenue question hanging over the buildout.
4. Chinese Labs Ship Three Fast Open Models in a Week
The open-weight release cadence out of China did not slow in late August. Zhipu AI (Z.ai) put out GLM-5.3-Flash and Alibaba's Qwen team shipped Qwen3.8-Flash-Next, both on August 26, while DeepSeek released the vision-capable DeepSeek-V4-Flash-Vision-Exp on August 21. All three target the low-latency, low-cost tier rather than raw frontier benchmark scores — the segment where open weights hurt closed APIs most.
5. Skild AI's S1 Learns a 10-Minute Task From One Human Video
Skild AI released S1, a robotics foundation model that executes roughly 10-minute tasks after watching a single human demonstration video, with no fine-tuning step. The company reports 66% success on unseen tasks versus 9% for competing systems. If the numbers hold up outside the lab, one-shot imitation removes the biggest cost line in robot deployment: per-task data collection.
6. XPENG's IRON Humanoid Pulls Record Physical-AI Funding
XPENG's robotics unit raised over $900 million at a $6.3 billion valuation to scale its IRON humanoid platform — a record round for the physical-AI category. The raise lands the same week as Skild's S1 and Nvidia's new Jetson Orin Nano 2 edge module for drones and robots, three signs that capital is rotating from chat interfaces toward embodiment.
7. Apple Ships Its First 2nm Chip and a Quad-Die M5 Ultra
Apple introduced the M6 built on 2-nanometer process technology with dual 16-core Neural Engines, alongside the quad-die M5 Ultra delivering 4.5x the AI compute of the M3 Ultra. The push keeps Apple's bet on local inference intact: large unified memory and on-package NPU capacity aimed at running models on the machine rather than in a datacenter.
8. Taiwan Indicts Nine Over Nvidia Server Smuggling to China
Taiwanese prosecutors charged nine individuals with illegally routing 74 Nvidia B300 servers into China via trans-shipment, in violation of US export controls. It is one of the larger enforcement actions to date on GPU diversion, and a reminder that compute controls are now being litigated criminally, not just administratively.
9. MIT Model Forecasts Extreme Weather Without Historical Disaster Data
MIT researchers published a method for predicting extreme weather events that does not rely on historical records of those events — addressing the core problem in disaster forecasting, which is that the rarest and most damaging events have the thinnest training data. The approach matters beyond climate: it is the same data-scarcity problem that blocks ML in fraud, failure prediction and rare disease.
10. AWS Shuts Down Mechanical Turk After Two Decades
Amazon told users it will close Mechanical Turk on September 30, 2026, ending the human-task marketplace launched in 2005 that trained a generation of ML datasets. The shutdown is a quiet milestone: the crowdsourced-human-labeling era that bootstrapped modern AI is being retired by the systems it built.
// KEY TAKEAWAYS
Three themes ran through this week. First, the compute layer is repricing itself — Nvidia raising system prices 15%+ while financing a quarter of its own forward demand, OpenAI taping out its own inference silicon, and Apple moving to 2nm all point at the same squeeze. Second, capital has rotated hard into embodiment: XPENG's $900M humanoid round and Skild's one-shot S1 model landed within days of each other. Third, the open-weight tier keeps commoditizing from below, with three fast Chinese models shipping inside a single week while export-control enforcement turns criminal in Taiwan.