Samwise Tech/AI/Robotics Newsletter
Monday, September 14, 2026
Positron raises $875M at $5B valuation to challenge Nvidia on inference
Startup Positron has raised $875 million at a $5 billion valuation — a fivefold jump from its Series B earlier this year — to build the Asimov ASIC, an inference accelerator on TSMC’s N3P process node. Unlike Nvidia’s H100 and B200, which depend on expensive high-bandwidth memory, Positron’s chip uses commodity LPDDR5X, dramatically lowering cost per token for batch inference workloads. Investors including NEA, Atreides Management, and Valor Capital participated, with Netscape co-founder Jim Clark anchoring a C-1 tranche. Positron claims comparable throughput to the H100 at approximately one-third the power draw and memory cost.
Sources: SiliconANGLE
OpenAI Agents API enters public beta, opens Codex harness to developers
OpenAI has opened its Agents API — the infrastructure underpinning its Codex software-engineering agent — to all developers in public beta. The API manages multi-turn sessions, tool coordination, error recovery, and step-level logging without additional cost beyond standard token pricing. It supports function calling, code execution, file search, and web browsing as first-class tool primitives. Launch partners include Cloudflare, DigitalOcean, and Vercel, all wiring the API into their development pipelines. Developers who previously built agentic workflows on the Assistants API can migrate directly. OpenAI says the Agents API is designed for production workloads requiring reliability, auditability, and consistent performance at scale.
Sources: MarkTechPost
OpenAI puts Pro subscriptions on hold due to GPT-6 Astra demand
OpenAI has temporarily halted new Pro subscriptions ($200/month) due to overwhelming demand for GPT-6 Astra, the company’s latest reasoning-and-tool-use flagship. Existing subscribers are unaffected, but prospective users are now placed on a waitlist. The pause reflects a hardware supply crunch — OpenAI is burning through inference compute faster than anticipated for Astra-tier requests. CEO Sam Altman acknowledged the bottleneck in a post on X, promising to restore signups “within days.” This marks the second time OpenAI has throttled subscription growth to protect service quality, after a similar pause in late 2024. The company did not indicate whether pricing would change upon reopening.
Sources: TechCrunch
YC Demo Day: Nori humanoid at $1,600, Lamb Labs inference chips, Dipole optical
Y Combinator’s Summer 2026 Demo Day surfaced a cluster of hardware bets. Nori Robotics debuted a humanoid priced at $1,600 — deliberately undercut to win market share — capable of walking, climbing stairs, and performing light industrial tasks. Lamb Labs unveiled an inference-optimized chip that promises three times the throughput of an A100 at one-quarter the cost for sub-10B model serving, targeting startups priced out of Nvidia clusters. Dipole Photonics demonstrated an optical interconnect cutting data-centre latency by 40 milliseconds per hop. Garry Tan noted that nearly 30 percent of this cohort is working on physical AI or custom silicon.
Sources: TechCrunch
Maven Robotics exits stealth with $100M Series A for warehouse robots
Maven Robotics has emerged from three years of stealth with a $100 million Series A led by RoboStrategy, alongside LocalGlobe and XTX Markets Ventures. Founded in 2024 by brothers Hamza and Khalid Derbas — Hamza brings autonomy engineering experience from Apple’s Special Projects team — the Santa Clara company builds wheeled, dual-arm robots for mixed-case palletizing and tote handling. Eight deployed units are already logging roughly 16-hour shifts at a Fortune 250 consumer-goods customer with 99-plus-percent uptime. The bots move at up to 10 mph and lift up to 30 kilograms via vacuum suction. Proceeds fund 250 third-generation units and a fourth-generation platform.
Sources: TechCrunch
Cohere releases North Small Translate: 218B MoE model for 50-plus languages
Cohere has released North Small Translate 1.0, a 218-billion-parameter mixture-of-experts model that activates just 25 billion parameters per forward pass and supports more than 50 languages. Licensed for research and non-commercial use under CC BY-NC 4.0, the model scored 83.60 on the WMT26 all-languages benchmark — beating DeepL, Google Translate, and open alternatives including Mistral Large 3 and GLM 5.2. An agentic multi-pass variant that reviews and self-corrects its output reached 84.36. Minimum hardware is one B200 or two H100s at W4A4 quantization. The model was co-developed with RWS, a translation technology firm whose Language Weaver team shaped real-world performance across all 50 languages.
Sources: MarkTechPost
Obama urges Democrats to develop clear plan for AI safeguards
Former President Barack Obama urged Democratic lawmakers to develop a “clear, unified plan” for AI governance during a private meeting with House members, according to multiple attendees. Obama expressed concern that AI development is advancing faster than regulatory frameworks can keep pace, and urged the party not to cede the policy debate to the technology industry. He reportedly endorsed a principles-based approach centred on transparency, liability standards, and sector-specific rules rather than a blanket AI act. The appearance marks Obama’s most direct engagement with AI policy and is expected to influence the Democratic Party’s positions heading into November’s midterm elections.
Sources: TechCrunch
⚡ Tech Pulse
Every newsletter preserved and searchable
Curated by JD · samwise.agency