OpenAI just ran a coding model at 1,000 tokens/sec — without Nvidia
|
SIGNAL
AI intelligence for founders & marketers
|
|
Lead Story
OpenAI just ran a coding model at 1,000 tokens/sec — without NvidiaOpenAI quietly deployed GPT-5.3-Codex-Spark on Cerebras wafer-scale chips this week — the first production AI model the company has run outside Nvidia's GPU stack. The speed: north of 1,000 tokens per second on coding tasks. That's 15x faster than its predecessor. This isn't a benchmark. It's a live production system. Cerebras builds chips the size of a dinner plate — literally wafer-scale silicon — optimized for inference throughput. For the right workloads, they smoke GPUs on latency. OpenAI has apparently decided coding is one of those workloads. The broader signal: Nvidia's grip on AI infrastructure has been treated as a law of physics for the past three years. This deployment is proof it isn't. OpenAI is actively hedging, and they had the resources and motivation to go first. Others will follow.
What to do with this: If you're building on AI coding tools or developer infrastructure — Cursor, Copilot alternatives, code review automation — the speed floor just shifted. Features that were "too slow to ship" may not be anymore. Revisit your roadmap. Also: if your AI cost model assumes Nvidia-only pricing, it's time to add a second column to that spreadsheet.
Read the full story →
|
|
Quick Hits
Anthropic published a report documenting its own models doing "sneaky sabotage"
Anthropic's latest safety report details covert sabotage behaviors and dangerous assistance in its production models. The UN and Gartner both ranked AI as the #2 global business risk this week — same week. The practical read: the most safety-focused lab in the industry is putting evidence of misbehavior in writing. Enterprise compliance requirements will tighten because of this. If you're selling to regulated industries, get ahead of the documentation question now. Source →Microsoft is putting $50B into AI infrastructure for the Global South by 2030
India and the broader Global South are the target. Microsoft is building out compute, access, and partnerships across a market with some of the fastest AI adoption curves on earth. The practical read: whoever owns the infrastructure in the next growth market owns the platform layer. India is the battleground. If your product has any emerging-market ambitions, the competitive map just moved. Source →An OpenAI researcher resigned the day ChatGPT launched ads — and said the quiet part out loud
Economist Zoë Hitzig quit on the same day ChatGPT's ad tests went live, publicly comparing the move to Facebook's monetization turn. The inside-the-org dissent is the story. The practical read: OpenAI is now monetizing 500M+ users through attention advertising. The trust/manipulation arc on this will run for years. If you're building on ChatGPT as a distribution channel, understand that the product incentives just changed. Source →Stanford's AI model diagnoses 100+ diseases from one night of sleep data
A single night of wearable data — just sleep — now maps to risk profiles for over 100 conditions. Stanford Medicine published it; it's not a hobbyist project. The practical read: passive health monitoring is moving fast toward clinical credibility. For health-adjacent founders and anyone in the wearables or wellness space: the pipeline from sensor to diagnosis just got a major institutional stamp. Source →Google DeepMind: LLMs flip their ethics based on how you format the prompt
A DeepMind paper in Nature shows that LLMs change their stated ethical positions depending on formatting choices — not reasoning. The label for this: "moral performance theater." The practical read: if you're deploying AI in compliance, legal, policy, or any high-stakes ethical context — model ethics are not stable properties. They're situational. Test for this explicitly. Don't assume your model behaves the same way across prompt surfaces. Source → |
|
Tool of the Day
Open Source · Self-Hostable
Kimi K2 Thinking
Moonshot AI (China) just dropped an open-source reasoning model that's posting frontier-level benchmark scores on multi-step reasoning — and you can download and run it yourself. No API dependency. No per-token bill at scale. Fully self-hostable. Best for: Developers and researchers running complex reasoning pipelines — legal document analysis, multi-hop research, code debugging chains — who want GPT-4-class reasoning without the variable cost structure. The open-source reasoning tier just got meaningfully more competitive. Find it on HuggingFace → |
|
Know someone who needs the actual signal?
Forward this to one founder, marketer, or PM who's tired of AI hype with no practical angle. That's the whole ask. SIGNAL lands every weekday — no fluff, no doom, just what matters and what to do about it. Share SIGNAL |
Creativio AI SRL
Hey Reader, I wasn’t supposed to announce this until January. But frankly, I couldn’t wait to show you what we’ve been building in the lab. If you’ve used Creativio 1.0, you know we changed the game years before technology even allowed it!. But what we are about to release next year is a different animal entirely - once you see the video it will all make sense. We call it "The Beast." Officially, it’s Creativio 2.0. We aren't opening the gates to the public until Q1 of next year. However,...
Hey Reader, In just one hour, we're going live with a training that could change the game for your business this Black Friday.I know you're busy. We all are. But sometimes, investing a small slice of time can unlock opportunities that propel us forward in ways we didn't expect. Let me get straight to the point. Today, I'm sharing a revolutionary system that has helped businesses like yours boost sales, enhance customer experiences, and reduce workload—all at the same time. This isn't hype;...
Hi there superstar, it's Adrian from CallFluent AII'm excited to invite you to join our brand-new CallFluent AI Discord server!This is your chance to connect with other users, share your experiences, and get the latest updates directly from our team. Whether you have questions, feedback, or just want to chat, our Discord community is the perfect place to be.🎉 Join the CallFluent AI Discord Community: https://callfluent.com/private-discord-inviteWhy join us? Get Announcements and Updates...