[AINews] Poolside gets $12B reverse-execuhire to NVIDIA; founders stay for $1B, employees go for $6B, Infraco scaling to 7GW neocloud

Less than a month ago we had just featured Poolside’s Model Factory with Eiso Kant on the pod (following our Paper Club coverage): It appears that Jensen really, really liked Poolside too, as he went from investor to doing licensing their factory and hiring 109 of their employees: Unless things changed drastically, this accounts for the overwhelming majority of the technical Poolside employees: Eiso Kant [01:52:31]: We are hiring on every possible role in applied research and engineering in the company, from training all the way to evals to post-training architecture. Like, we are still in a world where, individuals can have massive impact. And I think our pitch to join us —I think we are one of the places where it’s the highest ratio to individual to impact, Right? Less than 70 people built this model. Less than 115 between engineering and researchers, like, together did this effort, and that’s a very broad definition ‘cause I put myself in the 115 list. As the founders say, this is “not an acquisition and not an acquihire”: We’ve been calling the Windsurf-Google and Character-Google and Scale-Meta and Instacart-OpenAI deals execuhires because usually the executives go leaving the employees with a rich payout but holding the company remaining, but this is a first time it is happening the other way around. The action amounts to founders pivoting the company extremely hard to SOMETHING, and finding an EXTREMELY comfortable golden parachute for investors and employees to continue on with the original mission or stay aboard for the new pivot: For the last 3 1/2 years we’ve been directionally correct in a race where capital requirements went vertical. At the end of last year, we had a 6 week window in which to raise $2 billion dollars to pay for a 40,000 GB300 cluster coming online in January. We didn’t close it in time, and we lost the cluster. and: We also know that at 10,000-20,000 GB300s we would produce a great model that could rival the current frontier. But the scale of next year’s frontier models requires far more than an order of magnitude larger cluster. And for this the constraint today is not only capital, it is physical data center space and contracted compute. The compute needed to be at the frontier of the current model recipe is going vertical, and as the world accelerates along the axis of Recursive Self Improvement this will only become more evident. To this end, the PIC infraco, spun out in Jan 2026, is also interesting in its ambitions… We’re confused too, and the founders say they are “not ready to share the updated vision”, but everyone here is coming out with a lot of money so we’re just interested to see what’s next for everyone on the 3 different directions emerging from OG Poolside. The only hints left to us: We wholeheartedly believe that everything economically valuable, scientifically interesting and a lot of what will be personally meaningful is going to be underpinned by Al. The world has not yet reached 0.1% of this transition…. … We believe human level capabilities of intelligence will be fully commoditized by open source models, while super intelligence will likely not be. The world has two types of economically valuable problems, those that are intelligence bound, and those that are experiment bound. The first are problems which we can solve by scaling up intelligence e.g. building software, doing accounting, solving a math theorem. The second are ones that require real world experimentation to progress, and no amount of increased intelligence without experimental results will make progress. We could put 100,000 of the world’s brightest minds together to solve cancer but without a real world experimental feedback loop, they likely never will. Today’s model revenue is from coding and soon from all of knowledge work. In the future, companies who can go beyond human level capabilities will tap into revenue coming from scientific discoveries where there is a true data moat derived from real world experimentation. In our humble opinion, Al’s ultimate value will not derive from the first kind, that will become a low margin commodity, but it will from the second. Al will become the world’s most valuable scientific discovery engine. Fascinating. Sounds like we could not have timed our AI for Science podcast better. AI News for 8/19/2026-8/20/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews’ website lets you search all past issues. As a reminder, AINews is now a section of Latent Space. You can opt in/out of email frequencies! AI Twitter Recap OpenAI and Anthropic Expand the Agent Product Surface OpenAI pushed several desktop and builder features in one wave: @ChatGPT launched an Apple Messages plugin for ChatGPT Work/Codex on Mac, enabling message search, catch-up, drafting, and sending from the desktop app. @OpenAIDevs also added collaborative editing for ChatGPT Sites, with teammates sharing a project while Codex manages git/CI; shared read-only conversation links and PR-context sharing further push ChatGPT/Codex toward being a coordination surface, not just a chat UI. On the API side, transparent backgrounds in GPT-Image-2 are now in preview for reusable design assets. OpenAI’s desktop memory/workflow features continue rolling out geographically: @OpenAIDevs said Computer History and cross-app memory are now available in the EEA, UK, and Switzerland for Pro/Business/Enterprise Mac users, with Record & Replay also live there. Together, these features point to a product strategy of capturing user workflows on-device and turning repeated actions into reusable skills. Anthropic made its agent platform more composable and production-ready: @ClaudeDevs announced general availability for computer use, browser tool, Skills API, and Files API on the Claude Platform. The Skills API adds versioned reusable procedures; the Files API now supports expiration control, 5x higher rate limits to 500 RPM, and 1 TB/org. Anthropic also published an AG-UI adapter for Claude Managed Agents, mapping chat threads to managed sessions and streaming text, tool calls, and thinking into custom UIs. Model Economics, Usage Limits, and the Enterprise Shift Toward Open Models AT&T became the clearest public case study yet for hybrid routing: the most consequential enterprise datapoint in the set came via @Hesamation, summarizing AT&T’s internal AI deployment: 40% of employee AI usage already routes to open models, with a target of 60–70%; coding costs are down 56% for only a 2% quality drop, at 45B tokens/day. That supports the increasingly common view that frontier closed models remain reserved for the hardest tasks, while “good-enough” open models eat the broad middle of enterprise demand. @amir explicitly framed this as a warning sign for OpenAI/Anthropic’s enterprise moat, while @ollama welcomed AT&T to open models. Pricing pressure is intensifying across closed-model distribution: @eglyman announced GPT-5.6 Sol at 50% off through Router, and both @github and @code amplified the temporary discount for GitHub Copilot / VS Code users. At the same time, user sentiment suggests supply constraints are surfacing as usage caps rather than degraded quality: @bridgemindai complained that a $200/mo OpenAI Pro plan could be exhausted in a single heavy Codex day, and @theo noted it was possible to continue consuming substantial tokens after hitting the stated cap. The broader signal: labs are still searching for the right product boundary between high-end model access and economically sustainable agentic usage. Open-weight adoption and distribution continue to broaden: @ollama said Kimi K3 is now rolled out to over half its subscription base with US/EU hosting and zero data retention. On the open ecosystem side, @Google and @osanseviero highlighted Gemma surpassing 1B downloads, while @_philschmid launched an Awesome Gemma repo aggregating variants, deployment guides, and fine-tuning recipes. Multimodal and Agent Benchmarks: Muse Spark, GLM-5.3, Gemini 3.7 Flash Meta’s Muse Spark 1.2 had a strong benchmark day across multimodal/agentic evals: @AIatMeta presented demos spanning visual coding, robotics planning, and audio-visual understanding, and previewed WildArtifactBench, an internal eval using win rates and Elo from human/agentic judges for practical multimodal tasks. Third-party measurements were favorable: @arena reported +2.1% net improvement in Agent Arena, up from 0.9% in v1.1, with particularly strong Bash Recovery (+11.4%); @DesignArena placed Muse Spark 1.2 #1 for Video-to-Website, #2 for Image-to-HTML, and #3 for Image-to-Frontend, while noting it sits on the price-preference Pareto frontier. Zhipu’s GLM-5.3 keeps showing up in agentic/code evals: @AutoClawAIer announced GLM-5.3 integration into AutoClaw, Z.ai’s work agent. More importantly, @arena said GLM-5.3 Max shifts the Code Arena: WebDev Pareto frontier, projecting to #2 among open models and #8 overall at 1597 pts and $3.65/M. Separately, @ZixuanLi_ resurfaced SAO (Single-Rollout Asynchronous Optimization) as a key GLM-5.2/5.3 RL advance for stable asynchronous agentic RL. Gemini 3.7 Flash keeps accumulating “cheap and strong” evidence: @arcprize reported ARC-AGI-2: 84.6% at $0.25/task and ARC-AGI-1: 95.5% at $0.12/task, making Gemini 3.7 Flash stand out on cost-adjusted reasoning performance. @JonathanJarvis separately called it excellent for agentic vision tasks. Infra, Hardware, and Systems Work: Rubin, Cerebras, Linux Agents, Caching OpenAI’s next pretraining stack is moving onto Rubin: @udayruddarraju posted that OpenAI’s first NVIDIA Vera Rubin racks are now installed and running the training stack, explicitly tied to next-generation frontier pre-training. @gdb called it a major milestone in the OpenAI-NVIDIA partnership. Cerebras’ CS-4 drew attention for inference scaling without a node shrink: @kimmonismus summarized the launch as essentially doubling performance on the same 5nm wafer, 4T transistors, and 900k AI cores, via redesigned power delivery and cooling. Reported specs include 250 PFLOPs per WSE-3 Turbo, 43.2 PB/s memory bandwidth, and a 3-wafer CS-4 rack at 750 PFLOPs. The notable claim for practitioners: 4,400+ tok/s per user on GPT-OSS-120B, up to 30x faster than GPU-based systems. Agent runtime ergonomics are becoming a systems bottleneck: @theo argued that Linux materially outperforms macOS for agent workloads, especially on filesystem-heavy operations. @Qdrant_engine shared a practical semantic-caching writeup showing 57.1% hit rate, 55.7% fewer tokens, and ~15 ms hit latency. @MParakhin pushed gisting as an underused production technique, citing ~40% lower end-to-end latency and ~15% higher throughput with better results, and linked a Shopify engineering writeup. Agents, Memory, and Harness-Centric Learning Chroma launched a research preview of self-improving memory: @jeffreyhuber announced Foundation, Chroma’s approach to agent memory, built from prior agent sessions. This landed amid a broader shift from “single-shot agent” thinking toward persistent harnesses with accumulated state, skills, and memories. The most interesting agent research in the set was about harness evolution, not model weights: @omarsar0 highlighted a paper on harness continual learning, where prompts, memories, skills, and routing rules evolve independently of the model. The key failure mode is harness-level forgetting: improving one component can silently break previously reliable behavior. The proposed solution, guarded harness evolution, separates proposing updates from committing them, with reported >10% gains across textual, multimodal, and open-world tasks. Related negative results matter too: @dair_ai flagged a study showing that memory-based self-improving agents look worse once you control for task order effects and evaluation variance. @omarsar0 also summarized a paper arguing post-training agents tend to lock [truncated for AI cost control]