Nvidia reverse-execuhires Poolside for $6 billion

Nvidia's $6B move on Poolside redraws the model-building talent map, while fresh Ramp data shows the enterprise race whipsawing back toward OpenAI and Anthropic quietly retreating on its data-retention mandate. Open weights keep shipping — SenseNova's distilled image model, AntLing's Ling-3.0 base checkpoints — alongside real inference speedups and a Bun runtime that finally landed its Rust rewrite.

Nvidia pays $6B for Poolside's model factory and 109 engineers

Per an investor letter first reported by Newcomer, Nvidia is licensing Poolside's "Model Factory" — the pipeline behind its Laguna model — extending job offers to 109 of Poolside's roughly 115 technical staff, and investing $1B at a $12B pre-money valuation, while the three founders stay on. Poolside frames it as "not an acquisition and not an acquihire" and plans to distribute the $6B to investors by the end of next year. Latent Space calls it a reverse-execuhire: unlike the Windsurf, Character and Scale deals where executives left and staff stayed, here the founders keep the shell to pivot while employees and investors cash out.

Why it matters: Nvidia builds its own Nemotron open models, so it is now buying model-building capability from a startup it also invested in — another deal structured to lock in tech and talent without a full acquisition, following Groq ($20B) and Enfabrica.

OpenAI claws back enterprise ground as GPT-5.6 Sol drives revenue up 35%

New Ramp data on 70,000+ US businesses shows Anthropic still leads at nearly 44% of paying business users to OpenAI's nearly 40% as of July, but OpenAI is now growing faster this quarter, with API spend up 82% QoQ versus Anthropic's 76%. OpenAI says revenue is up 35% since GPT-5.6 Sol launched July 9, with enterprise revenue up more than 50%; its next model, Astra, is due in weeks. Ramp's economist credits Sol's growing developer preference and blames Fable 5's weak adoption on price plus data-retention requirements.

Why it matters: Just days after Anthropic overtook OpenAI on run rate, the lead is flipping again — a reminder that enterprise buyers switch on every model release, which should unsettle anyone betting on sticky AI revenue.

Anthropic moves Fable data retention into customers' own clouds

After enterprise pushback, Anthropic is reworking the policy that since June forced 30-day retention of all data from its Mythos, Fable and future flagship models on Anthropic's own servers for cyberattack detection. The 30-day window stays, but the data will now sit in the customer's cloud rather than with Anthropic; the company spent months building the system with 100+ regulated-industry customers and expects it to arrive this fall. OpenAI is testing a different content-control approach with Databricks and Microsoft.

Why it matters: The retention mandate was directly blamed for Fable's soft enterprise uptake, so relocating the data to customer clouds is Anthropic conceding the policy cost it deals — and shifting the forensic burden onto the buyer.

Ramp launches Router, its own OpenRouter rival

Ramp shipped Router, an API that routes across models from OpenAI, Anthropic, DeepSeek, Moonshot, Minimax, Nvidia, xAI and Z.ai, with strategies to pick a model by cost, provider flex tier, or up to three user-specified benchmarks. It is US-only, free through the rest of 2026 (you still pay inference) with a $26 launch credit, and gives you a dashboard for token spend, latency and fallbacks. Notably it defaults to one year of opt-out retention of inputs, outputs and tool calls, stripping PII before using the content to improve the product.

Why it matters: Another payments and expense-management player building an AI toll house after Stripe's OpenRouter buy — convenient for testing, but read the retention default before piping production traffic through it.

SenseNova U1.5-Lite trains expert models, then distills them into one

SenseNova released U1.5-Lite, an open image generation and editing model that trains task-specialized experts for text rendering, aesthetics and editing, then uses OPD distillation to fold them back into a single inference model — no router, no expert switching. Benchmarks improve over the preview (Qwen-Image-Bench 47.14 to 60.18 with prompt expansion, GEdit-Bench-EN to 8.26), with task-oriented RL for instruction adherence and edit fidelity plus native 4K generation. Weights and code are on HuggingFace and GitHub.

Why it matters: "Specialized in training, unified in delivery" is a clean sidestep of MoE serving overhead: expert-level quality from one model at inference time.

Liquid AI's DSpark drafts cut inference latency up to 3.2x, upstream on day one

Liquid AI released DSpark speculative-decoding draft models (~300M params) for its LFM2.5 line, reporting up to 3.18x throughput on an H100, 2.87x on-device on an M4 Max, and 57% lower function-calling latency for the 2.6B model. DSpark pairs a DFlash-style parallel backbone with a Markov sequential head and a confidence-scheduled verifier that prunes low-confidence suffixes; output is greedy-identical to the target by construction, so accuracy is unchanged. Checkpoints ship in Safetensors and GGUF with day-one llama.cpp and SGLang support.

Why it matters: Speculative decoding keeps eating the memory-bound decode tax, and shipping the drafts upstream on day one means you can run this without hand-patching your inference stack.

AntLing drops all six Ling-3.0 base checkpoints under MIT

AntLing released the full Ling-3.0 base matrix — two sizes (tiny, flash) across three training stages (pretrained, mid-trained, WSM-merged) — as six separate MIT-licensed HuggingFace repos. All are base checkpoints, none post-trained, aimed at continued pretraining, fine-tuning and research rather than chat or instruct use. The point of the release is letting builders choose where on the training trail to enter.

Why it matters: Publishing the mid-training checkpoints, not just the final base, is rare and genuinely useful if you do continued pretraining or want to study where capabilities emerge before quantization.

Bun 1.4 lands the Zig-to-Rust rewrite and a built-in WebView

Bun 1.4, the first stable release since the Rust rewrite, adds 1,517 Node.js compatibility tests, claims over 2,900 bug fixes, 5x lower idle CPU, up to 35% less memory and 50% faster Linux startup. New APIs include Bun.WebView (browser automation via macOS WebKit or a local Chromium over the Chrome DevTools Protocol), plus Bun.Image, Bun.markdown, Bun.cron and parallel test/run. Simon Willison built a shot-scraper-style JSON API on Bun.WebView, finding a full Chromium against complex pages needs a 192-256MB container.

Why it matters: A first-class WebView in the runtime means browser automation and scraping without dragging in Playwright — handy plumbing for agent tooling and screenshotting.

Also worth a look

Browse previous days →