Custom silicon comes for Nvidia

Custom inference silicon dominated the day: OpenAI taped out its first chip with Broadcom, Qualcomm announced a data-center processor and swallowed Modular, and IBM claimed a sub-1nm leap. Underneath the hardware, the agent stack kept consolidating — Gemini 3.5 Flash baked in computer use, Databricks open-sourced a meta-harness — while Anthropic escalated its distillation fight with Alibaba.

OpenAI and Broadcom tape out 'Jalapeño,' a custom LLM inference chip

OpenAI unveiled Jalapeño, its first custom accelerator (an 'Intelligence Processor') built with Broadcom specifically for LLM inference, with OpenAI doing chip design and Broadcom contributing silicon and Tomahawk networking. OpenAI claims design-to-tape-out took nine months — partly accelerated by its own models — and 'substantially better' performance per watt, though these are self-reported numbers with no technical report yet. Engineering samples are already running GPT-5.3-Codex-Spark in the lab; large-scale deployment is planned for late 2026 at gigawatt scale, with Microsoft reportedly committed to buying 40% of the first run. Community reverse-engineering pegs it as TPU-like, roughly 216GB HBM3E and ~10 PFLOPS FP4.

Why it matters: If the perf-per-watt claims hold, OpenAI gains leverage over inference economics and its Nvidia dependence — but until an independent technical report lands, treat the numbers as marketing.

Qualcomm enters the data center with Dragonfly C1000 and buys Modular for ~$4B

Qualcomm announced the Dragonfly C1000, a data-center processor optimized for AI agents and low power, with Meta planning to deploy it starting 2028. Alongside it, Qualcomm is acquiring Chris Lattner's Modular — maker of the cross-architecture Mojo/inference stack — for roughly $4 billion, with Modular saying Mojo open-sourcing stays on track. Qualcomm nearly doubled its non-smartphone revenue forecast to $40B by 2029 (targeting $15B from data centers); the stock jumped 15% after hours.

Why it matters: The Modular buy gives Qualcomm a serious CUDA-alternative software story to pair with its silicon — another front in the slow erosion of Nvidia's lock-in.

Gemini 3.5 Flash bakes computer use into the main model

Google made 'computer use' a built-in tool in Gemini 3.5 Flash, letting the model see and operate browsers, mobile, and desktop environments directly — previously this required a standalone Gemini 2.5 model. It scores 78.4 on OSWorld, ahead of Gemini 3 Flash (65.1) and GPT-5.4 mini (72.1) but behind GPT-5.5 (78.7) and Anthropic's Opus 4.8 (83.4). Google ships adversarial training plus two optional enterprise safeguards for prompt injection (action confirmation and auto-stop), and offers a Browserbase demo and GitHub reference implementation via the Gemini API.

Why it matters: Folding computer use into a fast, cheap general model lowers the barrier to building cross-environment agents — but the prompt-injection caveats are real, and Google still trails Anthropic on the benchmark.

Anthropic accuses Alibaba of large-scale Claude distillation

In a letter to the Senate Banking Committee, Anthropic accused operators affiliated with Alibaba and its Qwen lab of running the largest known distillation campaign against Claude: more than 28.8 million exchanges across roughly 25,000 fraudulent accounts between April 22 and June 5, 2026. Anthropic frames it as an effort to accelerate China toward its 'Mythos Preview' capabilities, following earlier accusations against DeepSeek, Moonshot, and MiniMax. The timing is fraught: days after the letter, Commerce restricted Anthropic's own Mythos and Fable models over military-misuse fears, forcing it to disable global access.

Why it matters: Distillation via API access is now a stated geopolitical and enforcement issue, not just a research-ethics footnote — and it cuts against the labs' own export-control headaches.

Meta-harness summer: Databricks open-sources Omnigent

Databricks open-sourced Omnigent, a pluggable 'meta-harness' that wraps coding and knowledge-work agents — Claude Code, Codex, Cursor, Pi, custom agents — behind one common API for sessions, files, tool calls, and cancellation, plus a server for sharing, history, and security. CTO Matei Zaharia emphasizes stateful, contextual security policies (e.g., block exfiltration after an agent reads many confidential docs) and per-session spend caps. swyx's AINews dubs this 'meta-harness summer,' noting the pattern is being independently reinvented across shops; Omnigent drew ~400 merged PRs within days of its Saturday launch.

Why it matters: If a standard agent-orchestration layer emerges the way MCP did, owning the harness and memory layer — rather than renting it from a model vendor — becomes the defensible position for enterprise teams.

OpenAI says Codex now generates 99.8% of its internal output tokens

An OpenAI economic-research paper claims agentic Codex has displaced ChatGPT as the company's primary internal AI tool: the average engineer now generates 99% of output tokens via Codex, and even Legal, Finance, and Recruiting crossed to majority Codex use around April 2026. By May, 70.2% of sampled individual users made at least one Codex request estimated to exceed an hour of human work, and 25.6% exceeded eight hours; non-developer adoption grew 137x for individuals since August 2025. Task-horizon figures rely on an LLM-as-judge over transcripts, so treat them as directional.

Why it matters: It's a vendor measuring its own dogfooding, but the directional signal — work shifting from short chats to delegated long-horizon agent runs — is the trend developers are being asked to plan around.

Baidu's MIT-licensed Unlimited-OCR transcribes dozens of pages in one pass

Baidu released Unlimited-OCR, an open (MIT) model built on DeepSeek-OCR that replaces the decoder's attention with Reference Sliding Window Attention (R-SWA): visual tokens stay fully visible to every generated token while the text only attends to a 128-token sliding window, avoiding the KV-cache blowup that makes page 20 cost far more than page 1. It inherits DeepSeek-OCR's encoder (a 1024x1024 page compressed to ~256 visual tokens) and MoE setup (3B total, 500M active). Baidu reports 93.92% on OmniDocBench v1.6 vs DeepSeek-OCR's 87.01% on v1.5 — vendor-reported and on different benchmark versions, so wait for independent evaluation.

Why it matters: Whole-document OCR in a single forward pass would simplify the chunk-and-stitch pipelines most PDF workflows rely on — and it's small, open, and permissively licensed enough to actually try.

Practitioners report MTP and vLLM quietly degrading output quality

Multiple local-inference users pushed back on the 'free speedup' framing of multi-token-prediction (MTP) speculative decoding. One found non-MTP Qwen 3.6 27B produced markedly better code reviews than the MTP variant (more findings, fewer tokens), with real-world agent runtime only ~20% faster despite 2x decode throughput. Separately, several report that the same model on vLLM feels 'lobotomized' versus llama.cpp — broken tool calls, lost context, blindness to messages — likely a mix of quantization, chat-template, and parser issues rather than a clean apples-to-apples win.

Why it matters: Speculative decoding is supposed to verify every drafted token at zero quality cost, so these reports point to config and serving-stack pitfalls worth benchmarking before you trust a throughput number for agentic work.

Also worth a look

Browse previous days →