The Harness

OpenAI owns the inference stack

Show Notes

OpenAI launched Jalapeño, its first custom inference chip built with Broadcom, completing a full-stack vertical integration play from silicon to product. Anthropic accused Alibaba of running 28.8 million distillation exchanges through 25,000 fraudulent accounts — the largest known attempt to extract frontier AI capabilities. Google shipped computer use natively into Gemini 3.5 Flash, making browser-to-desktop agent control a standard API feature across major providers.

What is The Harness?

A daily summary of what is interesting and happening in the AI industry, with a focus on what this means for people building harness experiences that are used.

Good morning, it's Thursday, June twenty-fifth.

In today's briefing we see OpenAI launching Jalapeño, its first custom inference chip completing a full-stack vertical integration, Anthropic reporting the largest known model distillation attack from Alibaba, and memory infrastructure reaching maturity as a service category.

First up - Today in the big model news;

Open AI
OpenAI unveiled Jalapeño, the company's first custom inference ASIC built with Broadcom. The design cycle took nine months from spec to tapeout, unusually fast for custom silicon. Specs: two hundred sixteen gigabytes of HBM3E memory, seven point one to seven point four terabytes per second bandwidth, approximately ten petaflops of FP4 compute. OpenAI now owns every layer between a user request and a model response: silicon, kernels, memory, networking, scheduling, and products. Inference is where AI companies spend and earn. Training still runs on NVIDIA H100s, but inference margin now runs on Jalapeño. The nine-month tapeout suggests OpenAI used AI in the chip design itself, a literal instance of the compounding-loop thesis applied to silicon. For product teams evaluating API providers, expect cost trajectories to diverge by how much silicon each provider owns, because companies that earn margin at every layer can structurally underprice competitors who rent inference capacity.

Google + Deepmind / Gemini
Google shipped computer use natively into Gemini 3.5 Flash, including browser, mobile, and desktop versions, via the Gemini API and the enterprise agent platform. This is a tool call in the standard model API, not a separate beta. Anthropic shipped computer use in late twenty twenty-four, OpenAI followed, and Google is now at parity. Computer use has moved from differentiator to table stakes across the major providers. For product builders shipping agent workflows, the competitive question shifts from whether to use computer use to which provider's runtime is most reliable in enterprise settings, because Google's adversarial training and prompt-injection auto-stop are now the relevant differentiators.

Anthropic
Anthropic reported to the US Senate Banking Committee the largest known model distillation attack. Between April twenty-second through June fifth, operators linked to Alibaba ran twenty-eight point eight million Claude exchanges through approximately twenty-five thousand fraudulent accounts, specifically targeting agentic reasoning and software engineering capabilities. Distillation harvests reasoning patterns from a frontier model by repeated prompting, then uses those outputs to train a weaker open-weight model. This February, Anthropic named DeepSeek in a similar campaign. The pattern is now documented and systematic: API-level distillation is the route under the export-control fence. The access-control regime was tightening from above while adversarial operators tunneled underneath it at API scale. For teams building frontier models, distillation is now a documented threat vector at scale, because API-level orchestration allows capability harvesting that evades traditional audit and control systems.

In AI Infra;

Memory is maturing as an infrastructure category. Weaviate's Engram reached general availability, treating memory as an asynchronous service layer where extraction, deduplication, and reconciliation run offline, separate from the context window. LangSmith's Context Hub adds sleep-time compute: trace analysis that generates persistent memories between sessions. The frame is clarifying. Memory is not a context-window problem. It is a data pipeline problem, and it is now getting data-infrastructure treatment. For teams building persistent agents, expect the primitives layer to arrive, because infrastructure for memory management is now available as a decoupled service.

In other news…

Mirendil closed a two hundred million dollar seed round to build what they call self-accelerating AI R&D systems. The founding team includes former researchers from Anthropic, Google DeepMind, and xAI. The thesis: an AI system that generates hypotheses, runs experiments, evaluates results, and controls its own compute, making progress on AI R&D without human intervention. This is the self-accelerating-R&D problem made into a product, not a research agenda. If the thesis lands, the compounding loop operates at the research layer: each AI-designed improvement generates training signal for the next iteration. For research teams building with frontier models, expect acceleration to compound at the research layer, because each cycle of AI-designed improvement generates training signal for faster progress.

The Chip Security Act is gaining industry support for hardware-level location tracking of AI accelerators, reflecting growing concern about uncontrolled compute deployment and supply chain security. For infrastructure builders and policy teams, expect hardware-level supply chain controls to move from proposal to implementation, because governments are locking down where accelerators run and who controls them.

That's the briefing. Have a great day.