Good morning, and welcome to today’s Techzip.

OpenAI's Jalapeño results show that inference economics are becoming a full-stack game. Perplexity is packaging an agent around local runtime, while Moonshot is reportedly seeking hyperscaler distribution for Kimi K3.

Around that core, cyber testing is moving closer to real deployment, Claude is making memory portable across products, WebMCP is turning websites into agent surfaces, and today's practical move is to make your context smaller before making your prompts longer.

OpenAI says Jalapeño, its first custom inference chip, delivers 1.5 to 1.9 times more AI work per watt and 1.7 to 3.6 times lower end-to-end latency on GPT-OSS 120B, DeepSeek R1 and Kimi K2.5. These are OpenAI-reported benchmarks, but the goal is clear: increase serving capacity without sacrificing responsiveness.

OpenAI is co-designing the chip, memory, network, software and workloads as one system. The company says Codex and GPT-Astra helped bring three open-weight models to high performance in two months, while AI-assisted attention and MoE blocks ran 1.5 to 1.8 times faster in internal results. A frontier lab may have a structural advantage here: it can use its own internal models to shorten the hardware-software loop, something harder for a chip vendor that only sells silicon.

Jalapeño is expected to enter OpenAI's infrastructure by year-end. The economics of ChatGPT, Codex and agents may increasingly be determined by the entire inference stack, from chip to software to workload, rather than by the model or silicon alone.

NVIDIA says Perplexity's Portable Computer can run its orchestrator, subagents and agent harness directly on DGX Spark with one-click local inference. Perplexity describes the local setup as cloud-independent, while users can still switch to cloud models when they need more capable systems.

The product is connected to work data through Google Drive, Gmail, Slack and GitHub. NVIDIA also says Perplexity has optimized a Qwen 3.8 27B model for DGX Spark, with a Nemotron 3.5 Lightning variant expected later. The claims come from the companies, but the product points to local agents moving beyond a model demo on a laptop.

Agents are starting to be packaged around runtime, not just model choice. For many teams, the question will be what should run locally for privacy, latency and control, and what is worth sending to the cloud.

Reuters reports that Moonshot is discussing revenue-sharing arrangements with Microsoft, Amazon and Google to host Kimi K3. Reuters says the talks are early, may not produce a deal, and could involve Moonshot giving up as much as 30% of revenue from related K3 services.

This is more than putting a model on another cloud marketplace. Hyperscalers own billing, account management, compliance, data controls and enterprise sales channels, the layers an open-weight model needs to move from a developer download to a product companies can buy and operate.

If the talks become a deal, Kimi K3 will show how distribution can matter as much as capability. A model can open its weights, but its route into enterprise still depends on who controls the cloud infrastructure and customer relationship.

Your pet lives like a king on your salary: organic treats, memory-foam beds, and endless toy drops. Isn't it time their high-maintenance lifestyle worked for you?  Meet Nibbles, the first credit card engineered specifically for pet parents. Earn 3x rewards on pet & vet expenses, from organic pet food to grooming, and get 1x rewards everywhere Mastercard is accepted. You can keep spoiling your highness without breaking the bank. Get the no annual fee card that actually rewards you for spoiling your pet.

Apply now.

Sponsored by Nibbles. Nibbles is not a bank. The Nibbles Card is issued by Lead Bank. Other fees and T&Cs apply. Rewards offered by Nibbles, not Lead Bank. See nibbles.com for details. Nibbles does not operate in CT, NY, KY, MN, WA, CO, HI, MA, or ME.

More context does not automatically make an agent better.

In an X Article, Thariq says Anthropic removed more than 80% of Claude Code's system prompt without measuring a decline on its coding evaluations. For teams with a long CLAUDE.md and overlapping rules, the move is to make context smaller, clearer and available only when needed.

Start with a familiar task the agent often gets wrong, such as fixing a bug or adding a small API.

How to do it:

This turns context engineering into a repeatable loop: start with the minimum context, disclose detail when needed, and expand only when a real failure proves that more guidance is necessary. If results remain inconsistent after the context is lean, inspect the tool interface and parameter handling. Context is only one part of the system.

This paper connects data from more than 1,500 organizations and 17 million messages to examine adoption by role and task. It asks a more useful enterprise-AI question: not whether a company uses AI, but where that use is actually landing and who is doing it.

Serious Insights argues that the center of gravity is moving from model novelty toward infrastructure, deployment, governance, work design and control. That lens helps connect Jalapeño, local runtime and enterprise agents in today's issue.

Paul Dix offers a provocative thesis: as agent throughput rises, the hard part may shift from reviewing every line of code to designing harnesses, verification and improvement loops. The operational question is simple: if agents write more software, what system proves that the code is ready to ship?

This position paper argues that context construction, tool interaction, orchestration and verification inside the harness can create more variance than changing the model on long-horizon agent tasks. It is a useful lens before trusting any agent leaderboard that publishes only the model score.