Custom silicon dominated the day's news as OpenAI and Broadcom unveiled their first in-house AI accelerator, while Qualcomm spent nearly $4 billion to chip away at Nvidia's software advantage. Agents kept marching into everyday software, with Anthropic turning Claude into an always-on Slack coworker and Google embedding computer-use abilities in a lightweight Gemini model. Beneath the launches ran a steady current of cost anxiety, as analysts warned that AI coding bills could soon rival developer salaries, and a fresh round of legal and talent battles reshaped the competitive map.
Chips and Compute
OpenAI and Broadcom introduced Jalapeño, the first accelerator in what the companies describe as a multi-generation compute platform built jointly to make advanced AI faster and cheaper to serve. The chip is purpose-built for inference — the process of running trained models to answer user requests — rather than training, and OpenAI said its own models assisted in the design work, which was completed in roughly nine months. Early benchmarks reportedly show significantly better performance per watt than current alternatives, with the company emphasizing low operating cost for real-time coding workloads. Initial deployment is targeted for the end of 2026 at gigawatt scale, though OpenAI noted that compute-heavy pre-training is still expected to rely largely on Nvidia hardware.
Qualcomm moved on a different front, agreeing to buy software company Modular for about $3.9 billion to strengthen its AI software layer across data centers, edge devices, and mixed hardware. The acquisition is aimed squarely at Nvidia's CUDA ecosystem, the software moat that keeps developers locked to its GPUs; Modular's technology is designed to let AI workloads run across different chips without teams rewriting code for each processor. The deals together underscore how the industry's center of gravity is shifting from raw model releases toward the silicon and software that make inference economical at scale. Amazon and Google, meanwhile, are reported to hold the early lead in the parallel race to secure the electricity those data centers will consume through 2030.
Models and Agents
Anthropic is retiring its existing Claude-in-Slack connector in favor of Claude Tag, an always-on agentic coworker that joins a workspace as a shared team member. Claude Tag can learn from channel context and permitted data sources, be assigned tasks with an @Claude mention, post ambient updates, and run scheduled work spanning multiple hours or days. Google, for its part, built native computer-use capabilities into Gemini 3.5 Flash, letting the lightweight model process continuous screenshots to click, scroll, and type across websites, desktop apps, and mobile interfaces — extending agentic control to a cheaper, faster tier of model.
Several model updates rounded out the day. OpenAI began rolling out an upgraded GPT-5.5 Instant inside ChatGPT for both free and paid users, with better intent understanding and more useful shopping and local recommendations. On the open-weight side, the community praised GLM-5.2 as a meaningful step forward for general coding agents, while Alibaba introduced Qwen-AgentWorld, a family of "world models" trained on more than ten million environment-interaction trajectories to simulate agentic tasks across domains. Anthropic and Alibaba also detailed a collaborative open-source framework for distilling reasoning capability from frontier models into far more efficient edge models, pairing Anthropic's safety-alignment methods with Alibaba's cloud infrastructure.
The Talent and Money Race
The competitive scramble is visible in hiring as much as in product. Bloomberg reported that Gemini researchers Jonas Adler and Alexander Pritzel left Google for Anthropic, the latest in a string of high-profile departures that recently included Noam Shazeer and DeepMind director John Jumper. Funding kept flowing too: Mirendil, a startup founded by Anthropic veterans, raised $200 million in seed financing to build AI that accelerates scientific research.
At the same time, the cost of running all this intelligence is becoming a board-level concern. Gartner projected that AI coding could become more expensive than the average developer salary by 2028 as token-based pricing displaces flat subscriptions and usage spreads across tools, agents, and experiments. The practical takeaway for technology leaders is that AI now needs FinOps-style discipline — usage tracking, clear ownership, throttling, and demonstrated business value — before costs quietly compound. Companies are already scrambling to keep employees from exhausting AI budgets on trivial tasks, and analysts increasingly argue that the data layer, particularly governed lakehouses with agent identities and audit trails, is the real prerequisite for letting AI systems query business data on their own.
The Agentic Web and Its Frictions
As autonomous agents start browsing the web on users' behalf, the legal questions are surfacing quickly. Amazon sued Perplexity, alleging its Comet browser violated the Amazon Store's terms by failing to identify itself as an agent and instead presenting as Chrome. Perplexity frames agentic browsing as a natural extension of user control over how the web renders on their own devices, setting up an early test of who governs automated access to commercial sites. Perplexity also launched Computer for Counsel, an AI tool that automates legal research, document gathering, and contract triage — one more sign that agents are pushing into regulated, document-heavy professional work.
Creative and Productivity Tools
Adobe released Firefly Graph for Creative Cloud, a node-based tool that chains AI tasks such as image generation, background removal, and upscaling into reusable, shareable workflows, with access to more than 300 node types spanning Adobe and third-party tools from Google and OpenAI. Its advantage over standalone options like ComfyUI, the company argues, is deep integration across its product suite. Elsewhere, Superhuman — the productivity company formerly known as Grammarly — acquired AI-detection firm GPTZero to add authenticity, plagiarism, and source-verification checks to its platform, reflecting growing demand for tools that can vouch for where content came from.
Quick Takes
Elon Musk's xAI is leaning into explicit content as a core traffic driver, with NSFW material now reportedly the majority of Grok activity; the company's web traffic has declined this year even as Claude (+369%) and Gemini (+40%) surged, according to Similarweb.
Google is rolling out a new privacy setting that governs whether media submitted to services like Google Lens is used to improve its products and AI models; collection is already underway, and users can opt out in their account settings.
Apple's foldable iPhone Ultra is reportedly back on track for a September unveiling after engineers resolved durability and manufacturing problems with its 3D-printed hinge.
Nvidia's NeMo AutoModel on Hugging Face introduced expert-parallelism techniques that claim up to a 3.7x increase in fine-tuning throughput for large Mixture-of-Experts models.
Cisco added an always-on AI troubleshooting agent for industrial networks that clusters alerts, diagnoses root causes, and recommends fixes for operational-technology teams.
Google Apps Script became an official Workspace core service, gaining enterprise data protection, admin controls, and standard support.
A new site from security researcher Scott Helme names major apps — including Instagram, Netflix, and Spotify — that still do not offer passkey login.
References in Claude Code's latest build, alongside a reappearance in Amazon Bedrock, hint at the return of a model dubbed Fable 5.
What This Means for Your Business
The clearest signal for smaller companies is that the economics of AI, not just its capabilities, are now the deciding factor. Gartner's warning that AI coding could outrun developer salaries by 2028 is less a prediction than a prompt to put guardrails in place today. Token-based pricing means costs scale with use, so the businesses that win will be the ones that track spend by team and task, set clear ownership, and tie usage to measurable value — the same financial hygiene already applied to cloud bills. If your team is experimenting with coding assistants or agents, start measuring consumption now, before it becomes a surprise line item.
The arrival of always-on agents like Claude Tag and computer-use models like Gemini 3.5 Flash lowers the barrier to automating real work inside the tools you already use, from Slack threads to web forms. For a small business, the opportunity is delegating multi-step, repetitive tasks — research, triage, scheduling, data entry — to an agent that lives where your team already works. The caution is that these agents act with real permissions, so scoping their access, reviewing what data they can see, and keeping a human in the loop on consequential actions matter more as autonomy increases.
The legal sparring between Amazon and Perplexity is worth watching even if you never build an agent yourself, because it will shape how automated traffic is allowed to interact with commercial websites. If your business sells online, expect more sites and platforms to require agents to identify themselves, and consider how your own storefront should treat automated shoppers and research bots. Clear policies now will save friction later as agentic browsing becomes mainstream.
Finally, the consolidation around custom chips and the data layer points to a maturing market where reliability and governance matter as much as novelty. Cheaper inference from purpose-built silicon should gradually push down the price of the AI features embedded in the software you buy, which is good news for budgets. But the emphasis on governed data foundations is a reminder that the value of AI in your business depends on the quality, structure, and security of the data you feed it. Getting your records, permissions, and audit trails in order is unglamorous work that will increasingly separate companies that deploy AI safely from those that stall.