AI Agents Get Production Tools | AI Field Notes #92

AnalysisAgents are getting their own silicon. At Hot Chips, the chip industry's annual technical conference, Nvidia detailed Vera, an 88-core CPU (central processor) it claims runs agentic workloads 1.8 times faster than rival x86 chips and compiles the Linux kernel 14 to 22% quicker than AMD's 96-core EPYC. The pitch is that agents, which chain many small steps and juggle tool calls, choke on memory bandwidth rather than raw math, so Vera dedicates hardware to each thread instead of sharing it. Vera anchors the Vera Rubin platform, which Nvidia says hits five times Blackwell's inference throughput. Orchestration is becoming a hardware problem.

AnalysisA company that barely existed five years ago is being lined up as the largest public offering in history. Anthropic's investors are targeting a $2 trillion valuation for an October listing, and the company has filed confidentially with regulators. The math they lean on: annualized revenue that passed $47 billion by May and a run rate backers expect to reach $100 to $120 billion by year-end, up from a $965 billion valuation in that same May round. The $2 trillion figure is investor talk, not a number Anthropic has committed to. Even so, an AI lab worth more than most national economies is now a filing away.

AnalysisThe AI buildout is being financed the way pipelines and toll roads are. Broadcom is in talks to raise as much as $100 billion in debt, roughly $60 to $70 billion of senior secured loans plus about $30 billion of junior financing, to fund custom silicon for customers including Anthropic. Apollo and Blackstone are arranging it through a special-purpose vehicle, with Broadcom guaranteeing part of the senior tranche. It builds on a $35 billion deal the same group struck in June. Structuring compute as tradeable debt turns AI capex, the spending on chips and buildings, into an asset class. Someone eventually has to service that debt with revenue.

AnalysisThe pieces to build a working software agent are now off the beta shelf. Anthropic moved four capabilities to general availability: computer use (Claude driving a desktop by screenshot), a distinct browser tool that reads a page's accessibility tree and fills forms directly, a Skills system for reusable procedures, and a Files API. Computer use took nearly two years to leave preview. The browser tool can take several actions per model call rather than one, cutting the round trips a task needs. Stitched together, an agent can read an intake form, follow a saved procedure, finish a job in a web app, and return a file.

AnalysisInference is being split in two. Nvidia has put its Groq 3 LPX accelerator into full production, a chip that does only the decode phase, the token-by-token generation a user actually waits on. It is not a GPU. Each rack packs 256 LPUs (language processing units) with 500 megabytes of on-chip SRAM (fast memory sitting on the die itself), and Nvidia says one rack pushes 3,400 tokens a second on a 31-billion-parameter model with a long context. The design sidesteps the memory-bandwidth wall that slows GPUs on generation. Cloud host Nebius signed on first. The economics of serving a chatbot just changed shape.

AnalysisOpen weights are quietly becoming the default substrate under a lot of software. Google says its Gemma family, small models it releases with the weights public, has passed a billion cumulative downloads, with developers publishing more than 100,000 variants built on top. The models are running in places that make the point: NASA and a few space startups use them for image analysis in orbit, and India's health authority folded Gemma into an app with over 100 million users. A billion downloads is a distribution number, not a quality one. But distribution is how a model becomes infrastructure, and Gemma is most of the way there.

AnalysisThe standards that let AI agents talk to each other are being pulled out of any one company's hands. Google donated A2A (Agent2Agent, its protocol for agents to coordinate across systems) to the Agentic AI Foundation, a Linux Foundation body that already governs Anthropic's Model Context Protocol. The foundation has grown from fewer than 40 members to more than 250 since December, including Amazon, Microsoft, OpenAI, and Shopify. Putting the two dominant agent protocols under neutral governance means neither Google nor Anthropic can quietly change the rules to favor its own products. Interoperability is winning, at least on paper.

AnalysisStealing AI secrets is easy to prove; proving you did it for a foreign government is not. A federal judge threw out seven economic-espionage convictions against Linwei Ding, a former Google engineer accused of taking thousands of pages on the company's AI chip and data-center designs, ruling prosecutors never showed he knew or intended the theft to benefit China. His seven convictions for stealing trade secrets stand. Ding was found guilty in January after an 11-day trial; the espionage counts each carried up to 15 years. He is sentenced on September 1. The ruling narrows a case the government had held up as a warning.