OpenAI Eyes 2027 IPO, Stripe Buys OpenRouter, and Google Deepens Its Custom-Chip Bet

OpenAI's IPO Clock Turns Frontier Competition Into a Public-Market Test

OpenAI chief financial officer Sarah Friar told employees that the company expects to be public in 2027, with an earlier debut possible if the business continues to accelerate. CNBC's detailed report on OpenAI's IPO timing and operating figures says the lab confidentially filed a prospectus in June. It is now preparing to explain an $852 billion valuation, heavy infrastructure requirements, and leadership turnover to public investors who will see more of the underlying economics.

The internal figures point to both scale and pressure. CNBC reported that OpenAI showed employees a 35% quarter-to-date increase in revenue run rate, a 50% rise in enterprise revenue run rate, and 20 million weekly active users across coding and work products. The same account said second-quarter revenue reached $6.7 billion, up 18% from the first quarter. Anthropic, meanwhile, reportedly reached a $65 billion annualized run rate by the end of July and preliminary second-quarter revenue of $11.5 billion.

“The IPO is not a finish line, it is a milestone, another fundraise.” — Sarah Friar, OpenAI CFO, quoted by CNBC

Stripe Buys OpenRouter and Claims the Economic Layer of Multi-Model AI

Stripe has agreed to acquire OpenRouter, moving beyond payments into the infrastructure that decides which model handles each AI request. Stripe's acquisition announcement says OpenRouter routes work across more than 400 models from over 80 providers using task complexity, price, speed, and reliability. Terms were not disclosed; CNBC's reporting on the OpenRouter transaction, citing The New York Times, put the price near $7.5 billion.

The scale behind the deal shows why routing has become its own control plane. OpenRouter's account of joining Stripe says its platform processes more than 10 trillion tokens per day for over 10 million developers and companies, with inference volume growing at least tenfold each year. Stripe already optimizes payments across authorization, fraud, and payment methods; it now wants to apply a similar decision engine to model cost and performance.

“We believe intelligence will be multi-model: no single model will be optimal for every task.” — Alex Atallah, OpenRouter cofounder and CEO

Google's Marvell Agreement Ties Custom Silicon to Long-Term Purchasing

Google has expanded its custom-chip partnership with Marvell through an agreement that could let Google buy up to $12.2 billion of Marvell shares. CNBC's report on the Google–Marvell chip agreement says the arrangement covers as many as 58,970,907 shares at $206.58 apiece and is tied to purchasing targets through fiscal 2033. Marvell shares rose nearly 10% after the disclosure.

The expanded work includes inference accelerators plus storage and network-interface controllers that connect to Google's tensor-processing-unit ecosystem. That breadth matters because AI performance is not determined by an accelerator alone. Memory movement, networking, storage, compiler support, and power efficiency shape usable throughput. Google and other hyperscalers are building custom silicon to reduce cost and dependence on Nvidia, while chip suppliers gain long-duration demand signals that justify specialized development.

The expanded agreement includes products that “attach to the tensor processing unit ecosystem.” — Marvell securities filing, quoted by CNBC

Ohio Makes Data-Center Consent an Election-Year Infrastructure Risk

AI data centers have moved from local permitting meetings into statewide election strategy. The Daily Signal's report on Ohio's data-center politics describes a private National Republican Senatorial Committee memo warning AI companies that opposition could influence a close Senate race. The report says Ohio already has more than 200 facilities, with as many as 80 additional sites possible by 2030, while residents raise concerns about electricity costs, water, noise, pollution, land use, and tax incentives.

Political positions are converging around a practical demand: communities should approve projects and operators should pay the infrastructure costs they create. Ohio paused new data-center tax exemptions in May, and candidates in both major parties have proposed moratoriums or conditions tied to utilities, environmental standards, local labor, and public consent. A planned Pike County campus backed by Nvidia and leased to OpenAI for 20 years ensures the debate will remain concrete rather than theoretical.

“The companies that need these projects built have to fix how Ohioans see them: who benefits, who pays, and why a community should want one.” — National Republican Senatorial Committee memo, quoted by The Daily Signal

“Model Hypnosis” Finds That Weak Prompt Cues Can Add Up to Strong Control

A new research preprint describes a prompt-level effect its authors call “model hypnosis”: individually weak, inconspicuous cues that combine to steer behavior strongly. The arXiv abstract and submission record for Model Hypnosis says the effect appears across model families and scales, including frontier reasoning systems, and can transfer between models. The paper was submitted August 17; its claims are preliminary and should be treated as research evidence to reproduce, not a settled security standard.

The reported mechanism is operationally uncomfortable because it does not require an obvious jailbreak phrase. Paraphrases, typos, stylistic patterns, or other low-salience choices may have little effect separately yet become forceful in combination. That challenges filters that score tokens or instructions one at a time. It also complicates interpretability: the visible intent of a prompt can look benign while the aggregate surface form shifts model behavior.

“Individually weak and seemingly irrelevant cues in the prompt can be systematically combined to strongly control model behavior.” — Enric Boix-Adserà and Benedict Tessler, Model Hypnosis abstract