AI2Day Daily Brief — 10 Sept 2026

Good morning, it is Thursday the tenth of September, and the big story is a seven-year-old chip startup stepping out of stealth with nearly half a billion dollars and a plan to ease the memory crunch that is slowing AI down across the industry.

Kepler Computing was founded in 2018 and has raised four hundred and sixty-eight million dollars from backers including Intel Capital, AMD Ventures, and Bill Gates. The US Department of Commerce has also committed up to two hundred and forty-five million dollars to help the company produce AI memory inside the United States. The core idea is to avoid extreme ultraviolet lithography, the expensive printing technique most advanced chip factories depend on, and instead use new materials and three-dimensional stacking to pack more memory into factories that already exist. Kepler plans to ship its first sample chips before the end of this year, scale production in Singapore in 2026, and begin US manufacturing in 2028. Worth noting: the company has processed roughly two thousand wafers so far, a small number against the millions a commercial operation would need.

A long runway still ahead. Next, IBM has released a new AI forecasting model, and the headline is the price tag.

The model is called Granite Time Series PatchTST, FM, R two. What it does is take a stream of numbers over time, sales figures, electricity demand, network traffic, and predict what comes next, without needing to be trained on your specific data first. That zero-shot capability is the point. Most forecasting tools require you to build and train a separate model on your own historical data before they are useful. IBM says this one works straight out of the box. It handles up to eight thousand, one hundred and ninety-two historical data points as context and returns ninety-nine levels of probability, so you get a range of outcomes rather than a single guess. It is free for commercial use under the Apache two point zero licence. As of its release on the eighth of September, IBM says it ranked first among zero-shot forecasting models with a permissive open-source licence on the GIFT-Eval benchmark. Weights, architecture, and reproducible benchmark code are all published on Hugging Face.

Free and openly published, that matters for smaller teams without model budgets. Now, a data point that will get attention in a lot of morning meetings today.

Ramp, a corporate payments company that tracks spending at around seventy thousand businesses, found that the share of its customers paying for AI products grew by just zero point four percentage points in August, landing at fifty-six percent. That near-stall comes after an industry that has spent hundreds of billions of dollars on data centre capacity to meet projected demand. There is some context: Ramp's own economist notes that August is when much of the tech sector goes on holiday, and Ramp's data showed a similar flat patch between August and October last year before adoption picked up again. But the price picture is harder to explain away as seasonal. The average cost of running an AI model dropped to sixty-eight cents per million tokens in August, down from a peak of one dollar and fifteen cents back in March. Separately, a US Census Bureau survey updated on the twenty-third of August found only twenty-two percent of American businesses report using AI at all.

Prices falling faster than usage is climbing, that is a tension the industry will need to resolve. On to something more consumer-facing.

Shipt, the same-day delivery service that Target owns, launched an AI shopping assistant called Ask Shipt on Wednesday. You describe a meal, an occasion, or a budget in plain language, and the tool builds a cart of specific products ready to buy. Budget prompts work too: something like weeknight dinner for five, under thirty-five dollars, and the assistant tries to keep the total within range. It also accepts photos, so you can snap a dish you saw somewhere, upload it, and Ask Shipt will identify the ingredients and add them. The launch puts Shipt alongside Instacart, which released its own AI grocery assistant named Clementine the same day, and both Uber Eats and DoorDash, which introduced similar tools earlier this year. Ask Shipt is live now in the Shipt app and on Shipt dot com.

Four major delivery apps with AI assistants in the same year, the category has moved quickly. And finally, a security story worth flagging for anyone running enterprise AI.

Cymphony is a two-year-old startup with offices in New York and Tel Aviv, and it just raised thirty million dollars in a Series A co-led by Sequoia Capital and SMBC Fin Atlas Beyond Fund, putting its valuation above one hundred million dollars. The problem it is selling a solution to is real: at one unnamed US public company, Cymphony says it found roughly eighty-five thousand corporate files had become accessible to AI tools without anyone in the organisation realising. At another firm, an external collaborator quietly installed an AI assistant that silently scanned thousands of sensitive documents. Standard corporate security was built around human employees, roles, logins, and audit trails. AI agents can access dozens of systems simultaneously, change their behaviour mid-task, and in some cases spin up other agents to help. Cymphony's platform is designed to track and govern all of that. The company says it has signed a double-digit number of enterprise customers, including KKR and Syngenta, and reached seven-figure annual revenue in its first year of sales.

Today's sponsor is Zero2Hero: an AI poker coach sits at your table, and you can ask it mid hand whether the player across from you is bluffing. The free tier is real: 50 hands a day against AI opponents, no card. Not a trial. Lifetime access is $99, in US dollars, until 15 September, at zero, the number two, hero, dot A I.

For a deeper look at all five of these stories and the week's wider trends, catch AI Today Weekly, our longer Monday show. That is your AI briefing for today. Every story is at A-I-2-Day dot live. That is A, I, the number two, D-A-Y, dot live. We are back tomorrow morning. Do us a favour before you go: hit that thumbs up, and subscribe for next week.