In partnership with
Reading time: 6 minutes
🗞️ In this edition
Nvidia's New Rack Beat Its Own Sales Pitch
Salesforce Stopped Renting A Frontier Model
Sponsored: Roku
OpenAI Bought A Camera Company
⚡ YOUR MOVE For people building the life they want, now.
In other AI news –
Superhuman Bought Its Notetaker
Apple Switched On The New Parental Controls
Teravolt Wants The Old Factories
4 must-try AI tools
Nvidia said its new rack would be three times better. Somebody finally measured it. Salesforce stopped renting a frontier model and trained its own. And OpenAI spent more than $300 million on a camera company.
Keep reading.
Source: SemiAnalysis
Jensen Huang told GTC to expect three times the work per megawatt.
WHAT HAPPENED
SemiAnalysis published measured results on Monday for Nvidia's Vera Rubin NVL72, running an agentic benchmark called AgentX on DeepSeek V4 Pro, a 1.6 trillion parameter model standing in for the size of model agents actually run on. Against GB300 the rack delivered up to seven times the token throughput per megawatt, where Huang projected three times at GTC 2026. At 170 tokens per second per user the analysts measured roughly 67 times the throughput per dollar of total cost of ownership, falling to between 1.4 and 3 times in the 60 to 100 tokens per second range where most products actually sit. The run covered Rubin, GB300, B200, H200 and AMD's MI355X across TensorRT-LLM, SGLang and vLLM. The same team notes Huang undersold the last generation too, promising 30 times Hopper for GB200 NVL72 and measuring 98.
WHY IT MATTERS
Every gigawatt contract signed this year was priced against the tokens per megawatt of hardware people could already see. If one generation moves that number by seven, the capacity booked for 2027 buys far more agent work than the spreadsheets behind those deals assumed, and the shortage everyone is planning around starts to look like a timing problem rather than a physical one. The gap inside this report is the part to hold onto. Sixty seven times is real and almost nobody operates there, while 1.4 to 3 times is the honest figure for a product with users waiting on it. Vendors will quote the first number and buyers will live inside the second.
OUR TAKE
Nvidia has now undersold its own silicon two generations running, and that reads as a choice rather than an accident. A launch number that outside testers beat in public is worth more than a keynote slide nobody can check, and it leaves room to price against the smaller claim.
Watch what the neoclouds do next. If published per token prices on Rubin capacity have not fallen by at least a third by the end of Q1 2027, the efficiency went into margin instead of the market, and the compute shortage stays exactly as useful to sellers as it is today.
When you don't click through on a story, what's usually the reason?
Source: TechCrunch
Koa is post trained from Nemotron 3 Super, Nvidia's 120 billion parameter open weight model, and tuned for sales, marketing and support work inside Agentforce. The two companies trained it on public and synthetic data only, simulating a customer service floor down to the irate callers, so no customer records went in. The technical report says it clears strong proprietary baselines on multi turn tool use while sitting below frontier models overall. Salesforce also announced ClaudeForce, which keeps Claude's work inside its own systems.
From Our Sponsor:
How Jennifer Aniston’s LolaVie brand grew sales 40% with CTV ads
For its first CTV campaign, Jennifer Aniston’s DTC haircare brand LolaVie had a few non-negotiables. The campaign had to be simple. It had to demonstrate measurable impact. And it had to be full-funnel.
LolaVie used Roku Ads Manager to test and optimize creatives — reaching millions of potential customers at all stages of their purchase journeys. Roku Ads Manager helped the brand convey LolaVie’s playful voice while helping drive omnichannel sales across both ecommerce and retail touchpoints.
The campaign included an Action Ad overlay that let viewers shop directly from their TVs by clicking OK on their Roku remote. This guided them to the website to buy LolaVie products.
Discover how Roku Ads Manager helped LolaVie drive big sales and customer growth with self-serve TV ads.
The DTC beauty category is crowded. To break through, Jennifer Aniston’s brand LolaVie, worked with Roku Ads Manager to easily set up, test, and optimize CTV ad creatives. The campaign helped drive a big lift in sales and customer growth, helping LolaVie break through in the crowded beauty category.
Source: BusinessStandard.com
Glass Imaging was founded in 2019 in Los Altos by Ziv Attar and Tom Bishop, two former Apple engineers who worked on the technology behind Portrait Mode. Its software cleans up a photograph while the shot is being taken rather than afterwards, learning the flaws of each camera system, and Honor has already shipped it on phones. The company had raised about $30 million. OpenAI has been buying hardware capability since it paid $6.5 billion for Jony Ive's io, and it still has not said what the first device is.
⚡️ YOUR MOVE
FOR PEOPLE BUILDING THE LIFE THEY WANT, NOW.
🎛️ ElevenLabs put voice, music, sound effects, image and video generation inside its MCP connector. It is live in Claude, ChatGPT and Cursor, on OAuth, with no API keys to paste. → Connect it today
🧠 Shanghai AI Lab's paper on Atria Dawn Preview landed today. The model is 744 billion parameters, MIT licensed, 256K context, with a free API preview and weights already on Hugging Face. → Grab the weights
🧾 Gartner told CIOs the AI vendors are not enterprise grade yet. 86% say risk is growing faster than value, and AI slop costs about $9 million a year for every thousand staff. → Take this to your vendor
🛡️ AIUC raised $40 million to audit AI agents, roughly 5,000 tests and a hundred page report per agent. Cursor, Lovable, Harvey and ElevenLabs are already paying for it. → See the standard
Superhuman Bought Its Notetaker – Fathom arrives with hundreds of thousands of users, and its meeting notes now feed Superhuman's email, calendar and docs. Terms undisclosed.
Apple Switched On The New Parental Controls – iOS 27, iPadOS 27 and macOS 27 add Ask to Browse, filtering for gore and violence in shared media and FaceTime, and a rebuilt Screen Time.
Teravolt Wants The Old Factories – The London company converts smelters and refineries into data centres, arguing compute earns $450 to $900 per megawatt hour against $170 to $190 for aluminium.
Cursor: Writes the feature while you describe it, inside the editor you already use.
Notion AI: Finds the decision buried in notes nobody has opened since March.
HeyGen: Record the video once and let it speak every language your customers do.
Framer: Design the page and put it live without waiting on a developer.
Three stories about paying less for the same work. Nvidia's new rack does several times more agent work per megawatt than its own keynote promised, Salesforce decided a 120 billion parameter model it tuned itself beats renting a frontier one, and OpenAI bought the part of a camera that is really software. The frontier is not where the savings showed up this week. They showed up one floor down, and everybody with a budget is moving there quietly.
Go down a floor.
Interested in featuring your services with us? Email us at [email protected]
