I priced every route to more memory, from 24 GB all the way to 128 GB, and almost every one has gone up. - 24 GB: the used 3090 is still the best memory for the money, but it climbed from about $1,200 to $1,300 in July to about $1,500 this fall. Test it hard before the return window closes. - 32 GB: the 5090 rung is out of reach for most of us. The same chip goes into the RTX Pro 5500, reportedly about $11,400, and Reuters reported orders for more than three million of them. - 48 to 96 GB: two used 3090s for about $3,000, or an older A6000. Nvidia's current workstation cards went up by two thirds to almost double. - 128 GB: Strix Halo from about $3,100, a Mac Studio at about $5,100, and the DGX Spark at about $6,950, up about 74% from launch. On Windows and macOS you get about 96 GB of that for graphics unless you raise the limit. - The card you already own is the one rung nobody can close. Offloading to system memory with tools like llama.cpp gets you further than you'd think, and renting covers the rest. Prices move every day, so check them yourself before you buy anything. Every source is in the video description. Any questions? Let me know. Cheers, Kawika

A used RTX 3090 runs $1,000 to $1,600 right now. Renting the same card runs about 10 to 50 cents an hour. So how many hours before buying actually saves you money? New video: I worked it out across the whole range. - Against a typical 20 cent marketplace rental, a used 3090 pays for itself somewhere between 8,000 and 13,000 GPU hours. At four hours a day, that's six to nine years. - Against a 50 cent data center machine, it's 2,400 to 3,800 hours, so owning wins much sooner. - Count the hours the GPU actually works, not the hours the PC is on. Most of us use a card in bursts. - Your power price moves it a lot. At California rates, running your own card barely beats renting at all. - If you rent, keep your models on a volume, and remember a stopped machine still bills for its disk. For most people starting out, renting is the cheaper way in. Prices move every day, so check them yourself before you buy anything. Any questions? Let me know. Cheers, Kawika

Studio 1.11.1 is live. The headline is secrets: everything Studio stores that could unlock something of yours is now encrypted in the database. ----------------------- IMPORTANT: This is a database upgrade so follow the upgrade steps below. ----------------------- Secrets are encrypted at rest Organization secrets, provider credentials, webhook tokens, secret fields in your nodes and secrets entered on a run form are all sealed in the database. Studio refuses to write a secret in plain text, so a new feature cannot quietly store one unencrypted. Members see secrets as configured, never as values Members see that a secret field or header is set, read only. The values stay with admins, and copying, importing or installing a workflow from the marketplace leaves webhook secrets behind. Steps are now nodes Everywhere you read it in Studio, a workflow step is now called a node, and every page uses one name for each thing. Also in this release - Chatterbox Turbo text to speech works. - Wait for Webhook checks the caller's API Key or JWT before it resumes the run. - Logs and error traces have secret values scrubbed out. - Notifications can be deleted or marked unread, and a rejected approval keeps its reason. - Approvals, manual nodes and reruns hand off to the next node the way they should. - Saving one section of Organization Settings leaves the others alone. - Tables and dialogs fit a phone screen. - 1.11.1 picks up the fix for the eMudhra/emSign certificate issue the Python community flagged this week, and Studio images no longer ship pip. ----------------------- To upgrade Back up your database first. On the first start, Studio encrypts the secrets you already have, and there is no going back to 1.10 after that. Then run `studio-console upgrade` and pick 1.11.1. If you run workers outside Docker with the `studio-workers` package, upgrade them to 1.11.2 as well (`uv tool upgrade studio-workers`), since Studio checks that worker and API versions match. Docker workers update with the rest of the install.

Something I’ve been curious about lately. From my experience working with different software projects, development can become one of the biggest costs for a growing company. But not every business has raised money. Some companies are bootstrapped. Some are already making revenue. Others are still trying to get the product into the market. So I wonder how this works in the US. Do most companies wait until they raise investment before building a proper development team? Or are profitable small businesses comfortable investing in developers from their own revenue? Curious to hear from people actually running companies there. How do you usually think about this?

I’m based in Tokyo and spend most of my time around software, AI, and business. Always interesting seeing what gets built in the US. Different speed, different market, sometimes a totally different way of doing things. For people here running a company in the US, what are you working on these days? Could be SaaS, AI, e-commerce, local business, anything really.