Sam Altman has a plan for the memory in your laptop, and he doesn't think it should be there. The same week he said so on a podcast, the people meant to pay for the data centers that would hold that memory gave OpenAI less money than it asked for. A federal judge let a lawsuit go ahead against one of the biggest of those data centers. And a chipmaker lost more than a quarter of its value because AI's appetite for memory is making ordinary PCs too expensive.
Read together, the weekend's news is one argument with two sides. On one side: put everything in the cloud and use it as hard as you can. On the other: money, power permits and plain customer demand are all starting to say "not so fast."
OpenAI asked for $30 billion and is settling for $20 billion#
The numbers, as TechCrunch pieced them together from the Financial Times and the Wall Street Journal:
September: OpenAI was reported to be raising $30 billion at a $1.4 trillion valuation.
October 22: the valuation was reported at $1.25 trillion, ahead of a planned IPO.
Friday: the FT says the target is now $20 billion, led by investors from the UAE, with a first close expected to be even lower. A later retail offering for wealthy individuals would help it get to $20 billion.
The WSJ says OpenAI has secured $11 billion so far, after what it called a "tepid" response from some large investors.
Twenty billion dollars is still an enormous round. But the direction matters more than the amount. A month ago this company could name its price. Now it's reportedly turning to wealthy individuals to finish a round that big institutions didn't fill. That follows Friday's forecast that Anthropic will get a junk credit rating after its IPO, and the broader spending jitters I wrote about yesterday. Investors aren't walking away from AI. They're starting to negotiate.
If you build on OpenAI's API, nothing changes this week. But a company raising less than it wanted has two levers: raise prices or slow the build-out. Neither one helps you. This is the second frontier lab in two days to send a funding signal, and I'd treat it as a reason to keep a second provider wired up, not as a curiosity.
Sam Altman would like your laptop's memory back#
On the Cheeky Pint podcast, Altman explained why OpenAI needs so much DRAM, and The Verge has the quote:
"From a global allocation perspective, actually, it's quite terrible to have everybody have a bunch of the world's limited supply of DRAM sitting in their own computer and being used five percent of the time." — Sam Altman
As a utilization argument, it holds up. As a pitch to the people paying more for laptops because Samsung and SK Hynix have long-term deals with cloud operators and projects like Stargate, it falls flat. The Verge notes those deals have reserved much of the world's RAM output for the next few years.
The bill is already showing up. Intel's stock fell about 28% on Friday after a weak fourth-quarter outlook. TechCrunch reports that the drop knocked about $12 billion off the U.S. government's 10% stake. CEO Lip-Bu Tan said customers "began signaling a PC slowdown based on the memory prices and availability." Intel's data center revenue grew 78%, but its PC business pays for AI's land grab. Altman's "maximum utilization" has a cost, and Intel's shareholders and anyone buying a PC this quarter are paying it.
Local inference isn't dead. It's just getting more expensive, and the man whose company is buying up the RAM would like you to stop asking for it.
A federal judge won't call xAI's turbines mobile#
The cloud also needs power. Judge Debra M. Brown denied xAI's motion to dismiss a Clean Air Act suit that the NAACP filed through the Southern Environmental Law Center, TechCrunch reports. The suit says xAI "installed and operated at least 35 combustion turbines" without air permits to power its Colossus data center outside Memphis. xAI argued the turbines were mobile sources exempt from permitting. The court disagreed, held that the plaintiffs had standing, and rejected xAI's argument that citizen suits are unconstitutional. A similar motion to dismiss filed by the Department of Justice was denied earlier in October.
This is a follow-up to the neighborhood suit against xAI from earlier in the week, and it's the more consequential ruling. "Bring your own gas turbines" has been the industry's way around a slow grid. If courts treat those turbines as stationary polluters that need permits, that workaround gets slower and more expensive everywhere.
The other side: smaller, cheaper and deliberately less#
Apple's CEO, John Ternus, told Wired's Steven Levy that the company doesn't want to ship "AI for AI's sake," and TechCrunch has the details. It's easy to dismiss as a line from a company that was late. But Apple keeps doing what it says. Apple Intelligence is coming to China in an iOS 27.3 beta on November 4, with a full release in December. It includes notification summaries, translation, writing tools and visual intelligence, plus Alibaba's Qwen in some experiences. It doesn't include the new Siri. Apple is shipping an incremental release that's mostly on-device, in a market it has been working toward since approval in July.
For builders, the more useful item is a VentureBeat guest post by Confluent's Sean Falconer on using open-weight models without running them yourself. He doesn't argue open versus closed. He argues for splitting the work. Use small, purpose-built open models for embeddings, classification and PII detection, where they can beat a frontier model at lower compute cost. Use a hosted inference provider such as Fireworks or Together AI instead of self-hosting. Save frontier models for the customer-facing paths where reliability is what you're paying for. He also has a warning worth repeating: "Open-weight isn't the same as open source." You get weights, not training data, and the safety and compliance work falls on you.
This is the practical answer to Altman's argument. You don't need RAM on your desk to stop depending on one lab's data center. A classifier on a commodity inference host does that.
Where the bargaining goes from here#
My bet: by the end of the year, at least one frontier lab announces a price increase or tighter rate limits and frames it as a capacity issue. When money, memory and megawatts all get scarce at once, the cost ends up with customers. This week I'm checking which of my high-volume classification and extraction calls could move to an open-weight model behind a hosted endpoint. That's not a statement about open versus closed. It's because negotiating leverage is useful when every supplier is under pressure.