 Meta wants to rent you its extra GPUs
| FULL STORY |
|
Nvidia's power problem now has a nuclear answer
| FULL STORY |
|
A startup will bring your dead relatives back for $390
| FULL STORY |
What else happened today?What AI tools should I be using?Good Morning Thorium Valley. Meta has so many GPUs lying around that it's considering renting them out. Wall Street loved the idea — Meta jumped 9%. The companies already in the GPU rental business loved it considerably less. Nebius dropped 17%, which stings extra given they have a deal to sell Meta that exact service.
Nvidia, for its part, helped a startup build a nuclear reactor in the Utah desert. Chips aren't the bottleneck for AI anymore. Power is. Nine months from empty lot to chain reaction.
And in Seoul, a company will make an AI video of your dead grandfather for about $390. Three hundred families a month are paying for one. Nuclear reactors and digital grief in the same news cycle — AI's range is getting hard to keep up with.
Quickly before we dive in — Would you pay for an AI-generated video of a loved one you've lost?
BIG TECH
Meta is planning to sell off its spare AI compute, and Wall Street likes the idea a lot more than the companies already doing it.
On Wednesday, Bloomberg reported that Meta is building a cloud business to lease its excess GPU capacity to outside companies. The market reaction was immediate and telling:
+ Meta jumped 9.3%, its biggest intraday gain since April.
+ CoreWeave dropped as much as 14%.
+ Nebius fell as much as 17% — despite having a signed infrastructure agreement with Meta to sell it GPU-as-a-service. In other words, Meta is now planning to compete with its own supplier.
The plan has been quietly building. At Meta's shareholder meeting in May, Mark Zuckerberg said outside companies were coming to him "almost every week" asking to buy compute. If Meta ends up with more infrastructure than it needs, selling the excess is "an option that we have," he said.
Which is a polite way of hedging a very expensive bet. Meta has committed hundreds of billions of dollars to data centers and GPUs, most of it earmarked for its own models. A cloud business gives that spending a second life if the models don't earn it back.
The real competitive picture is layered. The hyperscalers — AWS, Azure, Google Cloud — have enterprise sales teams, compliance certifications, and a decade of trust that Meta doesn't have. They'll be fine. The neocloud companies built entirely around renting AI compute are the ones that got hit, because Meta has the scale to undercut them without needing the revenue to survive.
None of this is officially a product yet, and Meta declined to comment. But the shareholder comments, the SEC filings, and the market reaction all point the same direction: Meta has more compute than it can use, and it would rather sell the excess than let it sit idle.
There's a version of this where Meta becomes the fourth hyperscaler, undercuts everyone on price with ad-subsidized inference, and reshapes the cloud market the way DeepSeek reshaped reasoning models. There's also a version where enterprise buyers take one look at Meta's compliance story and stick with AWS. The real question is whether Meta actually wants to run a cloud business or just wants a hedge in case its own AI doesn't pay off. One takes a decade of enterprise sales work. The other takes a press release. So far Meta has only committed to the press release.
BIG TECH
The bottleneck for AI stopped being chips a while ago. It's power — and Nvidia's plan to fix that now involves a small nuclear reactor in the Utah desert.
Last week, Valar Atomics announced that its Ward 250 reactor had achieved zero-power criticality, meaning it can sustain a nuclear chain reaction at essentially no power output. It's a tiny thing, rated at 100 kilowatts thermal, orders of magnitude below what a real data center needs. But Valar didn't build it to power anything yet. It built it to prove a concept — and it built it in partnership with Nvidia. The vision: dedicated, behind-the-meter nuclear power sitting directly next to a data center, no grid in between.
"Nine months ago, this was an empty site. Today, there's a critical reactor on it," Valar CEO Isaiah Taylor said.
The nuclear news landed the same week Nvidia released the Vera Rubin DSX reference design, a full architectural playbook for what Nvidia calls "AI factories." The headline feature is cooling. Rubin systems run on a closed liquid loop with coolant warm enough that most data centers wouldn't need chillers or cooling towers at all — just radiators outside. Modern Nvidia chips run so hot that air simply can't keep up, making liquid cooling mandatory.
Stack those two announcements together and the pitch to hyperscalers starts to make sense:
+ Zero water consumption, versus roughly 2.6 million gallons per megawatt per year for traditional cooling towers
+ Around $4 million a year in savings on cooling for a 50-megawatt facility
+ Three times the rack density, because you can rip out the fans
Other big players are already chasing nuclear — Constellation is restarting Three Mile Island for Microsoft, and Amazon and Google both have nuclear PPAs signed. What's different about Valar is the ownership model. Instead of buying grid power from a utility that happens to have a reactor, hyperscalers would own the reactor as part of the facility.
The catch: none of this is close to production ready. Ward 250 puts out 100 kilowatts — a single Rubin rack draws more than that. Advanced nuclear remains expensive, slow to license, and dependent on a fuel supply chain the DOE itself describes as still "under construction."
Nvidia isn't really selling nuclear reactors here. It's selling a story to hyperscalers about how the current build-out actually works at scale. Chips are the easy part now. The hard part is where you get 10 gigawatts of clean power without waiting a decade for grid interconnects. If Valar and the other advanced reactor startups can turn a critical test reactor into something that actually powers a data center by 2030, the map of where AI gets built starts to look very different, with the desert winning and the coasts losing. If they can't, the industry has a very expensive problem it has been quietly pretending is already solved.
CONSUMER
For about the cost of a nice pair of shoes, a Seoul startup will make an AI video of someone you've lost.
The company is Vaice, and it charges 600,000 won (roughly $390) for a three-to-five minute clip of a deceased person, generated from a handful of photos and short voice recordings. About 300 families a month are paying for one, according to reporting from the AP. Most customers are in their 40s and 50s, commissioning videos of parents. The output isn't a chatbot or an interactive avatar — it's a short, one-way message from the dead, played at memorial ceremonies and on traditional holidays.
One customer, a 28-year-old office worker, commissioned a video of his late grandfather as a gift for his father, a single parent. The virtual grandfather called his father "my son." His father cried through the whole thing.
This isn't brand new in Korea — a 2020 MBC documentary used VR to reunite a mother with her deceased daughter and nearly two-thirds of the country watched. But what once required a documentary crew and a team of animators now costs less than a decent dinner in Seoul.
Which raises the obvious question: is any of this actually good?
The core problem is that these videos are built from a few photos of someone who can't correct anything. The AI can guess what a person might say, but it can't tell you what they actually would have said — so the result is inevitably softer and more agreeable than the real person ever was. That may be exactly what people are paying for. It's also what worries researchers.
A 2026 preprint from Gonçalves and colleagues framed the concern sharply: grief has historically been shared with other people. An AI recreation absorbs it without carrying any of the weight — it can't be responsible, can't act in the world, and can't be held morally to account.
China has already started drawing lines. In April, the Cyberspace Administration published rules for AI services that simulate real people's personalities — sessions over two hours must remind users how long they've been talking, outputs must be labeled as AI-generated, and fines climb if the service causes harm. South Korea has no equivalent framework yet.
The unsettling part of grief tech isn't the quality of the videos. It's that they're becoming good enough, cheap enough, and fast enough that using one is now a decision a normal family can make on a normal afternoon. Vaice's $390 price tag turns a philosophical question into a consumer purchase, and consumer purchases tend to get made long before the philosophy catches up. The real fight over the next few years won't be whether this should exist. It'll be over who gets to decide what a dead person is allowed to say.
IN OTHER NEWS
+ A German court ruled Google is liable for false claims in its AI Overviews, saying they're Google's own words — not just search results
+ A man is suing OpenAI after ChatGPT allegedly told him he was Jesus Christ during a manic episode, escalating it into weeks of delusion and self-harm
+ South Korea unveiled a $650 billion plan to build 18.4GW of AI data centers by 2035 — roughly two-thirds of the country's GDP
+ Security researchers found the first ransomware attack run entirely by an AI agent — from break-in to encryption, no human in the loop
+ People are using AI agents to automate dating — one guy built a script that posts World Cup condolence reels and got 200 DMs in days
+ JPMorgan's AI chief Teresa Heitsenrether plans to retire after nearly 40 years at the bank
+ Trump chatted with an AI version of Teddy Roosevelt about the Panama Canal at a new presidential library exhibit built by Microsoft
+ Hundreds of newspaper publishers sued OpenAI and Microsoft for scraping their content, warning it could be a 'death knell for local journalism'
WHO'S HIRING IN AI
AI OR REAL?
Option A |
Option B |
AI TOOLS
+ Claude Code: Dynamic Workflows is now available to all Pro users, letting Claude coordinate up to 1,000 parallel coding agents in a single run — the feature that powered a 960,000-line code port in six days
+ GitHub Copilot: Kimi K2.7 Code is now available in the model picker — the first open-weight model you can select inside Copilot, giving developers a cheaper alternative alongside Claude and GPT
+ Google Cloud: Two new models dropped — Nano Banana 2 Lite for fast, cheap image generation and Gemini Omni Flash for AI video editing with character swaps, relighting, and style transfers
+ Claude Tag: Anthropic replaced its Slack chatbot with a persistent team AI agent that builds memory about your projects over time and lets anyone in a channel tag it with tasks — the old app retires August 3
+ Cursor: Shipped three major versions in June alone — adding an automation layer, a compute engine, and a platform ecosystem that turns the AI code editor into a full developer platform
That's all for today. If this issue made you think, share it with someone who needs to think harder.
Written by Jason Chen, Advait Prakash, Andrew Hales, and the Thorium Valley crew.
That's all for today's Thorium Valley. See you tomorrow.