Grab a coffee. Let's walk through it.
Meta Is Spending $135 Billion This Year and I Have Questions
Meta dropped a new model called Muse Spark, their first big release since the 14 billion dollar deal that brought Alexandr Wang into the fold. The pitch is smarter training, smaller models, same punch as the older Llama 4 family. Fine. That's the kind of engineering win you'd expect when you throw that much money and talent at a problem.
But here's the number that made me spit out my coffee. Meta said 2026 AI capital expenditures will land somewhere between 115 and 135 billion dollars. That's nearly double what they spent last year.
Why it matters: we are watching the largest infrastructure buildout in the history of software, and a lot of it is happening on vibes. Nobody has a clean answer for how this gets paid back. Zuck is betting the farm that owning the compute means owning the future.
My take: you don't spend 135 billion because you're confident. You spend it because you're scared somebody else is going to get there first. I'm not saying Meta is wrong to do it. I'm saying the panic is the real tell. And when the dust settles, half of that spend is going to look like a museum full of stranded silicon.
Google's Gemma 4 Runs on a Single GPU and Kicks
While Meta was bragging about spending, Google quietly rolled out Gemma 4. Four open source models, all small enough to run on a single 80GB H100, and Google claims they hit benchmark numbers comparable to models twenty times bigger.
That's a real number. Twenty times. If it holds up in practice and not just on the benchmark leaderboards, the economics of running AI just got kicked in the teeth.
Why it matters: everyone has been told the future of AI is giant models and giant data centers. Gemma 4 says the future might actually be a small model on a box under your desk. That changes who gets to build, who gets to deploy, and how much of this stuff you have to rent from a hyperscaler.
My take: this is the real story of the week and most people are going to miss it because it doesn't come with a flashy keynote. Small and efficient beats big and expensive nine times out of ten, and the tenth time it's because the engineers got lazy. Google is playing the long game here and they picked the right lane.
Anthropic Built an AI That Hacks So Well They Won't Ship It
Anthropic announced a preview model called Claude Mythos, and the headline is that it finds and exploits software vulnerabilities well enough that they flat refused to release it to the public. Instead they're handing it to about 40 companies, mostly the usual suspects. Apple, Amazon, Microsoft, the cybersecurity crowd.
Why it matters: this is the first time a major lab has basically said out loud, yeah, this one is too hot. The release notes aren't about features. They're about who gets access and who doesn't.
My take: I respect the caution. I also think we're about twelve months away from somebody with fewer scruples shipping the same capability on Hugging Face at 3 a.m. The genie doesn't stay in the bottle just because the nice folks in San Francisco are holding the cork. If you run infrastructure for a living, this week is the wake up call. Patch your stuff. Rotate your secrets. Assume the other team already has a copy.
Where This Leaves Us
Three stories, three different bets. Meta is betting on scale. Google is betting on efficiency. Anthropic is betting that someone needs to at least pretend to be responsible. Only one of those three is going to age well, and I'm telling you right now it's not the one with the 135 billion dollar price tag.
Catch you tomorrow.