Big week for AI, and it's only Thursday. DeepSeek shipped a stable release, Moonshot is about to open source their newest model, and Google just told Wall Street they're spending more money than most countries make in a year. Let's get into it.
DeepSeek V4 Goes Stable and the Price War Just Got Real
DeepSeek pushed out the stable release of V4 today, July 24. That sounds boring until you realize what it actually means. Preview builds scare off the companies that matter most, the ones running real production workloads who can't afford a model that changes underneath them every two weeks. A stable release removes that excuse.
And the pricing is the real story here. DeepSeek is running around $0.44 per million output tokens. That is not a typo. That's the number every other lab now has to explain when a customer asks why they're paying five or ten times more for similar quality.
Here's my take. I don't think the big US labs drop their prices to match this anytime soon, they're still selling on trust and enterprise support contracts. But every procurement team at every mid-size company is going to run this comparison in a spreadsheet by Friday. That pressure adds up over quarters, not days, and it's real.
Kimi K3 Weights Drop This Week Too
Moonshot AI launched Kimi K3 through their API back on July 16, and now they've promised to open source the weights by July 27. Stack that on top of DeepSeek's release and you've got what's probably the biggest pile of open weight model drops the industry has seen in one stretch.
Why does this matter beyond the headlines? Because open weights mean any company with the GPUs can run these models on their own infrastructure, tune them however they want, and skip the API middleman entirely. That's a very different business than renting access by the token.
My honest read is this is good for builders and bad for anybody whose whole pitch was "we have a proprietary model." When the frontier keeps getting given away for free, you better be selling something more than the model itself. Product, data, workflow, something.
Google Is Throwing $205 Billion at the Wall
Alphabet raised its 2026 capital spending forecast up to $205 billion, and they said flat out that demand keeps outpacing what they can build. Google Cloud grew 82 percent last quarter. That is not a typo either.
This is the clearest signal yet that the big spenders don't think this is a bubble that's about to pop, at least not one they're willing to bet against right now. When a company that size raises guidance mid-year instead of trimming it, they're telling you their internal demand numbers are still climbing.
My take, and I say this every time capex numbers like this come out: watch the utilization, not the announcement. Anybody can promise to spend $205 billion. What matters is whether that compute gets used at a price that makes sense, or whether we're all staring at a very expensive pile of GPUs in eighteen months. Right now the demand story backs up the spending, so I'm not worried yet.