curated · 2026-07-22

Sundar Pichai dropped Q2 numbers: overall revenue up 24%, Google Cloud up 82%. Gemini hit 950M MAU, API throughput at 22B tokens/min driven by Flash. Gemini Enterprise is in 90% of Fortune 100—solid distribution. The post doesn't disclose margins or capex. Cloud growing 82% sounds great, but if it's bought with heavy infrastructure spend, profitability could be thin. I'd wait for the full earnings release before calling this a win.

I clicked because the two numbers together are wild: a single campus committing $20B, potentially exceeding $30B at full build, while the six-year compute spend forecast jumps by $150B. 3.2 GW is roughly enough to power over two million homes, with first power not arriving until 2028 and full completion stretching to 2032. I'd discount the $750B figure a bit—the post doesn't spell out whether that includes partner spending or is OpenAI alone. If it's the latter, they're projecting an annual burn rate north of $100B on compute, which exceeds what most countries invest in their grids each year. The more practical takeaway is the timeline: no power until 2028 means OpenAI's near-term compute expansion still depends on leasing capacity from Microsoft and other cloud providers. This campus looks like a reserve for model training scale around 2030, not a fix for today's GPU shortage.

This episode packs a lot. A few things worth pulling out. Kimi K3 weights are promised for July 27, but the API is error-prone right now—Lambert had to get a $200 plan just to get stable access. They see real fine-tuning potential in K3, but loading the weights requires a full B300 node, so it's not exactly accessible. Qwen confirmed its next major model will be open-weight, and Xi Jinping's WAIC speech explicitly backed open source as a strategy. The signal is clear: Chinese labs are going to keep pushing open releases. The hosts push back on the "how many months behind closed models" framing. Their take: benchmark gaps vary wildly, but on agentic coding tasks, a few months' lag is a real competitive gap. The distillation discussion sounds sharp—they argue most takes miss the nuance—but the transcript doesn't spell out their specific argument, so you'd need to listen to the full episode for that.

The numbers here are worth a click: Auto Intelligence matches Fable on user satisfaction at ~60% lower team cost; Auto Balance beats Opus 4.8 satisfaction at ~36% lower cost. Early enterprise accounts routing everything to Opus 4.8 saw 30–50% savings with no quality drop. The logic isn't complicated—it's a request classifier that looks at your query, context, task complexity, and domain, then picks the model Cursor knows does that job best. UI tweaks go to the model with the best taste, long-horizon problems hit reasoning models, simple stuff hits cheap models. Trained on 600k+ live requests, validated across millions of online A/B requests. I'd discount this on two fronts. One, the post doesn't disclose routing latency or cache-hit specifics—it only says the classifier was trained accounting for cache-miss costs. Whether the extra routing layer adds noticeable lag in practice is an open question. Two, this is Cursor's own blog, not a third-party eval. The direction of the numbers is credible, but I'd take the magnitude conservatively. For teams, this beats manually switching models. The catch: it works best if your usage pattern looks like Cursor's training distribution. If you're deep in a niche language or non-standard tasks, the routing might be less sharp.