Two revelations this week have shaken the narrative in AI : Nvidia’s earnings & this tweet about Gemini.

The AI industry spent 2025 convinced that pre-training scaling laws had hit a wall. Models weren’t improving just from adding more compute during training.

Then Gemini 3 launched. The model has the same parameter count as Gemini 2.5, one trillion parameters, yet achieved massive performance improvements. It’s the first model to break 1500 Elo on LMArena & beat GPT-5.1 on 19 of 20 benchmarks.

70% of production AI teams use open source models. 72.5% connect agents to databases, not chat interfaces. This is what 375 technical builders actually ship - & it looks nothing like Twitter AI.

70% of teams use open source models in some capacity. 48% describe their strategy as mostly open. 22% commit to only open. Just 11% stay purely proprietary.

Ten months ago, DeepSeek collapsed AI training costs by 90% using distillation - transferring knowledge from larger models to smaller ones at a fraction of the cost.

Distillation works like a tutor training a student : a large model teaches a smaller one.¹ As we’ve shifted from knowledge retrieval to agentic systems, we wondered if there was a parallel technique for tool calling.²

Could a large model teach a smaller one to call the right tools?

By Monday lunch, I had burned through my Claude code credits. I’d been warned ; damn the budget, full prompting ahead.

When that still couldn’t solve the issue, I summoned Opus, the biggest & most expensive model, to solve it.

Datadog is becoming a platform company, & its Q3 2025 results underscore how successful this transition is. If nothing else, the consistency around 25% growth for the last 12 quarters exemplifies this point.

Net dollar retention underpins this growth, combined with accelerating new customer account acquisition. One of the biggest changes in the last five quarters is terrific cross-selling across an increasingly large product suite.

Just how much are we spending on AI?

Compared to other massive infrastructure projects, AI is the sixth largest in US history, so far.

World War II dwarfs everything else at 37.8% of GDP. World War I consumed 12.3%. The New Deal peaked at 7.7%. Railroads during the Gilded Age reached 6.0%.

AI infrastructure today sits at 1.6%, just above the telecom bubble’s 1.2% & well below the major historical mobilizations.

I wrote a post titled Congratulations, Robot. You’ve Been Promoted! in which OpenAI declared that their AI coders were no longer junior engineers but mid-level engineers.

The post triggered the largest unsubscription rate in this blog’s history. It was a 4-sigma event.

A Nassim Taleb black swan, this was something that should happen once every 700 years of a blog author’s career.

Clearly, the post struck a nerve.

In a job market 13% smaller for recent grads than recent years, a subtle fear persists that positive developments in AI accuracy & performance accelerate job losses. Stanford’s research found :

OpenAI has committed to spending $1.15 trillion on hardware & cloud infrastructure between 2025 & 2035.¹

The spending breaks down across seven major vendors: Broadcom ($350B), Oracle ($300B), Microsoft ($250B), Nvidia ($100B), AMD ($90B), Amazon AWS ($38B), & CoreWeave ($22B).²

Using some assumptions, we can generate a basic spending plan through contract completion.³

Across these vendors, estimated annual compute spending grows from $6B in 2025 to $173B in 2029, reaching $295B in 2030. We built a constrained allocation model with the boundary conditions defined in the appendix below, but this is just a guess. The actual growth rates are 124% (2027→2028), 54% (2028→2029), & 70% (2029→2030).

I sleep better knowing my agents work through the night. Less work for me in the morning.

My podcast processor transcribes & analyzes conversations. I started on my laptop, needed a little database to collect podcast data & metadata, & booted up a DuckDB instance.

But then the data started to grow, & I wanted the podcast processor to run by itself. I changed two little letters, & the database moved to the cloud :

Microsoft & Google both announced earnings yesterday, & the scale of AI adoption remains staggering. The infrastructure businesses are growing at accelerating growth rates that are the envy of businesses one-hundredth the size.

Capex

Capital expenditures were $34.9 billion, driven by growing demand for our Cloud and AI offerings. This quarter, roughly half of our spend was on short-lived assets, primarily GPUs and CPUs…. We will increase our total AI capacity by over 80% this year and roughly double our total data center footprint over the next 2 years, reflecting the demand signals we see. Just this quarter, we announced the world’s most powerful AI data center, Fairwater in Wisconsin, which will go online next year and scale to 2 gigawatts alone. - Microsoft