Perf-per-Watt

I am saying this with great embarrassment, I don't think I had fully internalized until today that big AI compute deals are announced... in MW or GW (GW being the new MW). It has probably been 6 months but it used to be that compute was bought by the GPU bucket. In the past couple of weeks alone: 1️⃣ NVIDIA and IREN (up to 5 GW): May 7, 2026, IREN (formerly a bitcoin miner) partnered with Nvidia to accelerate the deployment of up to 5 GW of next-generation AI infrastructure. 2️⃣ OpenAI and NVIDIA/Broadcom (10 GW goal): announcements this week indicate a "monumental" partnership where NVIDIA plans to invest up to $100 billion to support OpenAI in building data centers requiring 10 GW of power. Who's even counting the 4–5 million GPUs anymore? 3️⃣ Anthropic and Amazon/Google/Broadcom (Multiple Gigawatts): Anthropic announced massive expansions in April/May 2026, including an up to 5 GW agreement with Amazon (with ~1 GW new by late 2026) and a 5 GW agreement with Google/Broadcom starting in 2027. 4️⃣ Anthropic and SpaceX/Colossus 1 (300+ Megawatts): this week again, a "baby deal" by current standards, but one that captures the imagination for obvious reasons as it gives Anthropic full access to Colossus 1. A great Substack article by Michael Sanie lays out the reason why "perf-per-watt" is the new currency. Power has become the absolute bottleneck, accessing electricity from the grid can take years. But when I read that gigawatts provide a standardized, normalized metric to compare AI capability regardless of whether the system uses thousands of H100s, H200s, or future chips... I find this weird, as this is ignoring efficiency gains in software and architecture, including the vast differences in energy performance of available chips. So I like Perf-per-Watt, but just Watts?

https://sanieinstitute.substack.com/p/gigawatts-in-tokens-out-why-perf

Workflows-Schmorkflows

I haven't found anyone who has cracked the toughest problem of all: what should a workflow/process even look like to create the most synergistic human-AI value? Especially with super fast changing capabilities that can make a redesigned process obsolete before it is implemented. Clinical decision support, which I'm close to, is a great example: AI is getting great at diagnosis but in a way that is hard to integrate into a physician's workflow without disrupting many other things that are often even more important than an accurate diagnosis (see Graham Walker, MD). A job is not just a sequence of tasks, neither is a process, it is an integrated object. Reinventing processes in the age of AI demands a theory of centaurs that does not yet exist. The human in the loop approach gets it wrong.