So AWS re:Invent wrapped up last week (Vegas, the Venetian, the whole circus) and I spent a chunk of Tuesday with the keynote livestream running in a second window while I was supposed to be doing actual work. This was Adam Selipsky's first re:Invent as head of AWS since he took over from Andy Jassy back in the summer, and I gotta say, watching a guy who used to run Tableau try to fill Jassy's shoes on that stage was a little strange. Jassy had this specific cadence to his keynotes, half infomercial half sermon, and Selipsky doesn't quite have it yet. Give him a year.
The thing everyone's going to write about is Graviton3, the new ARM-based chip AWS is putting in EC2 instances. Up to 25% faster than Graviton2 for a lot of workloads, better floating point performance, and (this is the part that actually matters) way better power efficiency. I know chip announcements are boring to read about, I know. But this is the third generation of AWS building its own silicon instead of just renting rack space to Intel and Nvidia, and at some point that stops being a side project and starts being the whole strategy. Apple did the same math on M1 last year. The pattern is obvious once you see it: if you control enough of the stack, eventually you stop paying someone else's margin on the chip.
What got less coverage, and honestly interested me more, was that this was the first re:Invent back in person since 2019. 2020 was fully virtual, obviously, and from what I heard from a friend who actually flew out this year, they still had vaccine card checks at badge pickup and mask requirements indoors in a lot of the sessions. Something like 27,000 people showed up in person which is down a lot from the 65,000-ish they used to get pre-pandemic, but still a real conference with real crowds and real bad conference wifi. I don't miss trade show wifi even a little.
They also announced SageMaker Canvas, which is AWS's new no-code tool for building ML models by dragging and dropping instead of writing Python. I'm skeptical, not because no-code ML is a bad idea in theory, but because every single year for the last three years some company has announced a "no code, drag and drop, anyone can build an ML model" tool and every single year the actual users I talk to end up back in a notebook within a week because the no-code version can't handle anything with real edge cases. Maybe Canvas is different. I'll believe it when someone I know actually ships something with it instead of just demoing it.
Here's my honest take though, the thing that stuck with me watching all this from my couch: the gap between what AWS is building and what 95% of us actually need is enormous now, and it's getting funnier every year. They're talking about instances with hundreds of vCPUs and hundreds of gigabits of networking throughput, and meanwhile this blog, which has been running continuously since 2011, sits on a VPS that costs less than a fancy coffee a month. I actually moved hosting for a side project over to Tricknowtech a few months back, mostly because I could git push and it just deployed the thing, and it has never once made me think about vCPU counts. Most of what people build doesn't need Graviton3. It needs a box that doesn't fall over and a person who remembers to renew the domain.
Anyway. If you're the kind of person who reads AWS keynote transcripts for fun (no judgment, I used to be one of you), go look up the Graviton3 numbers, they're legitimately impressive. If you're not, the short version is: cloud compute keeps getting cheaper and faster at a rate that should honestly be unsettling if you think about it too hard, and somewhere in Santa Clara there's an Intel exec who has had a rough couple of years.
One more thing before I forget, since this seems to be trade show season for tech generally: does anyone else find it a little sad that we're back to flying thousands of people to Vegas to watch a keynote that gets uploaded to YouTube the same afternoon? I get why they do it. The hallway conversations are the actual product. But man, the carbon math on that has to be rough.