Earlier this week, as I was typing my update post about the midterms, I thought “Everyone wants to talk about Jensen and the spy, but I don’t have much to say, so I’m going to leave it alone.”
Then, he resigned.
So I’m going to do a double-post today. One about Jensen and one about the midterms.
Jensen Huang & NVIDIA
I, frankly, don’t care about Alice Huang (no relation to Jensen obviously), the Chinese national who was an executive at Singapore-based Megaspeed. I haven’t seen any compelling evidence that she’s actually a spy besides the U.S. government’s saying “trust us, she’s a spy.” I really, really, really don’t think that Jensen’s pillow talk with her is about chip architecture.
“What about violating export controls?” My personal opinion on NVIDIA export controls is irrelevant to the case. Megaspeed and its affiliates account for like $2B per year in NVIDIA revenue. For a company whose revenue was $215B last year, that’s nothing.
But, an inappropriate relationship with a VP is irresponsible and has clearly led to his resignation. This is a giant mess, but the world keeps spinning, so, let’s pick up the pieces and look briefly at how a few of the big companies are impacted.
Anthropic
Anthropic, in a deal a few years ago with Amazon, had to build its architecture to work on Amazon’s Trainium chip. It cost them in the near-term because NVIDIA chips (and software) were easiest to build for, but now they’re hardware-agnostic. They also use Google’s TPUs as well, so their model is not heavily reliant on NVIDIA GPUs.
Google’s, first of all, got TPUs, which are great chips. They can start shifting workloads in both GCP and for their own services. Google’s got the whole AI stack integrated, and now they’ve got a distracted competitor.
Open AI
Open AI’s architecture is much more dependent on NVIDIA chips. I say that they are only temporary losers because AI has gotten so good at writing code that it can probably assist Open AI in rewriting architecture that needs to be generalized away from NVIDIA.
The Dark Horse – Apple
Apple. Apple’s M-series chips are insane. Spend $45,000 and you can run a production Meta Muse Spark (816B parameters) which would be like running GPT-6 Luna. Considering the only power you need to run that is what you can get out of a standard wall outlet, that’s pretty amazing.
I would expect that NVIDIA’s innovation roadmap and order book are pretty set for the next few years, so not much will change near-term, but, come the 2028 election, I think that we’ll start to see changes, and Apple may end up in a pretty great position for local AI for your personal workflows.
Speaking of 2028…
The Midterms
Now that everyone is uninterested in talking about the midterms, I’m going provide my opinion.
AI was on the ballot and it lost nearly every time
40 of 41 local ballot initiatives for new data centers failed. Only North Dakota’s initiative passed.
Texas Governor Greg Abbott had already passed a moratorium on new data center until 2027. Texas has passed a new moratorium on data centers pending further review by the state legislature. The Texas constitution mandates that the legislature convene only every two years, if that review (the legislature meets in January 2027) doesn’t yield more approvals, Texas permitting may stall.
This is a big thing. Texas has lots of land and lots of power, and so “there’s always Texas” was the tech industry’s backup plan. Now that Texas is off the table, at least short-term, it’ll be interesting to see what happens with datacenters.
Datacenters will move abroad (and to space)
GPU racks can be popped in space by SpaceX (and, surely, they’ll do that in time). Google is building a datacenter in Malaysia. There’s lots of space in Brazil, Argentina, and Chile to pop up datacenters. I think it is a fool’s errand to believe that datacenters will stop being built. I think, most simply, they’ll move away. I haven’t mentioned the Middle East because of the war, but there’s plenty of land and oil over there for them as well.
These ballot initiatives are for new data centers, not previously-permitted ones
Datacenters will still be built into 2028 and 2029, so we can expect AI progress to continue for a few years barring Congress regulating AI or creating an FDA for AI.
This is the biggest thing. AI will be squarely on the ballot again in 2028 because progress will not appear to be slowing like the public thinks it may. Congress will have difficulty obtaining a supermajority to override the President’s veto on an AI FDA.
The people are giving the middle finger to big tech
AI has been the worst-marketed technology in history. Here’s the AI sales pitch from the Valley:
“We’re going to have taxpayers fund the development of our datacenters which are going to tear up the roads, make electricity prices rise, and be loud as hell. Then, we’re going to get hilariously rich building out the technology that will probably take their jobs.”
When the factories got moved overseas, the communities didn’t have a say. Now, they do. And they’re saying “absolutely not.”
Your AI initiatives can continue
Most businesses that we work with don’t need the most advanced, most GPU-intense AI. I think that adoption will continue through the incoming headwinds that arrived on Election Day. Cleaning up technologies and data and linking systems together and actually doing work can all continue.
I’m not talking my book here. They really can continue because, most of the time, the most advanced AI isn’t needed to get the job done. That’s the nice way of saying “most businesses are disorganized messes, and regulating AI doesn’t mean the house can’t get cleaned up.”
