I started with GTA VI’s frame-rate debate. I own a PS5, but I realized that “60 fps looks smoother” was about as far as my understanding went. I wanted to get past the arguments and understand the trade-off between smoother gameplay and a detailed open world.

Then there’s my laptop. I run Qwen through Ollama on an M1 MacBook with 16GB of memory, and I’ve been wondering how much further I can push it before an upgrade becomes necessary. That led me to Google’s TurboQuant research — and to a useful question: what’s actually possible on my machine, and what just sounds exciting in a headline?

I also wrote about the five hackathons I entered without winning. That part isn’t a technical deep dive. It’s just an honest account of wanting something, not getting it, and finding myself excited about another idea anyway.

Not an expert’s guide. A record of learning.

I don’t want this series to become a collection of polished announcements where everything works and every project has a happy ending.

I want room for the questions, the wrong assumptions, and the experiments that go nowhere. The human part of being interested in technology.

The first entry is now on my Substack, along with the pencil-sketch illustration that sets the mood: a small figure sitting on the floor with a laptop, quietly figuring things out.

Read the full first entry on Substack: https://officialjeff.substack.com/p/still-figuring-it-out-1-gta-vi-local?r=yn1pq&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

If that sounds like your kind of reading, subscribe there for a new entry every Thursday!

The catch: the M6 is a base chip, so the older M5 Pro still has more graphics cores and memory. Apple’s new OLED touchscreen MacBook Pro, which starts at $2,899, runs M5 Pro and M5 Max chips, not M6. And don’t wait for an M6 Pro. Bloomberg reports Apple is skipping it and going straight to M7 in 2027.

So the cheapest MacBook Pro has the newest chip, and the fancy new one is a generation behind. If you own an M5, skip the M6. If you’re coming from Intel or an M1, it’s a big jump.

The AI Sell-Off

On November 5, the S&P 500 fell about 4.4%, its worst day since April 2025, and Nvidia dropped about 9%, wiping out its 2026 gain.

Was it the election? Not really. The market rose the day after Democrats won. The real worries are data center rules, profit-taking after a huge AI year, and doubts about whether all the spending will pay off. Texas’s Republican governor has already paused data center permits, so it’s not one party’s issue.

We’ve heard “AI bubble” before, going back to 2025. The drop was real. “AI is dead” is not.

Want the full breakdowns? The numbers, sources and what I’m watching next are on my Substack: https://officialjeff.substack.com/i/224263740/part-1-the-macbook-pros

Not financial advice. Figures come from press reports and change quickly.

AI agents are programs that can do tasks on their own inside a company, like opening files and changing settings. On Tuesday, CrowdStrike launched an “AI Kill Switch” that lets security teams stop an agent that goes rogue, keep it from restarting, and cut off its network access. It also works with Anthropic’s Claude Cowork, and CrowdStrike agreed to buy two AI security startups for about $495 million.

Here’s the twist: Nvidia launched its own kill switch in September, and the two work very differently. CrowdStrike cuts the agent off from the AI model that gives it instructions, like cutting the phone line between worker and boss. Nvidia puts the security guard on a separate chip the agent can’t touch. One is software, the other is hardware. Okta, Straiker, and ServiceNow have launched their own versions too, so this is turning into a race.

Nvidia’s RTX 5080 Vanishes From Its Store

Nvidia quietly pulled the RTX 5080 graphics card from its online store, after the RTX 5090 disappeared earlier. Nvidia hasn’t said it’s discontinued, but the rumor is that a new 24GB version is coming in December or January. If you want a top-end card, expect tight supply and high prices until then.

Why a kill switch might not be enough, which cards to buy or wait on, and what CrowdStrike’s shopping spree says about where AI security is going: full breakdown on my Substack: https://substack.com/home/post/p-223578466

For years, buying an AI server meant two separate purchases: a GPU from Nvidia, which does the AI math, and a CPU from AMD or Intel, which manages and feeds data to the GPU. That split is breaking down, and it’s changing how these three companies compete.

Nvidia now sells the whole rack — CPU included

Nvidia’s newest platform, Vera Rubin, bundles the GPU, CPU, and networking together as one system, all built by Nvidia. The CPU is called Vera, and it connects to the GPU faster than a separately-purchased chip could.

The pitch to a buyer like Microsoft isn’t that Vera necessarily beats AMD or Intel on paper — it’s that there’s no longer a reason to buy a separate CPU at all when a fast one already comes with the GPU. That’s the real risk to AMD and Intel: not losing a spec-sheet fight, but losing the sale before it’s even considered. Nvidia has reportedly pointed to around $20 billion in server CPU demand already, compared to AMD’s entire data-center CPU business, which brings in about $5.8 billion a quarter.

Two separate CPU stories, not one

It’s easy to blend these together, so here’s the clean version:

Nvidia’s data-center CPU (Vera) is Nvidia’s own chip, built entirely in-house, and it doesn’t use x86 — the standard chip language that’s powered PCs and servers since the 1980s and that Intel and AMD both build on. Instead, Vera uses a different but common design called Arm. This is the CPU bundled into Nvidia’s AI racks.

The Intel-Nvidia partnership is a separate project. The two companies are jointly developing x86 chips — Intel’s traditional chip language — that combine Intel’s processors with Nvidia’s graphics technology. This is mainly aimed at PCs and laptops, not data centers. Nvidia invested $5 billion into Intel as part of this deal.

So Nvidia is competing with Intel and AMD in data centers with its own CPU, while at the same time working with Intel on a different product for a different market. Both things are happening at once, for different reasons.

AMD’s approach: staying open instead of locked-in

AMD’s strategy is built around two things that make it easier for customers to leave Nvidia’s ecosystem, or avoid getting fully locked into it in the first place.

Software: AI programs have to be written in a way a specific chip can run. Nvidia’s version of this is called CUDA, and it’s closed — Nvidia controls it, and it’s built to work best on Nvidia hardware. Over time, so much AI software was written specifically for CUDA that switching away from Nvidia became difficult even for companies that wanted to. AMD’s alternative is called ROCm, and it’s open-source, meaning anyone can view and build on the underlying code. The goal is to make it realistic for developers to write AI software that also runs well on AMD chips, not just Nvidia’s.

Hardware connections: AI data centers run thousands of GPUs wired together, and those chips need a way to communicate quickly. Nvidia’s connection technology, NVLink, only works between Nvidia parts. AMD, along with Intel, Google, and Meta, helped create an open standard called UALink that any company’s chips can use to connect. The idea is that a data center built on this open standard isn’t required to buy every component from a single vendor.

A quick example of what this looks like in practice

Say a company is building a data center and wants to use AMD’s GPUs, but they already have networking switches from Broadcom, a separate company that makes networking gear. Because AMD’s system runs on open standards, that combination works — AMD GPUs, Broadcom switches, no problem.

Try to do the same thing with Nvidia, and it’s a different story. Nvidia has reportedly been tightening its ecosystem specifically to keep outside networking gear, including Broadcom’s, from working smoothly with its GPUs. So a company that wants Nvidia’s GPUs is more likely to end up buying Nvidia’s networking too, whether that was the plan or not.

That’s the practical difference: AMD’s setup lets a buyer keep pieces of their existing infrastructure and add AMD GPUs into it. Nvidia’s setup tends to pull a buyer toward an all-Nvidia stack over time. This is AMD’s core pitch to potential customers: build with open standards now, and you keep the option to switch or mix vendors later instead of being fully dependent on one company. OpenAI has already committed to a large, multi-year deal with AMD partly on this basis. Nvidia’s head start with CUDA is still enormous, so it’s an unresolved competition, not a settled one.

Will Intel become a major foundry again, like TSMC?

This is still an open question. Intel’s new manufacturing process, called 18A, is its attempt to start building chips for other companies, the way TSMC does today. Intel has landed a few real external customers, including Amazon, Microsoft, and Japan’s Socionext, which is meaningful progress. But TSMC still manufactures around 95% of the world’s advanced AI chips, and there have been reports that Intel may scale back offering 18A to outside customers and shift focus to its next process instead. Intel isn’t likely to overtake TSMC in the near term. Its more realistic goal is proving its factories are reliable enough to keep attracting customers over time.

Bottom line

Nvidia is trying to make the GPU-versus-CPU decision unnecessary by selling a single bundled system. AMD is betting that openness, in both software and hardware connections, will matter enough to customers over time to earn it a real share of the market. Intel is pursuing a longer-term strategy: partnering with Nvidia in one part of the business while trying to prove its manufacturing can compete with TSMC in another. None of these outcomes are decided yet.

Written by: officaljeff