Cold War Escalation, OpenAI&Broadcom, & More

TLDR;

This week, we take a look at the latest grand deal between OpenAI and Broadcom, new interesting news from companies like Cisco, and the Netherlands’ unthinkable attack on China, as well as what might be the most exciting AI IPO ever.

Enjoy!

CHAT ASSISTANTS

Porn is Finally Coming to ChatGPT

In a quick announcement, Sam Altman has shared a new policy stating that they now feel they can age-restrict certain content and treat adults as such. Which is to say, porn is coming to ChatGPT (he literally mentioned ‘erotica’ as one of the things they were unlocking).

TheWhiteBox’s takeaway:

I’m happy to see age restrictions; AI slop is literally frying kids’ brains. Now, I’m all for treating adults as adults, but does society need a new wave of ever-horny, poorly-socialized, porn-obsessed adults in love with a digital file they took from chatgpt.com?

Am I weird for not understanding how this isn’t causing outrage? I think it’s important I share my biases too: I firmly believe porn is bad, very bad for you. Porn has been linked with several issues, like erectile dysfunction; it correlates quite literally with smaller brains, and severely decreases your appetite for real sex.

Now, picture this potentially unfathomable damage and a technology that generates it at scale, on demand, and adapted to your porn-addicting mind. What could go wrong?

HARDWARE

Cisco and the emerging importance of “Scale across”

Cisco has presented a new chip, the Silicon One P200, and a router to connect long-distance data centers (more than 1,000 km, or 625 miles) for AI workloads. As explained by the company itself, they already include several Hyperscalers as clients.

The programmable system can enable up to 3 Exabits per second of interconnect bandwidth, or three quadrillion bits per second. This scale sounds unfathomable, but it is precisely the level of communication speeds that AI workloads require.

But why is this a big deal?

TheWhiteBox’s takeaway:

In this newsletter, we have talked about scaling up and out several times.

Scaling up is increasing the number of accelerators (usually, GPUs) per server, which is what most providers like NVIDIA are doing (from 8 GPUs per server on the Hopper generation, to now having 72 GPUs per server with Blackwell, and pushing to 144 next year with Rubin and even more than 500 in 2027).

On the other hand, scaling out involves, once you reach the limits of scale-up, connecting several servers (becoming ‘nodes’) that, combined, build your data center. This turns a single server into something like xAI’s Colossus 2 in Memphis, where they’ll have 1.1 Gigawatts of GPUs interconnected.

At this point, we are still talking about connecting GPUs in one location. What Cisco is talking about here is scaling across locations (i.e., connecting entire data centers together).

This isn’t something entirely new; Google has been doing this for years now (it’s believed that Gemini 2.0 was already a multiple data center effort, which is why you can’t be sufficiently bullish about Google; they are incredibly…

Enjoyed the introduction? Subscribe or sign in to continue with this edition.