Tech Stories

Issue #66

Hello Reader,

Can't believe we're already midway through September and crossing into the final quarter of the year soon.

Today, let's talk about the continued rise of liquid cooling, Nvidia's plot twist in the data centre, and the gaming of social media.

The liquid cooling landscape expands

These days, any talk about liquid cooling almost always centres on direct-to-chip (DTC) technology, which pipes chilled water directly to cold plates attached to CPUs or GPUs. Liquid cooling isn't new technology, but the AI rush and its power-hungry GPUs have vastly accelerated adoption.

Single-phase DTC emerged as the technology of choice. It delivers solid cooling performance, offers quick time-to-market, and was easiest to build an ecosystem around. But other viable options are now entering the mainstream.

Last week, Firmus Technologies gained approval to build Australia's largest AI data centre. Their privately-funded Project Southgate will deliver 44MW of capacity in Stage 1a. What caught my attention isn't the scale - it's Firmus Technologies' unique approach to immersion cooling.

You see, everyone else doing immersion cooling uses tanks of dielectric liquid with individual power-guzzling pumps and piping serving each tank. Firmus Technologies arranges their tanks serially using clever pipework, with just one pump at the end. The efficiency of the system makes it literally the most energy efficient liquid cooling implementation that I know about.

Then there's two-phase DTC, which I finally got down to reading up about. Two-phase DTC, which uses refrigerant, offers multiple advantages: superior heat transfer performance, no risk of damage from leaks, and minimal concerns about corrosion or biological growth - all current headaches with single-phase DTC.

Single-phase DTC might have had first-mover advantage when it comes to AI, but the competition is heating up.

Nvidia's Rubin CPX plot twist

This week Nvidia pulled a plot twist with its new Rubin CPX chip. We've known that AI inference will overtake AI training at some point - every chart I've seen over the last two years shows this trajectory. And logically, it has to happen.

Well, it would appear we're approaching the inflection point where there's more money in inference than training. Or perhaps Nvidia is moving decisively before inference-only competitors such as Groq gain a substantial share of the market. At least, that's how it looks at first blush.

Rubin CPX represents a completely new class of GPU designed specifically for post-training workloads. It uses much cheaper GDDR7 RAM in higher quantities, features a monolithic die to lower costs, and maintains the 130kW-rack energy footprint of existing Blackwell GPUs.

What's clever here is Nvidia's new disaggregated inference architecture. The Rubin CPX handles the prefill and context phases of generative AI, reducing the load on more expensive, full-fledged GPUs which can then focus on the actual generation phase.

But here's where my mind went: What if the real reason for Rubin CPX is that Nvidia isn't confident the data centre industry will be ready for the second wave of Rubin Ultra GPUs with their 600kW rack requirements? That would make a lot more sense, wouldn't it?

The gaming of social media

Most, if not all of you, signed up for this newsletter after reading some of my LinkedIn posts. I'm still posting every day there, but here's the thing: it's been getting harder to be seen on LinkedIn for a while now. Sure, I'm still averaging 200 engagements per post (over 365 days), but it was 240 engagements per post and rising earlier this year.

What's happening? A conference session I attended on "Organic LinkedIn growth" in San Francisco last week crystallised the problem perfectly. I wrote about it in "The dirty truth about LinkedIn's top influencers," but here's the crux: The platform is drowning in black hat tactics and engagement pods. And I refuse to play that game.

Yes, I've always maintained that I write for myself. Because writing is thinking. And writing daily ensures I write more, not less. It's how I stay current on data centres, AI, cybersecurity, sustainability, and enterprise IT. All of that still rings true. But it doesn't feel right when recycled and one-dimensional content gets pushed to the front because 200 bots clicked on "like" in the first hour.

Many of you do take the effort to chime in on my LinkedIn posts, which influences the algorithm, and for which I am genuinely grateful. But many of you cannot because of your seniority or organisation - and I know because you either text me or tell me in person how much you enjoyed reading my posts despite rarely engaging. Others, sadly, simply don't my posts as often anymore due to the rampant gaming.

So how do I broaden reach whilst keeping things purely organic?

I've started diversifying beyond LinkedIn, including my Clearly Tech Substack, which crossed 200 subscribers yesterday (yay). When I started it, I've meant to rerun this commentary there... except it seemed to have taken on a life of its own.
This week, I've also launched a "Tech News" Channel on WhatsApp where I share new LinkedIn posts the moment they're published. It's anonymous - nobody can see your name or number. If you're interested, you can join here.

So, there you have it: I just wrote the world's most overwrought WhatsApp channel announcement.

As usual, you can reach me by hitting reply.

Regards,
Paul Mah.

You can now comment on stories published on techstories.co. If you received this newsletter from me, your account was created automatically.

All you need to do:

Scroll to Member discussion and sign in with the email used for this newsletter.
An email message with a magic link is sent immediately. Click to log in.
Comment away.

Spotlight