While “tokenmaxxing”, the social media trend of maximizing AI token consumption as a vanity metric, takes off online, the physical infrastructure behind AI is slamming into serious bottlenecks. In this Five-Minute Friday, Jon Krohn maps out the four overlapping supply-chain constraints choking AI compute: GPUs (with NVIDIA Blackwell sold out through mid-2026), high-bandwidth memory (quintupled demand since 2023, only three manufacturers worldwide), CPUs (agentic AI requires 12x more CPUs per GPU than chatbots), and electricity (Gartner projects power shortages will restrict 40% of AI data centres by 2027). Find out why the five biggest hyperscalers are on track to spend $725 billion on AI infrastructure in 2026, where the reasons for optimism lie, and why Jon says you should definitely not tokenmaxx.<br><br>Additional materials:⁠ ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠www.superdatascience.com/992⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠<br><br>Interested in sponsoring a SuperDataScience Podcast episode? Email natalie@superdatascience.com for sponsorship information.

992: Tokenmaxxing vs AI Hardware Bottlenecks

Super Data Science: ML & AI Podcast with Jon Krohn

# Full Transcript

**Podcast:** Super Data Science: ML &amp; AI Podcast with Jon Krohn

## Transcript

This is episode number 992 on tokenmaxxing and the AI Hardware Bottleneck. Welcome back to the Super Data Science Podcast. I'm your host, Jon Krohn. The hottest social media trend in AI today seems to be tokenmaxxing, wherein folks use agentic tools like CloudCode and OpenAI's Codex to maximize the number of tokens they consume.

I don't know why someone would think this is an actual indicator of productivity, but apparently it stems, this trend of tokenmaxxing stems from companies tracking AI usage via internal dashboards. I hope you and your company aren't engaged in this vanity metric nonsense.

And, relatedly, today's topic on the podcast is the AI Compute Crunch. How the breakneck rush to train and deploy ever larger language models with people using ever more tokens from them is now slamming into very real, very physical bottlenecks across the global supply chain. You might have noticed this personally over the past several months. If you're a heavy CloudCode user, you probably bumped up against the weekly rate limits Anthropic introduced last August.

Anthropic was unusually candid about why. They said that a small fraction of power users were running Cloud24x7 in the background and eating into capacity for everyone else. The company has been blunt that they are, in their own words, very compute-constrained. OpenAI also famously paused new Sora signups to redirect scarce GPUs. GitHub at one point stopped accepting new subscribers for its co-pilot bot.

And on the developer side, on-demand H100 access, a popular NVIDIA GPU, through AWS, Azure, or Google Cloud has become genuinely unreliable for any team without pre-reserved capacity. So, what's actually going on? In short, demand for AI compute is rising substantially faster than the underlying hardware supply chain can scale.

And the bottleneck is no longer any one thing. There are at least four overlapping strategies that have to be solved simultaneously, and solving any one of them doesn't actually unstick the system. Bottleneck number one is the most famous, GPUs themselves. Lead times on NVIDIA's data center GPUs are now running 36 to 52 weeks.

And Blackwell allocation, the next generation after those popular hopper-style H100s, Blackwell allocation is largely sold out through mid-2026, with reported backlogs in the millions of units. Even those older NVIDIA H100 chips, originally launched back in 2022, have actually gotten more expensive to rent on the spot market, somewhere around 30% more expensive since November, because customers can't get hold of newer hardware and are falling back on previous generations.

But here's the wrinkle. GPU fabrication itself isn't really the binding constraint anymore. The real choke point sits one layer up at TSMC, the world's largest dedicated independent semiconductor foundry, who have a monopoly on state-of-the-art chips. So yeah, the real chok...

992: Tokenmaxxing vs AI Hardware Bottlenecks

Transcript

Built for investment professionals

Every day you wait, your competitors are finding
edge in conversations you're not hearing.

992: Tokenmaxxing vs AI Hardware Bottlenecks

Transcript

Built for investment professionals

Every day you wait, your competitors are findingedge in conversations you're not hearing.

Every day you wait, your competitors are finding
edge in conversations you're not hearing.