Computing Power for AI: The Arms Race

If you want to win in AI - the hottest market ever - one hard truth stands: you need a ton of computing power.

Oracle's Billions-Dollar Bet

Larry Ellison isn't playing around. Oracle's co-founder and CTO announced a jaw-dropping plan: invest $100 billion over the next three years to supercharge computing power for AI models.

What's the game plan? Oracle aims to roll out autonomous systems. Imagine databases that handle security, tuning, and backups all by themselves.

There will be a deep integration of AI into their cloud offerings - automated analytics, business processes, and decision-making.

They're already building an 800-megawatt data center for their competitive LLM.

But why stop there? They've got an even bigger 1-gigawatt facility in the works, slated to be powered by three small nuclear reactors by 2030. Yes, they are going nuclear in the AI race.

Elon Musk’s xAI: problems in Memphis?

Meanwhile, Musk admits it's a "very difficult engineering problem" to train Grok 3 because the public power grid just isn't cutting it.

Having own nuclear reactors might be an essential move.

Colossus, xAI's data center in Memphis, currently housing 100,000 Nvidia H100s and set to double soon. That's 200 megawatts—a fifth of Oracle's planned capacity. (By 2030 xAI probably has doubled as well 1-3 times.)

But there's more trouble for Elon. Locals and city council members are worried about Colossus's environmental impact, especially its colossal water usage.

We're talking up to 1.5 million gallons (over 5.6 million liters) of water daily. Unless xAI addresses these concerns, they might face regulatory delays, skyrocketing costs, or be forced to change course.

Not exactly smooth sailing for Musk's AI ambitions.

Groq and Aramco: Building the Behemoth

Not to be outdone, Groq is teaming up with Aramco Digital to build the world's largest AI inferencing data center in Saudi Arabia, aligning with the Kingdom's Vision 2030 (deep pockets).

Using Groq's Language Processing Units (LPUs), which are essentially LLM-specific hardware, they're aiming to process 53 million tokens per second by 2025—that's 3,500 Lord of the Rings trilogies every second.

This center will offer AI capabilities through an 'as-a-service' model on Aramco Digital's marketplace, Nawat.

The New Generation of AI Models

All of this horsepower will yield new, stronger models. They will have a higher IQ than we have.

o1 by OpenAI

o1 is the first that surpassed the average human; by far with an IQ of 120:

OpenAI's new o1-preview model emphasizes advanced reasoning, allowing it to "think through" problems.

This advancement enables the AI to tackle very hard problems that require planning and iteration, such as novel math or science questions.

I tested it on a couple of unsolved math problems. Did it crack them? No. But the way it approached the problems was impressive.

It hypothesized, coded, tested, and iterated. No eureka moment, but the reasoning process was something to behold.

It probably also needed someone on its side who is much stronger in math than me. 😭 

o1 isn't perfect. It's not a better writer than GPT-4o and still has its fair share of hallucinations. But people are building awesome stuff:

A weather iOS app in under 10 min.

Or chess.

World Labs: Bringing Spatial Intelligence to AI

Then there's World Labs, founded by AI heavyweight Fei-Fei Li and a team of computer vision experts. They're zeroing in on spatial intelligence, creating AI that can understand and interact with 3D environments.

Their mission? Build Large World Models enabling AI to perceive, generate, and engage with 3D worlds—both virtual and real.

Potential applications include entertainment, urban planning, architecture, robotics, and more immersive VR experiences.

Backed by over $230 million from big-name investors like Andreessen Horowitz, NEA, Marc Benioff, and Eric Schmidt, they're tackling the overlooked frontier of spatial intelligence in AI.

As AI edges closer to higher intelligence (spatial intelligence is one step closer to it), figuring out how we collaborate with these systems becomes crucial—a puzzle that o1 and other models themselves can't solve.

The "Buttonization" Problem

But here's the kicker: Despite these leaps, AI's potential is being stifled by what Evan Armstrong calls "buttonization." 

We're cramming AI into existing workflows as simple, click-and-forget features—buttons—rather than letting it revolutionize how we work.

Sure, AI spruces up tools like email apps and word processors. But these are incremental improvements, not the shifts in productivity we were promised.

Slapping an AI feature onto old software is like putting a jet engine on a horsecart.

We need to rethink workflows from the ground up.

AI should automate or eliminate tasks, not just make them a tad quicker.

The current setup favors big players who can easily integrate AI into their existing products, making life tough for startups—unless they think outside the box.

Startups: Time to Break the Mold

So, how can startups compete?

1. Be Radically Faster: Co-develop products with AI like Claude AI or o1 to outpace incumbents. (Btw. batch 01 of Everyone can Code! ✨ is now permanently closed. Batch 02 might open end of year. Join Waitlist.)

1. Reinvent Workflows: Don't just add AI features; create entirely new workflows or business models that can't be easily copied.

Look at Replit's AI agent. It automates complex tasks by letting users build apps through simple prompts—it is an autonomous AI agent.

We need AI that doesn't just help us do tasks faster but takes over tasks altogether—like handling emails without human intervention.

The AI race isn't just about who has the most computing power or the smartest model (the basics). It's about who can break free from the shackles of traditional workflows and let AI reach its full potential.

And this can be done best in verticals. 🤔