From MIPS to exaflops in mere decades: Compute power is exploding, and it will transform AI

At the recent Nvidia GTC conference, the tech giant unveiled a groundbreaking achievement: the first single-rack system of servers capable of reaching one exaflop. That's a mind-boggling one billion billion floating-point operations (FLOPS) per second. This feat is powered by Nvidia's latest GB200 NVL72 system, featuring the cutting-edge Blackwell graphics processing units (GPUs). To put it into perspective, this system fits into a standard computer rack that's about 6 feet tall, a bit over 3 feet deep, and less than 2 feet wide.

Shrinking an Exaflop: From Frontier to Blackwell

The announcement got me thinking about how far we've come in just a few years. The world's first exaflop-capable computer, "Frontier," was installed at Oak Ridge National Laboratory in 2022. Built by HPE and powered by AMD GPUs and CPUs, it spanned a massive 74 racks. In contrast, Nvidia's new system packs 73 times more performance into a single rack. That's like tripling performance every year for three years straight! It's a testament to the incredible strides made in computing density, energy efficiency, and architectural design.

It's also worth noting that while both systems reach the exascale milestone, they're designed for different purposes. Nvidia's exaflop rating is based on lower-precision math—4-bit and 8-bit floating-point operations—ideal for AI workloads like training and running large language models (LLMs). These calculations favor speed over precision. On the other hand, Frontier's exaflop rating comes from 64-bit double-precision math, the gold standard for scientific simulations where accuracy is paramount.

We've Come a Long Way (Very Quickly)

The pace of this progress is almost hard to believe, especially when I think back to the early days of my career in computing. My first job was as a programmer on the DEC KL 1090, part of DEC's PDP-10 series of timeshare mainframes. That machine chugged along at a modest 1.8 million instructions per second (MIPS). It connected to cathode ray tube (CRT) displays via hardwired cables, with no graphics to speak of—just text on a screen. And of course, there was no Internet. Remote users had to connect over phone lines with modems that topped out at 1,200 bits per second.

500 Billion Times More Compute

While comparing MIPS to FLOPS isn't a direct apples-to-apples comparison, it gives us a sense of the incredible leap in computing power. MIPS measures integer processing speed, great for general-purpose computing and business applications. FLOPS, on the other hand, measures floating-point performance, crucial for scientific workloads and the heavy lifting behind modern AI, like the matrix math and linear algebra needed to train and run machine learning (ML) models.

Using these as a rough guide, the new Nvidia system is roughly 500 billion times more powerful than the old DEC machine. It's a stunning example of the exponential growth in computing power over a single career. It makes you wonder: If we've achieved this much in 40 years, what could the next five years bring?

Nvidia has given us some hints. At GTC, they shared a roadmap predicting that their next-generation full-rack system, based on the "Vera Rubin" Ultra architecture, will deliver 14 times the performance of the current Blackwell Ultra rack. We're looking at somewhere between 14 and 15 exaflops in AI-optimized work within the next year or two.

It's not just about raw power, either. Squeezing this performance into a single rack means less physical space, fewer materials, and potentially lower energy use per operation. Though let's not forget, the absolute power demands of these systems are still huge.

Does AI Really Need All That Compute Power?

While these performance gains are impressive, the AI industry is now wrestling with a big question: How much computing power do we really need, and at what cost? The race to build massive new AI data centers is driven by the demands of exascale computing and increasingly capable AI models.

The most ambitious project is the $500 billion Project Stargate, planning 20 data centers across the U.S., each half a million square feet. There's a wave of other hyperscale projects underway or in the works around the world as companies and countries rush to build the infrastructure for future AI workloads.

Some analysts are worried we might be overbuilding. Concerns grew after the release of R1, a reasoning model from China's DeepSeek that uses significantly less compute than its peers. Microsoft's recent cancellation of leases with multiple data center providers has fueled speculation that they might be rethinking their AI infrastructure needs.

However, The Register suggested that this pullback might be more about the planned data centers not being able to handle the power and cooling needs of next-gen AI systems. AI models are already pushing the limits of current infrastructure. MIT Technology Review reported that many data centers in China are struggling and failing because they were built to outdated specifications.

AI Inference Demands More FLOPs

Reasoning models do most of their heavy lifting at runtime through a process called inference. These models power some of the most advanced and resource-intensive applications, like deep research assistants and the emerging wave of agentic AI systems.

While DeepSeek-R1 initially led the industry to think future AI might need less computing power, Nvidia CEO Jensen Huang pushed back hard. In a CNBC interview, he argued the opposite: "It was the exact opposite conclusion that everybody had." He added that reasoning AI consumes 100 times more computing than non-reasoning AI.

As AI evolves from reasoning models to autonomous agents and beyond, the demand for computing power is likely to surge again. The next breakthroughs might come in areas like AI agent coordination, fusion simulations, or large-scale digital twins—all made possible by the kind of computing leap we've just seen.

Seemingly right on cue, OpenAI announced $40 billion in new funding, the largest private tech funding round on record. In a blog post, they said the funding "enables us to push the frontiers of AI research even further, scale our compute infrastructure and deliver increasingly powerful tools for the 500 million people who use ChatGPT every week."

Why is so much capital flowing into AI? The reasons range from competitiveness to national security, but one factor stands out: AI could increase corporate profits by $4.4 trillion a year, according to McKinsey.

What Comes Next? It's Anybody's Guess

At their core, information systems are about simplifying complexity, whether it's an emergency vehicle routing system I once wrote in Fortran, a student achievement reporting tool built in COBOL, or modern AI systems accelerating drug discovery. The goal has always been the same: To make sense of the world.

Now, with powerful AI on the horizon, we're crossing a threshold. For the first time, we might have the computing power and intelligence to tackle problems that were once beyond human reach.

New York Times columnist Kevin Roose captured this moment well: "Every week, I meet engineers and entrepreneurs working on AI who tell me that change—big change, world-shaking change, the kind of transformation we've never seen before—is just around the corner." And that's not even counting the breakthroughs that arrive each week.

Just in the past few days, we've seen OpenAI's GPT-4o generate nearly perfect images from text, Google release what may be the most advanced reasoning model yet in Gemini 2.5 Pro, and Runway unveil a video model with shot-to-shot character and scene consistency, something VentureBeat notes has eluded most AI video generators until now.

What comes next is truly anyone's guess. We don't know whether powerful AI will be a breakthrough or breakdown, whether it will help solve fusion energy or unleash new biological risks. But with ever more FLOPS coming online over the next five years, one thing seems certain: Innovation will come fast—and with force. As FLOPS scale, so must our conversations about responsibility, regulation, and restraint.

Gary Grossman is EVP of technology practice at Edelman and global lead of the Edelman AI Center of Excellence.