The Exponential Growth of AI Computing Power: From MIPS to Exaflops

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

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More At the recent Nvidia GTC conference, the company unveiled what it described as the first single-rack system of servers capable of one exaflop -- one billion billion, or a quintillion, floating-point operations (FLOPS) per second. This breakthrough is based on the latest GB200 NVL72 system, which incorporates Nvidia's latest Blackwell graphics processing units (GPUs). A standard computer rack is about 6 feet tall, a little more than 3 feet deep and less than 2 feet wide. Shrinking an exaflop: From Frontier to Blackwell A couple of things about the announcement struck me. First, the world's first exaflop-capable computer was installed only a few years ago, in 2022, at Oak Ridge National Laboratory. For comparison, the "Frontier" supercomputer built by HPE and powered by AMD GPUs and CPUs, originally consisted of 74 racks of servers. The new Nvidia system has achieved roughly 73X greater performance density in just three years, equivalent to a tripling of performance every year. This advancement reflects remarkable progress in computing density, energy efficiency and architectural design. Secondly, it needs to be said that while both systems hit the exascale milestone, they are built for different challenges, one optimized for speed, the other for precision. Nvidia's exaflop specification is based on lower-precision math -- specifically 4-bit and 8-bit floating-point operations -- considered optimal for AI workloads including tasks like training and running large language models (LLMs). These calculations prioritize speed over precision. By contrast, the exaflop rating for Frontier was achieved using 64-bit double-precision math, the gold standard for scientific simulations where accuracy is critical. We've come a long way (very quickly) This level of progress seems almost unbelievable, especially as I recall the state-of-the-art when I began my career in the computing industry. My first professional job was as a programmer on the DEC KL 1090. This machine, part of DEC's PDP-10 series of timeshare mainframes, offered 1.8 million instructions per second (MIPS). Aside from its CPU performance, the machine connected to cathode ray tube (CRT) displays via hardwired cables. There were no graphics capabilities, just light text on a dark background. And of course, no Internet. Remote users connected over phone lines using modems running at speeds up to 1,200 bits per second. 500 billion times more compute While comparing MIPS to FLOPS gives a general sense of progress, it is important to remember that these metrics measure different computing workloads. MIPS reflects integer processing speed, which is useful for general-purpose computing, particularly in business applications. FLOPS measures floating-point performance that is crucial for scientific workloads and the heavy number-crunching behind modern AI, such as the matrix math and linear algebra used to train and run machine learning (ML) models. While not a direct comparison, the sheer scale of the difference between MIPS then and FLOPS now provides a powerful illustration of the rapid growth in computing performance. Using these as a rough heuristic to measure work performed, the new Nvidia system is approximately 500 billion times more powerful than the DEC machine. That kind of leap exemplifies the exponential growth of computing power over a single professional career and raises the question: If this much progress is possible in 40 years, what might the next 5 bring? Nvidia, for its part, has offered some clues. At GTC, the company shared a roadmap predicting that its next-generation full-rack system based on the "Vera Rubin" Ultra architecture will deliver 14X the performance of the Blackwell Ultra rack shipping this year, reaching somewhere between 14 and 15 exaflops in AI-optimized work in the next year or two. Just as notable is the efficiency. Achieving this level of performance in a single rack means less physical space per unit of work, fewer materials and potentially lower energy use per operation, although the absolute power demands of these systems remain immense. Does AI really need all that compute power? While such performance gains are indeed impressive, the AI industry is now grappling with a fundamental question: How much computing power is truly necessary and at what cost? The race to build massive new AI data centers is being driven by the growing demands of exascale computing and ever-more capable AI models. The most ambitious effort is the $500 billion Project Stargate, which envisions 20 data centers across the U.S., each spanning half a million square feet. A wave of other hyperscale projects is either underway or in planning stages around the world, as companies and countries scramble to ensure they have the infrastructure to support the AI workloads of tomorrow. Some analysts now worry that we may be overbuilding AI data center capacity. Concern intensified after the release of R1, a reasoning model from China's DeepSeek that requires significantly less compute than many of its peers. Microsoft later canceled leases with multiple data center providers, sparking speculation that it might be recalibrating its expectations for future AI infrastructure demand. However, The Register suggested that this pullback may have more to do with some of the planned AI data centers not having sufficiently robust ability to support the power and cooling needs of next-gen AI systems. Already, AI models are pushing the limits of what present infrastructure can support. MIT Technology Review reported that this may be the reason many data centers in China are struggling and failing, having been built to specifications that are not optimal for the present need, let alone those of the next few years. AI inference demands more FLOPs Reasoning models perform most of their work at runtime through a process known as inference. These models power some of the most advanced and resource-intensive applications today, including deep research assistants and the emerging wave of agentic AI systems. While DeepSeek-R1 initially spooked the industry into thinking that future AI might require less computing power, Nvidia CEO Jensen Huang pushed back hard. Speaking to CNBC, he countered this perception: "It was the exact opposite conclusion that everybody had." He added that reasoning AI consumes 100X more computing than non-reasoning AI. As AI continues to evolve from reasoning models to autonomous agents and beyond, demand for computing is likely to surge once again. The next breakthroughs may come not just in language or vision, but in AI agent coordination, fusion simulations or even large-scale digital twins, each made possible by the kind of computing ability leap we have just witnessed. Seemingly right on cue, OpenAI just announced $40 billion in new funding, the largest private tech funding round on record. The company said in a blog post that 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. Although one particular factor stands out, as exemplified by a McKinsey headline: "AI could increase corporate profits by $4.4 trillion a year." What comes next? It's anybody's guess At their core, information systems are about abstracting complexity, whether through 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 greater sense of the world. Now, with powerful AI beginning to appear, we are crossing a threshold. For the first time, we may have the computing power and the intelligence to tackle problems that were once beyond human reach. New York Times columnist Kevin Roose recently 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 does not even count 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 a guess. We do not 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. It is clear, too, that as FLOPS scale, so must our conversations about responsibility, regulation and restraint.

Mapping Jensen's world: Forecasting AI in cloud, enterprise and robotics - SiliconANGLE

Mapping Jensen's world: Forecasting AI in cloud, enterprise and robotics We are in the midst of a fundamental transformation of computing architectures. We're moving from a world where we create data, store it, retrieve it, harmonize it and present it, so that we can make better decisions, to a world that creates content from knowledge using tokens as a new unit of value; and increasingly takes action in real time with or without human intervention, driving unprecedented increases in utility. What this means is every part of the computing stack -- silicon, infrastructure, security, middleware, development tools, applications and even services -- is changing. As with other waves in computing, consumer adoption leads us up the innovation curve where the value is clear, the volume is high and the velocity is accelerated, translating to lower costs and eventual adoption by and disruption of enterprise applications. Importantly, to do this work on today's data center infrastructure would be 10 times more expensive, trending toward 100 times by the end of the decade. As such, virtually everything is going to move to this new model of computing. In this Breaking Analysis, we quantify three vectors of AI opportunity laid out by Nvidia Corp. Chief Executive Jensen Huang (pictured) at this year's GTC conference: 1) AI in the cloud; 2) AI in the enterprise; and 3) AI in the real world. And we'll introduce a new forecasting methodology developed by theCUBE Research's David Floyer, to better understand and predict how disruptive markets evolve. IT decision makers pull back spending outlook to 3.4% for the year, down from 5.3% and below 2024 levels Markets remain under pressure, exacerbated by ongoing tariff concerns and back-to-back declines in key indices -- most notably the tech-heavy Nasdaq. Stocks have had their worst week since March 2020. Against this backdrop, the latest recent spending data from Enterprise Technology Research's quarterly survey of IT decision makers reflects a notable shift in spending expectations over the past year as shown below. Coming out of the pandemic, the "COVID spending spree" drove projected IT budget growth above 7%. As the Federal Reserve tightened monetary policy, that figure bottomed below 3%. In 2024, overall IT spending growth settled around 3.9%, and by January of 2025, survey respondents indicated a rise to 5.3% -- an improvement from 4.8% in the prior October survey. However, the April sentiment shows a concerning pullback, dropping from those optimistic levels down to 3.4%. Although ETR has not yet finalized these numbers, the downward revision is noteworthy. Not only does 3.4% fall short of the early-year projection, it also places IT spending below last year's baseline. This shift underscores the mounting uncertainty in today's macroeconomic climate and reveals a pronounced dip in confidence among enterprise technology buyers. Majority of ITDMs are staying the course on AI spending patterns ETR's latest drill-down data indicates that public policy pressures around AI -- ranging from tariffs to privacy to regulation -- are not yet significantly dampening AI spending plans. In a sample of approximately 500 respondents, just under half report maintaining or even accelerating their AI initiatives, largely to stay ahead competitively. Meanwhile, a sizable middle contingent plans to proceed at a steady pace, monitoring developments without making dramatic shifts. Notably, fewer than 10% are pumping the brakes on policy-related concerns (see below). Though broader macroeconomic headwinds are evident elsewhere in the data, this specific snapshot suggests that enterprise AI momentum remains largely intact. Still, a degree of caution persists. Some organizations appear to be grappling with the enormity of AI's potential disruption, hoping regulatory uncertainties may resolve -- or possibly fade. Despite recent market volatility, these survey findings underscore an ongoing drive to explore and implement AI initiatives, indicating that the technology's disruptive promise is outpacing immediate policy concerns. In his keynote at this past GTC, Jensen Huang laid out three primary AI opportunities, including: 1) AI in the public cloud; 2) AI in the enterprise - i.e. on-premises; and 3) AI in the real world - i.e. physical robots. Our goal today is to break down the data center forecast shown below into Jensen's first two opportunities. And we'll frame how we see AI in the real world evolving by applying a new forecasting methodology. The hyperscalers have emerged as pivotal enablers by providing massive compute capacity and skilled teams with the expertise to stand up large‐scale AI "factories." The cloud segment currently dominates AI infrastructure build‐outs, especially given the success and acceleration of consumer‐oriented services such as OpenAI, Meta Platforms, Apple, ByteDance and TikTok. More specifically, consumer return on investment is clear and dollars spent in consumer applications are paying dividends. The second vector -- AI in the enterprise -- encompasses on‐premises or private data center deployments. Though on‐prem stacks often encounter friction from data gravity and incomplete AI‐specific infrastructure, organizations increasingly seek to bring AI to their data rather than move data offsite. Several major original equipment manufacturers, including Dell and HPE, have introduced "AI factory" offerings tailored for enterprise data centers. However, many of these solutions remain heavily hardware‐centric. The emerging ecosystem of software components, AI‐optimized workflows and specialized talent is still taking shape, suggesting ongoing build‐out before enterprise AI reaches maturity. The third arena Huang highlighted is AI's role in the "real world," specifically physical robotics. From single‐purpose, task‐based machines to versatile humanoids, the potential is vast. Platforms such as Nvidia Isaac -- recently demonstrated via the Blue humanoid robot -- foreshadow a future where AI moves well beyond virtual interactions and enters factories, warehouses and ultimately everyday settings. This third area of opportunity is not included in our data center forecasts shown here. To underscore the magnitude of these shifts, our updated forecasts project a dramatic rise in accelerated computing or what we call Extreme Parallel Processing or EPP. Our analysis shows worldwide data center spending growing at a 16% compound annual growth rate over an 11‐year horizon, reaching a trillion dollars by the early 2030s. Within that total, EPP, or accelerated computing, is set to climb at a 23% CAGR, driving the market from roughly $43 billion to $180 billion in a single year (2024). This was the true beginning of the supercycle. By contrast, traditional computing, dominated by x86‐centric infrastructures, continue a slow decline -- reflecting the growing emphasis on graphics processing unit‐ and AI‐optimized architectures. In practical terms, this shift means that by 2035, overall IT spending could be ten times higher than it was in 2024, with much of that outlay dedicated to advanced AI data centers. While organizations may be cautious about transforming core applications and retooling legacy code, the trajectory of spending indicates a clear pivot. Over time, the center of innovation and budget allocation is poised to move decisively from legacy systems toward AI‐driven infrastructure and workflows. Our research shows accelerated computing moving from under 10% of total data center spending in 2020 to 85% by 2030. While a large portion of this rapid growth is being driven by public cloud hyperscalers, on‐premises enterprise deployments are also beginning to take shape. The next step is to break out the portion of spending that directly pertains to private data centers. Applying volume, value and velocity to forecast the AI opportunity To project these shifts effectively, a forecasting approach that uses but goes beyond classic Wright's Law is necessary. Wright's Law states that costs decline in a predictable manner as cumulative production doubles, but current market dynamics demand additional dimensions. Our methodology incorporates "Volume, Value, and Velocity" (3Vs) to capture how disruptive technologies like AI can undergo faster and deeper adoption: AI provides unprecedented value in both consumer (today) and enterprise (eventually) domains, creating a powerful incentive for organizations to accelerate their adoption cycles. Although on‐prem stacks have to address data gravity, software dependencies and skill gaps, the overall trajectory suggests a steady pivot toward AI‐optimized infrastructure. Our analysis indicates that x86‐centric data centers will remain in place for the foreseeable future, but the momentum behind extreme parallel processing is expected to reshape spending patterns and ecosystem investments in the years ahead. Our more granular forecast shown below, isolates just the AI portion of total data center spending, separating the market into public cloud environments and private on‐premises deployments. Cloud currently dominates thanks to hyperscalers' advanced tooling, specialized skill sets, and strong consumer‐driven use cases that demonstrate immediate ROI. Platforms such as Meta, Google, ByteDance and Apple are justifying substantial capital expenditures on AI, which accounts for the lion's share of near‐term growth. We also include platforms such as Grok (xAI) and so-called neoclouds (for example, Coreweave) in the cloud segment shown below in the dark blue bars. On‐premises enterprise infrastructure shown in the light blue above, shows a more gradual ramp, reaching steeper adoption curves around 2026 and accelerating into 2029. The relative delay reflects a number of factors, including a lack of fully evolved solution stacks, fewer in‐house AI experts and the added complexity of modernizing legacy data and applications. Although certain large institutions -- particularly in financial services -- can invest heavily in retooling data centers for liquid cooling, bringing in specialized hardware, building out their own software stacks and managing overall AI stack complexity, many organizations must wait for integrated solutions that address data gravity, governance, and existing transactional systems. Note: We include colocation facilities such as Equinix in the on-prem component of our forecasts. Our analysis projects a trillion‐dollar AI data center opportunity by approximately 2032, with around 20% of that spending in private enterprise environments. Jensen thanked the audience at GTC for adopting new architectures, such as disaggregated NVLink, liquid cooling and high‐power racks approaching 120 kilowatts each. Though he didn't say this specifically, this is happening in hyperscaler markets. By contrast, most enterprises still rely on air‐cooled facilities and stacks optimized for general‐purpose computing. Despite the slower start, the momentum around AI on‐prem is expected to gain speed as data harmonization, real‐time processes and "agentic AI" systems mature. This evolution will require broader access to legacy metadata, transactional platforms, and a more advanced software layer -- factors that point to a significant, but lengthier, transition path for traditional data centers. Our belief is that robotics is one of the most compelling frontiers for AI, with wide-ranging implications for both closed and open environments. Single-purpose robots -- such as those found in factories, transport fleets and specialized defense applications -- demonstrate significant near-term value because they automate well-defined tasks and offer predictable ROI. Closed-system deployments exhibit high velocity, as organizations can design workflows and facilities from scratch for maximum automation. This approach lowers costs, reduces errors and boosts adaptability, enabling newcomers to capture market share by building AI-native operations that can achieve revenue per employee at a scale few traditional companies can match. We see a distinct difference between single-purpose robots in factories, doing one job really well, and robots that mimic humans and perform a variety of tasks. Jensen wowed the audience with his demonstration of a humanoid on stage, but we see clear opportunities in the near term for single-purpose automation and believe mimicking humans has a much longer road ahead. Open-ended humanoid robots in our view face more complex adoption curves. Physical reality introduces an extensive range of edge cases, including unforeseen interactions with humans, other robots and the environment itself. These factors will temper the velocity of humanoid deployment, and we believe that broader adoption will likely take longer to unfold. Nonetheless, the long-term potential remains vast. Trade policy also plays a role, as tariffs can slow global adoption by reducing the incentives for automation and giving competitors time to close technological gaps. However, the opportunity to become a leading low-cost exporter of AI-driven goods remains an influential force. If companies or entire regions accelerate investment in robotics, they may replicate the kind of transformation once seen during the world's Industrial Revolution, where Britain was the low cost provider to the world. The U.S. could replicate that dynamic on a global, AI-fueled scale. But tariffs introduce dislocations to that vision and are backward-looking. Rather we'd like to see investments in automation where sensible, making the U.S. the world's low-cost producer at scale, and let other countries hide behind tariffs. In future Breaking Analysis episodes, we'll apply the 3Vs methodology and provide a forecast of these markets in more detail. A Wall Street analyst dubbed last year's GTC the "Woodstock of AI." Our John Furrier, in the weeks leading up to this year's GTC, called it the "Super Bowl of AI," a phrase Jensen used as well. It's appropriate, building on last year's watershed moment and highlighting a wholesale transformation of computing architectures that will continue each year like the Super Bowl. Here we highlight just a few of the many notable takeaways from GTC 2025. One major focal point is the shift from integrated, monolithic designs toward more disaggregated, distributed systems. These high‐density, liquid‐cooled racks can contain hundreds of thousands of components and deliver exaflop‐scale performance within a single rack -- far outstripping conventional data center footprints. The combination of hardware innovation and sophisticated software layers is creating an entirely new systems paradigm, rather than just another generation of GPUs. A second standout announcement is the release of Dynamo, described as the operating system for the AI factory. This new layer orchestrates inference at scale by managing resource allocation across multiple racks and GPU pools. The design aims to optimize workloads for minimal latency and maximum throughput, laying the foundation for next‐generation distributed AI environments. Over time, inference is projected to account for the largest proportion of AI spending, making a robust OS essential for integrating advanced accelerators, data pipelines and both x86 and Arm processors. These developments reinforce what we see as a "Wintel replacement strategy," where Nvidia appears positioned to capture the full stack -- much like Intel and Microsoft did for PCs. Nvidia is combining hardware and software in a way that leverages the high volume of consumer‐driven AI while also preparing for enterprise adoption. Historically, the volume advantage propelled x86 and Windows into market dominance. Today's AI revolution similarly benefits from massive consumer deployments in areas such as search, social media and targeted advertising, which feed breakthroughs and scale economies that ultimately migrate to corporate data centers. At the edge, however, adoption patterns are less certain. A proliferation of ultra‐low‐power chips -- such as those based on RISC‐V or alternative minimal‐footprint Arm designs -- may capture segments where small data sets, embedded applications and cost‐efficiency are paramount. Though Nvidia continues to expand its footprint in edge computing, the hardware and energy requirements and constraints differ substantially from those in high‐density AI data centers. As a result, multiple architectures are likely to coexist, especially in use cases that require thousands -- or millions -- of embedded, power‐constrained devices. Overall, the core message from GTC is that accelerated computing has moved beyond the realm of specialized GPUs into a fundamentally new model, where advanced fabrics, next‐level operating systems and massive consumer volume converge to push AI into every layer of technology -- from hyperscale clouds and enterprise data centers to the far edges of distributed environments. Below we summarize and highlight several areas we'll monitor with respect to our forecasts and predictions. Public policy remains a critical factor to watch, especially in light of ongoing tariffs, macroeconomic shifts and heightened regulatory conversations around AI. The key question is whether accelerated AI value will continue overshadowing cloud costs at scale, or whether on‐premises suppliers -- primarily Dell Technologies Inc., Hewlett Packard Enterprise Co., IBM Corp., Lenovo Group Ltd. and, to a certain extent Oracle Corp., and others in the ecosystem such as Pure Storage Inc., Vast Data Inc., WekaIO Inc., Data Direct Networks Inc. and the like -- will move rapidly enough to deliver AI‐optimized solutions that address data gravity and compliance requirements in a hybrid state. Many enterprise workloads in regulated sectors, such as financial services, healthcare, and manufacturing, have never fully transitioned to the public cloud, creating an opening for on‐prem incumbents to build specialized, high‐performance AI stacks. We believe a window of roughly 18 to 24 months appears to be the timeframe in which on‐prem vendors can prove they can deploy turnkey AI infrastructures at scale. A key question is will past migration trends repeat if public cloud offerings continue to innovate at a faster pace, capture more user volume and build ecosystem velocity. The opportunity exists for on‐prem vendors to avoid the scenario where cloud simply "washes over" enterprise workloads. Colocation and sovereign models are also playing a role, positioning hybrid‐cloud approaches as an alternative to pure hyperscalers - but much work needs to be done to attract startups and build full stack, opinionated AI solutions for hybrid/on-prem environments. As indicated, in the future, we'll quantify the single‐purpose versus general‐purpose robotics markets and further test the volume, value and velocity (3Vs) forecasting methodology. Early research findings indicate that consumer‐driven AI usage remains a catalyst, fueling investments in hyperscale platforms that could carry over into private data centers. The overall trajectory points toward the most impactful adoption curve in modern history, driven by generative AI breakthroughs that have propelled organizations to the steep portion of the S‐curve far sooner than most industry veterans anticipated. How do you see the future of AI? What are your organizations doing to apply AI to data that lives on premises? Are you investing in new data center infrastructure such as liquid cooling or are you more cautious and waiting for more proof?