The Agentic AI Super Cycle — PartsCube Global

The biggest technology revolutions of the past were driven by steam engines in the 1700s, railroads in the 1800s, electricity, cars, phones, planes, computers, and the internet in the 1900s. Now we are at the dawn of the Agentic AI revolution, which will be the biggest of them all.

A few years ago, Goldman Sachs predicted Generative AI would increase global GDP by about 7% over 10 years. McKinsey had a roughly similar projection. Today, more than half of USA GDP growth is due to AI infrastructure investment.

Agentic AI is much more powerful and more economically impactful than generative AI (and any previous industrial revolution) and requires much more compute investment for broad adoption. More than half of GDP in advanced economies is labor, with more than half of that labor being cognitive work that can be made much more productive with agentic AI.

So far, coding is the early adopter, with most coders now using agentic AI to generate and validate most code. That makes the coder more of an architect and manager. What’s astounding is that this happened in less than 2 years. Frontier models are now used extensively to generate code for the next frontier model. But how quickly will other fields extensively adopt Agentic AI, and will supply chain constraints be overcome to meet the demand?

Recently, Dario Amodei, CEO of Anthropic, called for a pacing of the development of frontier models to ensure everyone is building in the right controls and safeguards. AI leaders agree these controls are important. “We’ll make AI that is incredibly safe,” said Jensen Huang, Nvidia’s CEO.

Any pacing of model development won’t slow the pace of investment. Dario noted safeguards will involve more training. Sarah Friar, OpenAI’s CFO, noted on CNBC recently, “even if we stop today, the amount of intelligence that’s available in the world is massive,” and “there is so much opportunity to drive growth with good ROI (return on investment.)”

Data center AI CapEx (capital expenditure) will reach ~$1 trillion this year, and is projected to reach $3 trillion to $4 trillion by 2030 (Source: Jensen Huang, Goldman Sachs Technology Conference). By comparison, world GDP is expected to be $123 trillion this year. AI is expected to go from ~1% to ~3%.

Of the world’s top 10 highest market capitalization companies, 8 are major AI players: Nvidia, Alphabet, Microsoft, Amazon, TSMC, SpaceX, Meta, Broadcom. Collectively they make up almost 20% of the total market cap of the world’s publicly traded companies. Only a couple of these AI companies were in the top 10 a decade ago. Anthropic is expected to join the top 10 market capitalization club later this year.

Massive companies like Nvidia are projecting 70% growth in revenues next year at huge scale.

It’s natural to think this can’t keep going. But maybe it can. The key factors are:

If agentic AI has high ROI;
If it has a huge TAM (Total Adressable Market), and
If it continues to become more powerful and less costly for years to come.

If all those conditions are met, then we are in the biggest supercycle of our lives — and of history. The AI leader companies will get much bigger along with the US and world economies. There will be losers, there will be pauses/plateaus/consolidation, but this is a 10+ year super²cycle.

Data center AI investment offers very attractive ROI
The hyperscalers, Anthropic, and OpenAI are ramping AI CapEx because they see rapid demand growth, strong ROI, and because they don’t want to fall behind competition. This is a gold rush.

Morgan Stanley in July concluded the major hyperscalers had paths to 25% to 50% ROIC (return on invested capital) for GenAI. These are very attractive returns.

Dave Patterson, distinguished engineer and fellow at Google, said at the recent AI Infra Summit that Google recovers its cash outlay for a TPU in 1 year, and in 2 years for its custom Axion CPU. At Goldman Sachs Tech Conference, Thomas Kurian, CEO of Google Cloud, said TPUs paid back twice as fast as GPUs, so even GPUs recover their investment in 2 years. This is a fantastic ROI, although he might be leaving out land and building costs. The reason this isn’t increasing near-term profits already is because Google keeps increasing CapEx spend faster than revenue growth.

Amazon senior vice president Peter DeSantis talked at AI Infra Summit about how the company builds the data center and the racks in parallel with building the Trainiums. This accelerates cash recovery, enabling Trainium shipments from the end of the production line to the data center to be in operation (and making money) in a couple months. He contrasted this to most companies that build the chips, then the trays, then the racks, then the rows, then install. That keeps cash tied up much longer before revenue generation.

New Street forecasts that hyperscaler negative cash flow will bottom in 2028 at 20% of revenues, but then rapidly turns to a positive free cash flow of 10% of revenues by 2030. AI cloud revenue grows almost 10× from 2026 to 2030! And AI cloud revenue by 2030 exceeds non-AI.

TSMC builds >90% of all AI compute devices (GPUs and custom XPUs like TPU and Trainium). TSMC is an extremely well-run company that does its homework thoroughly. At the recent Goldman Sachs Technology Conference, TSMC’s CFO said the foundry thinks we are still in the very early stage of a megatrend. Token costs are dropping fast using advanced processes and packaging, but this leads to increased demand for AI compute. TSMC is building numerous huge fabs and advanced packaging factories to try to keep up.

We need a LOT more AI compute at much lower $/token
At the Goldman Sachs Tech Conference, Jeetu Patel, Cisco’s president and chief product officer, said that agent consumption of tokens now exceeds consumption of tokens by humans using GenAI, and <2% of humans are now using agents in a power user capacity. So the half of today’s tokens that are agentic could grow 50× when 100% of humans use agents in a power user capacity. And that’s for today’s models. Future models will grow exponentially in compute requirement, but costs are coming down exponentially at the same time.

Figure 1: Agentic token consumption has soared to >50% of total demand. Source: Cisco

Goldman Sachs in May projected token consumption will grow 40× from mid 2026 to mid 2030.

At the AI Infrastructure Summit, Ian Buck, Nvidia’s vice president of hyperscale/HPC computing, showed how Agentic AI is 100× more compute-intensive than chat – and this is with just four agents.

Figure 2: Agentic AI requires 100× more AI compute than chat. Source: Nvidia, AI Infra Summit

At AI Infra Summit last month, Pat Gelsinger, general partner at Playground Global, explained we need a 10,000× improvement across the AI infrastructure stack. His key points:

Reasoning models need 10× more computation;
Future models will need 10 to 100× more compute than reasoning models;
We are at most 1/10^th of the way to global adoption;
Most businesses/enterprises are just getting started.

Figure 3: AI Needs 10,000× cumulative improvement. Source: Playground Global

Nvidia’s Huang said in 2024 that AI computing power was seeing a fourfold increase annually that would lead to a million-fold increase over the next 10 years (4¹⁰ = 1,048,576).

At the AI Infrastructure Summit, Nvidia’s Ian Buck showed the performance boost of the latest Nvidia Vera Rubin NVL72 over GB300 NVL72: Up to 30× better at higher responsiveness, with up to 45× lower cost per token. The exact improvement depends on the nature of the workload. The improvements are less for batch AI, but they are huge for agentic AI, which needs high responsiveness in order to meet human desire for quick answers.

Figure 4: Nvidia’s Vera Rubin delivers 30× faster and 45× cheaper agentic AI. Source: Nvidia/AI Infra Summit

Epoch.AI data shows that cost/token has declined at 9×/year on the simplest tasks, and 900×/year on the most complex/challenging tasks. This is roughly consistent with the Nvidia data above.

Buck said that Nvidia will continue to deliver improvements on an annual cadence ,with Vera Rubin Ultra next year, then Feynman, then Feynman Ultra.

At the Goldman Sachs Technology Conference, Huang said that performance improvement is coming from a combination of factors — CMOS technology, more efficient numerical encoding, advanced packaging, CUDA software, disaggregation of prefill and decode, but especially networking. He said the GPU isn’t the chip, it’s the network of chips working together. Just a few years ago it was 4 tightly interconnected GPUs. Now it’s 72 (NVL72), and soon it will be NVL576: 8 racks of NVL72 with optical interconnect.

CNBC reported in late 2025 that Amin Vahdat, vice president of AI Infrastructure at Google, outlined at a company-wide all-hands meeting that it must double AI serving capacity every six months for a 1000× increase in 4 to 5 years.

Agentic AI TAM is huge
Recently, Morgan Stanley estimated that the global knowledge work TAM (work that can be augmented by AI) at $20 trillion to $30 trillion, with use cases spanning the economy. Software is just the start. The firm projects a modest ramp rate, reaching just 13% TAM penetration by 2033, which is $3.2 trillion of enterprise AI spend – almost double 2030 spend. Continuing growth in AI sales will require growth in AI infrastructure CapEx into the 2030s.

There certainly are barriers to adoption in large companies. But adoption could be much faster if pacesetter companies embrace agentic AI as a competitive edge, which would allow them to grow much faster than less nimble competitors. Rapid adoption by pacesetters is possible because the hyperscalers are making the huge investments that anyone can easily utilize.

Morgan Stanley also identified $30 trillion of consumer spend to be further digitized: e-commerce, agentic shopping, autonomous driving, restaurant delivery, advertising, logistics. It expects more rapid AI adoption in consumer spending and use.

Jevons Paradox
In 1865, William Jevons observed that more efficient use of coal by steam engines led to increased consumption of coal, not less. By making the economics of steam engines more attractive, they were used more widely.

As genAI became cheaper, more people used it. Now, more than a billion use GenAI on a monthly basis, and 100 million+ use it on a daily basis.

AI is improving AI
AI is assisting with improving AI, leading to a strong feedback loop accelerating AI improvement that includes:

Better models
Better agents
More automated software, design, research
Better chips, algorithms and infrastructure
More effective AI

In September, Anthropic said existing frontier models “lead” roughly 26% of the R&D work to build newer frontier models, working from a high -evel prompt under human supervision. No part of the frontier model development currently operates autonomously. The utilization of AI to design future models is likely to increase further.

OpenAI in September said it has developed an automated “research intern” that can carry out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days. The company is progressing toward an automated AI researcher by early 2028. Just this year, OpenAI researchers’ daily work has changed substantially, using coding agents throughout the day with total usage rapidly increasing.

At Hot Chips in August, OpenAI talked about taking nine months to tape out Jalapeño, its first chip, which may rival Vera Rubin. Company experts talked extensively about using OpenAI tools to do rapid evaluation of various architectures and rapid iteration to improve performance and features.

Almost all agentic AI today runs on versions of the Transformer architecture developed by Google a decade ago. Eventually, researchers with AI will find a more powerful model architecture that will use constrained resources much more efficiently.

Capital is abundant and willing
High-tech financing used to be a Silicon Valley venture capital cottage industry, with early IPOs by the major Wall Street firms where the big East Coast funds would buy in. Now, companies can grow to billion-dollar or even trillion-dollar valuations, staying private with the big East Coast funds buying into mid-stage rounds, along with Gulf sovereign wealth funds. AI infrastructure rounds are consistently oversubscribed at all levels, despite soaring valuations.

Trillion-dollar money managers like Blackstone are buying hundreds of billions of digital infrastructure assets and are now directly buying tens of billions of dollars of Google TPUs and jointly financing $500 billion of Nvidia GPUs.

The hyperscalers’ CapEx now exceeds their free cash flow, so they are issuing debt and finding that investors prefer Alphabet and Amazon notes over T-bills.

The huge inflow of capital has caused a tremendous ramp-up of early-stage data center AI startup valuations as investors look to get in on the next Anthropic.

The big players in data center AI have huge valuations, but they are justified by their financials. TSMC enjoys healthy >60% gross margins, as do NVDA, AVGO, and MSFT. Frontier model companies like Anthropic are rumored to have healthy ~80% gross margins. So the major suppliers of Agentic AI compute appear to be profitable and healthy with reasonable P/E ratios, given their tremendous growth, as long as the ROIC of AI CapEx remains positive into the future with continued agentic AI rapid demand growth.

Some likely implications of the AI Super²cycle over 5-10 years
Every hyperscaler and frontier model maker will build their own AI compute solutions. This is a very expensive endeavor, but their business size can justify it. Moreover, it gives them the ability to optimize for their needs and reduce costs.

This doesn’t mean GPUs go away, but they will need to compete with custom compute. Because of their superior supply chain and financial buys, Nvidia will stay the dominant supplier of AI compute at least through 2030, but will likely be more pressured after 2030. Nvidia is already working to create other customers like Neoclouds, to add value for non-hyperscalers by buying Hugging Face; to entice custom compute makers to use their infrastructure through NVLink. Nvidia may have to choose at some point to become more like a hyperscaler, lower prices to entice hyperscalers to buy GPUs, or focus on selling total solutions to non-hyperscalers.

ASIC suppliers will likely be acquired and integrated by the hyperscalers and frontier model makers as AI compute becomes a bigger and bigger internal business.

TSMC’s business will become majority AI, with most of its top 10 customers doing AI. TSMC today makes almost all of the world’s AI compute. Thus, the world economy is highly dependent on TSMC. No other company in AI is a “single point of failure.” This will provide an incentive for hyperscalers to try Intel Foundry and Samsung to create price competition and backup capacity.

Sarah Friar, OpenAI CFO, said at the Goldman Sachs Technology Conference that Astra was trained on 100,000 GPUs — the largest ever. The size of frontier models will keep increasing. Epoch.AI shows frontier training compute has historically grown by roughly 4× to 5× annually. GPT-2 was trained on ~2e21 FLOPs; GPT-4 was trained on ~2e25 FLOPs – a 10,000× increase across two generations. By 2030, Epoch.AI projects training runs can reach ~2e29 FLOPs – another 10,000× increase. Training is latency-sensitive, so to keep the very large number of GPUs required (1 million? More?) as close together as possible, data centers will need to go vertical.

Figure 5: Data centers like Oracle’s huge Abilene site will go from low-rise to high-rise. Source: Oracle image modified by ChatGPT

Conclusion
Looking out 5 to 10 years is very difficult because there are so many cross currents (will token cost come down faster than token demand ramps up?) and so many possible unpredictable developments (when does the next generation after Transformers come, and how efficient is it?). And even when the outcome in 10 years seems likely, the path to get there could be a straight line or a roller coaster.

Agentic AI looks very likely to be a Super²Cycle like none of us have ever experienced. It will have tremendous impact on the structure of the semiconductor and cloud computing industries, and of many other industries impacted by agentic AI productivity boosts. Understanding this can help semiconductor executives make better strategic decisions.

The post The Agentic AI Super Cycle appeared first on Semiconductor Engineering.