AI Is Compounding Faster Than Moore's Law: What It Means for Your Business

By MercConsulting · Published 2026-08-26 · Updated 2026-09-02

Moore's Law doubled chip capacity about every two years. AI's key measures now double every four to seven months, so what you saw AI do a year ago is not what it does today, and it is time to plan for capability you cannot yet see.

AI is improving several times faster than Moore's Law. Moore's Law described chip capacity doubling roughly every two years; the measures that matter for a business, meaning the compute behind AI models, the length of work a model can finish on its own, and the price of a fixed level of capability, now double every four to seven months. For an owner, that means the AI you evaluated a year ago is not the AI available today, and today's is not what you will be working with next year.

A technology that doubles every few months turns any plan built around its current limits into a plan built around limits that no longer exist.

What Moore's Law Actually Said

In 1965 Gordon Moore observed that the number of transistors on a chip was doubling about every year, and in 1975 he revised the pace to roughly every two years (the popular version says eighteen months). The eighteen-month figure came from Intel executive David House, who was estimating how quickly chip performance would double, not transistor count. Moore's own number was two years.

That two-year doubling ran for five decades and set the expectation that a serious technology improves by roughly half again each year. It is the right yardstick for AI not because the two are alike, but because it is the fastest sustained technology curve most owners have lived through.

The Three AI Curves

AI progress is not one number. Three measurements, tracked by independent research groups, each have their own doubling time, and all three are shorter than Moore's two years.

Training compute. The computing power used to train notable AI models has grown about 4.7x per year since 2010, a doubling roughly every six months (Epoch AI, dataset updated November 2025). For frontier language models the pace since 2020 has been about 5x per year, a doubling every five months. This is the input curve.

Task horizon. The research group METR measures the length of tasks, in human working time, that AI systems can complete on their own at 50 percent reliability. From 2019 to 2025 that horizon doubled about every seven months; METR's January 2026 update puts the doubling time since 2023 at about 131 days, roughly four months, and since 2024 at about 89 days, roughly three months. This is the output curve: how much finished work a system can carry.

Price for a fixed capability. The cost of buying a given level of AI performance has fallen between 9x and 900x per year depending on the task, with a median near 50x per year (Epoch AI, 2025). Stanford's AI Index 2025 measured a roughly 280-fold drop in about eighteen months, from late 2022 to late 2024, in the cost of the capability the public first met in ChatGPT. This is the cost curve: what yesterday's capability costs today.

METR's own caveat applies to the task-horizon figures: the trend is sensitive to which tasks are measured, and the newest models sit near the edge of what the test can measure. Use every four months or so as the working figure, with the seven-month long-run rate as the conservative case.

The Arithmetic of Compounding Over Two Years

Doubling times are easy to nod at and hard to feel. The way to feel them is to work the arithmetic over the two-year horizon most owners use for a serious systems decision: start each curve at one and count the doublings.

The point is not the exact multiple; it is the shape. A two-year doubling means the second year adds about as much as the first. A four-month doubling means the last four months of a two-year period add more capability than the first twenty months combined, which is why a capability appears to suddenly work on tasks that failed the last time anyone checked. Nothing sudden happened; the doubling kept doubling.

The Price Curve Is the Quiet One

Most attention goes to what AI can do, but the price curve decides what an ordinary business can afford to run all day. A fixed capability whose price falls 50x a year at the median goes from a research budget to a rounding error in about two years. The AI that was too expensive to run across every call, email, and document in your business a year ago is now cheap enough to do exactly that, and what is expensive today will be cheap by the time a well-run implementation finishes. Payroll rises a few percent a year; the cost of the capability behind an agent falls by a multiple.

Why an Owner Should Care

What you saw AI do a year ago is not what it does today. If you tried a tool in 2024 or 2025 and it fumbled a task, that test is stale. At a four-month doubling, a year is three doublings, roughly 8x in the length of work a system can carry on its own. The honest response to "I tried it and it did not work" is to ask when.

Plan for capability you cannot see yet. A systems decision made today will be in service for years, and if it is designed around what AI can do this quarter, it will be constraining the business by next year. The right design assumes the capability will keep doubling and asks where it should plug in when it arrives.

The lag is now the cost. The distance between what AI can do and what a typical business uses widens every month. That distance is the subject of our companion guide on the AI adoption gap, and it is where the advantage sits for the businesses that close it.

MercConsulting's view is that AI is the largest change in how businesses operate since the internet, and in our view larger than the computer or the internet. The computer changed how records were kept; the internet changed how businesses were reached. AI changes who, or what, does the work, on a curve neither of those technologies ever ran.

Build the Connection Layer, Not the Model

The natural reaction to a fast-moving technology is to try to pick the winner, and for a business owner that is the wrong problem. The models improve on someone else's budget, and whichever one leads this quarter will be matched or passed within months. Choosing a model is a decision that expires.

What does not expire is whether your business can be reached by a model at all. Most businesses run several disconnected tools, and an AI agent dropped into that environment can talk but cannot act; it is a chatbot with no hands.

The durable investment is the connection layer: one connected system, built by mapping the tools you have, consolidating them where needed, and adding custom software where needed, so an agent can reach records, phones, email, calendar, documents, accounting, and the CRM wherever it needs to. That layer does not go stale when the model improves; it is what the improvement flows through. Our guide to making your systems AI-ready walks through the mapping, and how we think about the speed of AI explains why we do it first.

Lease the Agent So Upgrades Arrive Without a Project

A business that builds its own AI capability owns a snapshot of the curve on the day the build finishes; a business that leases an agent from a firm whose job is to keep it current is subscribed to the curve itself. When the capability doubles, the leased agent improves, and nobody at the business ran a project to make that happen.

That is the model MercConsulting uses. We build the agents, own the technology, and lease each agent to the business for a monthly amount; the business does not build, host, prompt, train, or support anything. Laws and regulations, new case law, HR and state rules, and the business's own changing procedures are kept current by us on a cadence from daily to yearly as the subject requires, with all support and maintenance included.

Each agent is built for a role the owner chooses (CFO, bookkeeper, inside-counsel support in the areas of law the business needs, HR manager, reception and intake, sales, or reporting) and trained on the business's own data, documents, policies, pricing, and procedures. The owner works with it the way they would with an employee: by phone, from a button on the computer, by talk-to-text, in plain chat, or on a video call with a human presence. How leased AI agents work covers the mechanics.

On cost, our rule of thumb is that an agent leases for roughly 12 percent of what the business spends today on the comparable function, for comparable usage. Say a business spends $100,000 a year on outside real estate counsel; we work backward from typical hourly rates to the hours that bought, size the agent to carry a comparable volume of work, and the lease comes in around $12,000 a year. That is an estimate to start a conversation, not a quote; the real number comes from the business's actual usage. Because the price of capability keeps falling, the math behind cutting expenses this way tends to improve with time.

Where the law requires it, a licensed professional of the business's choosing reviews and signs; faster AI changes how much routine work reaches that professional already done, not the line itself. The free 30-minute discovery call is where this starts: we map your actual overhead and tell you which function to automate first, or whether you should not.

Book a free 30-minute discovery call

Frequently Asked Questions

How fast is AI improving compared to Moore's Law?

Moore's Law doubled transistor counts roughly every two years. AI training compute doubles about every six months, the length of work AI can finish on its own has doubled about every four months since 2023 (seven months over the long run), and the price of a fixed capability falls roughly 50x a year at the median. Over two years that is 2x for chips against roughly 8x to 64x for AI.

Should I wait for AI to settle down before adopting it?

No, because it is not going to settle down on any horizon a business plans around, and waiting means the gap between what AI can do and what your business uses keeps widening. The way to adopt a fast-moving technology is to connect your systems once, properly, and lease agents that are kept current by someone else, so each improvement arrives without a new project.

Will the AI I buy today be obsolete in a year?

The model behind it probably will be, which is why we do not recommend buying a model. The parts that hold their value are the connection layer that lets AI reach your records, phones, email, documents, and accounting, and a lease arrangement in which the agent is upgraded as the underlying capability improves. Those two decisions are designed to outlast any particular model.

What does task horizon mean in practical terms?

It is the length of a task, measured in the time a skilled person would need, that an AI system can finish on its own at a given reliability. A horizon of minutes means a tool that answers questions. A horizon of hours means a system that can carry a piece of work, such as reconciling a month of transactions or drafting and checking a contract, from start to finish.

MercConsulting is a business consulting firm, not a law firm or CPA firm. Estimates in this guide are illustrations, not quotes; licensed professionals review and sign where the law requires.