Lawrence Burns: Manager Review

As with any investment, your capital is at risk.

Agentic Dawn

In January 2025, the artificial intelligence company Anthropic had an annualised revenue run rate of $1bn – a measure that takes recent sales and projects them over a full year. Fifteen months later, it had surpassed $30bn. No company in recorded history has grown organic revenue at this scale and pace.

Anthropic, one of Scottish Mortgage’s private holdings, has helped usher in the third era of generative artificial intelligence. The first was the conversational era, which began with the launch of OpenAI’s ChatGPT in late 2022, when models became reliably capable of following natural-language instructions and holding back-and-forth exchanges with their users.

The second was the reasoning era, which began with OpenAI’s o1 in September 2024, when models learned to pause, think through problems step by step and produce considered, rather than instinctive, answers. This made them better at solving complex problems, particularly in mathematics, science and coding.

By late November 2025, with the release of Anthropic’s Claude Opus 4.5, the agentic era had become unmistakable. Models could now be given a goal and work towards it over many steps: planning, using tools, checking their work and producing useful outputs without constant human prompting. Each era has built on the last rather than replaced it. Today’s agents are reasoning models that have learned to act, just as reasoning models were conversational models that had learned to think.

The implications of the rise of agents reach across our portfolio into the structure of the software industry, the value of consumer businesses, the rise of a parallel Chinese AI ecosystem, and the physical supply chain that must be built to meet insatiable computational demands.

Software

The impact of agents has been felt first at scale in software. There the work is digital, the value is high, the goals are often clear, and feedback comes quickly.

If an application is not working properly, an agent can be asked to find the problem, write a fix and test the result before users receive the update.

Adoption has been rapid. Google’s chief executive has said that 75 percent of new code is now written by AI. The founder of one of our portfolio companies recently told us that it is spending more on AI coding tools for its engineers than it is paying them in compensation. Moreover, he claimed the company gets a better return on the tools than on the engineers themselves. At Anthropic itself, AI is now writing between 70 and 90 percent of all code.

Our own software holdings are skewed towards this infrastructure layer. Databricks and Snowflake organise the governed company data that agents need if they are to be useful inside enterprises. Cloudflare provides the network and security layer on which agent applications can run, while helping websites identify, control and charge AI agents for access. Adyen and Stripe provide the payment and trust infrastructure that allows agents to transact safely on behalf of customers and merchants. Stripe’s founders have been careful not to overstate the speed of change, arguing that agentic commerce is likely to arrive in small chunks rather than one sudden leap; but each chunk of autonomy still requires programmable, permissioned and trusted financial rails.

Far beyond software development

The impact of agents will not remain confined to software development. Most knowledge work, when stripped down to its components, is some combination of reading, writing, reasoning and using software as a tool to get things done. These are precisely the capabilities at which agents are now becoming proficient. Work that has long looked specialised: drafting a legal memo; synthesising a clinical trial; building a financial model; reviewing a patent application, is specialised at the level of expertise but generic at the level of cognitive operation. The consequence is that agents are not a tool for one industry but many.

The same logic applies to consumer businesses. Agents will become our personal shoppers, financial advisors and everyday assistants. This creates risk if horizontal AI assistants sit between customers and platforms. But the strongest platforms control assets an agent needs: trust, customer history, payments, credit, logistics, product catalogues and merchant networks. Amazon, MercadoLibre and Sea Limited are thus looking to build their own vertical agents to serve their platforms and enable shopping beyond them as well.

Nubank offers a similar possibility in finance. Long before agents were in vogue, its founder, David Vélez, told us that Nubank’s ambition was to give every customer a private banker in their pocket. Agents could make that ambition more practical: helping users manage bills, understand spending, choose when to borrow, build savings and find the cheapest rates on the market. Agents improve price transparency, tailor options to personal circumstances and reduce the friction to taking action. For Vélez, this is an opportunity. As a low-cost operator, Nubank is well placed to seize it.

Few companies will be left untouched by these developments. For some, agents will create new demand; for others, they will threaten existing profit pools; for many, they will do both at once. This will be a key challenge of growth investing in the years ahead. Scottish Mortgage is well placed to meet it because we invest across both public and private markets. Many of the companies shaping the AI frontier remain private, and our access to them gives us a broader view of how quickly the technology is improving, how it is being adopted, and where value may ultimately accrue.

Beyond the Valley

Meeting that challenge also requires geographic perspective. It is tempting to read the AI story as a Silicon Valley one. On questions of frontier model capability, that reading is broadly right. But it is incomplete. China is not merely a follower in artificial intelligence. It is developing different strengths under different conditions.

The first is physical AI. Simulation will be vital, but embodied intelligence improves fastest when virtual training is connected to real-world deployment. China’s manufacturing base matters because it provides the world’s largest deployment surface. This is reinforced by the largest installed stock of industrial robots, dense local supply chains, supportive policy, and an electric vehicle industry already combining software, hardware and cost-focused manufacturing. Horizon Robotics, one of our holdings, sits directly in this intersection between AI and the physical world, enabling autonomous cars with the ambition to extend this into broader robotics.

The second is cost-performance. Restricted access to the most advanced chips has pushed Chinese model companies to do more with less. This matters because the agentic era will be far more compute-intensive than the conversational era. If agents are to be widely adopted, the cost of useful intelligence must fall dramatically. MiniMax, one of our holdings, develops open-source models that approach frontier capability at a fraction of the training cost, part of a wider Chinese ecosystem pushing intelligence down the cost curve. Low-cost models do not need to win every benchmark to matter. They can win by making intelligence cheap enough to embed into software agents, consumer apps, enterprise workflows, robots and cars.

The three eras present compounding S-curves of compute demand, with none yet plateauing and each steeper than the last. The consequence is a sharp and continuing rise in demand for chips. It is this logic that underpins our holdings across the chip supply chain in TSMC and ASML, which are among our largest positions, and NVIDIA, to which we have been adding.

Investing in the supply chain is, in effect, a bet on the growth of AI itself, rather than a bet on which company will capture it. Whichever applications succeed and whichever frontier models prevail, the underlying compute demand runs through the same handful of companies. This is what makes investing in the supply chain such an unusually attractive way to own the growth of AI.

The pattern of revolutions

We are well aware that the history of revolutionary technology is also a history of market overshoot. Human and market psychology have a reliable capacity to misprice the path of even the most transformational innovation. The railway companies of the nineteenth century reshaped the modern economy. At their 1880s peak, they comprised roughly 60 percent of the entire US stock market, before a series of busts wiped out a great deal of capital. The canal buildout of the late eighteenth century and the fibre-optic buildout of the late 1990s followed a similar pattern: real technological progress, real economic impact, and real financial excess. We should expect the AI buildout to echo that history.

Yet the rational response is not to stand aside from a technological revolution. That is not the safe position it may appear to be. If AI disrupts most industries, then avoiding it doesn’t remove risk, it merely shifts it. You might still own businesses exposed to disruption, and not own the businesses in line for generational upside.

The harder task is to remain invested without becoming indiscriminate: to distinguish between durable value and temporary exuberance, between enabling infrastructure and fragile applications, and between companies that merely invoke AI and those that can turn it into enduring economic advantage.

The emergence of capable agents has made us more convinced that AI demand can keep expanding. But history argues for humility. There will be waste, disappointment and overbuilding along the way. Our job is not to believe every claim made for AI, but to own the exceptional companies that can benefit as intelligence becomes cheaper, more capable and more widely deployed.