Scottish Mortgage is Baillie Gifford's flagship investment trust (founded 1909, LSE ticker SMT), known for its maximalist growth style — long-term stakes in Tesla, Amazon and ASML plus bold allocations to private companies like SpaceX and ByteDance. It is the UK retail investor's flagship vehicle for global disruptive growth.
Scottish Mortgage's fund manager looks back at the past year and argues that AI has entered an 'agentic' era: instead of just answering questions, AI can be given a goal and work through it using tools on its own. That could keep chip demand rising, split software winners from losers, and give Chinese AI firms a different path through cheaper models and real-world robotics. For everyday investors, the lesson is not to treat all 'AI' as one bet, and to remember this is a manager talking up his own holdings. Worth reading for the warning that transformative technologies often bring hype and overinvestment.
Scottish Mortgage's research indicates that AI has entered the agentic era: Anthropic's annualized revenue has surged from $1bn at the start of the year to $47bn, setting a historic record. Following conversational AI (ChatGPT) and reasoning models (o1), Claude Opus 4.5 marks the arrival of models c
The author of this issue, Lawrence Burns, is a fund manager at Scottish Mortgage Investment Trust. Reviewing the past 12 months, he argues that AI has entered its third generation — the agentic era — with effects set to spread from software to consumer platforms, China's AI ecosystem, and the physical supply chain for compute. The most consequential judgment in the piece: "Until now, AI demand has been capped by human attention — a person can only ask a limited number of questions a day, and every answer waits for the next prompt; agents completely remove this ceiling" (Lawrence Burns). This underpins the core investment thesis of "three generations of compute demand forming stacked S-curves, with chip demand rising continuously and steeply." Note: this piece is a fund manager's affirmative case for his own holdings, and the conclusion explicitly positions the fund as "a front-row seat in the AI revolution." Readers should be aware of this holder's perspective.
Lawrence Burns argues: at the end of November 2025, Anthropic released Claude Opus 4.5, marking generative AI's entry into the "agentic era" — models are no longer merely capable of conversation and reasoning, but able to act autonomously.
The most direct evidence is Anthropic's revenue curve: at the start of 2025, annualized revenue run-rate was $1bn; 17 months later, it surpassed $47bn. Burns says "No company in recorded history has grown organic revenue at this scale and pace." Anthropic is one of Scottish Mortgage's private holdings.
The boundaries between the three eras are clear:
| Era | Starting Point | Landmark | Newly Acquired Capability |
|---|---|---|---|
| Conversational | Late 2022 | OpenAI's ChatGPT | Reliably following natural-language instructions and engaging in multi-turn dialogue with users |
| Reasoning | September 2024 | OpenAI's o1 | Pausing to think step by step, producing considered rather than intuitive responses, particularly strong in mathematics, science, and coding |
| Agentic | Late November 2025 | Anthropic's Claude Opus 4.5 | Once given a goal, advancing through multiple steps: planning, using tools, self-checking, without requiring continuous human prompting |
Burns stresses the key mechanic: each era is layered on top of the previous one rather than replacing it. The progression he offers: "today's agents are reasoning models that have learned to act, just as reasoning models are conversational models that have learned to think." From this, he extrapolates that the impact will span the entire portfolio: from software, to consumer business value, to China's parallel AI ecosystem, and to the physical supply chain that must be built to satisfy nearly insatiable demand for compute.
Burns argues: the market's repricing of the software sector is broadly rational in direction, but "indiscriminate"; agents hit software first, yet software is not all the same — product-style software built for human users is vulnerable to disruption, while infrastructure-style software underpinning digital activity actually benefits.
Adoption speed has been confirmed by data: Google's CEO says 75% of new code is already written by AI; the founder of one portfolio company told Burns that his spending on AI coding tools for engineers now exceeds his salary expense, and these AI tools deliver better returns than the engineers themselves. From this, Burns extrapolates: software is moving toward a future "increasingly built, operated, and used by agents."
The magnitude of the market response: the global software sector has lost roughly $2tn in market capitalization over the past 12 months. Burns concedes that part of the repricing is justified — starting valuations left no room for questions about "pricing power, competition, and value capture"; but he notes the repricing is collateral damage: "it has also been indiscriminate, and not all software is equal."
Mechanism breakdown:
Scottish Mortgage's software holdings are clearly tilted toward the infrastructure layer: Databricks and Snowflake organize enterprise governance data, a precondition for agents to function inside the enterprise; Cloudflare provides the network and security layer, helping websites identify, control, and charge AI agents for access; Adyen and Stripe provide payment and trust infrastructure, enabling agents to transact securely on behalf of customers and merchants.
Burns goes further: the impact of agents will not stop at software development. Strip away the components of most knowledge work and you have reading, writing, reasoning, and using software to complete tasks — "that is exactly what agents are learning to do." Agents are therefore a tool across industries, not a single trade. They will become personal shopping assistants, financial assistants, and everyday assistants. If AI assistants insert themselves between customers and platforms, that is disruptive; but the strongest platforms control the assets agents need: customer history, payments, credit, logistics, and merchant networks. That is why Amazon, MercadoLibre, and Sea Limited are building their own agents, both to serve their own platforms and to extend shopping scenarios beyond the platform. In Latin America, Nubank offers a similar possibility in finance: before the agent craze, founder David Vélez told Scottish Mortgage that his ambition was to give every customer "a private banker in their pocket"; agents make this ambition more practical — managing bills, understanding spending, and finding the lowest market rates for users, improving price transparency and lowering friction. Burns cautions that for many companies, the two effects happen simultaneously: agents both create new demand and threaten existing profit pools, and this will be the core challenge for growth investing over the coming years. His response logic is to position across public and private markets — many shapers of the AI frontier remain private companies, and the private channel lets the fund observe the pace of technological improvement, adoption patterns, and where value lands with greater completeness (a statement that also defends the fund's own structural advantage).
Burns argues: China is not a follower in AI, but has grown two distinct advantages under different constraints — "physical AI" and "cost-performance"; supply-chain investment is essentially a bet on "AI's overall growth"; and history shows revolutionary technology is invariably accompanied by financial excess, so avoidance is not a safe position.
Two threads of China's advantage:
1. Physical AI: simulation matters, but embodied intelligence progresses fastest when virtual training connects to real-world deployment. China's manufacturing base provides the world's largest deployment surface, combined with the largest installed base of industrial robots, dense local supply chains, and supportive policy. Portfolio company Horizon Robotics sits at this intersection, entering through autonomous driving with ambitions to expand into broader robotics.
2. Cost-performance: restricted access to advanced chips has forced Chinese model companies to "do more with less." The agentic era is far more compute-intensive than the conversational era, and for intelligence to spread widely, costs must fall sharply. Portfolio company MiniMax's open-source models approach frontier capability at a tiny fraction of the cost — "these models do not need to win every benchmark; they will win by making intelligence cheap enough." Cheap enough to embed in software agents, consumer applications, enterprise workflows, robots, and cars. Burns therefore argues that the next phase of AI will be shaped not only by companies with the biggest models, but also by companies that can make intelligence cheaper, useful, and deployable in the physical world.
The compounding logic of compute demand: each generation of generative AI layers a new demand on top of existing demand rather than removing the old layer; the agentic layer is the third and fastest-growing layer. Anthropic's own data shows: a single agent accomplishes about 4 times the work of an ordinary chat, and a team of cooperating agents about 15 times as much. The larger structural shift — the most critical assertion of Burns's entire piece — is that the cap human attention placed on AI demand has been removed: once given a goal, agents can loop through dozens of reasoning steps, work through the night while we sleep, and increasingly coordinate with other agents on the same task. The three generations of demand therefore form "stacked S-curves," none yet peaked, each steeper than the last; the consequence is sharp and sustained upward demand for chips, corresponding to holdings TSMC, Nvidia, ASML, and SK Hynix. Burns gives the core strategic rationale: "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." Regardless of which application succeeds or which frontier model wins, underlying compute demand flows through the same set of supply-chain companies — that is what makes the supply chain "unusually attractive."
Sober about history: the history of revolutionary technology is also the history of market overshoot. Burns cites examples: 19th-century railroad companies reshaped the modern economy and at their peak accounted for roughly 60% of the entire U.S. stock market, before a series of crashes consumed vast amounts of capital; canal construction at the end of the 18th century and fiber-optic construction in the 1990s followed the same pattern — "real technological progress, real economic impact, real financial excess" — and AI buildout should repeat it. But he argues that the rational response is not to stand aside: "avoidance does not eliminate risk; it merely transfers risk." (That is, avoiding AI is not as safe as it appears; you may still own disrupted businesses while missing the generational winners.) The harder task is to remain invested without becoming indiscriminate: distinguishing durable value from temporary mania, enabling infrastructure from fragile applications, and companies that merely "cite AI" from companies that can turn AI into a durable competitive advantage. The author concedes that "history demands humility": there will be waste, disappointment, and overbuilding along the way; his response is not to believe every claim about AI, but to own the exceptional companies that benefit as intelligence becomes cheaper, more capable, and more widely deployed.
| Company | Author's Stance | Key Data / Role |
|---|---|---|
| Anthropic | Bullish | Annualized revenue from $1bn to $47bn in 17 months; Claude Opus 4.5 opens the agentic era |
| Databricks | Bullish | Enterprise governance-data infrastructure; necessary precondition for agents to operate inside the enterprise |
| Snowflake | Bullish | Same as above |
| Cloudflare | Bullish | Network and security layer; helps websites identify, control, and charge AI agents for access |
| Adyen | Bullish | Payment infrastructure required for agent transactions |
| Stripe | Bullish | Trust infrastructure required for agent transactions |
| Amazon | Bullish (holding status not disclosed) | Owns assets agents need: customer history, payments, credit, logistics, and merchant networks; building its own agents |
| MercadoLibre | Bullish (holding status not disclosed) | Same as above |
| Sea Limited | Bullish (holding status not disclosed) | Same as above |
| Nubank | Bullish (holding status not disclosed) | Founder David Vélez's "private banker in the pocket" vision; agents make it practical |
| Horizon Robotics | Bullish | China's physical AI: autonomous driving extending into broader robotics |
| MiniMax | Bullish | Open-source models approach the frontier at a fraction of the cost; winning by "making intelligence cheap" |
| TSMC | Bullish | Compute supply chain; core beneficiary of stacked S-curves across three generations of demand |
| Nvidia | Bullish | Compute supply chain |
| ASML | Bullish | Compute supply chain |
| SK Hynix | Bullish | Compute supply chain |
1. Lawrence Burns (Scottish Mortgage): Agentic AI has removed the ceiling human attention placed on compute demand — a single agent does roughly 4 times the work of an ordinary chat, an agent team about 15 times, and the three generations of demand form stacked S-curves, so chip demand will continue rising steeply; investing in the supply chain is therefore a bet on "AI's overall growth," rather than on a single winner.
2. Lawrence Burns (Scottish Mortgage): The history of revolutionary technologies is inevitably accompanied by financial excess (railroads collapsed after peaking at roughly 60% of the U.S. stock market), but avoiding AI is not a safe position — it merely transfers risk into "holding disrupted companies and missing generational upside"; the real task is to distinguish durable value from temporary mania, infrastructure from fragile applications.