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.
This manager report from Scottish Mortgage, a global growth fund, argues that AI has entered the 'agent era': programs that can plan and act on their own, not just chat. That means much more computing power, so the fund wants to own AI infrastructure like chipmakers and data platforms. It has done well lately, but the report also notes software stocks have lost about $2 trillion in market value, so the boom carries real bubble risk. The manager admits overbuilding may happen, but says avoiding AI just swaps one risk for another. Concrete examples, like Anthropic's revenue surge and China's AI progress, make it worth reading.
Generative AI has entered its third era, centered on AI agents: from ChatGPT's conversational approach and o1's reasoning approach, to Anthropic Claude Opus 4.5's agentic approach. Anthropic's annualized revenue jumped from $1 billion in January 2025 to $30 billion 15 months later—an unprecedented g
The author believes the "agent era" has arrived, and AI compute demand will expand in three overlapping layers—training/inference/agents—so one should take heavy positions rather than avoid it — [Optimistic]
The author uses Anthropic's revenue growth to argue that the "agent era" has arrived: its annualized revenue run rate jumped from $1 billion to over $30 billion in 15 months, and states that no company in recorded history has achieved such organic revenue growth at a comparable scale and speed. The author explicitly notes that Anthropic is one of Scottish Mortgage's private holdings.
| Milestone | Annualized Revenue Run Rate |
|---|---|
| January 2025 | $1 billion |
| 15 months later | Over $30 billion |
The author divides generative AI into three progressive eras: the conversational era began in late 2022, when OpenAI launched ChatGPT, and models could reliably follow natural-language instructions and interact back and forth; the reasoning era began in September 2024, when OpenAI released o1, and models learned to pause and think step by step, performing better on complex problems in mathematics, science, and programming; the agent era has become impossible to ignore since Anthropic released Claude Opus 4.5 in late November 2025 — models can be assigned a goal and advance it in multiple steps: planning, using tools, checking their work, and producing useful results without frequent human prompting. The author emphasizes that each era builds on the previous one rather than replacing it, and summarizes the relationship as: agents are reasoning models that act, and reasoning models are conversational models that think. The author believes that the rise of agents will have implications across its portfolio, reaching the structure of the software industry, the value of consumer businesses, the parallel development of China's AI ecosystem, and the physical supply chain that must be built to satisfy insatiable demand for computing power.
The author judges that agents will scale first in their impact on the software industry, identifying three structural effects — weakening seat-based pricing, intensifying competition and eroding moats, and value accrual shifting to the "intelligence layer"; the global software sector has seen about $2 trillion in market value evaporate over the past 12 months, but the author considers the sell-off "indiscriminate." Software is the first to bear the brunt because its work is entirely digital, high-value, clearly scoped, and fast to provide feedback; the author gives an example: when an application has a problem, an agent can be directed to locate the issue, write a fix, test it, and then deliver it.
The adoption-speed data is striking: Google's CEO says 75% of new code is written by AI; the author mentions that the founder of an unnamed portfolio company said the company's spending on AI coding tools has exceeded what it pays engineers, claiming the tools deliver a higher return than the engineers themselves; 70% to 90% of Anthropic's own code is generated by AI. The author believes software will increasingly be built, operated, and used by agents in the future, with three consequences: first, it weakens the link between software value and "seats" (licensed users) within organizations, which underlies most pricing in the software sector; second, it lowers the barrier to creating software, raises competitive intensity, and erodes moats built on accumulated code and complexity; third, it raises a deeper question of where value accrues — to the software application itself, or to the "intelligence layer" that understands the task, calls on data, and directs the work.
The market reaction has already appeared: the global software sector has shed about $2 trillion in market value over the past 12 months. The author acknowledges that part of this repricing is justified, as starting valuations left almost no room for questions of pricing power, competition, and value capture; but the author also believes the sell-off is indiscriminate — in the author's own words, "Not all software is equal." The author distinguishes between "product-style software intended mainly for human use" and "infrastructure-style software on which other digital activities depend": the former may be more exposed to disruption as agents change how humans interact with software, while the latter may benefit as agents lift underlying demand (data queries, security checks, compute loads, payments, and identity). Readers should note that the fund's own software holdings lean toward the infrastructure layer, and this "indiscriminate sell-off" argument carries the perspective of a holder.
The author discloses that its software holdings lean toward the "infrastructure layer," mapping five companies to the data, network, and payments foundations that agents require. The author states that its software holdings are structurally weighted toward this infrastructure layer:
| Company | Direction | Core Logic (One Sentence) | Key Data |
|---|---|---|---|
| Databricks | Held for observation | Organizing governed corporate data within enterprises is a prerequisite for agents to function inside companies | The original text does not disclose financial data |
| Snowflake | Held for observation | Organizing governed enterprise data to support agents' deployment within companies | The original text does not disclose financial data |
| Cloudflare | Held for observation | Provides the network and security layer required for agent applications to run, and helps websites identify, control, and charge AI agents | The original text does not disclose financial data |
| Adyen | Held for observation | Payments and trust infrastructure that allows agents to complete transactions safely on behalf of customers and merchants | The original text does not disclose financial data |
| Stripe | Held for observation | Also a payments and trust foundation; its founder cautions that agent commerce will arrive in "small pieces" rather than a sudden leap | The original text does not disclose financial data |
The "held for observation" label in the table is based on the original text's explicit listing of these companies as software holdings; the original text does not indicate any buy or sell actions. The Stripe founder's judgment is that agent commerce will most likely arrive in small increments rather than one big leap; but each increment of autonomy still requires programmable, permissioned, and trustworthy financial rails.
The author judges that the impact of agents will not stop at software development, but will spread to knowledge work and consumer domains: Amazon, MercadoLibre, and Sea Limited plan to build vertical agents of their own, while Nubank has the potential to realize a "private banker in your pocket" vision on a low-cost model (the fund's direction of action on the latter four is not stated in the original text). Breaking down knowledge work, the author argues that drafting legal memos, synthesizing clinical trials, building financial models, and reviewing patent applications differ at the professional level but are common at the level of cognitive operations — reading, writing, reasoning, and completing transactions with software tools — precisely the capabilities in which agents are getting stronger. In the author's own words, "agents are not a tool for one industry but many."
On the consumer side, agents will act as personal shoppers, financial advisors, and everyday assistants; the author also flags a risk: if horizontal AI assistants interpose themselves between customers and platforms, platform value could be siphoned off. But the strongest platforms hold the assets agents need — trust, customer history, payments, credit, logistics, product catalogs, and merchant networks — which is why Amazon, MercadoLibre, and Sea Limited are planning to build their own vertical agents, serving their own platforms while also extending shopping scenarios beyond them. Nubank offers a similar possibility in finance: its founder, David Vélez, articulated the vision of "giving every customer a private banker in their pocket" long before agents became popular, and agents bring that goal closer to reality — helping users manage bills, understand spending, time borrowing decisions, build savings, and find the lowest rates in the market. The author believes agents will improve price transparency, tailor options to individual circumstances, and reduce the friction of taking action, and that Nubank, as a low-cost operator, is favorably positioned.
In closing, the author offers a forward-looking assessment and self-positioning: few companies will be entirely immune; for some, agents create new demand; for others, they threaten existing profit pools; and for many, both will happen at once — this will be the key challenge for growth investing in the years ahead. The author claims Scottish Mortgage is able to cope because it invests in both public and private markets, and many companies shaping the AI frontier remain private, an exposure that affords a broader perspective. This is a self-promotional statement by the fund — "able to cope" is a holder's judgment rather than third-party verification, and readers should be mindful of the fund's standpoint.
The author believes China is not an AI follower but has formed differentiated advantages across three areas: physical AI, cost-performance, and productization; these judgments correspond directly to companies already held in the portfolio.
The author concedes that the Silicon Valley narrative is largely correct on frontier model capability, but considers it incomplete. China's first advantage is physical AI: embodied intelligence progresses fastest when virtual training connects to real-world deployment; China's manufacturing provides the largest global deployment scenarios, together with the largest installed base of industrial robots, a dense local supply chain, supportive policies, and an EV industry that has already combined software, hardware, and cost-oriented manufacturing. Horizon Robotics sits at the intersection of AI and the physical world, applied to autonomous driving with plans to extend into broader robotics.
The second advantage is cost-performance. Restricted access to advanced chips has forced Chinese model companies to do more with fewer resources; and the agent era will depend on compute even more than the conversational era, so the cost of useful intelligence must come down substantially. MiniMax's open-source models approach frontier capabilities at a small fraction of the training cost, part of a Chinese ecosystem that is pushing intelligence down the cost curve. Low-cost models do not need to win every benchmark, as long as they make intelligence cheap enough to embed in software agents, consumer applications, enterprise processes, robots, and automobiles.
The third advantage is productization. The author believes that when generative AI is combined with companies skilled at recommendation, interface design, and viral distribution, it can quickly become a mass consumer habit; ByteDance's Doubao is exactly this, with monthly active users exceeding 226 million, leading the Chinese market. The next phase of AI will be shaped by companies that make intelligence cheaper, more useful, more physical, and more habitual.
Position Moves:
| Ticker | Direction | Core Logic (One Sentence) | Key Data |
|---|---|---|---|
| Horizon Robotics | Hold / Monitor | At the intersection of AI and the physical world, focused on autonomous driving and extending into robotics | Largest global deployment scenarios and largest installed base of industrial robots (no specific figure given in the original) |
| MiniMax | Hold / Monitor | Open-source models approach frontier capabilities at extremely low training cost | Training cost is only a small fraction of frontier models (no specific figure given in the original) |
| ByteDance | Hold / Monitor | Doubao shows generative AI can rapidly become a mass habit | Monthly active users exceed 226 million, ranking first in the Chinese market |
| TSMC | Hold / Monitor | Core of the chip supply chain, benefiting from the three-layer stacking of compute demand | One of the largest positions (no percentage given in the original) |
| ASML | Hold / Monitor | Core of the chip supply chain, benefiting from continuously rising compute demand | One of the largest positions (no percentage given in the original) |
| NVIDIA | Add | Leader of the chip supply chain; the author is continuously increasing the allocation | No specific data given in the original |
Here the author uses China's advantages to justify the portfolio's logic; readers should note that this is the perspective of a position holder.
The author judges that generative AI's compute demand is stacked from three layers — training, inference, and agents — with the agent layer growing fastest and continuing to amplify because the ceiling of human attention has been removed; the stance on AI demand is optimistic.
Training was the original driver of the conversational era and is still expanding; the inference era added a second layer — models spending more compute checking logic and considering alternatives before answering; the agent era adds a third layer without removing the first two, and it is growing fastest. Anthropic data shows that a single agent consumes roughly 4x the compute of a single conversation, and multi-agent systems roughly 15x. The bigger change is this: AI demand was previously constrained by human attention — a person can only ask a limited number of questions in a day; agents remove that ceiling, able to loop through dozens of reasoning cycles, run autonomously while humans sleep, and collaborate with other agents. The author's original words: "The three eras present compounding S-curves of compute demand, with none yet plateauing and each steeper than the last," meaning: "The three eras present compounding, stacked S-curves of compute demand, none of which has yet entered a plateau, and each is steeper than the one before."
The author defines chip supply chain positions such as TSMC, ASML, and NVIDIA as a bet on overall AI growth, rather than a bet on any single application or model winner.
The direct result of continuously rising compute demand is strongly rising chip demand. TSMC and ASML are among the largest positions, while NVIDIA has been consistently added to. The author believes that regardless of which applications succeed or which frontier models win, the underlying compute demand will flow through the same set of companies, making the supply chain an unusually attractive way to own AI growth.
The author acknowledges that the AI buildout may repeat the overinvestment seen in railways, canals, and fiber optics, but believes that avoiding a technological revolution is not a safe position; the stance is cautious on market valuations and constructive on AI participation.
Human and market psychology have a reliable capacity to misprice the most transformative innovations. Nineteenth-century railway companies reshaped the modern economy, accounting for roughly 60% of total US stock market value at their 1880s peak, before a series of crashes destroyed enormous amounts of capital; canal construction in the late 18th century and fiber optic buildout in the late 1990s followed the same pattern: real technological progress, real economic impact, real financial excess. The author expects the AI buildout to echo this history but believes that avoiding it is not safe. The author's original words: "If AI disrupts most industries, then avoiding it doesn't remove risk, it merely shifts it," meaning: "If AI disrupts most industries, then avoiding it does not eliminate risk — it merely shifts it." The emergence of agents makes the author more confident that AI demand can keep expanding, but history demands humility: there will be waste, disappointment, and overbuilding along the way. The harder task is to remain invested without being indiscriminate: distinguishing durable value from temporary frenzy, enabling infrastructure from fragile applications, and companies that merely mention AI from those that can convert it into long-term economic advantage. The author says the job is not to believe every AI claim, but to own the exceptional companies that benefit as intelligence becomes cheaper, more powerful, and more pervasive. Readers should note that this is a position holder's argument for the rationale of continuing to hold.
In the year ended March 31, 2026, the share price and NAV outperformed the benchmark by 8.8 and 9.4 percentage points, respectively; the original text provides no attribution for the relative performance.
Performance comparison: The following is Scottish Mortgage's performance for the years ended March 31 each year (GBP total return; NAV calculated at fair value after deducting borrowings; benchmark is FTSE All World Index (GBP) TR):
| Year | Share Price Return (%) | NAV (%) | Benchmark (%) |
|---|---|---|---|
| 2022 | -9.5 | -13.1 | 12.8 |
| 2023 | -33.5 | -17.8 | -0.9 |
| 2024 | 32.5 | 11.5 | 21.0 |
| 2025 | 6.0 | 11.2 | 5.5 |
| 2026 | 26.8 | 27.4 | 18.0 |
Across the five years, FY2022 and FY2023 significantly underperformed the benchmark; in FY2024 the share price outperformed but NAV lagged; in FY2025 and FY2026 both outperformed. Past performance does not represent future returns.