Baillie Gifford is an Edinburgh investment partnership founded in 1908, famous for ultra-long-horizon, high-conviction growth investing — its early stakes in Amazon, Tesla and NIO are classics. Its "actual investors" philosophy holds world-changing companies on 5-10 year views; AUM is around $120bn. The Insights column carries its managers' investment views and thematic research.
This article is a fund manager's defense after AI-related stocks fell in July. Baillie Gifford argues the selloff was about crowded positions and panic, not a broken long-term story. Demand for AI computing power remains strong, so it is keeping its semiconductor and AI infrastructure holdings. Three companies stand out: Astera Labs, whose chips speed data flow between GPUs, saw revenue double from a year earlier; BESI, which makes equipment for connecting chips and memory, grew quarterly revenue 69%; IREN, which turns cheap electricity into computing power, expects over $4 billion in annualized AI cloud revenue by 2026, with about 85% already contracted. The manager sees these as key bottlenecks.
The author believes the July AI infrastructure sell-off was a positioning and sentiment issue, not a reversal of the long-term thesis; AI compute demand remains robust, and the portfolio deliberately maintains exposure to semiconductors and AI infrastructure [Optimistic]
At the end of July, the market turned sharply against everything tied to AI infrastructure buildout. Given the rapid run-up in share prices, elevated expectations, and converging investor positioning, the author acknowledges there is reason to suspect some stocks had run ahead of company fundamentals; but the more critical question is whether the long-term opportunity has changed, and the author believes recent events do not prove that. The triggers are threefold: First, Moonshot AI released the open-weight model Kimi K3, with training parameters publicly disclosed and performance (especially in coding) approaching top-tier models, shaking the business-model narrative of leading model vendors; its architecture handles long sequences efficiently by caching already-processed information rather than recomputing each time. Second, Alphabet again raised its 2026 capital expenditure guidance, and media reports suggest OpenAI and NVIDIA have large-scale future AI campus investments. Third, banks reportedly made margin calls on hedge funds levered up on AI bets; the author notes that forced selling says nothing about long-term value but does amplify short-term share price volatility.
Kimi K3's memory efficiency has raised concerns, because memory is one of the most expensive and constrained elements of running AI models. But the author explicitly disagrees with the inference that "more memory-efficient models = lower demand." The author's exact words: "Efficiency may affect the shape and pace of demand, but it does not make it disappear." — that is, efficiency may affect the shape and pace of demand, but it does not make demand disappear. The reasoning: Kimi K3 is still a large model capable of processing vast amounts of information simultaneously, and running a model of this scale still requires substantial compute and memory. The article acknowledges that efficiency will change the shape and pace of the demand curve, but it does not change the existence of demand itself.
The market's long-standing unease about the largest cloud vendors' spending is another source of the sell-off sentiment. Alphabet again raised its 2026 capital expenditure guidance, and combined with reports of large-scale AI campus investment involving OpenAI and NVIDIA, this intensified the debate over "whether this money is worth it." The author offers a two-sided reading: the same spending both makes investors worry about future returns and shows that today's tech giants still see enough demand to justify building capacity at scale. Massive capital expenditure is itself evidence that demand exists.
The fund's portfolio underperformed its benchmark in the short term due to declines in semiconductor positions (roughly 15% of assets) and broader AI infrastructure names such as IREN. The author acknowledges this is uncomfortable but not entirely surprising, because the past two years involved deliberately building this kind of exposure. In the second quarter these positions were the major positive contributors; although part of those gains has since been given back, they remain net positive contributors over the past 12 months. The article emphasizes: the investment logic of these companies does not depend on AI infrastructure spending continuing to grow at the recent high rate every year. The firm looks for companies solving important problems — as AI systems scale up, these technologies should become more important, and even if overall buildout slows, company revenue can still grow faster than the broader market; recent results have already demonstrated real demand, so the logic does not rest on future growth alone.
The three cases together point to the AI bottleneck areas the author favors: data movement, chip-to-memory connectivity, and low-cost power.
Astera Labs: solves the data-movement bottleneck when GPUs coordinate at massive scale. AI systems increasingly have large numbers of GPUs working together across entire racks; the challenge is not just the number of chips but the speed of data movement between them. Its connectivity chips and software are purpose-built for this. Latest quarter revenue was $392m, up 104% year over year, and the company guides for continued revenue growth next quarter. Author's stance: clearly bullish.
BESI: solves chip-to-memory connection efficiency. As it has become harder to improve single-chip performance, the industry is shifting toward better interconnect solutions; hybrid bonding offers higher density with better speed and thermal performance. The article cites discussions with semiconductor research institute imec and argues that hybrid bonding is likely to become unavoidable in high-bandwidth memory over the long term. BESI's second-quarter revenue was up 69% year over year, and management expects continued growth in the third quarter; the article also acknowledges adoption is still uneven, but BESI has already benefited from the advanced packaging trend. Author's stance: bullish.
IREN: provides the capability to convert power into compute; it is an AI infrastructure company but not a semiconductor equipment maker. The constraint is cost — customers care both about whether they can obtain GPUs and about who delivers them at the lowest sustainable cost. Its Bitcoin mining history has forced discipline in power procurement, site selection, data center design, and operational efficiency, and this cost discipline should help it compete. Scale figures: annualized AI cloud revenue expected to exceed $4bn by end of 2026, of which roughly 85% is already contracted; AI cloud capacity has grown from 3 MW at the start of this year to 480 MW, potentially reaching 1.2 GW by 2027. The company's CEO recently wrote that "capacity (demand) exceeds everything we can build" — i.e., demand capacity exceeds everything we can build. Author's stance: bullish.
| Company | Bottleneck Position | Latest Data | Author's Stance |
|---|---|---|---|
| Astera Labs | Data flow between GPUs | Last quarter revenue $392m, +104% YoY | Bullish |
| BESI | Chip-to-memory connectivity (hybrid bonding) | Q2 revenue +69% YoY | Bullish |
| IREN | Low-cost power to compute | 2026 AI cloud annualized revenue >$4bn, 85% contracted | Bullish |
The author has not changed the long-term view on AI infrastructure because of the July sell-off, and continues to selectively support companies whose technologies can become key parts of AI infrastructure and whose recent results have validated demand; the portfolio's exposure to semiconductors and AI infrastructure is deliberately maintained. Note that this is a self-defense from the holder's perspective — the "not surprised, not concerned" language in the article itself carries the purpose of reassuring investors, and readers should view the explanation of the sell-off in light of the author's own position.
| Position | Direction | Author's Stance in One Sentence | Key Data |
|---|---|---|---|
| Astera Labs | Hold — observe | Clearly bullish, positioned at the data-flow bottleneck during large-scale GPU co-processing | Last quarter revenue $392m, +104% YoY; next quarter expected to continue growing |
| BESI | Hold — observe | Bullish on hybrid bonding technology; long term it may become inevitable in high-bandwidth memory | Q2 revenue +69% YoY; management expects Q3 to continue growing |
| IREN | Hold — observe | Bullish on its cost discipline and ability to convert power into compute; demand exceeds buildable capacity | 2026 AI cloud annualized revenue expectation >$4bn, ~85% already contracted; capacity 3MW→480MW |
| Moonshot AI | Not disclosed | Kimi K3, with its efficient long-context architecture, challenges the narrative of leading model makers and is the sell-off trigger | Open-weight model, performance close to top models (especially coding), uses caching to process long sequences |
| Alphabet | Not disclosed | Raising capital expenditure raises return concerns, but the author believes massive spending in fact proves strong demand | 2026 capital expenditure guidance raised again |