Please refer to our disclaimers, which can be found in the footnote of this page and here. Critically, we make no guarantees of future performance, and our return projections should not be relied upon.

5 December 2025

Contents

• Contents
• Overview
• Where do we have Conviction?
• Portfolio Company Views
• Meta Platforms
• Alphabet
• Microsoft
• Salesforce

Overview

Our portfolio has meaningful AI exposure through Alphabet, Microsoft, Meta Platforms, Alibaba, and Salesforce. These companies are large consumers of AI compute (Search, YouTube, Gemini, Facebook, Instagram, Microsoft Copilot, Taobao/Tmall, Agentforce) and also providers of AI cloud compute (Google Cloud, Azure, AliCloud).

Discussions of a potential AI bubble have increased sharply over the last six weeks. You might ask: how can there be a bubble if everyone is talking about it? Yet there was also plenty of "bubble" talk in the run-up to the dotcom crash. We have been considering and monitoring this since our portfolio companies signaled significant increases in capital investment starting in 2022. We offer some thoughts below.

We are not macro pundits and don't claim to predict bubbles or their timing. Instead, we manage risk through business quality and margin of safety. We focus on how quality and long-term returns might be affected across plausible scenarios, examining the bottoms-up valuations of our companies.

Overall, we see Alphabet's valuation as stretched but not extreme, Meta Platforms as slightly stretched, Microsoft as reasonable, and Salesforce as cheap. This could indicate an emerging bubble, but we make decisions on a bottoms-up basis. We have no view on valuations elsewhere, though we would not be surprised to see more "bubble" signs in venture-backed businesses.

Where do we have Conviction?

We cannot predict an AI bubble, but we do have high conviction on several more general points:

• Generative AI ("gen-AI") productivity gains are real and growing: We are already seeing material productivity improvements, especially in coding, digital advertising, and customer service. CIO surveys and management commentary point to significant engineering labour productivity gains. For example, the CEOs of Microsoft and Salesforce claim 50% and 30% engineering productivity increases respectively when using coding copilots. There are also real gains in information retrieval and creative work.

The recent MIT report “The GenAI Divide” is often cited in AI bubble debates. It notes: “Despite $30–40 billion in enterprise investment into GenAI, this report uncovers a surprising result in that 95% of organizations are getting zero return.” This is frequently misread as “AI has no value.” The actual conclusion is that the 5% who are succeeding are doing specific things worth emulating. As with prior technologies, it takes time for diffusion and for organisations to adapt their processes.

• Real end demand already exists: User numbers across ChatGPT, Gemini, Google AI Overviews / AI Mode, Meta AI, Github Copilot, Anthropic, and others demonstrate this clearly. Adoption has been faster than any previous technology wave. Genuine end-customer revenue exists today. We estimate tens of billions of dollars in current revenue from model subscriptions, plus a similar amount from gen-AI uplift to existing businesses — primarily Meta Platforms, Google Search, and YouTube. Enterprise CIO surveys we track also show planned increases in AI spending.

Note: Nvidia and AI cloud service revenue is not end demand.

• Current unit economics are unreliable for extrapolation: Some argue AI cloud services require more capex per dollar of revenue than traditional cloud. This is true at today's prices, but the market is too young to reveal customers' true willingness to pay. If AI proves to be a strong labour substitute, willingness to pay could be very high. Anchoring on current pricing and early-stage utilization is misleading.
• A lot of capital is chasing the opportunity: Based on current announcements, we estimate $3–5 trillion in capex over the next 5–7 years (around 0.7% of global GDP at the midpoint; higher if you only consider US GDP). For comparison, the dotcom internet build-out peaked at around 1.2% of US GDP.
• This capital must earn a return to be sustainable: The most cited figure at present comes from David Cahn’s AI’s $600B Question. He estimates the industry needs 4x consensus Nvidia GPU revenues to be sustainable (GPU costs being 50% of the infrastructure and AI software providers needing a 50% gross margin). However, this assumes an Nvidia-monopoly status quo. Further, it ignores that the largest end-users today are Meta and Alphabet in their existing businesses, where the additional 50% margin requirement is too high. It also assumes that the capex level is perpetual in nature. If capital intensity falls after the build out phase, the long-term sustainable revenue could be lower. Whether this can indeed occur depends on what the steady-state infrastructure and supply-demand balance looks like. Overall, we think the sustainable number is lower but it is still considerably more than current revenue generation. We discuss this from a bottoms-up view below.
• Depreciation will have a material impact on cloud hyperscaler profit margins: This has slowly become consensus, though we were surprised it took the market two years to appreciate. After all, it's fairly mechanical. We've done extensive work on depreciation schedules that we believe match the economic reality of our companies. The lack of full disclosure makes this tricky, and companies that use finance leases at scale blur the picture further. The hyperscalers are also playing games with asset useful lives, but we use our own assumptions. The depreciation increases for Meta Platforms, Alphabet, and Microsoft are shown below. Meta Platforms plans to spend the most capital relative to its revenue and is therefore hit the hardest.

• There will be over-investment: Firms are racing for first-mover advantage. History suggests investment will overshoot adoption, with a subsequent correction. The severity and timing are unknowable but a correction of some sort is likely.
• An oversupply phase benefits end users: Over-building ultimately makes compute cheaper for consumers of compute. Some argue this is less valuable than prior investment bubbles because GPUs depreciate rapidly. That view ignores the long-lived components: data centres, power infrastructure, foundries, and accumulated intellectual property. That said, while lower pricing could be beneficial in the long-term for consumers of compute, the short-to-medium term would still be painful after a severe bubble.
• Higher-quality AI companies will weather the storm better: Business resilience is core to our business quality philosophy. Free-cash-flow generative businesses with strong balance sheets can absorb macro shocks and may even emerge stronger. Companies reliant on leverage or ongoing venture funding may not survive. Businesses like Microsoft, Alphabet, and Meta Platforms are investing in AI from a position of strength. This doesn't make them immune to an AI bubble, but it does give them flexibility in any downturn that may occur. In contrast, companies like OpenAI have no such flexibility.
• The first major AI setback could be supply-led, not demand-led: Most AI bubble analogies point to the dotcom bubble: huge over-investment, disappointing demand. That could occur here. But a different path is also plausible: a reset driven by supply constraints. It will likely take longer to build out the necessary infrastructure, especially power, than current capex projections imply.

In a supply-led scenario, bottlenecks would drive higher prices, compressing downstream margins. The ability to pass on higher costs would depend on pricing power at each layer. Returns on capital would fall; investment would slow; growth expectations would be revised downwards, and multiples would de-rate. This would be painful, but likely shorter in duration than a classic demand bust.

In a demand-led bust, by contrast, capital projects would be written off, leveraged and FCF-negative AI businesses would be heavily impaired, and there would be a broader risk-off shift. Over the medium to long term, cheaper compute would benefit end users. Where that benefit accrues would again depend on pricing power. Some of today’s perceived “AI losers” (e.g. traditional SaaS) could even come back into favour.

Portfolio Company Views

As mentioned above, we don’t make top-down predictions. Instead, we focus on the bottoms-up fundamentals of each company.

Meta Platforms

Management has disclosed that AI-driven recommendations have delivered:

• mid-single-digit percentage-point increases in engagement,
• low-to-mid single-digit improvements in ad conversions, and
• roughly 20% improvements in advertiser return on ad spend.

Meta also now has $60b of annualised revenue running end-to-end through its AI-powered ad tools. Adjusting for revenue that would likely have occurred anyway, we estimate gen-AI has boosted Meta’s revenue by roughly $15-20b in 2025. Meta has spent about $60b of capex cumulatively above its pre-gen-AI run rate through FY2025, which means these early gains represent a decent return on investment.

Meta will likely spend a further $290b on AI by FY2029, representing 12% of current market cap in present value terms. This is a very large investment and future returns will depend on whether AI-driven engagement and conversion gains can be sustained. In a “Goldilocks” upside, Meta could also capture more of the ad-creative layer with gen-AI ad content integrated into the ad targeting and optimisation algorithm.

While the future is uncertain, we can at least ask what's needed to produce a return on this capital. Let's assume capex and revenue reach steady state in 5 years, then grow in perpetuity at a standard economic growth rate (unlikely, but it's a useful exercise so bear with me). Depreciation takes another 15 years to fully normalize under such assumptions, given some assets have long useful lives. Let's also assume the discount rate is 15%, given the risk. We can now do a net present value ("NPV") calculation to work out what revenue run rate and margin is required. For example, at a 30% margin, Meta's AI revenues need to reach $170b to beat a 15% risk hurdle (see chart below for NPV calculations expressed as a % of current market cap for various revenue and margin combinations). The downside is capped at -12% of current market cap because management can always cut spending if the gains aren't materializing.

These are big numbers! For context, Meta's 2025 revenue will likely be around $200b, so the required additional revenue is 0.85x existing revenue or 52% of FY2029 revenue. The picture improves if run-rate capex declines after this 5-year investment phase. For example, if run-rate capex fell to 50% of the FY2029 level, only $75b would be required at a 30% margin (23% of FY2029 revenue). This is more achievable given current revenue progress, and a 30% margin roughly matches Meta's current margin less the depreciation impact. It's hard to know what the post-investment-phase capex will look like. It depends on market depth, the cost of AI compute assets once supply and demand normalize, the useful life of those assets, and many other factors.

We don't think management has a better crystal ball. Rather, they're betting that sacrificing 12% of the company's value is worth avoiding being left behind in AI. A key area of concern is that Zuckerberg appears somewhat ideological on AI. We may be wrong that management will cut spending if AI returns don’t materialise.

What's priced in? Our base case projects low double-digit IRRs at a $673 stock price. This would rise to mid-teens if management stopped AI spending today, and remain in the low double digits even if management wrote off all AI investments in five years. In other words, our base case assumes Meta's AI investments are NPV negative — success is the upside case. The valuation is slightly stretched, but not in bubble territory.

How would a bubble impact Meta? A supply-led AI bubble is probably worse for Meta than a demand-led one. In a supply crunch, input costs rise, putting pressure on ROI. In a demand-led bust, Meta’s share price would suffer in the short term, but its own proven use cases would benefit from cheaper compute over time.

Alphabet

Alphabet monetises AI through its existing businesses, Google Search and YouTube, and via Google Cloud and Gemini. Management claims its AI-driven ad products have increased conversions by 14%, and that AI Overviews drive 10% more queries when shown. There is clear potential for strong ROI within the core business alone. In addition, gen-AI search is structurally superior to legacy search, so some level of heavy investment is effectively existential for Google.

Google One subscriptions (with Gemini included in higher tiers) have surpassed 150m users, though the revenue attributable specifically to Gemini is unclear. We estimate Google Cloud (ex-Workspace) revenue at over $50b in FY2025, with about $7b from AI services. Going forward there could also be upside from future third-party TPU sales.

We estimate Alphabet will have spent $66b cumulatively above its pre-gen-AI capex run rate by FY2025, and another $275b by FY2029 (about 5% of current market cap in present value terms). This is similar in absolute terms to Meta, but on a larger revenue base. Comparable incremental gains therefore yield higher ROI, with AI cloud revenue as a further tailwind.

We ran the same NPV analysis for Alphabet as we did for Meta. The results are shown below. Given the similar spending levels, the required revenue and margin combinations are nearly identical. The key difference: in the downside case, only 5% of market cap is at risk. Alphabet has four ways to monetize its AI investments — existing businesses, Gemini, Cloud, and TPUs — while Meta has only one to date (existing businesses). This gives Alphabet a higher probability of success.

What's priced in? After the recent stock price appreciation, our base case projects mid-single-digit IRRs, assuming the AI investments are NPV neutral. Returns improve if the AI investments succeed. Since the capital spend is a much smaller percentage of market cap than Meta's, writing off the investments in five years would have limited impact. On the base case alone, this valuation is stretched.

How would a bubble impact Alphabet? A supply-led AI bubble would pressure Alphabet via higher input costs. A demand-led bust would likely force some write-downs, given a portion of capacity is being built for third-party demand that needs to materialise. However, the $270b in incremental AI capex through FY2029 is only around 5% of market cap, which we see as a measured commitment. Actual “wasted” capital should be lower unless management misses a clear demand slowdown for several years.

As discussed in prior letters, the larger strategic risk for Alphabet remains the transition to gen-AI search and whether it can defend its monopoly position.

Microsoft

Microsoft’s AI exposure is via its Copilots and AI cloud services. Earlier this year, Microsoft disclosed:

• AI revenue has surpassed a $13b run rate, and
• AI contributed 13 percentage points of Azure’s 31% growth rate.

We estimate Microsoft will have spent roughly $70b of capex above its pre-gen-AI run rate by FY2025, and a further $430b by FY2029 (8% of market cap in present value terms). To earn an adequate return, revenue must rise meaningfully and persistently. Given the greater planned investment spending, more AI revenue/margin is required. For example, at a 30% margin, a $280b run rate is required. See below for our NPV analysis on Microsoft’s AI investments.

What's priced in? Our base case projects low-to-mid teens IRRs, assuming the AI investments are NPV neutral. If the investments are written off in 5 years, returns fall to low double digits. Compared with Meta and Alphabet, Microsoft's future AI revenue mix is more heavily skewed toward cloud services than its existing businesses. We think this represents greater risk because the economics are less proven and success depends on third-party demand. Another key risk is Microsoft's technology and revenue dependence on OpenAI, which has made commitments of around $1.4 trillion against only ~$20b of run-rate revenue.

How would a bubble impact Microsoft? A supply-led AI bubble would hurt Microsoft through higher input costs and make adequate ROI harder to achieve. A demand-led bust would drive material capital impairment. Nonetheless, AI capex through FY2029 is ~8% of market cap — not trivial, but not “betting the farm” — and Microsoft retains flexibility to slow spend if demand disappoints.

Salesforce

Salesforce is primarily a consumer rather than a supplier of AI compute. At its recent investor day, the company disclosed a $440m run rate for Agentforce revenue as of July 2025. A supply-led bubble could slow growth by raising costs for end customers. A demand-led bubble would not be catastrophic in the long term, though the share price would still fall short-term. It would mean slower overall AI revenue growth. It would also mean cheaper compute, which would improve the Agentforce value proposition. Unlike the hyperscalers, Salesforce is not putting large amounts of capital at risk. Salesforce is also in the “AI loser” bucket, which may change if AI ceases to be a reality.

As discussed in our last letter, we think the valuation of Salesforce is cheap today and not representative of a bubble.