I argue that Spotify’s moat is much wider, built on personalisation, user habit, and building an ecosystem. It is a business which is critical to the music industry, getting more relevant to more users globally, has inherent pricing power, and an attractive earnings profile which I see compounding at a high-teens rate over the next five years.

The Case Against Spotify

The bear thesis centres upon three core assertions:

1. Spotify is an effective distribution system for the music industry, but it doesn’t own the underlying content, much of which is commoditised anyways.

2. It is competing with well capitalised players who bundle music for free.

3. Its growth is increasingly being driven by lower quality customers in emerging markets.

Combined, these three factors suggest the business has a narrow moat, which makes it inherently difficult to underwrite several years of earnings growth with significant conviction.

A good parallel presented was that of Dropbox (NASDAQ: DBX). Much like Spotify, Dropbox built a superior access and organisation layer on top of content it did not own, it competed with the bundling capabilities of Microsoft and Google, had a margin inflection story and attempted to expand its TAM beyond its core product.

Dropbox grew for many years, maintained strong brand loyalty and reported healthy user metrics well into 2019-2020. However, the decay was gradual. New user acquisition slowed, churn crept up at the margins and growth flatlined. Paying users for Dropbox have now stagnated at around 18m. As a result, the stock has effectively done nothing since the company’s IPO in early 2018 with the multiple continuing to compress despite growing earnings.

These are all valid arguments and more or less true to a certain extent. The question, therefore, is does Dropbox serve as a cautionary tale for Spotify or is there more nuance to the story here?

Curation at Scale

The bear case somewhat reduces Spotify to a pipe that delivers music, and much like a pipe, it assumes that one is no better than another. This may have been true in the early days of the company, but I don’t believe that to be true today.

Notably, Spotify has built the most sophisticated recommendation engine in audio. This is an advantage that continues to compound with time. Every stream, skip, and playlist feeds into a personalisation engine built on 750m+ users and nearly two decades of listening data. This data asset is far from a commodity and almost impossible to replicate. It is the core reason the user experience on Spotify is markedly better than Apple Music, Prime Music or YouTube Music.

Features such as Discover Weekly, Release Radar, AI DJ, and Daily Mixes are powered by the personalisation engine, and this translates to measurable user behaviour. Here are a few statistics:

• Users who engage with Daily Mix stream roughly twice as long as users who do not.

• One in six (or 48m) Premium users use the AI DJ feature weekly.

• Average daily listening time continues to increase, now at 148 minutes per user up from 114 minutes per user.

• Recommendation accuracy has improved such that skip rate is down to 28%.

Better personalisation drives higher engagement. Higher engagement results in higher retention. Higher retention drives customer lifetime value and ultimately enterprise value for Spotify. Excess profits are then reinvested back into improving the personalisation of the product and the flywheel continues.

Two decades of learning, iteration, and investment results in the product millions of users around the world enjoy today. And it continues to get better. AI is unlocking even more opportunity for the business to drive engagement on its platform through new features like Prompted Playlists or Personalised Podcasts.

I recognise that product moats erode, and the best product doesn’t always win. The competitive environment also feels especially acute in this case. Apple, Amazon, and Google all want a slice of this market, each brings deep technical talent, ample capital and the ability to bundle music at a loss. But for all three, music is a non-core line and a rounding error against group earnings. It’s hard to outcompete a scaled incumbent like Spotify, that pours every resource into improving its only product.

The results bear this out. Apple has bundled music into Apple One since 2020, Amazon gives Prime Music away with Prime, and YouTube Music comes free with YouTube Premium. Yet Spotify’s share of global music streaming sits at 33% and keeps climbing. It has roughly 3x Apple’s paid subscribers, despite Apple owning the device ecosystem, pre-installing Apple Music, and throwing in a free three-month trial with every new device.

Listeners choose Spotify not because of its price or catalogue but because of the experience. That choice is underpinned by the superior personalisation which is a moat that compounds over time.

Building The Audio Ecosystem

I am also of the view that Spotify plays a more important role than of a mere middleman for how we consume audio. Now, being a middleman isn’t a bad business, in and of itself, especially when you wrap value around the product you distribute.

But I’d go further. Spotify is far more consequential to the audio industry than even a value-added distributor. It has built an ecosystem and serves as a critical piece of infrastructure atop which the entire audio streaming industry operates.

A brief detour through history makes the point. The recorded music industry has passed through three structural eras.

1. Physical Era – In the physical era, labels pressed CDs, shipped them to retailers who then sold it to customers. No single store held pricing power or informational advantage, and industry sales peaked in 2000 at $22bn globally.

2. Digital Era – Then followed the digital era, bringing along piracy which gutted the industry. Apple’s iTunes partially stabilised things by offering $0.99 digital downloads but this remained a simple transaction with no insight into listening behaviour. By 2014, the industry had bottomed out at $13.1bn of global sales.

3. Streaming Era – Spotify introduced the third model of music consumption – unfettered access. It shifted consumer behaviour from ownership of specific content to on-demand streaming of all content. In doing so it breathed life back into the industry with global recorded music revenues growing 6% and reaching $31.7bn as of 2025.

Beyond reviving the industry, Spotify’s ecosystem delivers more value to everyone in it. Let’s just focus on the two main participants here – the listeners and the creators. What does each actually get?

• For Listeners – A near-endless library of music, podcasts and audiobooks in one app, surfaced by a discovery engine which curates content specifically to their taste. Even the free tier is a genuine upgrade to radio. The result is an experience that is multiples better than any form the music industry has taken before and more comprehensive than that of its peers.

• For Creators (and their labels) – Analytics on who’s listening, where and how. Tools to promote releases, identify super fans, sell tickets and merch. Podcasters get ways to monetise their content, connect with advertisers and now deliver video content. Most importantly, Spotify offers a single front-door to 750m+ listeners and an ability for creators to perpetually monetise their back catalogue.

What’s especially interesting is that Spotify has brought down the cost of discovery and listening to new content to effectively zero. Under previous models, the requirement to purchase new content created friction that prompted listeners to only stick with what they knew or were familiar with. Today, for $12.99 per month, a user has access to 100m+ songs, thousands of podcasts, and audiobooks, which encourages exploration. This is not only compelling value from the user’s perspective but highly beneficial for creators looking to expand their audiences.

This ecosystem advantage compounds and the value each side gets increases as Spotify grows. In the physical and digital era, distribution was a commodity and simply facilitated a transaction. In the streaming era, Spotify manages the relationships between creators and their audiences.

The Habit You Won’t Break

Consuming audio content is largely habitual. My own day is a good example: a workout most mornings accompanied by music, a podcast on the commute to and from the office, and some sort of ambient, low-key music while I’m at my desk. All in, I’m consuming between three to five hours a day of audio content on Spotify. The app is therefore deeply entrenched in my daily routine such that I cannot imagine a day without it.

Human nature is generally averse to change and when people architect their daily lives around a product or service, the incentive costs are higher. This is why people rarely switch bank accounts (when it’s easy to do so) even though they are being subject to low interest rates on balances and high fees.

For many people across the globe, and particularly in markets such as the US, Australia, UK, Spotify is the default interface through which they interact with music and audio content. The more they use it, the stickier it is. Spotify’s first mover advantage in these large developed markets has been real in establishing a loyal base of customers.

In the same vein, the company is expanding aggressively in nascent markets like Brazil and India. The idea here being that they acquire customers through a compelling free tier, build habits and then monetise that relationship further through premium conversions over time. These may be low value customers today, where churn rates are higher and ARPU is lower. But it represents a huge funnel of potentially valuable customers five to ten years out.

Another important point, related to habit is social cohesion. Adding social features to its platform such as Spotify Wrapped, shared playlists, Jams, Spotify Blends, and the Family Plan, encourages adoption within tightknit social circles. People tend to follow the advice of those closest to them and generally don’t like being outsiders. As such, being outside the Spotify ecosystem, when your peers are within it, levies social costs that most people are unwilling to shoulder.

It might be challenging to underwrite human psychology into an EPS forecast, but it’s a powerful force which creates real hurdles to well-capitalised peers gaining share.

Conclusion

So, will Spotify’s fate mirror that of Dropbox? I clearly don’t believe so.

Dropbox’s access layer was ultimately replicable. A folder in the cloud is a folder in the cloud, and once Microsoft and Google bundled the same thing in for free, the reason to pay evaporated. Spotify is a different animal. Its personalisation engine compounds with every stream, its ecosystem sits between creators and their audiences rather than merely shuttling content between them, and habit has quietly woven the app into hundreds of millions of daily routines. None of those are things a competitor can bundle their way past.

Trading at roughly 32x forward earnings, Spotify isn’t exactly cheap. However, if we assume the business continues to compound its earnings at a high-teens rate, the stock should roughly double by 2030 with no help from the multiple at all. There may be some risk with valuation, but investors are being well compensated with the quality of the franchise you receive for the price. Whilst many see a narrow moat around a simple app, I see one of the widest moats in the media industry, getting wider. We can settle the debate over lunch in 2030.

If you’d like to get in touch with me please email me at moatsandmultiples@gmail.com

The headline for me wasn’t a stock pick, it was Chris Hohn admitting he’s “scared” about what he’s seeing in equities today. This is coming from a man who has delivered 7% net alpha over 22 years. In accordance with this, he has quietly reshaped his portfolio towards more defensive businesses featuring hard assets whilst leaning into weakness for some purported AI losers.

Two names did draw much of the room’s attention during the day:

1. Microsoft – which TCI has now fully exited

2. S&P Global – which they have added to

I have strong views on both businesses and if you haven’t read my write up on Microsoft, start here:

Where is Chris Hohn Hiding?

If you haven’t heard Chris Hohn present, his gentle monotone can even put an insomniac to sleep. However, his exceptional track record suggests it’s probably worth listening when he speaks.

Chris kicked off the event by providing an update on the portfolio and performance. Notable changes include a full exit from their long-standing position in Microsoft (read on to find out why) and an addition to the portfolio in Deutsche Borse.

Nearly half of his portfolio’s assets are concentrated in three aerospace names – Safran, GE Aerospace, and Airbus. This reflects his view that these businesses are highly protected from disruption (a common theme throughout the day) and have decades long backlogs which provides unusual earnings visibility.

He is also very underweight technology relative to the benchmark and has almost zero exposure to the semiconductor supply chain which has driven much of the recent market gains. This naturally means that the fund has underperformed by quite some margin year-to-date. However, his view is that under the paradigm of AI, previously established moats across software and professional service companies are being eroded and that innovation is now more important than ever which is making it harder to pick between winners and losers. Geopolitical uncertainty and rising competition are also keeping him largely on the sidelines with semis.

I largely echo his sentiment and there’s nothing wrong with putting things in the “too hard basket” especially when staying true to process. But I would argue there are several businesses within the software/semi complex which fit within his definition of durable compounders with pricing power that could warrant consideration – ASML being a good example.

Within the technology sphere, TCI maintains positions in Google and SAP. Whilst they didn’t really discuss SAP, they seemed quite bullish on Google and its prospects in AI. Notably, they pointed out that unlike its peers, Google has a strong presence across every layer of the AI stack and that AI was a self-reinforcing competitive advantage in its other businesses e.g., AI has made the search business even better for users and customers. TCI project Google Cloud to grow at a 45% p.a. rate to 2030 and see it as one of the big winners in AI.

The rest of the conference featured presentations from analysts discussing their thesis in businesses like Ferrovial, Vinci, Airbus as well as presentations from management at Visa, S&P Global and Aena.

There was quite a bit of discussion around businesses in what TCI labels financial markets infrastructure – Visa, S&P Global, Moody’s, Deutsche Borse. The common seam through these discussions was the standards/protocol nature of these businesses which makes them near impossible to disrupt. All the while, most of these companies seem to be trading at decades low valuations on the fear of AI disruption from the market. TCI’s conviction here was evidenced by the increased allocation of capital to these businesses which the firm hopes will generate a high-teens net IRR over the next five years.

Finally, the conference concluded with a panel of the entire investment team discussing AI and disruption. I’ll pull a few interesting snapshots from the discussion:

• Rapid technological change is making competitive advantages harder to predict. This is why TCI prefers asset-heavy industries where the risk of disruption is very low. The investment team mentioned that Aena represents a “perfect Chris Hohn stock” and its one they never lose sleep over.

• Previously, TCI described its investable universe comprising of 200 stocks. That may be shrinking. Chris noted there is a real scarcity value of great assets and they may choose to hold existing positions for longer. Their average holding period is 9 years.

• Chris said that most companies he speaks to state they are at “peak employment”, meaning they don’t need to add headcount to grow and if anything will rationalise headcount from here on.

• Higher index concentration, momentum driven investing and elevated equity valuations is creating a more cautious investing environment for TCI. Chris stated, “Investors are complacent about the risks in equities today – I am scared about what I am seeing in the market, it feels quite dangerous”.

Why TCI Sold Microsoft (And Why I Don’t Agree)

As I mentioned, TCI fully exited its position in Microsoft on three primary concerns:

1. Increasing competitiveness across Office and Azure franchises.

2. Net reduction in knowledge workforce impacting seat growth.

3. Lack of confidence in their ability to innovate and stay relevant.

On competitive pressure across Office and Azure, the investment team noted that AI is augmenting workflows for businesses and see the Office franchise in particular under threat from competition with OpenAI and Anthropic. As for Azure, they noted the business is becoming less special as it becomes more of a GPU rental business – I spoke to this at length in my write-up on Microsoft.

Whilst I would agree with their point on Azure, I challenge their assertion on Office being disrupted by competitive pressure.

First, I believe Office (like TCI’s other investments) represents more of a standard. It is the agreed upon global platform for knowledge work and despite competent free products from Google, Office seats and usage continues to grow. Let’s also not forget that many enterprise software businesses over the past two decades have been built around the premise of replacing “spreadsheets” or simplifying workflows within an organisation. Despite that, Office has continued to grow at healthy levels.

Second, OpenAI and Anthropic seek to capture value within the Office ecosystem not in spite of it. They have launched plug-ins which integrate with Excel, Word, and PowerPoint and I don’t necessarily see the value or incremental ROI in reinventing the wheel. The modus operandi for the LLM companies is to increase token usage. Doing that within Microsoft’s enormous distribution network seems to be far more advantageous than convincing millions of businesses to switch off their Office licenses for a potential product suite that is at most incrementally better than what is currently on offer.

Finally, I believe TCI are undervaluing the platform nature of Office. It comprises workflow tools, communications applications, and security all at an incredibly compelling price point. For decades, enterprises have built their businesses on the back of these applications and to unwind that today will take decades still.

Yes I agree that Office growth may be pressured if there is little growth in knowledge workers. I also agree that Microsoft has missed the mark when it comes to innovation, after all it should’ve never been Claude that lives within my Excel spreadsheets. But to sell the position on the premise that somehow OpenAI and Anthropic will replace or disrupt the tools which represent the gold standard for all knowledge work seems a bit shortsighted.

S&P Global: Standards Don’t Get Disrupted

S&P Global has been a victim of the AI loser trade with many believing that its products are prone to being disrupted. TCI does not believe this to be the case and Martina Cheung, CEO of S&P Global, was also there in person to drive this point.

Much of S&P’s business represents standards or benchmarks upon which global markets operate. Take the rating agency which is critical for global debt markets or the Platt’s business which is written into to almost every physical oil contract traded. AI can help improve but not supersede these offerings.

TCI made the rather salient point that throughout human history, rarely has a standard or benchmark been disrupted. Rather they tend to become less relevant measures and in turn lose importance e.g., the Dow Jones Industrial Average index is less meaningful today than it was 50 years ago but is still widely quoted. Market standards work in reverse to most moats which erode over time – trust and liquidity reinforce the standard, and time makes it harder to replace.

Moreover, over 80% of S&P’s profits are generated from the three franchises which TCI believe face no AI disruption risk – Ratings, Indices, Platt’s. The market is therefore concerned about the Market Intelligence division, which by revenues is the largest business. However, the mental shortcut made by many investors is that CapIQ (S&P’s Bloomberg equivalent product) represents the bulk of this business and it largely feeds publicly available data to its customers.

That is a false assertion, in fact Martina Cheung noted that in an absolute worst-case scenario, 5% of S&P’s revenue is at risk from AI and the proportion of profits is even lower. CapIQ therefore is a small part of an otherwise important Market Intelligence division. The majority of revenues within the division are tied to datasets which are either proprietary, contributory, or hard to source (physical records).

Thus far, S&P has only made a small portion of its data available to the LLM companies, most of this being public data. The strategy is to use the LLMs as an additional distribution arm for its data but there is no intention of letting them train on or acquire the data. Martina also noted that there hasn’t been a single customer who has found an AI alternative for their products.

What I found most interesting is just how compelling (and undemanding) the valuation of the business is right now. At 19x forward earnings, the business is trading at not only its decade low valuation but at the largest spread to its closest competitor Moody’s. In fact, if you take the current multiple for Moody’s and apply it to S&P’s ratings business and you take the current multiple of MSCI and apply it to S&P’s index business, then you effectively receive the rest of S&P for free.

For a business that has highly favourable risk characteristics, is growing at MSD-HSD, expanding its margins and delivering over 85% of FCF back to shareholders, I agree with TCI in believing this might be a unique buying opportunity.

Conclusion

If there was one thread running through the day, its that TCI is buying standards and selling stories. Airports, exchanges, rating agencies are businesses exempt from running in the AI race and TCI seem to prefer getting a full eight hours of sleep each night than fret over competitive displacement of their investments. I admire the conviction to stick to process even when performance is lacking. After all, it’s easy to compromise on your principles when profits are at stake. Whether they are right or wrong, is yet to be seen. But Hohn’s portfolio is a bet that in an age of increasing uncertainty the most valuable thing a business can be is irreplaceable.

Over the years, Microsoft has repeatedly deployed a successful playbook of building platforms to become more relevant to more people. This success has been rightly reflected in the company’s over $3tn market capitalisation today and spectacular long-term returns.

Today, the business trades at a 21.5x forward P/E multiple, only slightly above the S&P500’s multiple of 20.5x. Yet it is expected to grow earnings at a rate of 18% p.a. over the next two years and is far higher quality relative to the average S&P500 business. Seems like a no-brainer right?

My view on the long-term trajectory of Microsoft’s moat has actually soured over the past few months. It might seem like heresy to say this especially when prominent investors such as Bill Ackman have taken significant long positions in the business. But I think the calculus around risk/reward is changing at the margins such that I am less bullish on the business than I once was. Below I outline three reasons for why I believe this to be the case.

From capital-light to capital-hungry

The graph above shows Microsoft’s annual capex over the past five fiscal years, alongside the projected spending over the next three (note we are already three quarters in FY26). Analysts expect the company to spend close to $500bn between FY26 and FY28. In FY27 alone, Microsoft could spend roughly as much as it did across FY21 to FY25 combined. This represents a staggering pace and quantum of investments for what has historically been a relatively capital light business.

Of course, the spend itself is not the central issue. What matters is the return Microsoft can generate on that investment. Thus far, there’s little evidence to suggest this is a poor use of capital. Azure continues to grow at a 40% rate and likely earns very healthy operating margins (despite some gross margin pressure from the infrastructure build-out). The useful life of these assets may also exceed their six-year depreciating schedule, while some of the infrastructure should remain fungible across customers and workloads.

However, several questions arise as we look out further into the future:

1. Is Microsoft overbuilding capacity?

2. Are investments being made near the top of the cycle?

3. How quickly will capex continue to grow and what impact does this have on FCF?

4. What return can investments in these assets sustain over a decade?

5. Can Microsoft continue to largely self-fund its capex or will it become more reliant on capital markets?

These questions are all open to debate and much has already been written on them. The broader point here is that they were largely absent from my original thesis for Microsoft. Outside Xbox and Windows, which both retain some exposure to hardware cycles, the business was neither particularly capital intensive nor especially cyclical. That appears to be changing rapidly.

The cyclicality may not yet be visible, but it would be imprudent to assume that demand for tokens will follow a smooth exponential curve indefinitely. AI demand may ultimately justify the investment, but the path is unlikely to be linear. Periods of constrained supply could be followed by excess capacity, falling compute costs or slower workload growth. The risk is not necessarily that demand disappears, but that supply is built ahead of demand and returns temporarily compress.

This does not mean Microsoft is misallocating capital. The company may be rationally investing ahead of a large and enduring opportunity. However, the range of possible outcomes has widened. Investors must now assess not only the growth of demand, but also utilisation rates, asset obsolescence, customer concentration and the durability of returns on each new dollar invested.

That is a materially different exercise from valuing a capital-light software company with recurring revenue and limited reinvestment needs. Microsoft may still deserve a premium valuation, but the composition of its earnings is changing. As more value is generated through infrastructure, the appropriate multiple should increasingly reflect the risks that come with owning infrastructure.

Renting GPUs vs. running the enterprise

This brings me to my second point: what is the right multiple to pay for a business that is more capital intensive and potentially more cyclical?

My mental shortcut for Microsoft has always been to equate the company with software. Perhaps that thinking is a bit antiquated, considering the decade-plus investment the business has made in scaling its Azure. Broadly, however, we can think about Microsoft providing enterprises with large, deeply embedded IT platforms. Windows is a platform for personal and enterprise computing, Office is a platform for productivity and collaboration, Azure is a platform for computing and data storage.

The important point is that these platforms are not interchangeable pieces of infrastructure. They sit inside the daily operating fabric of a business. Enterprises build workflows around them, train employees on them, integrate third-party tools into them and embed security, compliance and identity management around them. That makes Microsoft’s revenue base unusually durable because the customer is not merely buying compute or storage. It is buying continuity, familiarity, interoperability and reduced operational risk.

This is what has historically made Microsoft such a valuable business. A platform that becomes the system of work, benefits from high switching costs and deep customer dependency. The economic value is not just in the product itself, but in the ecosystem that forms around it. Windows, Office and Azure have each benefited from this dynamic in different ways. They are not commodities because customers are not making purchasing decisions purely on price.

The question is whether AI infrastructure revenue carries the same characteristics. With growing share of Azure’s revenue coming from LLM companies consuming enormous amounts of GPU capacity, the nature of the business may be subtly different. In that case, Microsoft is not necessarily providing a deeply embedded enterprise platform. It is providing scarce infrastructure to customers whose primary attachment is to compute availability, price, performance and scale. Those are valuable attributes, but they are less clearly unique.

This distinction matters because customers such as OpenAI, Anthropic or other AI labs are not necessarily locked into Azure in the same way a large enterprise is locked into Microsoft’s productivity, identity, security and cloud estate. In theory, model companies can move workloads across hyperscalers if the economics, capacity or technical performance are better elsewhere. That does not make the revenue unattractive, but it may make it less platform-like. It is closer to being a provider of critical infrastructure than a provider of an irreplaceable application layer.

That creates a tension in how the market should value Microsoft. The company may still be one of the biggest beneficiaries of AI, but the quality of that incremental revenue is arguably worse. Revenue generated from scarce infrastructure can be high growth, but it is also more capital intensive, potentially more competitive and more exposed to supply and demand cycles. Ultimately, this will be reflected in the multiple investors are willing to pay.

Betting the thesis on two labs

A key pillar of my original investment thesis was the breadth of earnings streams Microsoft enjoyed. That seems to be unravelling especially as a greater proportion of earnings are derived from the Azure business.

Now this isn’t necessarily a surprise given Azure has grown at an average rate of 40% p.a. over the past five years whilst the rest of the business has grown slower. Mathematically, it must constitute a larger proportion of the base over time. However, as I mention above, the quality of those revenues is quickly augmenting in a way that I would be more hesitant to apply the same multiple in prior years.

As of the end of FY25, revenues across Microsoft’s three reportable segments are fairly balanced with the majority of earnings stemming from core platforms such as Office, Dynamics and LinkedIn. However, if we project out to 2028 (I am using consensus estimates here), we can see that over 50% of revenues will be generated by the cloud segment alone.

Double clicking into the composition of growth within the cloud segment, we observe that a significant amount is being generated from AI workloads. In its most recent Q3 call, management called out that AI workflows have surpassed a $37bn ARR. Most sell-side analyst estimates peg that number to be in the mid-to-high $20bn range for FY26 and growing close to 100%. At this pace of growth, in FY28 that will be over $80bn or between 15-20% of total group revenues.

Again, this isn’t necessarily bad, but let’s consider the prior makeup of cloud revenues to the incremental dollar that is flowing through. Previously, Azure growth was driven by thousands of corporations shifting on-premise compute to Microsoft’s public cloud. They architected their business around Microsoft’s offerings and no single enterprise represented a meaningful share of those revenues. Compare that to the AI workloads which are primarily driven by the inference and training requirements of two companies – OpenAI and Anthropic.

Now even in FY28, it would be a far cry to suggest Microsoft is suffering from massive customer concentration. OpenAI and Anthropic would at most make up 5-10% of total group revenues each and these are high-margin, high growth streams of revenue. But let’s assume AI continues to follow its exponential curve and take a minute to do the mental exercise of considering Microsoft’s revenue mix in 2035 and beyond. That picture looks very different to what we see today.

Crucially, the relationship between frontier model developers and hyperscalers differs from Microsoft’s relationship with its core enterprise customers. These companies do not rely on Azure as their system of work, nor are they necessarily wedded to its broader software architecture. What they need above all is access to enormous amounts of compute, which in turn requires capital, power, chips, networking capacity and considerable expertise in building and operating large-scale data centres.

In this sense, Microsoft’s role is closer to that of an infrastructure sponsor than a traditional technology platform. It absorbs much of the upfront capital burden and provides scarce compute capacity at a scale few model developers could initially finance or operate themselves. That relationship is still valuable, but the source of dependency is different. The model developer relies on Microsoft because replicating the infrastructure is difficult and expensive, not because its organisation has been built around Azure.

Over time, however, that calculation may change. As model developers become larger, better funded and more confident in their future utilisation, the economics of owning or directly controlling infrastructure become more attractive. At sufficient scale, the margin paid to a hyperscaler may exceed the benefits of outsourcing. xAI’s development of Colossus illustrates the direction of travel.

This does not mean that model developers will become entirely independent. They will still depend on chipmakers, utilities, data-centre operators and external capital, while hyperscalers retain meaningful advantages in procurement and execution. The more credible risk is that the largest customers gradually diversify providers, build dedicated capacity and negotiate away some of the hyperscaler’s economics. Microsoft may therefore retain the workload without retaining the same degree of customer dependency or pricing power that characterises its traditional enterprise platforms.

As the Microsoft investment thesis becomes increasingly dependent on the returns generated from AI workloads, a new concentration risk is emerging. That greater uncertainty should command a higher risk premium, and all else equal, a lower multiple on future earnings.

Conclusion

I don’t mean to sound alarmist or suggest liquidating your shares in Microsoft (indeed I still have a position myself). I also certainly don’t mean to suggest Microsoft is a poor business or that its capital is being misallocated. It remains one of the highest-quality franchises in the world and a genuine beneficiary of the shift to AI.

The nuance here is that the character of the business is changing. A company once defined by capital-light, recurring software revenue and deeply embedded platforms is becoming more capital intensive, more cyclical and more reliant on a narrower set of infrastructure-led earnings whose customers are not necessarily locked in the way enterprises are.

Each of these shifts pulls in the same direction. Widening the range of outcomes and weakening the durability that justified Microsoft’s premium. The investment may still work, but it is no longer the same investment. As I see it, the windows are not closing, but they are narrowing.