For professional investors only - not for retail distribution

This note sets out how we are thinking about investing into the data centre build out within the Goodhart Global Future Leaders Fund (GFL). GFL holds 36 positions and looks very different from any  index, because it is built to own the companies we think will lead the next cycle, not the ones that led the last one. I have spent almost 30 years on both sides of this industry, first building data centres and then investing in them, so a fair amount of this note is history. The rest is the best guide I have to what may happen next.

1. My History in the Data Centre Industry

It started with a Phillips screwdriver.

The business began the way a lot of small technology businesses began in the late 1990s, by solving the problem immediately in front of us. We built PCs for friends and family. That turned into designing websites for local businesses. Once we had built a handful of websites, we quickly noticed the valuable recurring revenue for hosting was going to somebody else. So we insourced it. That decision was the whole of DediPower's founding strategy, and it was made by three people, Craig Martin, Spencer Tarring and myself, with no industry experience and roughly £15k of startup capital. We were 18 and ready to become dot com millionaires. But as with all good stories there were many twists and turns to come.

The first real data centre was built in 2000, in an office we had taken in a Grade II listed Tudor house in Wokingham near Reading. The kit went under the stairs. Connectivity came from a 2Mb leased line costing, from memory, around £15k pa, which makes me laugh now at how far connectivity prices have fallen. The servers were built by hand, incorporating AMD's original Athlon, the K7, which was the chip that had just made the company competitive with Intel on performance rather than price, and which therefore found its way into a lot of businesses that cared more about output per $ than brand.

We were building into the back half of the dot com boom, and then straight into the bust. The lesson has stayed with me, and it was not the one most people took at the time. The applications were not fake. Online retail, banking and media all happened, at a scale beyond what was promised in 1999. What was wrong was the timing. Consumers adopted far more slowly than the capital assumed, and a great deal of infrastructure was built for demand that arrived five or ten years later. Capacity built ahead of adoption isn't a long-term problem, but it is a very real short-term one for whoever is funding the gap.

From there we grew, and the easiest way to describe it is through the buildings. Out of the room under the stairs into space above a tyre shop, from there into converted warehouses, then purpose-built facilities and eventually multiple sites. Every move was forced by the same three constraints, and they still define this industry a quarter of a century later.

The first is power: how much you can get, how reliably, at what price, and whether the substation serving your building has any headroom left. The second is connectivity, because two buildings of identical specification in different locations are not the same asset. The third is redundancy: paying for backup power and cooling you hope never to use. It's expensive, and it's usually where operators who fail find out they skimped.

We sold DediPower in 2011, eleven years after building our first data centre. AWS was starting to bring down the cost of hosting online applications, and we needed to consolidate to keep competitive. I set this out not for nostalgia but because it is the basis on which I would ask you to weigh what follows. When the debate turns to gigawatts added, power availability, cooling architecture and the difference between a contracted lease and a productive asset, these are problems I have had to solve. That does not make my conclusions right. It does mean that when the numbers in a hyperscaler presentation do not match the physical reality of building the thing, I take note.

2. Data centres in the Public Markets

I joined Fidelity in 2005, and the operating experience translated into the investment seat faster than I expected, because the supply and demand picture in European colocation was visibly improving and almost nobody was looking at it. From a peak in Q3 2002, available data centre space in Europe had fallen steadily. Yes, supply can fall in data centres.

The first name I recommended at Fidelity was IXEurope. It was an AIM listed operator with multiple data centres across Europe. The thesis was uncomplicated. Demand for quality colocation in the major European metros was growing faster than anyone was adding supply, the assets were difficult to replicate, and the market was valuing the business as though it were a commodity hosting company. The stock performed well and was then taken out by Equinix. More usefully for me, it established some credibility internally at a point when I badly needed it early in my career.

Citing research from the time, Equinix noted that global demand had outpaced supply by nearly three to one over the preceding twelve months, and that the European imbalance was more than double that. Shortage has been the norm in this industry for as long as I have worked in it, and it is ingrained in how people think about it. The one period of oversupply is a distant memory.

The listed universe broadened from there with Fidelity backing the Telecity IPO.  By 2013 the sector had a proper cohort of listed names, Equinix, Digital Realty, Interxion and Telecity all come to mind. Many of those names have since disappeared from public markets due to consolidation.

I draw two conclusions from this. The first is that scarcity of quality assets here is real and durable, and private capital has repeatedly paid well above public market valuations to own it. The second is more relevant to today. Most of that consolidation was large operators buying well connected, high quality assets they could not easily replicate, then running them with more scale and a lower cost of capital than the sellers could manage. Over time, returns came to depend as much on scale as on service. Smaller facilities outside the major metros found that harder to compete with, until AI training created demand for capacity that does not need to sit close to the user. That is part of why CoreWeave and the other neoclouds are seeing strong demand. How long it lasts depends on the handful of companies funding the next generation of models.

That is the lesson I carry into the current cycle. I am less interested in whether AI demand is real than in which part of the supply chain keeps the economics once capacity catches up.

3. The Drive for Efficiency: Virtualisation, Cloud and Kubernetes

Three waves of efficiency have passed through this industry since we built our first data centre, and each one arrived with a credible argument that it would reduce the amount of physical infrastructure the world needed. Each one was wrong in the same direction. I don’t want to make the same mistake, and I think many bearish commentators are making it now.

The first wave was virtualisation. Before VMware's technology became standard practice, the operating assumption was one application to one physical server. That model was enormously wasteful. Typical utilisation on a dedicated box sat in the low single to double digits, which meant customers were buying, powering and cooling machines that spent the overwhelming majority of their lives idle. Virtualisation broke that link by letting many workloads share one physical machine safely. Consolidation ratios of ten to one, and later considerably more, became routine. In pure arithmetic, that should have collapsed demand for servers, and with it demand for the space and power to house them.

The second wave was public cloud. AWS made compute and storage available on demand, and Azure and Google Cloud followed, converting a capital decision with a three-to-five-year commitment into an operating line that flexed with usage.

The third was containers and orchestration, Docker and then Kubernetes, which took the granularity of deployment down another level and allowed far denser packing. Both were read at the time as deflationary for the industry. If everybody shares one enormous pool of capacity, run at higher utilisation by better operators, the world plainly needs fewer machines.

What actually happened, in every case, was that total consumption of compute rose. The mechanism is the one William Stanley Jevons described in 1865 when he observed that improvements in the efficiency of coal use pushed total coal consumption up, not down. Making a resource cheaper to use expands the set of things worth doing with it, and if that set expands faster than the efficiency gain, total consumption rises.

Virtualisation made it economic to run applications that could never have justified a dedicated server. Cloud made it possible to start a business without buying any infrastructure, producing a generation of software companies that would not otherwise have existed, each of them a customer for compute. Containers made it economic to decompose applications into hundreds of small services, which in aggregate consumes more infrastructure, not less.

The pattern is not a guarantee. Jevons needs elastic demand and a large reservoir of latent uses waiting to become economic. Where those conditions fail, efficiency gains simply reduce consumption. The chart below points to a lot of latent demand in AI. The question is how quickly it is adopted.

This is where the bull and bear cases diverge. Inference costs per token have been falling very fast. The bull case holds that cheaper tokens will unlock more demand than the price decline destroys. The bear case holds that adoption follows a classic S curve, and that the risk sits at the start of it. At low penetration, below around 5%, growth can moderate before the steep part of the curve arrives at 20% and above. Capacity built for the steep part then must wait through the flat part, which is what happened after 1999. On the chart above, paying AI users are around 1% of the world’s population. We will return to this in section seven.

4. COVID and the End of the Digitalisation Cycle

For the two decades before 2020, the demand driver for data centres was fundamentally human. People used cloud services, made video calls, streamed media, shopped online and stored photographs. Every incremental unit of capacity largely traced back to a person doing something on a screen. That defined the ceiling of that cycle, and explains why it ended.

COVID compressed roughly a decade of that adoption into about eighteen months. Offices emptied and video conferencing went from a tool used by distributed teams to the default mechanism of professional life. Retail moved online at a pace that would otherwise have taken years. Streaming consumption rose sharply because there was very little else to do. Collaboration software, remote access infrastructure and cloud storage all saw step changes in usage instead of the smooth growth curves the industry had been planning around.

Operators and their customers made sensible decisions on the information they had. Many of those decisions turned out to be wrong. Capacity was added aggressively. Headcount was added aggressively. Growth rates observed during a period of enforced confinement were extrapolated as if lockdown behaviour would last forever, and valuations were set accordingly. 

Then everybody went outside again, and the hangover through 2022 and into 2023 was severe and broad. Companies that had hired for a demand curve that did not materialise conducted the largest wave of technology redundancies in twenty years. Enterprises that had migrated in a hurry began optimising spend, which showed up directly in decelerating hyperscaler growth. Capacity contracted on the assumption of continued acceleration took longer to fill than underwritten.

Because the AI cycle arrived so quickly afterwards, the two periods have blurred together in most people's memory. But they were distinct cycles, and the first one is over.

The human-driven demand for data centre capacity is now a mature, GDP-like growth business. More people will stream more video and make more video calls, and that will grow, but it will grow at the rate at which populations and incomes and screen time grow, which is to say modestly. The step change in per person consumption of digital services has substantially happened. Whatever growth premium exists in this industry today does not come from that source.  New AI agents/applications may even cannibalise some of the demand.

It matters for how you underwrite the build out in two ways. First, there is no longer a second engine. Through the 2010s an investor in data centre capacity was underwriting a diversified set of end demands, enterprise migration, consumer internet, streaming, gaming and e-commerce, and if one disappointed the others carried the load. Today the marginal gigawatt is being planned for AI and that diversification is largely gone. Although admittedly AI will span so many new services and functions that calling it AI may end up looking like a misnomer.

Second, 2021 to 2022 is the most recent and most relevant evidence we have about how this industry behaves when demand assumptions prove optimistic. It was not a catastrophe. It was a painful but survivable adjustment, in which most assets retained value and eventually filled, though some were mothballed, but capital deployed at the peak earned poor returns for several years. That is a possible template for a disappointing AI outcome, and rather more useful than either the utopian case or the dot com analogy reached for too readily.

5. The Rise of AI

This cycle differs from every one before it in one important respect. Every previous wave of data centre demand was about moving existing human activity onto shared infrastructure. This one is about producing cognitive work that did not previously exist, at a marginal cost that is falling.

The physical consequences arrived faster than the industry could accommodate. For twenty years facilities were designed around racks drawing five to ten kilowatts, air cooled. AI racks now draw multiples of that, into the tens and in the newest configurations beyond a hundred kilowatts, and air cooling does not remove that much heat. Liquid cooling has moved from a high-end option to a design requirement, which means a large share of the world's existing estate cannot host the workloads driving demand growth without substantial reconstruction. Ordinarily new demand fills existing capacity. Here, a great deal of it cannot.

The capital response has been extraordinary. Combined hyperscaler capital expenditure commitments for 2026 sit in the region of $700-750bn*. To put that in context, it exceeds global upstream oil and gas investment, which the IEA put at roughly $570bn in 2025, and it is being committed by a handful of companies against a demand curve that is roughly three years old.

The industry plans in gigawatts, not dollars. JLL has global data centre capacity roughly doubling from 103GW in 2025 to 200GW by 2030. Gartner works from power demand and gets to around 290GW on the same date, having put 2026 at 132GW. That is a spread of some 90GW between two research houses on a five-year view, which is roughly the entire installed base of 2019. Nobody knows how much capacity this ends up needing, including us.

The contracting behaviour underneath those forecasts is running hotter still. US technology companies committed over $850bn of data centre leases in the first quarter of 2026, a massive increase on the year before, led by Meta, Microsoft and Oracle. Those are multi-year contracted commitments rather than annual spend, so they are not directly comparable with the capex figure above, but the direction is unambiguous. Looking further out, Goldman Sachs projects $7.6tn of AI infrastructure capex between 2026 and 2031, and broader estimates put total data centre and related investment at $10.3tn from 2025 to 2032, which would make this the largest economic bet in US history, ahead of the railways and the interstate highway system.

Supply is constrained by more than money. TSMC warned in June that global chip supply will fall short of AI demand for years, with capacity a key bottleneck even as new US fabs come online. SK Hynix is doubling memory capacity over five years against a shortage it expects to persist until 2030, and Wiwynn has flagged shortages developing in data centre components beyond memory. This is why the generational tightness described below hasn't been arbitraged away within a quarter or two.

My co PM Andrew Heap found a good indicator of real time demand: GPU rental pricing. Everybody quotes the capex number, but few look carefully at what it costs to rent a chip. The rental market is the closest thing this industry has to a spot price for compute, and it gives you an indication of where scarcity actually sits.

The shortage rolls from one chip generation to the next. A100 rental rates have been grinding steadily lower for two years, which is what you would expect of an ageing chip that nobody is queuing for. H100 rates held broadly flat through most of 2025 and have firmed through 2026 without returning to their old highs. The newest Blackwell generation is where the scarcity has concentrated, and it has not eased. B200 rental rates have just reached an all-time high of $7.88 per GPU hour, up 68% over the year, which is a remarkable thing for a chip to do a year into its life.

That pattern is an important signal in this cycle. Compute is not short in aggregate, but the current generation is. Each generation passes through its own tightness cycle as the installed base catches up with demand, while the previous generation softens underneath it. So the signal to watch is whether each new chip generation continues to command a premium, more than total gigawatts added or capex committed. The first generation that fails to command one will tell you the build out has caught up with monetisable demand, and it should tell you before any capex announcement does. The current data does not suggest we are there yet, although this is only one of several indicators we monitor, and for now somebody is making very good money renting out Blackwell chips.

There is a second, subtler signal in the same data. USD measures of token expenditure have been falling through 2026 even as rental rates for frontier chips have held firm. It looks contradictory, but it isn't. Volumes have risen very fast while price per token has fallen faster, largely because open weight models, many of them Chinese, have taken a substantial share of inference. Volume up, unit price down, dollars flat or lower, and physical demand for the newest silicon still rising. Any analysis using dollar spend on tokens as a proxy for compute demand will get this cycle wrong.

Andrew and I read this differently. I have the memories from previous cycles; he is looking at it with a fresh set of eyes.

6. Power, the Binding Constraint

Of the three constraints I described in section one, power has gone from an engineering problem to the thing that decides how fast the whole industry can grow. Capital is abundant and chips are being made. Electricity, and more precisely the right to connect to it, is what now determines whether announced capacity becomes real capacity.

There is already plenty of evidence that this is biting. Kimmeridge Energy Management suggests up to half of planned US data centres may face delay or cancellation on a combination of physical infrastructure constraints and political resistance. In Europe the gap between announced and deliverable is starker still. Spain and Portugal have announced 10.5GW of capacity, of which only around 4.5GW is considered active, with grid capacity, network access and permitting the binding constraints. The UK has just provided a live example. Microsoft has pulled out of anchoring Nscale’s 50MW site at Loughton in Essex, after grid officials said power would not arrive until the 2030s. The site was due to open early next year and was one of the deals announced during President Trump’s state visit in September 2025. Nscale says it is still committed to the site and is looking at alternative power, but grid connection queues in Britain are now delaying some projects by as long as 13 years.

The same constraint is showing up at the largest buyers. Business Insider reports that AWS has started taking ownership of Nvidia GB300 racks before its data centres are ready for them, with internal emails projecting around $5bn of deliveries in August and roughly $3bn a month after that. Some sites had the space but not the power to switch the racks on, so finished equipment has been staged in warehouses. Amazon says this is standard practice.

Regulators are starting to push back. Denmark has published an emergency grid law placing most data centres last in the queue for connection, which is a template other countries will study. Ofgem is proposing upfront fees and tougher project milestones to clear speculative projects out of the UK connection queue. In the US, PJM has accelerated plans to pair data centres directly with generation, which moved power stocks when it was announced. The industry’s response has been to bring its own electricity. Microsoft has signed a 20 year agreement to take the output of the restarted Three Mile Island Unit 1, now the Crane Clean Energy Center, and Meta has signed multi gigawatt nuclear agreements. Bring your own energy has shifted from a differentiator to a requirement.

This cuts both ways. A constrained grid supports pricing for whoever already holds capacity, which is bullish for incumbents and for anyone with energised land. It is also the mechanism by which a demand disappointment could be disguised for years. If half of announced capacity is never built, rental rates can stay firm on restricted supply even as underlying demand softens, and the signal we set most store by would keep reading bullish while the thing it is supposed to measure deteriorates.

The way to separate the two is to watch utilisation alongside pricing. Tight supply with high utilisation is real demand. Tight supply with falling utilisation, or with lease terms quietly softening on renewals, is a shortage of electricity masking a shortage of customers. I would also watch the politics. Voter anger over utility bills is already testing incumbents in Ohio, and House Democrats have pressed the Interior Department for detail on data centre projects proposed for federal land. Social acceptance is not usually in a supply and demand model. Let us hope we do not end up back where the Swing rioters of 1830 did, smashing the threshing machines that had taken their work.

7. The Outlook for AI

Bear (ish) Case – Written by Russell Champion

My concern is not that AI is unreal or that demand will disappear. It is that the capital being committed is underwritten against an assumption of continuous compounding demand growth, and that the structure of who is committing it has become more fragile than it looks. A few things concern me.

The first is where the risk now sits. The large platform companies are funding this out of operating cash flow, from businesses that were profitable before AI existed and would remain so if it disappointed. They can absorb a poor outcome. The risk is still there, but it has concentrated onto a much smaller number of entities, most obviously OpenAI, which has committed to infrastructure spending on a scale that requires revenue growth it has not yet demonstrated and funding it has not yet secured. A very large portion of the forward demand underpinning this build out traces back through commitments of that kind. That is a materially different risk profile from a diversified group of cash generative buyers, and it is not being priced as such.

The second is the monetisation gap, and the circularity now bridging it. Capacity is contracted years ahead of the revenue expected to fill it, which is normal in infrastructure. What is unusual is the ratio, and what is filling it. Commitments have grown considerably faster than demonstrated end customer revenue, and a pattern has developed in which the suppliers of compute hardware invest in, lend to, or take commercial positions in the customers buying that hardware, who in turn commit to purchases that support the supplier's forward revenue. Each arrangement can be defended on its merits. In aggregate the effect is that demand appears more durable and more independent than it is, because a good part of it is being funded by the party recognising the revenue. I saw an attenuated version of this in telecommunications equipment vendor financing in 1999 and 2000. It flattered growth on the way up and accelerated the unwind on the way down. When gaps like this close, they close quickly, usually because someone stops funding the next tranche.

The third is depreciation, and it is the one I think is most underappreciated. The economics of this build-out depend heavily on how long a GPU stays productive. Useful lives have generally been extended in recent years, which reduces the annual charge and flatters reported earnings, and that is defensible if hardware remains productive for five or six years. The rental data in section five suggests something more nuanced. Each generation retains value, but at a materially lower rate as the next arrives. If the newest generation consistently commands double the rate of its predecessor, the economic life of a GPU is considerably shorter than its physical life, and a good deal of reported profitability across this ecosystem comes from an accounting assumption, not cash economics.

Oracle's Project Jupiter is the clearest live illustration of all three risks converging. The 1,400 acre site in Dona Ana County, New Mexico underpins Oracle's $300bn compute contract with OpenAI. According to FT reporting, Oracle has issued a force majeure notice over the power supply after permitting delays and local opposition, yet remains liable for a carry cost covering debtholder interest in full plus an equity return to Blue Owl for up to three years if the site is not ready by the first half of 2027. More than twenty banks led by BNP Paribas, Goldman Sachs, MUFG and SMBC provided $18bn of construction debt, efforts to syndicate it stalled on concerns about Oracle's creditworthiness, and the loans have been quoted below 90 cents on the dollar. Oracle says the project remains on schedule. Whatever the outcome, the structure tells you where the risk sits, which is on the party that guaranteed the payments rather than on the investors who funded them.

None of this makes me an all out seller of the theme. It makes me insistent that the price paid for exposure should allow for a two-to-three-year digestion period, not just an uninterrupted ascent.

I'm actually a big user of AI, I would never have been able to write this note as quickly otherwise, and I get that same exciting feeling I did in 1999 when I knew the internet would have applications beyond what we could see at the time. I am a long-term bull trapped inside 20 years of experience investing into technology and data centres.

Bull (ish) Case – Written by Andrew Heap

My case begins from a different starting point, which is not whether the capital is well underwritten but whether the output being produced is valuable enough that the demand question resolves itself.

The core of it is a comparison. A software engineer on a six-figure salary can do a certain finite quantity of work. Hiring a second engineer can double output, or spending the same amount on tokens can increase output drastically more. If that is true even approximately, the value being created per dollar of compute is so far above the cost of the compute that the question is not whether demand exists but how quickly organisations can restructure themselves to capture it.

So long as the entire value chain can generate the required rates of return on their investment without subsidising adoption, and the value exceeds the cost, more tokens will be consumed, and more data centres will be needed. With improvements in models combined with data centre performance, this will only compound as intelligence per dollar will continue to outgrow human productivity. Key will be how fast industries can adopt and if output (GDP growth) can accelerate with new tasks solved or whether it will be mainly a labour substitution question, and ultimately the question becomes how many tokens will be needed in the future.

The demand trends support this. Token volume on OpenRouter has risen from around 7.5 trillion tokens per week in January 2026 to roughly 130 trillion per week in September 2026. Average revenue per token has fallen sharply over the same period, substantially because open weight models, many from Chinese labs, took more than half of routed volume having been negligible eighteen months earlier. The bearish reading is that the product is commoditising. The bullish reading, and the one the physical data supports, is that this is the Jevons mechanism described earlier, with cheaper tokens expanding the range of tasks worth doing faster than the price decline reduces the value of each one. Ultimately the concentration of economic profits is being dispersed amongst consumers rather than being concentrated at Anthropic and OpenAI.

The margin structure matters enormously here and is frequently overlooked whilst the labs remain capital hungry and in growth mode. Labs are currently compute-constrained but the marginal compute coming online is increasingly allocated to inference rather than training, and inference economics are fundamentally different. Depending on the model and the utilisation achieved, incremental margins on inference workloads can run north of 80%.  Labs can still pay a premium for compute (e.g. the recent Anthropic deal with SpaceX) and retain attractive margins, which allows data-centre owners to generate returns above their cost of capital, given so much value is generate and can be shared across the value chain. Questioning lab economics whilst framing against historic numbers given such rapid growth, at a time when inference allocation was lower remains dubious.

On adoption, the striking observation though is how early this remains, despite combined ARR of OpenAI and Anthropic topping $100bn. Corporate spending data from Ramp shows the median business is spending $15.70 per employee per month on AI tools as of September 2026, a rounding error against salaries. Salesforce quotes a $300m AI budget in 2026, equating to roughly 3.8% of developer salaries, again a small number. Compare this to the most advanced firms, who are spending just shy of $10,000 per employee per month according to Ramp, showing where the direction of travel is going. If the median moves even part of the way towards where leading firms are today or even a year ago, implied demand is orders of magnitude above current capacity.

Which brings the argument back to physical supply. Capacity is short today, and the B200 chart above is what that looks like. The bear risk requires a scenario in which gigawatt additions keep rising, then plateau, and even fall, as the industry discovers post 2030 that it has built enough capacity to satisfy demand. That is possible. It also requires believing that the enormous quantity of work currently done by people, which AI can already do at a fraction of the cost, will simply not be handed over or new workloads aren’t created. Given how little of that potential has been monetised so far, worrying that we are building too much looks like the wrong risk to spend time on. The more likely error, on this view, is building too slowly and ceding the returns to whoever builds faster. Ultimately, how the value created is shared amongst the value chain is the most important question to answer.

8. How GFL Is Investing In This Cycle

Portfolio holdings and weights are as at 30 September 2026. Holdings and allocations are subject to change and should not be considered a recommendation to buy or sell any security.

Having set out both cases, here is how the portfolio is positioned: to participate without underwriting either extreme. We do not run this fund against the benchmark. Position sizes reflect conviction in individual businesses rather than a view on how far to deviate from an index weight, and the data centre theme is expressed through roughly half the book, split between the companies supplying the build out and the software businesses living with its consequences.

Around 22% of the fund sits in the semiconductor and semiconductor equipment names that supply this build out. Soitec is the largest at 6.2%, held on silicon photonics and its exposure to AI spend rather than as a capital equipment proxy. The two behave very differently through a cycle. ASML is 4.3% and ams-OSRAM 5.0%. All have growing exposure to data centres and market leading positions in particular technology niches. PDF Solutions at 3.5% and ACM Research at 2.7% are smaller positions on the same cycle, beneficiaries of rising complexity or geographical hot spots in semiconductor manufacturing, but they aren't central to the case.

Over 30% of the fund sits in software, in names including GitLab, Zeta Global, nCino, Kinaxis, Dynatrace, NICE, Money Forward and Twilio. The fund owns the suppliers of the build out and it owns a larger position in software businesses whose economics are, on some views, threatened by the same technology. We have deliberately spread exposure across the sector, because we do not claim to know how this debate will resolve.

Equally important is what the fund does not own. There is no position in Nvidia, Micron, AMD, Intel, KLA, Applied Materials or Lam Research. Nvidia alone is 5.9% of the Bloomberg Developed Markets Large & Mid Cap index (30^th September 2026), so one of the largest decisions in the strategy is arguably the decision not to own the most obvious beneficiaries of the theme. That reflects a focus on investing through the lifecycle, and a view that the risk reward in the most heavily owned names does not compensate for the concentration risk and potential demand air pockets described in the bear case. Memory, semiconductor capital equipment beyond ASML, probe card manufacturers and the merchant logic names remain under active research, but no positions have been taken yet. The logic names, Nvidia, AMD and Broadcom, are a particular focus. The difficulty there is not whether the end market grows, it is that owning them requires a view on market share between a small number of competitors and their own customers' internal silicon programmes, which is a considerably harder call than the one we are making in the supply chain.  ASML has 100% share in advanced Lithography.  Soitec apparently has 90% share in supplying wafers into Silicon Photonics.

There is also no exposure to Utilities, Energy, Real Estate or Materials, which may look odd for a fund with this much exposure to the theme. The power and property layers are where much of the physical bottleneck sits, and they have attracted substantial capital for that reason. The history in section two is why we have not gone there. Assets at the infrastructure layer have repeatedly been competed down to commodity returns unless they sit on scarce connectivity or power, and identifying which ones do from the outside is harder than it looks.

We won't say what we would buy or sell, but we can say what we are watching. An important signal is whether each new chip generation continues to command a rental premium over the one before it, read alongside utilisation for the reason given in section six. The second is funding behaviour at the concentrated buyers, because that is how the monetisation gap closes.

On the software book, we have set out our thinking separately in “AI, SaaS and the mispricing of durability”, which addresses the durability question in more depth than this note can.

I have come a long way in data centres, from two megabits under the stairs to an industry planning in hundreds of gigawatts, and I am looking forward to finding the opportunities in the years to come.

* Consensus estimates as at 18 September 2026 for Amazon ($221.5bn), Alphabet ($201.8bn), Microsoft ($158.4bn) and Meta ($139.4bn), combining to $721.0bn. Microsoft's figure covers the fiscal year ending June 2026, the others are calendar year. These are total company capital expenditure figures rather than data centre capital expenditure alone, so they overstate compute spending at Amazon in particular. They also exclude Oracle, the neoclouds and the Chinese platforms, so the aggregate understates the total build out.

Sources:

Section 1

AMD Athlon (K7) launched August 1999, first AMD chip to outperform Intel's Pentium III. AMD product history.

2Mb leased line at around £15k a year. Author's recollection.

Section 2

European available data centre space falling from a Q3 2002 peak. CBRE, 2007 (chart in note).

Equinix acquisition of IXEurope, announced June 2007, completed September 2007. Equinix press release, 2007.

Global demand outpacing supply by nearly three to one, Europe more than double. Equinix press release, June 2007, citing Tier 1 Research.

Sector cohort and consolidation. Jefferies, Data Centers initiation, July 2013. Subsequent deals: Equinix/Telecity (2015), Digital Realty/DuPont Fabros (2017), Digital Realty/Interxion (2020), Blackstone/QTS, American Tower/CoreSite and KKR with GIP/CyrusOne (all 2021).

Section 3

AWS on demand compute and storage from 2006. Amazon (S3 March 2006, EC2 August 2006).

Jevons paradox. W.S. Jevons, The Coal Question, 1865.

Server consolidation chart. VMware, 2016 (chart in note).

Section 5

2026 hyperscaler capex of $721.0bn. Bloomberg consensus estimates, 18 September 2026 (footnote in note).

Global upstream oil and gas investment around $570bn in 2025. IEA, World Energy Investment 2025.

Global data centre capacity 103GW (2025) to 200GW (2030). JLL, 2026 Global Data Center Outlook.

Capacity of 104GW (2025), 132GW (2026) and around 290GW (2030). Gartner, 2026.

Over $850bn of US data centre leases in Q1 2026, up 204%. Bloomberg News, "Data Center Leases Top $850 Billion as Meta, Microsoft, Oracle Lead Record AI Infrastructure Spending", 5 July 2026.

$10.3tn of data centre and related investment 2025 to 2032, larger than railways and highways. Bloomberg News, "U.S.'s AI Build-Out Is on Track As Biggest-Ever Economic Bet", 24 September 2026.

Goldman Sachs projection of $7.6tn AI infrastructure capex 2026 to 2031. Cited in Bloomberg Law, "Data Center Deals on Tribal Lands Write Playbook", 15 September 2026.

TSMC warning on chip supply. Bloomberg News, "TSMC Warns Chip Supply Won't Meet AI-Fueled Demand for Years", 4 June 2026.

SK Hynix doubling capacity, shortage to 2030. Bloomberg News, "SK Hynix Plans to Double Capacity to Ease Memory Chip Crunch", 2 June 2026.

Wiwynn component shortages. Bloomberg News, "Nvidia Server Maker Wiwynn Sees AI Bottlenecks Beyond Memory", 28 May 2026.

A100 and H100 rental rates, token expenditure index. Silicon Data rental indices via Bloomberg, to 17 September 2026.

B200 at $7.88 per GPU hour, up 67.7% over the year. Ornn B200 Index, 23 September 2025 to 22 September 2026 (chart in note).

Open weight and Chinese models taking a majority of routed inference volume. OpenRouter, openrouter.ai/rankings.

Section 6

Up to half of planned US data centres delayed or cancelled. Bloomberg News, "Half of Planned US Data Centers May Face Delays, Kimmeridge Says", 26 August 2026.

Spain and Portugal, 10.5GW announced, 4.5GW active. Bloomberg News, 17 July 2026.

Denmark emergency grid law. Bloomberg News, 20 August 2026.

Ofgem fees for speculative projects. Bloomberg News, 28 July 2026.

PJM pairing data centres with generation. Bloomberg News, 20 May 2026.

Meta multi gigawatt nuclear deals. Bloomberg News, 9 January 2026.

Mitsui LNG. Bloomberg News, 28 May 2026.

Bring your own energy becoming a requirement. Bloomberg Tax, 3 September 2026.

Ohio utility bill backlash. Bloomberg News, 27 July 2026.

House Democrats and the Interior Department. Bloomberg News, 24 September 2026.

Swing Riots, 1830. Historical record.

Section 7

Oracle Project Jupiter, force majeure, carry cost, $18bn construction debt, loans below 90 cents, $300bn OpenAI contract. Financial Times, Michelle Chan, "Oracle on the hook to pay data centre investors even if site has no electricity", September 2026. Bloomberg first reported the force majeure notice.

Token volume roughly tripling through 2026. OpenRouter.

Among firms paying for AI, the top 1% spent a median of $9,567.53 per employee per month on AI, compared with $15.70 for the median firm." Ramp AI Index, ramp.com/data/ai-index.

Section 8

Fund weights and holdings. Goodhart, GFL portfolio data as at 30 September 2026.

Nvidia at 5.9% of the index. Same source, benchmark constituent weights.

Software approach. Goodhart Partners, "AI, SaaS and the mispricing of durability", goodhartpartners.com.