Every gigawatt in this report has to be paid for. The build-out is the largest coordinated capital cycle technology has run, and the cost has inverted: the silicon, not the building, is now the asset. This is the money layer, who funds it, how the stack is built, and where the risk is hiding.
The other readings in Strata place the markets, rank the conditions, and follow the power and the silicon. Capital reads across all of them and asks the question every announced gigawatt eventually faces: where does the money come from. Three things make this cycle different. The sums are unprecedented, with credible estimates clustering between $3tn and $7tn by the end of the decade. The cost structure has inverted, so that in an AI hall the accelerators are roughly sixty per cent of the build and the real estate is a minority. And roughly half of the spend cannot be funded from hyperscaler cash flow, which has pulled private credit, securitisation and sovereign capital into a market that used to be plain real-estate debt.
Two waterfalls, one toggle. Cost to build stacks the layers of a single AI facility, from the land up to the silicon, and shows the inversion: the cheapest layer gates the project, the most expensive one depreciates fastest. How it is funded stacks the macro sources of the spend through 2028, and shows the gap that hyperscaler cash flow leaves behind. Hover or tap any block to read it.
Each tile carries one of his six points as the heading, and beneath it an independently sourced figure that Entelligencia has put to the same question. The evidence is ours, not his: he was not shown these numbers and did not cite them. Open any tile to read his point in full, verbatim, alongside the data and its grade.
Appetite for opportunistic and development strategies fell in 2025, the first pullback in four years, even as fundraising hit records.
Grid lead times in the slowest FLAP-D markets run five times longer than the 18 to 24 months needed to build the shell.
Down from 16.9% in 2021, with 83% of the European pipeline pre-let before it completes. Scarcity is what lets the best platforms price.
A Brussels retail colocation carve-out against a Paris hyperscale one. Same year, same continent, nearly double the multiple.
Up from under 8 billion in 2024. GPU-backed paper prices at SOFR plus 225 basis points, against AAA-rated hyperscaler credit.
A record year for data-centre M&A, and private equity now funds roughly four out of five deals by value.
He answered twelve questions for the Capital chapter. Six are set out here, each paired with an independently sourced figure that Entelligencia has put to the same question. The evidence is ours, not his. Open any tile for his answer in full, verbatim.
He scores it 88 out of 100 toward a different game entirely. The capex curve behind that call has risen roughly two and a half times in four years.
His line is that power, not demand, now limits growth. Texas alone saw its large-load queue nearly quadruple inside a single year.
His investment method is to name the future buyer first and tune the asset to them. The exit market has grown enough to make that arithmetic work.
He argues PE is uniquely placed to fund this because it can raise continuously. Its share of data-centre deal value has moved from about half to about four fifths.
He names the neo-cloud cohort as the screen. The paper prices exactly as he describes: hyperscaler credit at AAA, GPU-backed neo-cloud debt hundreds of basis points wider.
His warning to LatAm founders. A peso-denominated return stream lost roughly a third of its dollar value between the long-run average rate and the 2025 peak.
A capital allocator’s read, contributed to The Next Hotspot
Milan Radia runs ConnectedCompute and sits on the data-centre strategy side of a roughly $2bn Saudi programme. He is the only contributor to score this report’s delivery gap below the median and then argue the measurement is wrong. His case is that capacity is arriving but the wrong kind of it, that build-to-suit took the capital while colocation was left behind, and that inference is about to make that a problem. Eight files, six of them with independent evidence beside them.
His correction to this report’s central frame, and he scores the delivery gap at 60 out of 100, lower than most contributors. The question, on his account, is not whether a headline capacity gap exists but whether the right type of capacity is arriving. Accelerating workload change has left much of what was delivered in recent years out of sync with the rack density, structural loading and latency requirements that AI deployments actually need.
Average server rack density remains below 8 kW, and most facilities have no racks above 30 kW at all. An NVIDIA GB200 NVL72 rack draws 132 kW.
Supports him precisely. Uptime found 29 per cent of operators upgrading existing halls and 33 per cent building new high-density capacity, which is the retrofit problem he describes showing up in the survey. Retrofit runs $3–8m per MW against $8–15m greenfield.
The mechanism behind where the money went. Almost all recent global data-centre investment has gone into build-to-suit, with developers raising debt against ten to fifteen year contracts from hyperscale platforms while the platforms themselves issue bonds in immense quantity to fund the GPU compute. Colocation was left behind on smaller scale, the need to nurture ecosystems and communities of interest, and a lower proportion of pre-commitment.
Hyperscaler bond issuance in 2025, against roughly $20bn in 2024 and a 2020–2024 average near $28bn a year. Meta’s $30bn print in October 2025 was the largest non-acquisition investment-grade corporate bond sale on record.
And the capacity split moves with it. Hyperscale reached 48 per cent of world capacity by the end of 2025 across 1,360 facilities, up from 44 per cent at the start of the year, with non-hyperscale colocation at 22 per cent. Synergy projects hyperscale at 67 per cent by 2031.
“Colocation today feels like a lost art, but it is arguably more relevant than ever when considered in the context of AI Inference.”
The line worth keeping. Colocation today feels like a lost art, and is arguably more relevant than ever when read against AI inference. Training runs are centralised and campaign-based. Inference is continuous, latency-bound and closer to the user, which is the workload profile colocation was built for and hyperscale campus economics were not.
AI inference demand, growing at a 35 per cent compound rate and overtaking non-AI demand by 2029. Training grows from 23.1 GW to 62.2 GW over the same period, at 22 per cent.
Inference is projected at two thirds of AI compute in 2026, from one third in 2023. These are projections rather than realised figures and are printed as such, but the direction is the one his argument depends on.
What his projects across Africa, Europe and the Middle East hold in common: carrier and cloud neutrality, density of connectivity, and latency characteristics fit for AI inference, content delivery, online gaming and mission-critical enterprise work. These read as obvious and are missing across much of the world, and their absence is itself an impediment to hyperscale investment and deployment.
Equinix monthly recurring revenue churn, on 507,000 interconnections including 437,700 physical cross-connects, against FY2025 revenue of $9.2bn.
The clearest available proxy for what interconnection density is worth. Cross-connects price at $100 to $300 a month against near-zero variable cost, which is why the ecosystem, not the floor space, is the margin.
To understand the most successful models globally you have to be a student of history. Telehouse London and Science Park in Amsterdam remain the sought-after venues for mission-critical and latency-sensitive work, command the highest pricing, see minimal churn and return exceptionally for their owners. The primary barrier to building that kind of infrastructure is a failure to understand how it differs from a conventional data centre in an industrial park fifty kilometres from the population it serves.
Telehouse London Docklands opened in 1990 as Europe’s first carrier-neutral facility and is the primary home of LINX. Amsterdam Science Park hosts 170 to 210 carriers and was the first home of AMS-IX, plus NL-ix, NDIX and NetherLight.
Both examples check out as the archetypes he names. What no public dataset gives is a clean yield or churn series for carrier hotels against conventional facilities, so the pricing and returns claim rests on his seat and Equinix’s disclosed churn.
The most actionable thing in the submission. A government anchor-tenant contract can bring an African project to life, unlocking funding and drawing in financial services, energy and other sectors that previously had no credible third-party capacity. Where it goes wrong is precise: if the state ends up with an ownership stake in the operating company, which deters cloud and content platforms from deploying, or if it proves reluctant to follow through on the accompanying deregulation and licensing reform.
The Microsoft and G42 geothermal project in Kenya, announced May 2024, suspended in May 2026 after the government declined to guarantee roughly a gigawatt of power and the payment offtake behind it.
The clearest live illustration of his second failure mode: the state not following through. On his first mechanism, that a government equity stake deters the platforms, no public case names a deal killed that way. It is graded as his judgement and is printed as such.
“Chaos ensues.”
It is extraordinary how many projects run dramatic cost and schedule overruns to the detriment of every stakeholder. His diagnosis is a lack of educated oversight, leaving project silos disjointed and producing expensive rework: a civil contractor with little or no data-centre experience misreads seemingly ancillary elements, and nobody discovers it until the civil works are declared almost complete and the mechanical and electrical teams arrive on site.
Data centres commit 60 to 75 per cent of construction budget to mechanical, electrical and plumbing, against 30 to 40 per cent for typical commercial buildings. Poor design coordination accounts for up to 52 per cent of rework.
The MEP concentration supports why his failure mode is expensive when it happens. His claim is scoped to the execution phase, civil works and MEP implementation, rather than to project delay overall. Power procurement and equipment lead times are known before execution begins and sit outside that scope, so they are not a counter-case to it. Corrected at the contributor’s request, 3 August 2026.
His read on the market his own fund targets, and the reason he gives for it. Saudi Arabia has sovereign projects and build-to-suit at scale, and very little genuine carrier-neutral colocation. What is missing, in his words elsewhere, is the connectivity ecosystem model that made Equinix, Telecity and Interxion successful in Europe, and nobody in the region has yet built a true community-of-interest environment.
Saudi live IT load, with roughly 760 MW more projected by 2030 at a 29 per cent compound rate, across 34 operational colocation facilities as of December 2024.
Supports the shape of his argument. The Kingdom’s scaffolding is real, from the 2020 open-access initiative through the 2023 Cloud Computing Special Economic Zone to the June 2025 National Data Centre Strategy targeting 1.5 GW. The capacity is arriving; his point is about what kind.
He is building the thing he says is missing. Carrier-neutral colocation in Saudi Arabia is his commercial position, and the reader should have that while reading a case for carrier-neutral colocation in Saudi Arabia.
Two things make the file worth publishing anyway. The first is that he scores the delivery gap lower than the report does, which is not the convenient direction for someone raising against it. The second is that his central argument, that the industry is measuring the wrong thing, is testable, and Uptime’s density survey tests it in his favour.
On the Saudi programme itself: the desk records it as a memorandum of understanding signed in November 2025 between Taranis Capital and Emaar Executive Company, endorsed by the Saudi Investment Promotion Authority, at a target value near $2bn across 40 to 50 MW campuses. It is not closed or financed capital, and it is graded Announced rather than Verified for that reason.
Contributed via the Entelligencia briefing survey, 1 August 2026, attributed by name with firm and title at the contributor’s request. Seven questions were put; all seven were answered, and the contributor noted that several questions did not align cleanly with the answers he wanted to give. That is recorded here rather than resolved by rewriting him: where a response addresses a different question from the one asked, it is published under the argument he made rather than the question we put. His positions are reported in the desk’s house style except where quotation marks appear. His delivery-gap score, his market judgements and his causal readings are graded Contributed; the Saudi programme is graded Announced.
This layer has read the capital from the fund side, the analyst side and the bank side. Mike Niekoop reads it from inside the documents, where the risk actually has to be allocated. His account of what lenders now demand before money moves runs from power diligence and borrowing-base tests to what a GPU is assumed to be worth in year four. Eleven files, including the €3.3bn financing that split one European platform into two balance sheets.
Contributed by Mike Niekoop, Partner, Norton Rose Fulbright, 11 August 2026, named with firm and title at the contributor’s request · passages inside quotation marks are reproduced from his written submission; his other positions are reported in the desk’s house style · supporting evidence and grading are Entelligencia’s and were gathered independently · he advised on the DATA4 financing described in file 02 while at Clifford Chance, and that is stated on the file rather than implied
The headline estimates differ by trillions, partly because they measure different things over different horizons. We line them up here on a common scale, in trillions of dollars, with each one's horizon labelled. Treat all of them as projections, not realised spend. The realised figures, at the foot, are an order of magnitude smaller, and growing fast.
Scale is trillions of dollars. Horizons differ: Brookfield and McKinsey run to the end of the decade and count the wider value chain or all data-centre build; Morgan Stanley's is a tighter 2025 to 2028 window. McKinsey's $5.2tn is the AI-specific slice of its $6.7tn total.
Those forecasts count dollars. The layer that actually constrains them is silicon. Accelerated hardware has gone from about 40 per cent of AI-infrastructure spend in 2023 to roughly 65 per cent on IDC’s read, as CPU-only builds give way to GPU clusters, and IDC puts cumulative AI-infrastructure spend near USD 500 billion through 2028. The unit story underneath is a supply ramp: global GPU shipments roughly tripled, from 1.2 million units in 2022 to 3.8 million in 2024 on Omdia and Jon Peddie data, and that volume pulled the H100-equivalent average selling price down from about USD 25,000 toward USD 18,000.
The ramp created a new asset class: the GPU-native cloud. The market for GPU-accelerated cloud reached roughly USD 62 billion in 2024, up from about USD 15 billion in 2019, and specialists such as CoreWeave now sit alongside the hyperscalers, financing dense silicon on private credit with the chips themselves as collateral. The mismatch that runs through this page holds here too: the building depreciates over twenty to thirty years, the silicon over three to five, which is why the capital, not the concrete, sets the clock.
The external half of the build is being underwritten by a short list of very large private-capital houses, increasingly alongside sovereign funds and chipmakers as anchor investors. The figures below mix committed capital, holdings and headline deals, so read them as scale markers rather than like-for-like. The single largest private-credit transaction on record, Meta's $27bn Hyperion facility, sits across two of these names.
Behind every earmarked gigawatt sits a supply chain, and most of it trades on public markets. We map it in five layers, from the silicon up to the power that runs it, with market value, 52-week return and analyst consensus for 28 listed names, then set that listed world against the private valuations climbing behind it. Market data as at 18 June 2026; figures are shown for analysis, not as investment advice.
Scale is billions of dollars. CoreWeave is now listed; the rest are private marks. Nscale, London-based, roughly doubled from about $7bn to $14.6bn in six months, anchored by NVIDIA, Aker and 8090 Industries, with a 1.3 GW pipeline across the UK, Norway and the US. It is the British name-check in the government’s AI strategy, and the one whose curve most closely traces CoreWeave’s pre-IPO path.
The thesis across the supply chain is overwhelmingly bullish, but valuation discipline is the whole game: the build is real, the suppliers are mostly listed, and the question the rest of this reading asks of every announced gigawatt applies here too. Sources: market and consensus data as at 18 June 2026 (Dell’Oro, JLL, company filings, CNBC, Reuters); private marks from Pitchbook, CNBC and Reuters. For analysis only, not investment advice.
The same houses, read as geography. Each node is a named commitment to a specific build, sized by headline value and colour-coded by how firm it is: verified and financed, announced, or early-stage. The United States holds the largest single bets; the rest is a short list of national programmes from France to the Gulf. Filter by region, or open a node for the deal behind it.
The numbers above describe how much capital the build needs. They do not describe how it gets arranged. That part is engineered deal by deal, by the people selling into it, and it rarely appears in a capex table. This section collects those structures as they are described to us, attributed and graded as the contributor’s own reading rather than as measured record. One supplier, one cost desk: how the resilience gets financed, and why the number it is financed against moves.
“Power is the new differentiator. Standby isn’t backup anymore, it’s strategic infrastructure integrated with on-site power.”
Standby capacity is financed together with the prime on-site power system rather than sitting on the customer’s balance sheet as a separate capital line. The resilience kit stops being a discrete purchase and becomes part of a supplied power service.
If it holds at scale it moves on-site power from a capital decision to an operating one. That changes who can afford to build without a balance sheet the size of a hyperscaler’s, and it puts the supplier, not the lender, at the centre of the resilience question.
“LATAM projects win funding when the numbers are believable, not just competitive.”
Row 04 above prices an AI-grade build at $30 to 40m per MW. Castro Vila is the seat that decides whether a number like that survives contact with a market. His argument is that schedule remains a real risk on any Latin American build, and that cost uncertainty is usually the greater one: tax complexity, currency movement, labour regulation and market pricing can wreck a project whose dates were always achievable. He scores it 70 out of 100 toward cost being the risk.
Overruns, on his account, are rarely estimating errors alone. They come from engineering development and design change, gaps in scope definition, procurement assumptions that do not hold, and project requirements that evolve after commitment, compounded by how the contingency was built in the first place. Asked what a realistic contingency looks like, he moved the question from the percentage to the method behind it. Scrutiny at funding concentrates on five lines: contingency, escalation, imported equipment, currency exposure and contractor capability. Imported plant is not simply more expensive. Currency, customs, logistics and where the risk sits in the contract mean its effect falls on cost certainty rather than cost alone, and the harder problem is that contractors price those risks inconsistently, which turns bid evaluation into a technical exercise rather than a commercial one.
A percentage contingency is drawn from historical benchmarks and applied one-size-fits-all. He accepts it as useful early in a project, and weaker as a basis for a decision, because it describes developments that already happened rather than the one being priced. The question, on his account, is less the percentage than the methodology used to derive it. What he wants instead is a deterministic contingency built from an integrated risk assessment: the team identifies each risk, evaluates it, and quantifies it against probability and impact. The number that falls out is tailored to that project’s actual risk profile rather than to the average of a benchmark set.
Read against this report’s central argument, it locates the optimism precisely. The gap between announced and delivered capacity does not open because someone estimated badly. It opens because a percentage carried forward from other projects was never a statement about this one.
It supplies the mechanism this report has argued without: how announced capacity fails to land without anyone acting in bad faith. A global benchmark applied without local calibration is not a lie. It is just wrong, and it is wrong before a spade goes in the ground.
Chitan co-heads CMS’s global Communications practice and has spent more than two decades structuring the debt and equity behind digital infrastructure, across data centres, fibre, towers, subsea cable and satellite. On the compute side that includes the Ark Data Centres financings in the UK and a pan-European financing of Atlas Edge described as a first-of-its-kind one-stop solution; on the connectivity side, CityFibre’s acquisition of FibreNation and Macquarie’s purchase of rural fibre covering 1.1 million Spanish households. The read that gives her is cross-estate: what actually underwrites is not a single hall but a platform.
Her through-line from recent TMT Finance panels is that platform creation, not asset assembly, is what earns a premium valuation, and that AI is reshaping the fibre routes, metro connectivity and edge that feed the data centres, not just the data centres themselves. The consent bundle, planning, grid, wayleaves and spectrum, now sits on one financing timeline. The demand, in her telling, is starting to be financed as a system, compute plus network plus edge, rather than asset by asset.
Where Banzon reads the chip cycle, Szlezak writes the equity cheques behind it. He sizes global data-centre spending at around $250bn a year, and reads US live capacity at roughly 16 to 18 GW against about 6 GW each in Europe and Asia, a base he expects to double or triple. The structural shift he points to is on power: operators have moved from a just-in-time approach to guaranteeing access up front, which is precisely the premium the waterfall's power layer is starting to price.
In June 2026 KKR put that thesis into a vehicle, launching Helix, a hyperscale platform seeded with more than $10bn of long-duration capital, anchored by KKR alongside Nvidia, the Kuwait Investment Authority and Vistra, with Szlezak as chief investment officer. It sits inside a KKR infrastructure platform of more than $100bn.
The senior layer of the financing waterfall is where the bank desks sit, and ING has arranged more than two hundred data-centre financings for operators including Equinix, Vantage, EdgeConneX and Aligned. That gives Boomsma's desk a long read on what actually underwrites, and his public emphasis is consistent: power is the gating factor, and in Europe the underwriting is sustainability-led, with operators aligning their infrastructure to EU environmental goals.
On a 2023 European financing he noted that even as real-estate and leveraged-loan markets tightened, there was still significant liquidity available from infrastructure-focused lenders, the through-line being bankability: which projects clear an infrastructure lender's test, and where underwriting tightens first.
A financing structure is only as sound as its weakest assumption. Four sit underneath this one, tagged by how contested they are. The pattern across them is the same tension that runs through the whole report: the gap between what is announced and what can actually be delivered, financed and run.
On the data. The capital-stack and funding waterfalls are Entelligencia syntheses of published research and should be read as indicative, not audited. The build-cost split draws on McKinsey, JLL and component-level estimates; the funding split follows Morgan Stanley's framing of roughly $2.9tn of spend through 2028, about half self-funded, with the external gap carried by private credit, investment-grade bonds, securitisation and a sovereign and bank catch-all. Capex totals are projections with wide scenario ranges and differing horizons, and are tagged accordingly. Contributor sections quote only public, attributable remarks; where a financier's deeper capital-stack views are not yet on the public record, that is stated. Headline figures for capital providers mix committed capital, holdings and named deals and are scale markers, not like-for-like. Depreciation, circular-financing and refinancing items are tagged Contested or Estimated. Sources include McKinsey, Morgan Stanley, Moody's, KKR, Blackstone, Apollo, Brookfield, Blue Owl, TAM Asset Management, ING, Quinn Emanuel, S&P Global and the named company disclosures. See the full method →
Where the build actually happens, and what is in the way.
