Where cloud goes next.
What the cloud providers, the university research labs and the world's leading research houses say is coming over the next ten to twenty years, and what it means for the cost of technology in financial services.
Written for a mixed audience: chief executive, finance director, technology officer, and the engineer in the room. Findings cross-checked against McKinsey, Gartner, Deloitte, Goldman Sachs, JPMorgan and the International Energy Agency.
The same story, through four lenses
This chapter makes one argument: the economics of cloud computing are about to change more in the next ten years than they did in the last twenty, and managing the cost and value of technology is moving from a back-office report to a board-level discipline. Here is what that means depending on the seat you sit in.
The four biggest cloud providers alone are spending roughly USD 725 billion this year building AI infrastructure, up 77 percent in a single year, and McKinsey expects nearly USD 7 trillion of data centre investment worldwide by 2030. Whether they are right or wrong, your technology bill is exposed to their bet. If they are right, capacity will be scarce and prices firm. If they are wrong, a correction follows and the firms with cost discipline buy capability cheaply.
Cloud prices fell for twenty years and most finance teams quietly relied on that. That assumption is now unsafe. The providers are funding the build-out with debt and re-investing roughly half of every pound of revenue back into infrastructure, ratios normally seen at utilities. They will reward customers who commit to multi-year spend and charge a premium to those who do not.
Three shifts will reshape your estate. AI agents are moving into operations; Google already reports 75 percent of its new code is AI-written and engineer-approved. Electricity has replaced chips as the limiting factor on growth, which puts region choice, scheduling and carbon inside the cost model. And the research labs are working to make the cloud providers interchangeable.
Your job changes but it gets bigger, not smaller. The manual parts of cost optimisation will be done by software agents within five years. What grows is the work above it: setting the policies those agents follow, auditing what they did, and explaining to the board what the estate costs and earns.
One discipline connects all four seats: knowing what your technology costs, what it is worth, and who is accountable for the gap. The next twenty years make it more important, not less.
Plain English first
This chapter avoids jargon where it can. Where a technical term is unavoidable, here is what it means. No prior cloud knowledge is assumed.
- Hyperscaler
- One of the giant cloud providers: Amazon, Microsoft, Google, plus Meta and Oracle in the infrastructure race.
- Capex
- Capital expenditure. Money spent building long-lived assets, in this case data centres, chips and power supply.
- Token
- The unit AI models charge by, roughly three-quarters of a word. Every question to an AI and every answer back is metered, like minutes on an old phone contract.
- AI agent
- Software that does not just answer questions but takes actions on its own to reach a goal, within rules a human sets. A junior employee with a defined mandate, working at machine speed.
- FOCUS
- An open industry standard that puts every provider's bill into one common format. A universal bank statement for technology spend, maintained under the Linux Foundation.
- SMR
- Small modular reactor. A factory-built nuclear reactor small enough to power a single data centre campus.
- Fault tolerance
- A quantum computer that can correct its own errors fast enough to run long, useful calculations. The milestone everyone is racing toward.
- Intercloud broker
- A service that takes your workload, compares every provider's price and performance, and runs it on the best combination automatically.
The money: the largest infrastructure build-out in business history
Start with the headline numbers, because everything else follows from them. Amazon, Microsoft, Alphabet and Meta have guided to roughly USD 725 billion of capital spending for 2026, up 77 percent on the USD 410 billion they spent in 2025, with Oracle adding approximately USD 50 billion beyond that. It is the largest annual increase on record, and every revision through 2026 has moved the figure upward rather than down. Around three quarters of the money goes directly to AI: chips, data centres, networking and power.
The independent research houses say this is not a blip. McKinsey projects USD 6.7 trillion of cumulative data centre investment worldwide by 2030, with global demand nearly tripling to around 220 gigawatts. Gartner forecasts total worldwide AI spending of USD 2.59 trillion in 2026 alone. When McKinsey, Gartner, JPMorgan and the providers' own audited accounts all point the same direction, the direction is settled; only the magnitude is debated.
| Provider | 2026 capex | Context |
|---|---|---|
| Amazon | ~USD 200bn | Up from ~USD 125bn in 2025; in-house chip business at a ~USD 20bn revenue run rate |
| Alphabet | ~USD 175–185bn | Google Cloud order backlog above USD 460bn; eighth-generation TPU chips built for AI agents |
| Microsoft | ~USD 110–120bn | Says it will remain capacity-constrained through at least 2026; component price inflation adds materially to the plan |
| Meta | ~USD 115–135bn | Guidance raised twice in a year; cites competition for land, power and skilled labour |
| Oracle | ~USD 50bn | Anchored by the Stargate programme with OpenAI and SoftBank |
Why a CFO should care: the cheap-cloud era is ending
For twenty years cloud prices mostly fell, so doing nothing was a passable cost strategy. The providers are now reinvesting roughly 45 to 57 pence of every revenue pound into infrastructure, ratios normally seen at utilities, and funding the gap with debt: over USD 100bn of bonds in 2025, with projections of up to USD 1.5 trillion of issuance to come. A company carrying that much debt does not cut prices voluntarily. Expect firmer pricing, strong incentives to sign multi-year commitments, and genuine scarcity in premium AI capacity.
Roughly USD 725bn of 2026 capital spending from the top four providers alone, validated by McKinsey's USD 6.7tn-by-2030 projection and Gartner's USD 2.59tn AI forecast. The era of assumed price deflation is over. Cost management shifts from harvesting falling prices to actively managing commitments, scarcity and provider risk.
The power wall: electricity is the new limit on growth
The constraint on cloud growth has moved from chips to electricity, and on this point the research houses speak with one voice. Goldman Sachs forecasts global data centre power demand rising 165 percent by 2030 against 2023 levels, a figure its analysts have since revised toward 220 percent. The International Energy Agency projects global data centre electricity use reaching 945 terawatt-hours by 2030, roughly Japan's entire current consumption, and 1,200 TWh by 2035. Deloitte lands in the same range at 1,065 TWh.
The practical symptoms are already visible. Grid connection queues in key regions run five to ten years. Utilities in Northern Virginia, Oregon and Texas have warned that substations and transmission lines are near their limits. Ireland has restricted new data centre construction around Dublin. A single large AI training facility can draw 100 megawatts, the consumption of a small city.
The nuclear answer, explained simply
All three hyperscalers have signed nuclear power deals because nuclear is the only carbon-free source that runs constantly at the scale they need. Microsoft is paying to restart the Three Mile Island plant, 835 MW expected by 2027. The next step is small modular reactors, starting around 10 MW for a single campus and scaling toward 300 MW designs. Two caveats keep this honest: SMRs take five to ten years to permit and build, so the near-term gap is being filled by natural gas, awkward for net-zero pledges; and hyperscalers are already signing long-term uranium supply deals, which tells you how seriously they take the constraint.
What it does to your bill
Sites with guaranteed power already command lease premiums of 15 to 25 percent, and that premium will surface in cloud pricing as differences between regions. Expect stronger time-of-day and carbon-linked price signals, and expect your sustainability report and your cloud bill to become the same dataset. The FinOps Foundation formalised this convergence in its 2026 framework. For a UK financial institution with net-zero disclosure obligations, power-aware cost management stops being optional within two reporting cycles.
Goldman Sachs, the IEA and Deloitte independently confirm the same constraint. Electricity now sets the ceiling on cloud growth, and energy economics are entering the cloud cost model whether firms plan for it or not.
The agentic turn: software starts running the shop
2026 is the year AI agents moved from demonstrations into production. Gartner forecasts spending on agentic AI within software rising 141 percent in 2026 to roughly USD 202bn, predicts 40 percent of enterprise applications will contain task-specific agents by the end of 2026, up from under 5 percent a year earlier, and projects that by 2028 at least 15 percent of day-to-day work decisions will be made autonomously. Deloitte expects as many as 75 percent of companies to be investing in agentic AI by the end of this year. The providers' own behaviour matches the forecasts: Google reports 75 percent of its new internal code is now AI-generated and engineer-approved.
The Gartner warning, and why it strengthens the case
Anyone presenting this material will be challenged with Gartner's other headline: over 40 percent of agentic AI projects will be cancelled by the end of 2027. Quote it back with its reasons, because they make the argument for discipline. Gartner attributes the failures to escalating costs, unclear business value and inadequate risk controls, and notes that of the thousands of vendors claiming agentic capability, only around 130 are genuine. McKinsey adds that while two-thirds of enterprises have experimented with AI agents, fewer than 10 percent have scaled them to measurable value. The projects fail for exactly the reasons a cost-and-value discipline exists: nobody priced the work, nobody defined the value, and nobody set the controls.
For a regulated UK institution there is a question coming. An autonomous agent that moves workloads or renegotiates capacity is making an operational decision. Expect the FCA and PRA operational resilience regimes to ask, within a few years, exactly how such agents are bounded, logged and overridden. Firms that design the control framework now will answer that letter in a paragraph. Firms that do not will answer it in a programme.
Gartner, Deloitte and McKinsey converge on the same picture: agents are arriving fast, and roughly half the early projects will fail for governance and value reasons, not technical ones. The winning posture is neither enthusiasm nor scepticism. It is controls-first adoption with cost and value measured from day one.
Inside the research wings
Product announcements tell you about next year. To see the next two decades, look at where the providers and the universities are placing long research bets. Genuinely secret internal R&D is unknowable by definition, but the layer just above it is public and remarkably revealing.
Bet one: make the cloud providers interchangeable
UC Berkeley runs five-year flagship computing labs with an unusual record: previous labs produced the technology behind Spark and the company Databricks. The current Sky Computing Lab, led by Databricks co-founder Ion Stoica, exists to do one thing: turn cloud computing into an interchangeable commodity. The mechanism is an intercloud broker, software that takes your workload, compares every provider's price and performance, and runs it on the best combination automatically. Its open-source SkyPilot tool is already used in production for training AI across providers.
The detail worth repeating in any boardroom is the sponsor list: Google, IBM, Intel, Samsung, VMware and SAP fund the lab. Several of the world's largest technology firms are paying a university to erode the pricing power of AWS and Azure. If the work succeeds even partially, switching providers becomes cheap and continuous. If it fails, the likely cause is data-exit fees and proprietary AI services keeping customers locked in, and regulators in the UK and EU are already examining exactly those fees. Either branch raises the value of knowing your costs in a provider-neutral format.
Bet two: change the physics of computing
All three hyperscalers run their quantum programmes through university-linked labs, each betting on different physics. The table is deliberately plain about status: nothing here is commercially useful yet, and the honest independent view is more cautious than the corporate roadmaps.
| Lab | Bet | Roadmap and honest status |
|---|---|---|
| Google Quantum AI | Superconducting chips, recently adding a second track | Willow chip hit a key error-correction milestone; leadership targets useful machines around 2029. Running two tracks suggests no approach has clearly won. |
| AWS / Caltech | Cat qubit chips with error correction built into hardware | Ocelot claims up to 90 percent less error-correction overhead, potentially pulling timelines forward by five years. Promising; unproven at scale. |
| Microsoft | Topological qubits, more exotic physics with a bigger payoff | Majorana 1 on a roadmap toward a million qubits; Microsoft projects quantum machines in data centres by 2029. The underlying physics is still scientifically contested. |
| IBM Quantum | Superconducting, modular systems | Publicly expects verified quantum advantage by end of 2026 and fault tolerance by 2029. The most specific public timeline in the industry. |
The providers' roadmaps cluster around 2029; independent assessments put industrially useful quantum five to ten years out. It will arrive as a cloud service with its own metered pricing, not as a machine you buy. And one consequence is certain regardless of timing: financial regulators will mandate quantum-safe encryption before the threatening machines exist, because migrating cryptography across a bank takes years. That migration is a plannable, budgetable programme and it belongs in multi-year investment plans today.
Bet three: remove the heat
If electricity is the constraint, the obvious research response is computing that produces less heat. Photonic computing performs AI's core mathematics with light instead of electrons, with near-zero heat loss; the academic consensus is that it moves from laboratory to core infrastructure around 2030. The hyperscalers' heavy investment in optical networking is the on-ramp to the same technology family.
The labs reveal a coherent long-term plan: make the cloud layer interchangeable, change the physics of compute, and remove the energy ceiling. Every one of those outcomes creates new pricing models and new arbitrage, and therefore more need for cost-and-value management, not less.
The orbital frontier
This sounds like science fiction, so lead with why serious money disagrees. In the right orbit, a solar panel generates up to eight times more energy per year than the same panel on Earth, almost continuously, with no land purchase, no water for cooling and no ten-year grid queue. Space is a direct answer to the power wall.
Google's Project Suncatcher was announced with a published research paper rather than a press release. Google has radiation-tested its own AI chips for a five-year orbital mission, partnered with satellite firm Planet Labs, demonstrated the laser links needed to connect satellites, and launches two prototype satellites by early 2027. Its analysis finds that if launch costs fall below USD 200 per kilogram by the mid-2030s, an orbital data centre becomes cheaper than building one on Earth. Separately, Starcloud, backed by Nvidia and Google, launched its first satellite carrying a top-end AI chip and has filed for an 88,000-satellite computing constellation.
Now the sceptics, with equal weight. Rival space firm Varda calculates orbital computing at roughly three times the cost per watt of ground facilities at today's launch prices. Cooling dense hardware in a vacuum is brutally hard, satellites cannot be repaired once launched, and ground links face weather constraints. The sober reading: orbital compute is a 2030s capacity class for specific batch workloads, not a replacement for ground regions.
Treat orbital data centres the way one should have treated GPUs in 2016: an apparently fringe capacity class that serious capital is quietly funding. The one number to watch is launch cost per kilogram. Below USD 200 the economics flip.
Predictions, with confidence ratings
These are this guide's own calls, built from the evidence above. Confidence is rated high (would be surprising if wrong), medium (balance of evidence), or speculative (directionally argued, timing genuinely uncertain). Where a major research house has published a comparable view, it is named.
| Prediction | Confidence and external support |
|---|---|
| Token-level AI cost management becomes a standard enterprise capability; FOCUS becomes the default data format across cloud, SaaS and AI spend | High Gartner: agentic software spend +141% in 2026. FinOps Foundation: 68% of USD 100m+ spenders already on FOCUS. |
| AI agents take over routine cost optimisation under human-set policy | High Gartner: 15% of daily work decisions autonomous by 2028. Deloitte: 75% of firms investing by end 2026. |
| Roughly half of early agent projects fail on cost, value or controls, creating a governance consulting wave | High Gartner: 40%+ cancelled by end 2027. McKinsey: under 10% scaled to measurable value. |
| Power-driven regional price differences appear; carbon and cost reporting converge into one dataset | High Goldman Sachs, IEA and Deloitte power forecasts all imply it; FinOps Framework 2026 formalises it. |
| Nuclear supplies 5+ GW of dedicated data centre power by 2030; at least one major jurisdiction restricts new construction | Medium Three Mile Island restart contracted; Ireland precedent set. |
| A capex digestion phase hits provider pricing behaviour before 2030, in one direction or the other | Medium McKinsey flags stranded-asset risk; bond markets already pricing protection. |
| Prediction | Confidence and external support |
|---|---|
| The cost discipline formally splits into agent governance and value architecture | High Direct extrapolation of the FinOps Foundation's published agentic maturity model. |
| Intercloud brokerage becomes commercially real for AI workloads; data-exit fees are restructured under regulatory pressure | Medium Berkeley Sky trajectory plus active UK and EU scrutiny of egress fees. |
| Early fault-tolerant quantum machines sell as premium cloud services; quantum-safe migration becomes a regulator-driven programme | Medium IBM and Microsoft both target 2029; regulators will move ahead of the physics. |
| Photonic accelerators enter production data centres for AI inference, bending the energy-per-token curve | Medium Academic consensus centres on ~2030; inference is already two-thirds of AI compute. |
| First commercially meaningful orbital compute capacity is sold as a service | Speculative Prototypes fly 2026–27; economics hinge on the launch-cost curve. |
| Prediction | Confidence and external support |
|---|---|
| Compute is priced and traded like energy: spot, futures and capacity markets spanning providers and orbital tiers, navigated mostly by agents | Medium Every component is independently in motion. |
| The infrastructure layer is substantially commoditised; provider differentiation lives in models, data and agents | Medium The explicit goal of the Berkeley programme and its corporate sponsors. |
| Orbital data centres are a normal capacity class | Speculative Requires sub-USD 200/kg launch costs. |
| Quantum, photonic and conventional chips coexist as workload-routed options, each with distinct unit economics | Speculative Multiple physics maturing on overlapping timelines. |
| The discipline now called FinOps reports to the board as a peer of treasury, likely under a broader name | Medium The Foundation already changed its mission from value of cloud to value of technology in 2026. |
The validation matrix
Every load-bearing claim above is cross-checked against at least one tier-one research house or primary source. Nothing rests on our analysis alone.
| Our claim | Who else says it | Their number |
|---|---|---|
| Historic build-out underway | McKinsey | USD 6.7tn cumulative data centre capex by 2030; demand triples to ~220 GW |
| 2026 is the inflection year | Gartner; JPMorgan | Worldwide AI spend USD 2.59tn in 2026, +47%; top-four cloud capex up 77% year on year |
| Electricity is the binding constraint | Goldman Sachs; IEA | Power demand +165% by 2030, since revised toward +220%; IEA 945 TWh by 2030 |
| Energy costs flow into bills | Goldman Sachs; Deloitte | US utilities need ~USD 50bn of new generation; US share of electricity heading to 8–11% by 2030 |
| Inference, not training, drives load | Deloitte | Inference is two-thirds of all AI compute in 2026 |
| Agents are arriving now | Gartner; Deloitte | Agentic software spend +141% to ~USD 202bn; 40% of enterprise apps with agents by end 2026 |
| Half of early agent projects fail | Gartner; McKinsey | 40%+ cancelled by end 2027; under 10% of enterprises scaled to measurable value |
| Cost data is standardising | FinOps Foundation | FOCUS 1.4 ratified 2026; 68% of USD 100m+ spenders using or trialling it |
| Quantum clusters at end of decade | IBM; Microsoft; Google | IBM: fault tolerance 2029. Microsoft: data centre quantum by 2029 |
| Space compute is funded, not fantasy | Google; Starcloud | Two Suncatcher prototypes by early 2027; sub-USD 200/kg crossover by mid-2030s |
The six questions you will get in the room
A forward-looking document earns trust by surviving challenge. These are the pushbacks a sharp CFO, CTO or non-executive will raise, and the grounded answers.
"Isn't this just an AI bubble? Why build a discipline on it?"+
Partly, perhaps. McKinsey itself flags stranded-asset risk and the bond market is pricing protection. But the discipline is bet-neutral. If the boom holds, you manage scarcity and commitments. If it corrects, you capture the renegotiation. It is the only part of the AI story that does not require the optimists to be right.
"Gartner says 40 percent of agent projects will fail. Why invest?"+
Read Gartner's reasons: escalating costs, unclear business value, inadequate risk controls. Those are not technology failures; they are the absence of exactly this discipline. The 40 percent statistic is the business case for governance-first adoption, not against adoption.
"Our cloud bill has been flat. Why would prices rise now?"+
Because the providers' economics changed. They are reinvesting roughly half of revenue into infrastructure and borrowing to do it. Microsoft has told investors it will be capacity-constrained through 2026. Debt-funded, capacity-constrained suppliers do not cut prices. They reward commitment and charge for flexibility.
"Energy is the data centre operator's problem, not ours."+
It was. Goldman Sachs puts demand growth at 165 to 220 percent by 2030 and the IEA at near-Japan-scale consumption; power-certain sites already carry 15 to 25 percent premiums. Those costs pass through to regional pricing, and for any firm with net-zero disclosures the carbon side lands on your report. Energy enters your cost model through two doors at once.
"Quantum and space data centres: seriously?"+
Seriously, with dated humility. Nothing in either field changes a 2026 budget. What changes plans now is quantum-safe encryption, because regulators will mandate migration before the machines work and a bank-wide cryptography migration takes years. And the launch-cost curve is a single observable number. Track it annually and ignore the noise.
"Why act now rather than wait for this to settle?"+
Three of the shifts are already invoiced: AI spend, agent deployment and the FOCUS data standard. The data foundations and governance frameworks are cheap to establish early and expensive to retrofit after the spend has scaled. Waiting does not avoid the work; it moves it to a worse price.
Every answer above cites a tier-one source, not our opinion. That is the standard this chapter holds itself to, and the standard any adviser should be held to.
What to do, by seat
Closing where the summary opened: the concrete moves, in plain terms.
| Role | The move | When |
|---|---|---|
| CEO | Put technology value on the board agenda as a standing item, owned by a named executive, with one page connecting spend to business outcomes | Next board cycle |
| CFO | Treat multi-year cloud commitments as treasury-grade decisions; commission an AI unit-cost baseline before AI spend scales | 2026 |
| CTO / CIO | Stand up the agent governance framework — boundaries, logging, override, audit — before agents touch production; add power and carbon to region selection | 2026–2027 |
| FinOps lead | Adopt FOCUS-formatted cost data as the foundation; build token-level AI cost allocation; run the first bounded automation | Now |
| CISO / CRO | Start the quantum-safe cryptography inventory and migration plan; fold autonomous agents into the operational resilience framework | 2026–2028 |
The next twenty years automate the tasks of cost management and elevate its questions. The work moves from reports to operating models, from optimisation to governance, from cloud cost to technology value. Position there now, with the validation matrix in hand.
Sections on capital expenditure and agentic adoption move fastest and are reviewed every six months. Figures were verified in June 2026 and the capital and agentic numbers re-checked in September 2026.
Independent means independent. Nivaan sells no cloud, no tooling, no licences and no managed services, and takes no commission from any vendor. What that means. Nothing above is a recommendation to buy anything, from us or anyone else.
This is one chapter of a longer reference.
The full guide covers the operating side: cost taxonomy across AWS, Azure and GCP, accountability, tooling, and a ninety-day plan. Fifty-eight pages, sent by email the same working day.
Request the guideOr start with the three short guides, which download directly.