Where cloud goes next.

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.

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.

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.

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.

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.

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.

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 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.

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.

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.

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.

What to do, by seat

Closing where the summary opened: the concrete moves, in plain terms.

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.