The Productivity Story Is Over — the Power Story Has Begun

A thought leadership essay on cloud concentration, hybrid reality, and why the next decade is about Data Battles (and the risk of GenAI “junkification”).

Let’s stop pretending we’re still debating “cloud adoption.” That debate ended years ago. What’s happening now is more consequential: cloud has become a concentrated utility, and most enterprises are quietly becoming tenants in three privately-owned empires.

In Q4 2025, global cloud infrastructure services reached $119B in quarterly revenue, up about 30% year-on-year—and the growth engine wasn’t migration, it was GenAI demand and AI-specific consumption.
And as that market accelerates, it’s not spreading out—it’s tightening at the top: AWS, Microsoft, and Google captured roughly two-thirds of global cloud infrastructure spend.

This is not an accident.

This is what mature platform markets do: they stop competing on “features” and start competing on gravity.

The game has moved from adoption to power dynamics.

1) The Cloud Market Is a Three-Horse Race

Cloud adoption didn’t “hit a wall.” It hit a ceiling of easy productivity.

For most enterprises, cloud has largely plateaued in terms of migration-driven gains. The first wave delivered obvious wins—provisioning speed, elastic capacity, reduced data-centre friction. That part is done. What remains is the expensive, unglamorous work: operational discipline, governance, FinOps, platform engineering, and measurable outcomes—not more lift-and-shift theatre.

And while productivity gains get harder, cloud power gets simpler: three vendors dominate the infrastructure substrate of modern enterprise IT.

Market tracking for Q4 2025 shows the “Big Three” holding roughly two-thirds of cloud infrastructure services: AWS at about 28%, Microsoft around 21%, and Google Cloud around 14%—with the remainder split across Alibaba, Oracle, IBM, and the long tail.
Synergy’s reporting emphasises the structural picture: AWS + Microsoft + Google together represent about two-thirds of enterprise spend, and GenAI has pushed the market “into overdrive.”

That concentration changes the meaning of “cloud strategy.” It’s no longer “Should we adopt cloud?” It’s Which empire do we standardise on—and what do we lose by doing so?

When two-thirds of the market sits in three hands, “cloud strategy” becomes less about architecture and more about how—and how deeply—you accept vendor gravity and lock‑in.

2) Supercloud Is a Nice Idea. Hyperscalers Prefer Gravity.

You’ll see Supercloud referenced in industry literature: an abstraction layer above providers, promising portability, unified control, and the comforting illusion of neutrality. Nice idea. Wrong incentive structure.

Hyperscalers aren’t neutral roads. They are destination economies—and they behave like it.

In a market this concentrated, providers don’t win by enabling easy exit. They win by increasing switching costs and making your next decision feel “obvious” inside their ecosystem—especially when GenAI is accelerating spend and tightening dependence on platform-native services.

The lock‑in playbook doesn’t even need to be sinister. It can be mundane, incremental, and spreadsheet-approved:

Reservations and committed-use discounts that reward multi‑year dependency and punish migrations.
Managed services that are “too good” relative to self-managed options—until your architecture becomes inseparable from proprietary control planes.
Integrated identity, security, observability, and data platforms that collapse toolchains into one vendor.
Industry clouds and vertical solutions that embed “best practice” directly into vendor-specific primitives—so your compliance and process maturity become provider-shaped.

These aren’t edge cases. They’re standard operating procedure—because they work.

And they work because the harsh truth is economic: it often costs more to avoid lock‑in than it takes to justify a positive multi-cloud business case. If portability is the goal, you pay for it in overhead created by the abstraction layer, duplicate tooling, operational overhead, constrained architectures, or lowest-common-denominator designs that nobody loves and few teams sustain.

So most organisations do the rational thing. They take the discounts. They adopt the native services. They lean into the ecosystem. And they accept the dependency—because the spreadsheet says it’s responsible and the market is structured to reward exactly that behaviour.

This isn’t poor architecture. It’s the market working as designed.

3) “On‑Prem Is Dead” Is Lazy Thinking. But So Is “Cloud Is a Scam.”

Cloud can dominate a market and still fail to dominate physics, regulators, latency, and predictable unit economics—which is exactly why on‑prem refuses to die.

The data doesn’t support either extreme. Uptime Institute’s Global Data Centre Survey reports 55% of workloads are now off‑premise, implying roughly ~45% remain on-prem/private environments, depending on how respondents classify “off‑prem.”
At the infrastructure-capacity level, Synergy shows on‑prem data centres down to ~37% of global capacity as hyperscale and colocation expand—dramatically lower than historical levels, but still nowhere near “gone.”

So yes: on‑prem is shrinking as a share. But no: it isn’t collapsing as a category.
And the reason is simple: enterprises aren’t sentimental. They keep what works where it must.

Why On‑Prem Persists (Because Reality Has Requirements)

Security & compliance (control beats abstraction when risk is asymmetric)
Private cloud is increasingly treated as a strategic equal to public cloud, driven by cost predictability, GenAI requirements, and trust in security/compliance delivery.
Latency & edge reality (round‑trip time is not a negotiation)
2026 outlooks describe hybrid portfolios as the default: enterprises blend on‑prem, colocation, hyperscale and edge because sensitive and latency-critical workloads don’t tolerate distant regions.
Cost predictability (variable pricing is a feature until it becomes a board problem)
Private cloud is increasingly preferred for AI model training/tuning/inference, where consumption volatility and governance risk make “cloud-first” financially fragile.
Legacy & risk concentration (“lift-and-shift” often just relocates the invoice)
Gartner explicitly flags cloud dissatisfaction from unrealistic expectations and suboptimal implementation—exactly what you get when legacy systems are moved without genuine modernisation.

The part people miss: “On‑prem” in 2026 often means Private Cloud, not “Old Servers”

Modern on‑prem is increasingly private cloud with cloud-like operating models—automation, self-service, consumption-style procurement—not a museum of ageing hardware.
Even Synergy’s capacity analysis reinforces that while on‑prem share is shrinking, absolute enterprise capacity can remain relatively stable because the total pie is growing—hyperscale grows faster, but enterprise footprints don’t evaporate overnight.

So the honest framing isn’t “cloud vs on‑prem.” It’s this: enterprises are becoming hybrid by design—placing workloads where constraints dictate, not where vendor marketing points.

And this is where the narrative turns: if cloud strategy is now platform dependency and infrastructure strategy is constraint management, the real fight isn’t compute.

It’s data—what it’s worth, where it can live, and who gets to touch it.

4) GenAI Arrived: Overpromised, Undermeasured, and Now Non‑Optional

GenAI didn’t arrive—it detonated. And it arrived wrapped in overpromised narratives: chatbots sold as transformation and demos mistaken for strategy.

But hype cooling doesn’t mean retreat. It means reality has arrived: governance, cost, workflow friction—and the uncomfortable question nobody wants to answer: where is the measured productivity?

HBR’s point is blunt: companies rarely realise significant value immediately. Many see an initial productivity dip before gains—a productivity J‑curve—and the only way through it is disciplined organisational experimentation, not vibes-based rollout theatre.
Translation: most organisations are deploying GenAI faster than they can measure it.

And leadership still doesn’t automate.

In HBR’s interview with McKinsey’s Bob Sternfels, he highlights what GenAI isn’t good at: aspiration, judgment, and truly novel thinking. He reinforces the same McKinsey thesis: GenAI can draft, code, summarise and accelerate—but it can’t set aspiration, make tough calls, build trust, or generate truly new ideas.

So yes—we’re in the slope of disillusionment. But no—this doesn’t go away. It becomes baseline. Then it becomes infrastructure. And then agentic patterns push it from “assistive” to “active,” whether governance is ready or not.

Which brings us to the part most technology narratives skip: how the big models were created—and what happens when the fuel runs out.

5) Scaling Laws Built the Big Models. Scarcity Will Decide What Comes Next.

Let’s step back.

Big language models were created by throwing enormous quantities of compute and enormous quantities of data at training runs, guided by empirical scaling laws. Scaling laws made brute force rational: more compute + more data = more capability—until the inputs stop scaling.

That’s why AI has become a data-centre supercycle: 2026 outlooks describe power constraints, new facility designs, and AI-driven demand reshaping infrastructure economics.

The new trinity is unavoidable: hardware, energy, data.

And here’s the quiet constraint that changes the endgame: high-quality training data is finite.

Epoch AI’s analysis (and the associated paper) estimates the effective stock of high-quality public human text and argues that—if trends continue—frontier training will approach fully utilising that stock between 2026 and 2032 (earlier under more aggressive assumptions).
Nature has highlighted the same looming constraint: developers are “picking the Internet clean,” and rights-holder crackdowns tighten access further.

Could we just repeat the data? Only for a while.

Research on data-constrained scaling shows repetition can yield diminishing gains up to a point, but beyond that, the marginal value of additional compute collapses toward zero.

So what happens when “the Kraken” runs out of new data to ingest?

Two things:

Synthetic data becomes unavoidable, and
Advantage shifts away from “who has the biggest model” to who has the best data, rights, feedback loops, and governance—the things competitors cannot download.

And that takes us to the future state.

6) The Future State: Data Battles

This is where the next decade actually lives.

As foundation models diffuse and capability becomes commoditised, the durable advantage is no longer model weights. It’s proprietary data you legally control; governance, lineage and quality at scale; feedback loops that convert operations into learning; and the ability to deploy AI where the data is—across hybrid, edge and cloud. [arxiv.org], [deloitte.com]

Gartner’s cloud trend work warns of growing dissatisfaction driven by unrealistic expectations and uncontrolled costs—signals that the “just move to cloud” era is over.
Meanwhile, cloud-market acceleration is being driven by GenAI workloads, intensifying competition among the top providers and deepening platform dependence. [hyscaler.com], [statista.com] [statista.com], [asianfin.com]

Put plainly: the old battle was infrastructure. The new battle is control of data gravity.

Who controls the data?
Where is it allowed to live?
Who is permitted to train on it?
Who pays to move it?
Who carries liability when it leaks or drifts?
Who can prove provenance when regulators and courts come asking?

In a world where the Big Three already hold about two-thirds of cloud infrastructure spend, the strategic question becomes less about vendor features and more about where your data becomes trapped—economically, legally, and operationally.

That is the real lock‑in of the AI era.

Not compute. Not Kubernetes. Data.

7) The Next Curve After Data Battles: GenAI Commoditisation → Junk

Now the uncomfortable part.

When the marginal cost of generating “content” collapses, the first-order outcome is not a renaissance. It is oversupply—and oversupply destroys signal.

Music is the early warning. Platforms are already dealing with floods of AI-generated tracks, spam, impersonation, and discovery systems that users no longer trust.
Reports describe massive removals of suspected AI spam and deceptive uploads as content farms collide with streaming incentives.

Translate that pattern to enterprise knowledge work, and the hazard is obvious:

more documentation, less reliability
more summaries, less understanding
more “ideas,” fewer original insights
more automation, more noise

And there’s a second-order problem: as the world fills with machine-generated text, models trained on that output risk feeding on their own exhaust—recursive imitation that degrades signal over time. Data scarcity pressures already push us towards synthetic data; the risk is that synthetic becomes the dominant substrate.

The endpoint is not “smarter everything.” It’s the commoditisation of generation and the devaluation of meaning.

In other words: junk.

Not because the models are weak—but because incentives reward volume, not truth; speed, not judgment; output, not understanding.

8) The Real Risk: Not That AI Gets Smarter—That People Get Lazier

We’re still amazed by GenAI capabilities and still not fully sure how to evaluate them properly. HBR’s call for disciplined experimentation exists for a reason: organisational impact is hard to measure, and early productivity dips are common.
And leadership remains stubbornly human: aspiration, judgment and truly novel thinking don’t come from predicting the next likely token.

So the long-run danger is not simply “AI replaces jobs.” It’s that cognitive offloading becomes default, and scarce human skills atrophy—making it harder to generate truly new ideas, harder to reason about causality, and harder to set direction when the map is unclear.

Some people will keep sharpening skills. Most won’t. That’s how tools reshape populations.

Closing Thesis

Cloud adoption hit a plateau in migration-driven productivity because migration became routine.
GenAI will commoditise because capability will diffuse and the data commons will tighten

So the next durable battleground is neither cloud nor models.

It is data—and the right to use it, the ability to govern it, and the power to keep it from becoming trapped inside someone else’s ecosystem.

The winners of the next decade won’t be those who “picked the right cloud.” They’ll be the ones who understood early that the future is a set of Data Battles—and built the discipline to extract real advantage before the market drowns in cheap, automated output.

Because the trajectory is clear:

Data Battles first. GenAI commoditisation second. Junk eventually.

References

Q4 2025 cloud infra revenue ($119B), ~30% YoY growth; Big Three share & GenAI-driven acceleration (Synergy data reported by CRN). CRN — “Global Cloud Market Share Q4 2025; Google Grows, AWS’ Lead Narrows” [statista.com]
Alternate summary of Synergy Q4 2025 shares (AWS ~28%, Azure ~21%, Google ~14%). AsianFin summary of Synergy release [crn.com]
Synergy Research: cloud infrastructure services tracking programme (method/coverage). Synergy Research — Cloud Infrastructure Services [asianfin.com]
Gartner trend: cloud dissatisfaction; realism gap & cost/control issues driving dissatisfaction through 2028. Gartner press release — Top Trends Shaping the Future of Cloud [hyscaler.com]
HBR: productivity J-curve and disciplined organisational experimentation for GenAI adoption. HBR — “A Systematic Approach to Experimenting with Gen AI” [research.a…ltiple.com], [hbr.org]
HBR interview: Sternfels on what GenAI can’t do well (aspiration, judgment, novel thinking). HBR — “We Want to Make Ourselves Better” [gartner.com]
McKinsey: leadership work GenAI can’t do; human skills remain central. McKinsey — “Building leaders in the age of AI” [linkedin.com]
On-prem/off-prem workloads: Uptime Institute Global Data Center Survey 2024 (55% off-prem). Uptime Institute — Global Data Center Survey 2024 (PDF) [goldmansachs.com]
On-prem capacity share: Synergy (on-prem ~37% of capacity in 2024; projected decline). Synergy press release — capacity trends [1624046.fs…nt-na1.net]
Hybrid as default + on-prem footprints shrink but sensitive workloads remain on-site. JLL — 2026 Global Data Center Outlook [deloitte.com]
Private cloud “cloud reset” + GenAI drivers for private cloud preference. Broadcom — Private Cloud Outlook 2025 [marketplac…gement.net]
Data scarcity window: approaching limits of high-quality public text between 2026–2032. Epoch AI blog; arXiv paper [arxiv.org], [forbes.com]
Nature: “AI running out of data” and shrinking data commons / access tightening. Nature news feature [marktechpost.com], [arxiv.org]
Data repetition diminishing returns in constrained regimes. JMLR — Scaling Data‑Constrained Language Models [aclanthology.org], [arxiv.org]
AI music flood / trust erosion examples. TechRadar (Jan 2026); ZME Science (Oct 2025) [thediscourse.com], [audionerdz.net]

GenAI‑Curated Content Disclaimer

This piece was written with the support of generative AI as a drafting and refinement tool. The ideas, structure, reasoning, and conclusions are my own, formed through independent analysis and professional experience. All views expressed are personal and do not reflect those of my employer or any organisation with which I am associated.