Hardware Is King Again

There is a pattern in technology economics that survives every generation of jargon: value accrues to whatever is scarce. Everything else, the licensing models, the delivery models, the acronyms ending in -aaS – is downstream of that single fact. Over 30 years the scarce resource has migrated four times, and each migration redrew the industry’s map, redistributed its profits, and rewrote what a technology budget looks like.

The migration has now happened again, in the opposite direction to the one most organisations spent a decade preparing for. It is worth walking the arc, because the through-line is the point.

The Machine Was The Product

In the era of the mainframe and the minicomputer, computing power was the constraint. Machines were capital assets, depreciated over years, and software largely arrived as an accessory to the metal — written for the box, priced within the box, and inseparable from it. The commercial logic was simple: you sold the scarce thing, and the scarce thing was machine cycles.

The formal beginning of the end came in 1969, when IBM unbundled software and services from hardware pricing – a decision made under antitrust pressure that created, almost incidentally, the possibility of an independent software industry. It took two more decades to fully express itself, but the direction was set.

Software Takes The Margin

By the 1990s the constraint had moved. Standardised architectures and a competitive components market made computing power abundant and cheap; assembling a box became a logistics and working-capital discipline rather than an engineering moat, and hardware margins compressed accordingly. What remained scarce was functionality — and functionality had an economic property hardware could never match: once written, it copies for nothing.

That asymmetry defined the decade. Software companies enjoyed near-zero marginal cost and pricing power tied to value rather than materials, while hardware vendors fought over points of gross margin. The enterprise budget shifted with it: licences, maintenance contracts, and the annual true-up became the line items that mattered.

The Scarce Thing Becomes Making It Work

By the turn of the century, core enterprise software had matured. Rival products converged on similar functionality; the differentiator moved from what the software could do to whether an organisation could make it deliver. Large-scale implementations – enterprise resource planning above all – routinely cost multiples of the licences that triggered them, and failed often enough that implementation risk became a board-level concern.

Scarcity had migrated to integration skill. That is the era in which IBM reoriented decisively toward services, and in which India’s technology majors built global businesses on the same insight. It is also, not coincidentally, the era in which the corporate data centre lost its glamour. Its strategic status faded to that of a cost centre to be optimised, consolidated and – through virtualisation – squeezed. The ambition was fewer racks, not more.

The Cloud, And Software’s Return In A New Wrapper

Then the constraint moved again — to elasticity. What organisations could not easily buy was the ability to scale capacity up and down without owning the underlying iron, and cloud infrastructure sold precisely that. The financial consequence was the defining CFO conversation of the 2010s: capital expenditure converted into operating expenditure, ownership traded for consumption.

Software returned to primacy, but on entirely different terms. Perpetual licences gave way to subscription; the vendor’s incentive shifted from closing a sale to preventing churn; and the -as-a-service suffix spread from applications to platforms to almost everything. Recurring revenue became the metric on which technology companies were valued.

One nuance in the popular telling deserves correcting, because it matters for reading the present. Data centres did not die in this period – ownership of them migrated. Enterprises exited the business of running their own halls while a small number of operators industrialised the same function at unprecedented scale. The facility stopped being a strategic asset on your balance sheet and became a strategic asset on someone else’s.

The Physical World Reasserts Itself

Then came AI at scale, and with it a constraint no software abstraction can dissolve: the availability of compute, and of the electricity and silicon that produce it. The scale of the response is without precedent in corporate history. The largest US cloud and AI infrastructure providers have collectively guided to somewhere in the region of $600–725 billion of capital expenditure in 2026 – trackers differ on the exact aggregate, but all place it roughly 60 to 77 per cent above 2025’s own record of about $390–410 billion. Goldman Sachs has projected multi-trillion-dollar cumulative capex for the largest four through the end of the decade.

More telling than the totals is where the bottleneck now sits. Microsoft has disclosed an Azure order backlog in the region of $80 billion that it cannot fulfil, with GPUs waiting on power rather than the reverse. That is the sentence that best captures the new era: the binding constraint has become electricity, land, transformers and cooling – inputs measured in years of permitting, not sprints of engineering.

The repricing has run right down the component stack. Memory, the archetypal commodity, has become a strategic asset: contract DRAM prices rose by roughly 50 to 60 percent quarter-on-quarter through the first half of 2026 on TrendForce’s survey, moderating to a still-substantial 13 to 18 percent for server DRAM in the third quarter as manufacturers reallocated capacity toward high-bandwidth memory for AI servers. Micron has posted record revenue. IDC expects average PC selling prices to rise by as much as 8 per cent this year. And in the detail that matters most for procurement teams, TrendForce notes that cloud providers holding multi-year long-term agreements are insulated from much of the increase — meaning the rises fall disproportionately on buyers without them.

Hardware, in other words, is not merely relevant again. For the first time in a generation, it is appreciating in real terms while you own it.

The Rule Beneath The Eras

Read in sequence, the arc is not a series of fashions. It is one mechanism repeating: scarcity moves, and value follows. Machine cycles were scarce, then functionality, then integration skill, then elasticity, and now energy and silicon. Each era’s winners were the firms positioned at the constraint; each era ended when the previous constraint became abundant — usually because the industry had spent a decade attacking it.

Two properties of the current constraint make it different from its predecessors, and both are consequential.

First, it is physical, and therefore slow. Software scarcity is relieved by writing more software; compute scarcity is relieved by building substations, securing turbines, laying transmission and waiting for planning approval. You cannot compile a power station.

Second, it is capital-intensive at a scale that reconcentrates the industry. Elasticity was purchasable in small increments; frontier compute is not. That pushes bargaining power back toward whoever can commit capital years in advance — which is precisely what the long-term agreement data shows.

What This Means For The People Who Buy Technology

The practical implications are ordinary, immediate, and mostly about contracts and planning horizons.

The OPEX settlement is partly reversing. A decade of financial orthodoxy held that owning infrastructure was a strategic error. AI workloads are pulling in the other direction — toward reserved capacity, committed spend, long-term supply agreements and, for some organisations, owned or colocated hardware. This is not a return to 1998; it is a recognition that when a resource is scarce, the buyer who commits early pays less than the buyer who arrives spot.

Procurement leverage now sits in the contract, not the negotiation. The single clearest lesson of the memory market is that price outcomes are being determined by who holds multi-year agreements. Organisations planning significant refreshes should be securing terms across cycles, not quarters.

Refresh economics have inverted. Planning models built on the assumption that equivalent hardware gets cheaper each year are, for now, wrong. Deferring a refresh in the hope of a better price is a materially different bet than it was three years ago.

Energy strategy has become IT strategy. Where compute physically sits is now a question about power availability, tariff structure and grid reliability. This is why the storage and generation stories – including the African Development Bank’s financing of the continent’s first integrated battery gigafactory in Morocco – belong on the same page as the cloud stories.

Location is a legal question as well as a physical one. The arrival of sovereign cloud regions, including Oracle’s forthcoming Nairobi region hosted at iXAfrica, reflects a market in which jurisdiction over data has become a purchasable specification. That, too, is a scarcity: the right to keep regulated workloads at home.

Architect for the swing, not the moment. The consistent error across all five eras is optimising an organisation for where value sits today. Model-agnostic architectures, portable data, and contracts that permit exit are the practical expression of humility about which layer will be scarce in five years.

The Honest Uncertainty

Anyone claiming to know how long the current configuration holds is guessing, and the recent record punishes that confidence. When DeepSeek released a reasoning model in January 2025 claiming a training cost of a few million dollars, roughly $590 billion was wiped from Nvidia’s market capitalisation in a single session – a reminder that algorithmic efficiency can reprice a hardware boom overnight.

So the useful posture is not prediction but instrumentation. Three markers will tell the story faster than any forecast: whether the enormous capex converts into commensurate AI revenue, which the providers’ own disclosures will show; whether power, rather than chips, remains the reported constraint into 2027, as the backlog data currently suggests; and whether efficiency gains in models and silicon begin to outrun demand growth, which would move scarcity again — most likely back up the stack, toward whoever can turn abundant compute into business outcomes.

Because that is the pattern’s final implication. If compute does become abundant, the scarce resource will be what it has always been at such moments: the capability to apply it. Every previous swing ended that way, and the organisations that came through each one best were not the ones that guessed the next era correctly. They were the ones that had not built themselves so tightly around the last one that they could not move.