Ask anyone who has priced a server in the last six months. Quotes that used to hold for thirty days are now valid for a week, sometimes forty-eight hours. Configurations get re-quoted between the proposal and the purchase order. This is not normal market noise. Something structural has changed.
The “something” is a memory. DRAM contract prices rose 90 to 95 per cent in the first quarter of this year, the steepest quarterly increase TrendForce has ever recorded, and another 58 to 63 per cent in the second. NAND flash is now accelerating past it. Dell’s chief operating officer has pointed out that DRAM which cost around 43 US cents a gigabit in mid-2025, recently crossed $2.39. Memory is now roughly a third of the bill of materials on a new machine; a year ago it was a fifth. Every OEM has re-priced. Some have re-priced twice.
It is worth being precise about why, because it changes what a sensible buyer does next. This is not a pandemic-style shock that corrects when factories reopen. High-bandwidth memory for AI accelerators consumes roughly three wafers of capacity for every one of conventional DDR5, and the hyperscalers building AI data centres have placed open-ended orders at whatever price the market asks. SK Hynix declared its 2026 output essentially sold out before 2026 had even begun. What remains is what the rest of us bid for. Analysts have started calling the premium the AI tax, and the name is fair. Every enterprise in India buying a server, a workstation or a laptop today is part-funding someone else’s AI build-out. And new fabrication capacity does not arrive in volume before late 2027, so this is the market for at least the next eighteen months.
Here is the uncomfortable part for the “just move it to the cloud” reflex: the cloud is made of the same memory. The hyperscalers are the buyers driving the shortage, and memory is swallowing an unprecedented share of their own data-centre spend. Their costs are rising too. Whether and when they pass those costs through is the only open question. The cloud remains brilliant at what it was built for, which is elasticity. Spiky demand, experiments, capacity needed by tomorrow morning. Nothing else comes close. But a great deal of enterprise computing is not spiky. It is steady, predictable, and runs around the clock, and paying an elasticity premium for steady work made weak sense before the AI tax. It makes less sense now.
Which is why I keep returning to hybrid. Not hybrid as a compromise between two camps, or a transition phase on the way to something purer. Hybrid as a permanent architecture, decided workload by workload. The mature enterprise of the next decade will not have a cloud strategy or an on-premises strategy. It will have a workload strategy.
But there is a second axis to hybrid that receives far less attention, and in this market, I would argue it matters more. Call it hybrid ownership. The question is no longer just “cloud or on premises?” It is “for this workload, should we buy new, rent, refurbish, or extend what we already have?”
Run that question through today’s prices. A five-year-old server you already own has its memory bought at 2021 rates; keeping it alive with the right spares and support for two more years is, in effect, arbitrage against the shortage. A certified refurbished machine carries yesterday’s component costs into today’s budget. Renting converts a capital shock into a predictable monthly line and hands the residual-value risk to someone else. None of these were exotic decisions before. At current DRAM prices, they are simply what a disciplined CFO would ask for.
I should declare the obvious interest here. Comprint has spent more than three decades across exactly these businesses, procurement through rentals, maintenance, refurbishment and recycling, so of course I would say this. But the arithmetic does not care who states it. Total cost of ownership, measured across the whole life of an asset, purchase price, maintenance, downtime risk, financing, upgrade cycles and eventual retirement, is becoming the real unit of infrastructure decision-making. The sharper organisations we work with no longer begin with “what should we buy?” They begin with “what does this workload need, and what will it cost over its whole life?”
Regulation is adding its own weight. India’s data protection rules were notified last November and the substantive obligations land in May 2027. The law does not force blanket localisation, but the categories of data that significant fiduciaries must keep onshore are still to be notified, and sectoral regulators, the RBI most prominently, already require it for payment data. Enterprises are therefore designing 2027 architectures under 2026 uncertainty. Uncertainty argues for control: knowing exactly where your data lives, on infrastructure you can point to.
So I expect three shifts to define the next few years. Workload-level economics will become standard discipline; IT budgets will be interrogated line by line, not approved as a single number. Refresh-by-default will fade; lifecycle extension, certified refurbishment and structured asset retirement will become normal boardroom answers rather than exceptions that need defending. And compliance will move from afterthought to architecture; where data lives will be decided as much by regulation as by engineering.
For India’s mid-market, the thousands of companies that never make technology headlines but run the country’s everyday economy, none of this is a strategy-deck concept. It is survival math: which workloads justify cloud spend at today’s rates, which servers justify replacement at today’s prices, and which can serve reliably for two more years with proper support behind them.
After more than three decades in this business, one lesson has held through every cycle: infrastructure decisions age well when they are made on economics and accountability, not fashion. The memory market has just made that lesson expensive to ignore.
Kunal Sancheti is a Director at Comprint Tech Solutions, a Mumbai-headquartered enterprise IT company that has managed the complete technology lifecycle for more than three decades: procurement, rentals, maintenance, refurbishment and certified recycling. Views are personal.