Michael Dell, CEO of Dell Technologies, has said the memory shortage will become more severe in 2027. In an interview published October 16 by The Wall Street Journal (WSJ), he explained that building a semiconductor fab takes about four years, so supply cannot keep pace with growing AI demand. Dell is sitting on a $95 billion backlog of AI server orders, while Micron, a memory supplier, is expanding production but cannot yet say when supply and demand will come back into balance. Turning that backlog into revenue requires securing the right types of memory at the right time. Dell's production model, which keeps inventory to a minimum, and the long-term contracts memory makers use to lock in supply years ahead are both entering a new phase as AI demand surges.
A $95 billion backlog and $16.4 billion in quarterly revenue
According to Dell's earnings call for the second quarter of fiscal 2027, released September 1, AI-related orders totaled $60.9 billion, AI server revenue reached $16.4 billion, and the AI backlog at quarter-end stood at $95 billion. Investment in AI infrastructure is expanding rapidly, but a large time lag has opened up between collecting orders and actually delivering products and recognizing the revenue.
Dividing the $95 billion backlog by this quarter's $16.4 billion in AI server revenue gives roughly 5.8 quarters' worth of sales. However, this is merely a comparison of the scale of the backlog against revenue; it does not mean it will take about 5.8 quarters to deliver every order. The actual pace of revenue recognition will vary with component availability and product mix, and new orders will keep arriving in the meantime. Nor can we conclude that the entire $95 billion backlog is waiting for delivery because of the memory shortage.
Changes in when customers place orders are also contributing to the growth. On the same call, Chief Operating Officer Jeff Clarke explained that customers trying to secure future supply are now ordering against demand further out than before. Some large enterprises want to receive products earlier than scheduled if supply allows. Public-sector organizations and small businesses with fixed budgets, on the other hand, have to reduce the quantities they buy as prices rise. In other words, the backlog reflects not only growth in AI demand itself but also customers pulling procurement forward out of concern about future shortages. Deliveries are being delayed by supply constraints while customers move future orders earlier, so the backlog alone cannot accurately capture the relationship between real demand and delivery timing.
Why an HBM shortage affects entire AI servers
In an interview with the WSJ, asked when memory supply constraints might ease, Michael Dell said, "Based on what we can see right now, I think it gets worse in 2027." This is his own forecast and does not mean that supply of semiconductors in general will deteriorate across the board. The problem is whether the specific types of memory that AI servers need can be secured in sufficient quantities as demand grows.
In AI servers, HBM (high-bandwidth memory), mounted close to the GPU, and DDR-family memory, used as main memory on the CPU side, play different roles. HBM is also a type of DRAM, but loading up on ordinary server DDR memory cannot simply substitute for it. The two differ in connection method, memory bandwidth, and packaging, and systems are designed around the characteristics of each. To complete an AI server, therefore, a manufacturer must secure not only GPUs but also the HBM and DDR memory that match the design, in the required quantities.
Micron's white paper on AI agents and HBM describes a test environment combining an NVIDIA Blackwell GPU with 192GB of HBM3e and 2TB of DDR5 system memory. It is a concrete example of high-speed memory on the GPU side and large-capacity memory on the host side serving different purposes. The company also confirmed experimentally that as the number of requests processed simultaneously in AI inference rises, the capacity of the KV cache, which holds information mid-processing, becomes a performance constraint. Results may not be the same for every AI workload, but the findings show that GPU compute performance alone does not determine overall system throughput; memory capacity and bandwidth matter as well.
Michael Dell also explained that delivering an AI system to a customer requires combining GPUs and memory with liquid cooling equipment, networking, storage, and more, and then getting it running at the customer's site. Even if high-performance GPUs can be procured, a shortage of the necessary memory means a finished AI server cannot be delivered. When demand is expanding this quickly, a shortage of even one component can delay delivery of the entire system and constrain a manufacturer's revenue growth.
Micron to expand from 2027, but no balance in sight
Memory makers are racing to expand supply capacity, but whether they can keep up with demand growth remains uncertain. On its September 30 earnings call, Micron said the memory and storage market in calendar 2027 and 2028 will be tighter than in 2026. It also published updated schedules for its capacity expansion projects in the United States and Singapore. Expansion begins in 2027, but no date was given for when supply will catch up with demand and the market return to balance.
According to pages 2–3 of Micron's earnings materials published September 30, the start dates for its main production sites are as follows. Each site handles different manufacturing processes, and start times differ by several years.
| Site / project | Planned initial production | Process |
|---|---|---|
| HBM advanced packaging facility in Singapore | Early 2027 | HBM advanced packaging |
| ID1 fab in Idaho, US | Mid-2027 | Wafer production |
| ID2 fab in Idaho, US | Late 2028 | Wafer production |
| First fab in New York, US | 2030 | Wafer production |
These are all plans announced by Micron, and there is no guarantee that a given volume of finished product will ship from the stated dates. Moreover, advanced packaging, which completes HBM, and wafer fabrication, which produces the base for DRAM, are different processes, so their capacities cannot simply be added together. Micron explains that after initial production begins at new DRAM and NAND fabs, it takes several quarters before supply capacity expands in earnest, with output ramping up in stages.
This does not mean memory output will see no increase through 2029 or 2030. Production at new facilities starts in 2027, and supply is expected to grow gradually. Even so, the shortage is considered hard to resolve because demand, driven mainly by AI, keeps rising too, and new fabs cannot deliver maximum capacity immediately upon starting up. Progress on expansion plans and a balanced market overall need to be considered separately.
On page 5 of the same materials, Micron says that even including its planned cleanroom additions for DRAM, it cannot foresee when supply and demand will return to balance given the strong demand. Michael Dell also pointed out that the spread of AI agents is changing how the technology is used, and memory demand is growing faster than supply capacity. What matters going forward is not only when new fabs begin construction. The key to resolving the shortage will be how quickly the volume of memory products that can actually ship increases, and whether that supply growth outpaces the rise in demand.
Dell's build-to-order model protects capital efficiency, but delivery challenges remain
As memory prices climb and shortages persist, Dell is trying to maintain its business model centered on build-to-order production. In the WSJ interview, Michael Dell said the company continues, as far as possible, to assemble products in response to customer demand and keep inventory to a minimum, and that its cash conversion cycle remains negative. The more component prices rise, the heavier the burden of cash tied up in inventory that does not lead to sales. Dell's approach of producing what is needed in response to orders has the advantage of limiting that burden.
The cash conversion cycle is a metric calculated by adding the days inventory is held to the days it takes to collect receivables, then subtracting the days until payables are due. A negative value means that, on average, the company collects revenue from customers before paying suppliers for components. A negative cash conversion cycle does not, however, mean Dell receives full payment before manufacturing every AI server. It is simply an indicator of working capital efficiency across the business as a whole.
Meanwhile, strong inventory management and cash collection do not make it possible to secure memory that is in short supply. Avoiding losses from excess inventory and procuring necessary parts to meet delivery deadlines are separate challenges. If rising prices for memory and other components cannot be adequately passed on to selling prices, profit margins could also come under pressure. Dell's build-to-order model is effective at limiting inventory risk from demand swings, but it does not unconditionally guarantee product delivery or profitability.
Against this backdrop, contracts are spreading among memory makers on the supply side that require customers to take on long-term purchase obligations and financial commitments in exchange for securing future volumes. On its September 30 earnings call, Micron disclosed that it has signed 26 multiyear contracts involving purchase and payment obligations. Customer financial commitments total $32 billion, most of it in the form of cash deposits. For Micron, this makes it easier to expand capacity with visibility into future demand; for customers, it makes it easier to secure the memory they need.
However, it has not been disclosed that these contracts are with Dell. Even so, the fact that securing long-term memory supply comes with deposits and purchase obligations shows how AI demand growth is changing the semiconductor procurement environment. Buyers are no longer just purchasing the required quantities when needed; they are being asked to forecast demand years ahead and decide how much volume they can take on purchase responsibility for. Micron also says it has already signed contracts for most of its calendar 2027 HBM supply, with prices expected to rise substantially from the previous year.
To turn its $95 billion AI server backlog into sustainable growth, Dell must not only deliver products on schedule and recognize the revenue but also protect profit while absorbing rising component costs. Customers, for their part, need to determine when they can receive products with the memory configurations they want, and what financial burdens or purchase obligations are involved in reserving supply. The focus going forward is whether the memory expansion planned for 2027 translates into actual supply growth and how far it can meet robust AI demand. What will shape the growth of AI server makers such as Dell is not only expanding orders but whether they can secure the necessary components and move AI infrastructure from order to actual operation.