The question this piece answers is simple: who is paying for the buildout, and when do they get their money back? Between August and November 2026, Dell, Supermicro and HPE each reported a quarter in which AI infrastructure demand exceeded their ability to ship it. All three said the same sentence in different words — supply, not demand, is the constraint. What differed was what happened to cash, and the differences are large enough to change how you read every "AI server backlog" headline you will see over the next year. Every number below is taken from the company's own press release, its SEC filing, or its earnings call, and each is cross-checked against at least one independent source; where the three companies' metrics are not comparable, the piece says so instead of averaging them.
What did the three builders actually report?
Start with the raw quarter, because the interpretation only works if the numbers are on the table first. The three fiscal calendars do not align — Dell's and HPE's quarters both ended 31 July 2026, Supermicro's ended 30 September 2026 — so treat this as a season, not a single period.
| Metric | Dell (Q2 FY27) | Supermicro (Q1 FY27) | HPE (Q3 FY26) |
|---|---|---|---|
| Revenue | $46.97B, +58% YoY | $19.946B, vs $5.018B a year earlier | $12.2B, +34% YoY |
| AI-specific revenue | AI-optimized servers $16.40B, +100% | Not broken out by product | AI systems ~$1.6B; Cloud & AI segment $9.0B |
| Gross margin | 20.9% (from 18.3%) | 12.7% GAAP (from 17.5% QoQ, 9.3% YoY) | 40.4% non-GAAP (+1,050 bp YoY) |
| Segment profitability | ISG operating income $4.78B = 15.0% of ISG revenue (from 8.8%) | Operating income $2.064B = 10.3% | Cloud & AI 17.0% OM (from 7.0%); Networking 22.0% OM |
| Orders / backlog signal | $60.9B AI orders; $95B AI backlog; >6,500 AI customers | Q2 guide $19.5–21.5B; FY27 $72–82B | AI orders $2.4B, +30% QoQ; backlog +14% QoQ to a record |
| Cash | FCF $986M, −47% YoY; OCF $2.225B, −13% | Operating cash flow −$434M; capex $47M | OCF $1.6B; FCF $958M |
One caution before the analysis: those three margin numbers are not the same metric. Supermicro's 12.7% and Dell's 20.9% are GAAP company-wide gross margins; HPE's 40.4% is non-GAAP and includes a networking business that grew 74.9% after the Juniper acquisition. Anyone who ranks these three on "margin" without noticing that HPE sells switches at a 22% operating margin and Supermicro mostly sells racks is comparing a restaurant to a grocery store. That difference is not noise. It is the story.
Why did margins go up while cash went down?
Because in a shortage, the margin is rent on allocation. When HBM, DDR5 and GPUs are rationed, a builder with supply commitments can choose which orders to accept and at what price — and every one of the three said so, in their own vocabulary. Dell booked a record $60.9 billion of AI orders in a quarter where it could only recognise $16.4 billion of revenue. HPE's CFO attributed its record gross margin to "disciplined traditional-server pricing and a richer networking mix," and its CEO said supply availability, not demand, is what limits order conversion. Supermicro's CEO described a transformation toward a "Total AI and IT Datacenter Solution provider" — which is the same claim in a different direction: the way out of a 12.7% margin is to sell the software, services and networking wrapped around the box.
Now the cash. Dell's revenue grew 58% and its free cash flow fell 47% to $986 million; for the fiscal half, revenue rose 71% while free cash flow moved by $8 million — $4.104 billion against $4.096 billion. That is not a rounding artefact, it is the business model becoming visible: the growth is sitting in inventory and receivables, and on Dell's balance sheet the financing receivables line went from $14.28 billion to $20.43 billion in two quarters as Dell Financial Services originations ran at about $2.8 billion a quarter against $1.6 billion a year earlier. Supermicro's version is starker because it has less non-AI business to hide behind: net income of $1.594 billion and operating cash flow of negative $434 million, with accounts receivable up $3.56 billion and inventories up $2.48 billion in the quarter.
Then the line that makes the model work: Supermicro's deferred revenue rose $3.05 billion in the same quarter. Deferred revenue is money customers paid before the company shipped. In plain terms, a business earning 12.7% gross margin carried $15.05 billion of inventory and $9.69 billion of receivables — 1.24× a full quarter's revenue — partly on its customers' prepayments. That is not a scandal; prepayment is exactly what a buyer should demand from a supplier in a shortage. It is a structural fact worth naming, because it means the balance sheet has a second customer, and that customer can stop paying forward (ours).
Where does the margin actually come from?
Not from the box. Supermicro's quarter is the cleanest read on what a GPU rack is worth as a product: revenue grew 79% quarter over quarter, and gross margin fell 480 basis points, from 17.5% to 12.7%. Run the counterfactual (ours): at the prior quarter's 17.5%, $19.946 billion of revenue would have produced $3.49 billion of gross profit. It produced $2.54 billion. That is roughly $950 million of gross profit given up to win the volume — a company buying share with price at exactly the moment its order book was fullest. Year over year the margin did rise, from 9.3% to 12.7%, which tells you the 17.5% was itself an unusual quarter, not a baseline. The honest reading is that a rack assembled from other companies' parts clears somewhere in the low-to-mid teens, and the swings around that number are mix and pricing power, not capability.
The same arithmetic explains HPE's 40.4%. Strip out networking — $2.9 billion of revenue at a 22.0% operating margin, up 74.9% with Juniper inside — and the remaining Cloud & AI segment earned 17.0% operating margin on $9.0 billion, of which AI systems were only about $1.6 billion. HPE's own CFO told analysts gross margin "should moderate toward historical levels as AI systems become a larger revenue mix." That sentence is the most useful disclosure in the whole season: the highest margin in the group is being earned by the part of the business that is not AI servers, and management expects the mix to dilute it.
And above all three sits the supplier. NVIDIA's fiscal Q2 2027 (ended 26 July 2026) reported $96.221 billion of revenue at a 75.0% gross margin, with data center at $89.0 billion, up 117%. Put the two numbers side by side and the shape of the industry is obvious: NVIDIA's data-center revenue alone — $89.0 billion — is larger than Dell's entire revenue ($46.97 billion) plus Supermicro's ($19.95 billion) plus HPE's Cloud & AI segment ($9.0 billion), which is $75.9 billion combined, or $78.8 billion once HPE's $2.9 billion networking segment is added. All four quarters ended within ten weeks of each other. The builders are not the profit pool; they are the pipe the profit pool flows through.
Why is the backlog a short position on memory?
Because a backlog is a price agreed in the past and a cost paid in the future. Dell's $95 billion AI backlog is 5.8 quarters of its current AI revenue rate and 1.28× its own fiscal-2027 AI revenue guidance — the order book now extends past the planning horizon in a way that no hardware backlog normally does. Every one of those orders carries a price. The bill of materials behind them does not. And the bill of materials is the part that has been moving: all three companies named memory as the binding constraint in the same season, with Dell's management calling out DRAM and NAND as the most severe shortages alongside CPUs, drives, optics and substrates, and Supermicro's revenue jumping from a $14.5–15.5 billion guide to $19.9 billion after a mid-quarter update.
This is why the margin expansion and the backlog should be read as opposite signals. A gross margin earned on shipped revenue reflects today's prices for today's parts. A backlog reflects yesterday's prices for tomorrow's parts. If memory prices keep rising, backlog margin compresses at delivery and the compression is invisible until it happens; if memory prices fall, the backlog becomes more valuable and the current margin was the peak. Dell's CFO said the quiet part in the call — that the quarter benefited from several favourable factors coming together and that not all of them should be expected to persist. HPE's CFO said the mix version of the same thing. Two companies, two vocabularies, one warning.
There is a second exposure that gets less attention: a backlog is a delivery obligation with a penalty structure nobody publishes. Both Dell and HPE described multi-year supply agreements and record purchase commitments as the way they secured allocation. Purchase commitments are the mirror image of a backlog — a promise to buy at agreed terms, valuable when demand holds and expensive when it does not. When a customer pushes a delivery date because the data center shell is late — and shells are late, as our modular-construction piece documented — the builder holds the parts and the commitment, not the revenue (ours).
What does this mean if you are the one buying compute?
Three questions, and they are the same three whether you are buying a rack, a colocation contract or a cloud reservation. First: who carries the working capital, and for how long? If your supplier is financing inventory on its own balance sheet at 12.7% gross margin, it will eventually price that carry into the next order — the question is whether it happens before or after your contract renews. Second: is your price locked or your delivery date? A price locked without a delivery date is a short position on the parts; a delivery date locked without a price is a floating-rate bet. Get both in writing or accept that you are trading one risk for the other. Third: what is the supplier's exposure to the memory curve? A builder with a long backlog and a short memory position has an incentive to ship late and an incentive to renegotiate; both show up as your project slipping, not as a line item you can dispute.
And for anyone modelling a compute business on "hardware margin plus utilisation," the quarter contains a blunt warning. Supermicro earned $1.594 billion of net income and consumed $434 million of operating cash in the same three months. A model that treats net income as available cash would have called that quarter healthy; the balance sheet says it was a quarter in which the company bought its own growth. The same test applies to every GPU cloud, neocloud and colocation operator you are evaluating: compare operating cash flow to net income before you believe the growth rate. Our utilisation piece made the same point from the demand side; this is the supply side of it.
What are the caveats?
Fiscal calendars differ. Dell's and HPE's quarters ended 31 July 2026; Supermicro's ended 30 September 2026; NVIDIA's ended 26 July 2026. The comparisons are structural, not same-period. Metric bases differ. Dell's 20.9% and Supermicro's 12.7% are GAAP company-wide gross margins; HPE's 40.4% is non-GAAP and includes a networking segment enlarged by the Juniper acquisition; segment operating margins are not gross margins. Deferred revenue is not a red flag. Customer prepayment is normal and rational in a shortage, and Supermicro's deferred revenue increase is disclosed, not hidden — the point is that it exists and can reverse, not that it is improper. Some detail is second-hand. Supermicro's press release gives revenue, margin, earnings and operating cash flow; the receivables, inventory and deferred-revenue movements come from its cash-flow statement as reported and analysed by two independent outlets, which agree to the dollar. And nothing here is investment advice — it is a reading of three public filings and two earnings calls, offered because the same arithmetic decides what a rack, a megawatt or an hour of compute should cost you.
Comments