Home โ€บ Blog โ€บ The hardware vendors
AI Factory as a Service ยท Part 4 ยท The hardware vendors

The Arms Dealers: How Dell, HPE, Lenovo and Supermicro Sell AI Factories

NVIDIA designs the AI factory; the server makers build it, ship it, install it and increasingly run it. Dell alone has a $95 billion backlog of AI server orders. Under the "AI factory" label sit two very different businesses: a few enormous rack-scale orders from neoclouds and AI labs, and thousands of smaller packaged systems for enterprises. This part follows the money, the margins and one company's legal trouble.

FL
FrontierAGI Team
โš ๏ธ How this was researched (as of 26 Oct 2026) Dell figures come from its SEC earnings filings and results coverage; HPE's from its Q3 FY2026 results and coverage by Constellation Research and others; Supermicro's from earnings coverage and DCD, Fortune and court reporting on the export case; Lenovo's from press reports. Each company reports different metrics (orders, backlog, pipeline), so they are not directly comparable. Customer counts are company claims. This is not investment advice. See Reading the numbers.

1. Case files

Four vendors

Tap a vendor. Each reports a different headline metric.

2. Timeline

The server makers' AI boom, 2024 โ†’ 2026

One lane per vendor. Solid dots are confirmed by company filings or results, hollow dots come from the press, dashed dots are planned. Click a dot for its source.

Click a milestone to see where it comes from.

3. Two businesses under one label

Dell says more than 6,500 customers buy its AI Factory products, and it also took a record $60.9 billion of AI server orders in a single quarter. Those two facts describe different businesses. A handful of buyers, neoclouds such as those in Case Study 7, AI labs and sovereign projects, place orders worth billions of dollars for complete racks. Thousands of enterprises buy a few racks or a packaged AI factory with software and services. Supermicro disclosed about $39 billion of orders from more than 20 customers, a sign of how concentrated the big orders are โš ๏ธ.

Few big buyers, many small ones

An illustration of the shape of the market, not company data. Each dot is a customer: horizontal position is order size (log scale), vertical is how many customers place orders that size.

On the left, thousands of enterprises with orders of a few hundred thousand to a few million dollars, where vendors earn more on software and services. On the right, a few dozen rack-scale buyers with orders of hundreds of millions to billions, where volumes are vast and margins thin.

4. Dell's backlog

Dell's AI server backlog and orders

Billions of dollars, from Dell's results. Toggle the measure.

Dell shipped more than $25 billion of AI servers in fiscal 2026 and entered fiscal 2027 with a $43 billion backlog. That rose to a record $51.3 billion after the first quarter and $95 billion after the second, when it booked $60.9 billion of orders and $16.4 billion of AI server revenue. In September 2026 it raised its fiscal 2027 AI server revenue forecast from about $60 billion to $74 billion, and its total revenue outlook to $192 billion.

5. What's in the box

Build a vendor bundle

Tick what you buy from the server maker. The tool shows who supplies the rest. Simplified: real bundles vary by vendor and deal.

Data Center Knowledge describes the vendors shifting "from silicon to services": racks arrive pre-integrated and liquid-cooled, the vendor installs and tests them, and increasingly offers to run them under a managed contract. HPE's Private Cloud AI, launched in 2024, packages NVIDIA hardware and software as a turnkey private cloud; Dell's AI Factory, also launched in 2024, combines servers, storage, networking and services; Supermicro sells "Data Center Building Block Solutions" for whole halls.

6. Margins

The boom has a cost. Deloitte notes that AI is sending hardware revenue surging while pressure on profitability rises: GPUs are expensive pass-through components, big buyers negotiate hard, and liquid-cooled rack integration takes capital and labour. Supermicro's reported gross margin of 17.6% for fiscal 2026 โš ๏ธ shows how thin rack-scale margins can be. Services, software and smaller enterprise deals are where vendors hope to earn more.

Revenue up, margins down

An illustration of the trade-off as the share of rack-scale AI orders in a vendor's mix rises. Not any one company's data.

7. Supermicro

Supermicro became one of the fastest suppliers of liquid-cooled GPU racks to neoclouds and AI labs. It reported fiscal 2026 revenue of about $39.1 billion, up 78%, and guided $65โ€“72 billion for fiscal 2027 โš ๏ธ (single source). In March 2026, US prosecutors charged co-founder Yih-Shyan "Wally" Liaw and two others over an alleged scheme to divert about $2.5 billion of servers containing NVIDIA GPUs to China, in violation of export controls. Liaw has pleaded not guilty. Supermicro says it is not a defendant or a target of the grand jury investigation, and has launched a board-led independent investigation. Futurum asked whether the case could shift market share to Dell and HPE.

โš–๏ธ Export controls reach the server makersCase Study 8 showed how China sources NVIDIA hardware despite export controls. This case is a reminder that much of that hardware moves inside servers built by third parties, which puts the server makers under scrutiny too. The allegations are unproven.

8. Compared

DellHPESupermicroLenovo
Headline AI metric$95B AI backlog (Q2 FY27)$7.6B AI backlog, $6.8B AI systems (Q3 FY26)~$39.1B FY26 revenue โš ๏ธ; ~$39B orders from 20+ customers โš ๏ธAI server pipeline $15.5B or $21B (reports differ) โš ๏ธ
Enterprise packageDell AI Factory (2024)Private Cloud AI (2024)DCBBS building blocksHybrid AI factory offerings
Customer claim6,500+ AI Factory customersPrivate Cloud AI orders up triple digitsโ€”โ€”
Big deal$60.9B orders in one quarter$3.5B inference deal, unnamed hyperscaler โš ๏ธNeocloud and lab racksโ€”
WatchConcentration in a few buyersSmaller AI scaleExport case, marginsLimited disclosure
$95BDell AI backlog, Q2 FY27
6,500+Dell AI Factory customers (claim)
Thin marginsGPUs are pass-through cost
Export riskSupermicro co-founder charged
VerdictThe server makers are the factory floor of the AI boom. Their order books are enormous but concentrated in a few rack-scale buyers with thin margins; their enterprise "AI factory" packages reach thousands of customers and carry more services. Dell has pulled ahead on scale, HPE is smaller but growing, Supermicro is fast but under a legal cloud, and Lenovo discloses least. Part 5 turns to the buyers themselves: who these thousands of customers are and what they run.

9. Reading the numbers

Dell (filings)
FY26: >$25B AI servers shipped, $43B backlog entering FY27; Q1 FY27 backlog $51.3B; Q2 FY27: $60.9B AI orders, $16.4B AI server revenue, $95B backlog, $46.97B total revenue; FY27 AI server guidance ~$60B โ†’ $74B; 6,500+ AI customers (company claim).
HPE
Q3 FY26: record $7.6B AI backlog, AI systems backlog $6.8B (+14%), AI systems orders $2.4B; earlier ~$5B (Q1) and $5.9B (entering Q3). Private Cloud AI orders up triple digits; $3.5B hyperscaler inference deal (as reported โš ๏ธ).
Supermicro
Q3 FY26 sales $10.2B, >80% AI GPU platforms; FY26 $39.1B, gross margin 17.6%, FY27 guidance $65โ€“72B (single source โš ๏ธ); ~$39B orders from 20+ customers (reported โš ๏ธ). Export case: charges in March 2026; not-guilty plea; company not a defendant.
Lenovo
AI server pipeline reported as $15.5B and $21B by different sources โš ๏ธ.
Illustrations
The scatter, bundle builder and margin slider are simplified.