Supermicro posts $11.1 billion quarter as AI systems reach 60 percent of revenue
Server makers turning GPUs into racks are riding the same demand wave as chipmakers, with liquid cooling now a headline capacity metric.
Server maker Supermicro reported revenue of $11.12 billion for its fiscal fourth quarter ended June 2026, up 93.2% year on year, according to an analysis published by The Next Platform on 13 August. Operating income reached $1.49 billion, a 6.6-fold increase, and net income was $1.18 billion, roughly six times the prior-year figure and equal to 10.6% of revenue.
For the full fiscal year, revenue came to $39.06 billion, a 77.8% rise, with net income of $2.23 billion. Guidance for fiscal 2027 has a midpoint of $68.5 billion, which would be another increase of about 78%.
AI systems generated $6.69 billion in the quarter, up 64.2%, and now account for 60% of Supermicro's revenue, most of it conventional GPU-accelerated machines, with the rest split between agentic AI sandboxing and edge systems and traditional servers and storage. Customer concentration is high: a single customer, which the article identifies as xAI's Colossus 2 build, represented 28% of fiscal 2026 revenue, and nine customers spending over $1 billion each made up about half of the total.
On capacity, Supermicro operates nearly 4 million square feet of design and factory space in the United States plus facilities in Taiwan, Malaysia and the Netherlands, and can currently produce 3,000 direct liquid cooled racks a month, at up to 240 kilowatts per rack, alongside 3,000 air-cooled racks. The article contrasts this with Dell, which it projects will book around $60 billion in AI system sales in fiscal 2027 within roughly $120 billion of data centre revenue.
Why it matters
The money flowing into AI is not stopping at chip vendors; the companies that turn GPUs into finished, cooled racks are growing almost as fast. Rack output in kilowatts is becoming the capacity metric that matters, pulling thermal engineering into the centre of the server business. Reliance on a few large customers means revenue moves with a handful of build decisions. The Dell comparison suggests room for several large integrators.
The money flowing into AI is not stopping at chip vendors; the companies that turn GPUs into finished, cooled racks are growing almost as fast.
- Q4 FY2026 revenue growth, year on year+93.2%
- AI systems revenue growth, year on year+64.2%
- Full-year FY2026 revenue growth+77.8%
- FY2027 guidance midpoint versus FY2026~+78%
Figures: Figures reported by The Next Platform from Supermicro's results and guidance
What you can learn from this
- Rack-scale integration. A modern AI server is no longer a box but a rack: dozens of GPUs, the CPUs that host them, high-speed interconnect switches, power shelves and cooling manifolds delivered as one tested unit. Integrators buy chips, boards and switches from suppliers, then design the chassis, power delivery and cooling that hold them together. Customers pay for this because validation at rack level catches interconnect and thermal problems that individual servers hide. The margin sits in engineering and testing, not in the silicon.
- Direct liquid cooling and the 240 kW rack. Air can only carry away so much heat per rack before fan power and airflow become impractical, typically a few tens of kilowatts. Direct liquid cooling pipes coolant through cold plates bolted onto GPUs and CPUs, then exchanges the heat to facility water. Because liquid has far higher heat capacity than air, a single rack can draw well over 100 kW. The facility water then goes to a cooling tower or chiller, so the building's plumbing becomes part of the compute design. That is why rack output figures now come with a kilowatt rating and why liquid-cooled capacity is reported separately.
- Customer concentration risk. When one buyer supplies a quarter or more of revenue, the supplier's forecasts depend on that buyer's build schedule, financing and even site permitting. A delayed data centre or a change of hardware vendor shows up immediately as a revenue miss, and a dominant buyer also weakens the supplier's pricing power in negotiations. Companies mitigate this by diversifying across customers and product lines, but AI infrastructure demand is inherently lumpy because it arrives in campus-sized orders. Reading the customer mix is as important as reading the headline growth.
- Fiscal years and guidance. Companies choose their own fiscal calendar, so a year ending in June means the fourth quarter covers April through June. Guidance is management's forward estimate, usually given as a range, and the midpoint is what analysts plug into models. A guidance midpoint that implies 78% growth is a statement about backlog and factory capacity as much as demand. Comparing guidance to actual output capacity, in racks per month, is a useful sanity check.
- Reading the product mix. A split between GPU machines, edge and sandbox systems, and traditional servers tells you which parts of a business are cyclical and which are steady. GPU systems carry the growth but also the concentration; traditional servers have thinner margins but broader customer bases. Segment disclosures like these let outsiders estimate exposure to a single technology wave. The habit of decomposing a revenue number is transferable to any company you evaluate.
We teach this
How to use this in practice
- Size the power and heat of one rack. Take a public specification sheet for a current GPU server, note its rated power, and multiply by the number of units that fit in a 42U rack. Convert the total to kilowatts, then estimate the airflow a room would need to remove it using any published rule of thumb for cubic feet per minute per kilowatt. Compare the result with the 240 kW liquid-cooled figure in the article. Done means a short calculation showing why air cooling stops being viable somewhere on the way to that number.
- Draw the rack as a system. Sketch a rack with compute nodes, interconnect switch, power distribution and cooling manifold, and draw the flow of electricity in and heat out. Label the point where heat moves from chip to liquid to facility water, and mark which components an integrator designs versus buys in. Done means someone unfamiliar with data centres can follow the diagram from wall socket to cooling tower and see where the integrator's work sits.
- Decompose a revenue figure. Using the numbers in this article, build a one-sheet model: total revenue, AI systems share, the sub-segments as percentage ranges, and the single largest customer, with the source reference in a cell beside each input. Add a row that recomputes revenue if that customer halves its spending, and another where the guidance midpoint is missed by ten percent. Done means a sensitivity table you could rebuild for any hardware company from its filings.
- Compare two integrators on the same basis. Take the article's Supermicro and Dell figures and put them in a two-column table: AI system revenue, total data centre revenue, AI share. Note which numbers are actuals and which are projections, and which fiscal calendar each company uses, then align both to the same calendar period before comparing growth rates. Done means a table where every cell carries a source and a date, and no projection is mixed in with a reported result.
Sources
- The GenAI Boom Will Lift Supermicro, But It Will Lift Others, Too — The Next Platform
Our reporting is an original summary; full coverage is at the links above.
Don't just read about it — build it.
Square 1 teaches the skills behind the headlines, with every line of your work graded by AI. Find your starting point in 3 minutes.
Get your free skill report