The number that matters more than the record

Wistron, the Taiwan-based contract electronics maker that assembles AI servers for the world's biggest cloud providers, closed its second quarter of 2026 with revenue of NT$895.4 billion, net profit of NT$14.8 billion and earnings per share of NT$4.72 — all-time highs on every measure 3. First-half EPS reached NT$7.78, up roughly 94% from a year earlier, and revenue for the first seven months of 2026 hit NT$2.05 trillion, up 88.2% year over year 4. AI servers now account for more than 95% of Wistron's total server revenue, up from roughly 70% at the end of 2025 3 6.

The headline growth is not the interesting part. Gross margin for the quarter came in at 5.66%, up 0.45 percentage points from the first quarter and 1.22 points from a year earlier — an improvement, but still a margin any industrial company outside contract electronics would consider unworkable 6. That is the real story sitting underneath the record: Asia's AI-server assemblers are capturing an enormous and growing share of global tech capex spending, but they are capturing it at single-digit margins, while the memory and packaging suppliers further up the chain keep pricing power. Wistron's quarter is the clearest recent data point on how that split is evolving.

New logos, same concentration problem

Until this year, Wistron's AI server book leaned heavily on a small number of hyperscale cloud customers. That is starting to change. Digitimes reported this week that Wistron has landed new AI server business from Hewlett Packard Enterprise and Lenovo, plus a cloud-services order widely attributed to Oracle, with management guiding that the new accounts will lift fourth-quarter 2026 results specifically rather than simply pulling forward existing demand 1 2. July revenue alone was NT$308.2 billion, down slightly from June but still up 60.8% from a year earlier, and management told investors AI server shipments should keep growing sequentially through the second half with no gap between the outgoing VR platform and the incoming B2 platform in the third quarter 1 5.

That diversification matters for a structural reason, not just a growth one. A supplier whose AI server revenue sits on two or three hyperscaler relationships is one contract renegotiation away from a very bad quarter. Adding HPE, Lenovo and a large cloud customer spreads that risk across enterprise OEM channels as well as cloud-native ones, and it signals that AI infrastructure demand has broadened past the handful of US hyperscalers that dominated the first two years of the buildout. For a stock that trades on the durability of AI capex spending as much as its current size, a broader customer base is arguably worth more to the multiple than the extra revenue itself.

Why the margin will not fix itself

Management's own explanation for the still-thin margin is rising material costs, chiefly memory, colliding with higher unit prices on next-generation AI racks 6. That is a direct pass-through of the DRAM and HBM price spike that has been reshaping the chip supply chain all year — the same dynamic lifting SK Hynix's and Samsung's earnings is squeezing the assemblers that have to buy that memory to build finished racks. Asked how sensitive AI server demand is to further memory inflation, Wistron executives told investors the demand is effectively inelastic: cloud providers will keep buying racks even at higher component cost, so the impact on Wistron's overall business should stay limited even if the specific gross margin line stays compressed 6.

That is a defensible bet, but it is a bet on volume compounding forever, not on margin expansion. Wistron is responding by leaning further into scale rather than waiting for pricing relief: the board approved a combined NT$10.5 billion in new capital expenditure at its Hsinchu and Kaohsiung plants, plus roughly $93 million in fresh capital for its two US subsidiaries and its Vietnam site, and nearly NT$4 billion to build an AI computing center with National Yang Ming Chiao Tung University in Tainan 4. None of that spending is aimed at improving unit economics; all of it is aimed at making sure Wistron can physically build more racks as orders keep arriving.

The read for investors

Wistron's quarter is a useful proxy for the whole Taiwan ODM tier — Quanta, Foxconn and Inventec are running the same playbook of chasing volume at compressed margins while capacity remains the binding constraint on the industry, not customer demand. The bull case is that absolute profit dollars keep scaling as revenue compounds at 60-90% year over year, the way Foxconn's smartphone-assembly margins were also thin but generated enormous cash at scale. The bear case is that these companies have effectively no pricing leverage against either their customers, who can multi-source assembly, or their suppliers, who control the scarce memory and packaging capacity everyone needs. New customer wins like HPE, Lenovo and Oracle reduce concentration risk and extend the growth runway into 2027, but they do not by themselves change where the margin sits in the value chain. Until memory supply loosens or assemblers gain more pricing power on next-generation platforms, the most important number in Wistron's results will keep being the one under 6%, not the one hitting a record.