MediaTek Confirms a Second Cloud ASIC Customer for 2028 and Says Data Center Is Now Accretive to Operating Margin
Q3 2026 earnings call, October 30, 2026: Q4 guidance implies revenue growth of 40% to 48% year-over-year as the first AI accelerator ramps
A second hyperscale-class customer is confirmed, with revenue starting in 2028
The most consequential new disclosure came in the Q&A. Responding to UBS, CFO David Ku confirmed that MediaTek has a second cloud ASIC customer, distinct from the lead customer behind its first two accelerator programs. The project has already kicked off, with first revenue expected in 2028. Ku said, "We do have a second customer," and, addressing market speculation, added that the customer "is not coming from China." Management declined to say whether this is an AI lab or a hyperscaler, whether it is tied to the NVLink Fusion partnership, or how large the program might be. Ku said it is "too early to talk the detail" and that 2028 supply is still being finalized.
The disclosure matters because the main bear argument on MediaTek's ASIC story has been concentration in a single customer and a single program lineage. A second customer does not resolve that, since there is no revenue or margin framework yet, but it shows the first program was not a one-off. Morgan Stanley pressed on whether the second customer's first-year revenue could exceed this year's roughly $2 billion of ASIC sales. Management would not commit, saying only that customer demand "looks good" but that supply is the other half of the equation.
First accelerator ramp: $2.5 billion in Q4, and the high end of the $16 billion range for 2027
Initial shipments of the first AI accelerator began in September. CEO Rick Tsai said MediaTek is "confident of delivering USD 2.5 billion revenue in the fourth quarter," up from prior guidance of slightly more than $2 billion for the year. For 2027, the company now expects this program to reach the high end of its approximately $16 billion target range, which Tsai put at around 20% of the total AI accelerator ASIC market. Asked by Bank of America what drove the upside, Tsai attributed it to robust customer demand and to the supply chain team's success in lining up suppliers, including substrates, which he called "usually one of the most difficult things to do." Neither executive broke out unit versus price contribution.
Group revenue guidance reflects the scale of this ramp. Q4 revenue is guided to TWD 210.1 billion to TWD 221.8 billion, up 25% to 32% sequentially, at an exchange rate of TWD 31.7 to USD 1. Management said the underlying growth would have been stronger still without seasonal softness in consumer electronics. Full-year 2026 USD revenue growth is now expected to exceed the earlier target of high single digits. On the implied math, the accelerator would account for roughly 30% of Q4 revenue.
Margins: gross margin dilutive, operating margin accretive, and 20%-plus operating margin for 2027
The mix shift toward data center is changing the margin structure in the way management has said it would. Q4 gross margin is guided to 45% plus or minus 1.5 percentage points, down from 46.3% in Q3. Ku said data center gross margin "is still below the corporate average," and that this explains the one-point step down, alongside price increases passed through to reflect supply chain cost inflation. The operating expense ratio is guided to 26% plus or minus 2 points, which implies a Q4 operating margin of around 19%, versus 16.1% in Q3.
Ku confirmed that data center operating margin turns accretive to the company average starting in Q4, though he declined to quantify the premium. Morgan Stanley's back-of-envelope estimate was a roughly 10-point premium, and Ku would say only that it is higher. He described Q4 as "the starting point" for improving operating margin and guided to 2027 operating margin of "20% plus, if not slightly higher." He tied this to the model management laid out about 1.5 years ago, in which gross margin is slightly dilutive but operating margin is the payoff. Investors should note that the gross margin guide is lower and that the improvement is entirely an operating leverage story, which depends on volume and on the supply chain delivering.
Second-generation accelerator: tapeout noise addressed, packaging yield is the real risk
JPMorgan raised market chatter about design and tapeout delays on the second-generation accelerator, which is targeted for high-volume production in Q1 2028. Tsai said the program is "on track, on schedule." He said a couple of the dies on the multi-die package have already taped out, and that the largest and most complex die will follow on schedule, taking "a couple or 3 weeks longer than the other dies" because of its size. He declined to say how many dies are on the chip. Management described the chip as delivering multiple times higher performance per watt and significantly lower total cost of ownership. Ku was more combative on competitive positioning, saying that competing projects "are actually not even close from the overall performance perspective."
The more candid discussion concerned manufacturing. The second-generation chip relies on a U.S. IDM packaging process, widely understood to be Intel's, and Tsai called yield improvement there the company's "absolute priority." MediaTek has built an internal manufacturing team led by experienced retired operating executives. It is working with the IDM, Japanese substrate suppliers and Taiwanese suppliers. Tsai, who has a manufacturing background, said he has seen "clear identification of the key difficulties" and has "very high confidence that the yield will go up very quickly" once two or three key problems are fixed. That is confidence rather than data, and no yield figures were offered. Ku said that on current yield profiles MediaTek feels comfortable delivering what the customer wants for Gen 2, but the customer is seeking upside volume and MediaTek is still working with suppliers to provide it.
Deutsche Bank asked whether Intel could force MediaTek into its front-end foundry as a condition of packaging. Tsai rejected the premise of dependence, stating "we are not dependent on one supplier" and calling TSMC "our most important partner." He framed packaging as "a major redefining technology for this AI era" in which MediaTek is investing money, people and ecosystem relationships. He did not directly address the leverage scenario.
On 2028, Ku declined to provide a number, saying there are "so many moving pieces on the supply chain situation," but said to expect "another year of strong growth" over 2027.
Strategic capital: $3.9 billion convertible with NVIDIA and Alphabet participation
During the quarter MediaTek completed a $3.9 billion overseas convertible bond offering with participation from NVIDIA and Alphabet. Tsai said it strengthens the relationships with key AI infrastructure partners and enhances the company's ability to secure capacity for critical components. In effect, the proceeds help secure scarce supply such as substrates and advanced packaging.
Deutsche Bank asked whether MediaTek should engage with loss-making frontier model companies when a larger competitor is offering aggressive vendor financing. Ku said that if the market requires it and the customer is the right one, "we will consider as well." He pointed to improving revenue return on CapEx over the last six quarters and to at least one frontier model company preparing an IPO. This is a notable shift in tone. MediaTek is signaling that it may put its balance sheet behind customers, which increases both the addressable opportunity and the credit exposure.
NVLink Fusion and the edge portfolio
Tsai described NVLink Fusion as a business model with strong technical content, designed to cut customers' time to market and "time to money." MediaTek and NVIDIA build a chassis into which customers can more easily fit their own XPU or CPU designs and then sell the resulting rack. Ku said the companies are in discussions with several potential customers, that some "like it," and that he hopes an official customer announcement will come "in the near term." No customer has been named. MediaTek also supports UALink and Ethernet-based rack architectures. Asked about the 448G SerDes, Tsai said it has taped out, is undergoing measurement and verification, and will not be on the critical path.
On the barrier to entry, Tsai was measured. He cited 448G SerDes, packaging, HBM5 and low power, but argued the real moat is the ability to integrate these into manufacturable, scalable products. He added that MediaTek does not expect to be far ahead of strong competitors: "We always have a very tough fight against our competitors." He offered little on the third-generation program, which is likely to ramp in late 2029, saying the definition is still being finalized with a customer that is "very aggressive" on performance and TCO.
On the edge side, the RTX Spark AI PC using the N1X chip, which pairs NVIDIA GPUs with MediaTek SoCs, became commercially available in mid-October, and a second-generation SoC is already in development. The flagship Dimensity 9500 was featured in Samsung's Galaxy Tab S12 series.
Third-quarter results and segment mix
Q3 revenue of TWD 168.1 billion rose 10.4% sequentially and 18.3% year-over-year, above guidance. Gross margin was 46.3%, operating margin 16.1% and EPS TWD 17.50 (non-TIFRS TWD 17.92). Operating expenses rose to TWD 50.7 billion from TWD 47.4 billion in Q2.
The segment mix shows how much growth is coming from outside smartphones. Smart Edge grew 20% sequentially and 63% year-over-year to 58% of revenue, driven by volume production of a 5G modem for a new smartphone customer, AI tablets, AI PCs and automotive cockpit, plus higher DRAM content in TV SoCs. Mobile phone revenue fell 1% sequentially and 17% year-over-year to 37% of revenue, though flagship outperformed, including the first 2-nanometer Dimensity 9600 Pro. Power IC revenue rose 1% sequentially and 18% year-over-year to 5% of revenue on data center share gains.
Demand durability: management acknowledges market jitters
Morgan Stanley asked about investor concerns over neocloud financing and AI infrastructure deceleration. Tsai acknowledged "quite a bit of a market movement" in recent weeks but said AI compute demand will continue to grow over a one- to three-year horizon, though "the rate of which may vary somewhat." He argued that in a tougher market the most efficient chips win, and said MediaTek will "come out better and stronger." This is a view rather than a forecast, and it rests on the assumption that MediaTek's ASICs are among the winners in any shift in hyperscaler budgets.