For most of the AI buildout, the binding constraint on a hyperscaler’s roadmap has been how many Nvidia GPUs it could get. Increasingly, the more interesting question is how many it can avoid buying.
MediaTek — the Taiwanese fabless chipmaker best known for the Arm-based silicon inside mid-range Android phones, Chromebooks and Wi-Fi gear — is pushing hard into that gap. DIGITIMES reported Tuesday that the company is widening its custom ASIC services business as hyperscalers and large technology firms shop for workload-specific accelerators that cut their dependence on general-purpose AI processors. It is a strategic pivot with real money behind it, and it has begun showing up in MediaTek’s numbers.
On the company’s second-quarter earnings call on July 31, CEO Rick Tsai told analysts that MediaTek expects data center revenue to exceed $2 billion in 2026 — roughly double the $1 billion it guided in February — with its first AI accelerator ASIC entering production in the fourth quarter of this year, built in partnership with what Tsai described only as “a major US CSP customer.” He also raised the 2027 target: MediaTek now pegs the serviceable market for AI accelerator silicon at $80 billion next year and is chasing 15% to 20% of it, up from the 10% to 15% guided a quarter earlier.
“Recently, the release of increasingly capable frontier AI models, together with the rapid transition toward agentic AI, has become a key driver of compute demand across both cloud and edge AI,” Tsai said. “Agentic AI, which executes a series of actions including planning, reasoning, execution, and self-correction, further increases workloads. This represents a compelling growth opportunity for MediaTek.”
The core business is paying for the bet
The urgency behind the pivot is not hard to read in the financials. MediaTek’s Q2 revenue was NT$152.2 billion (about $4.68 billion), up just 1.2% year over year. Operating income fell 22.2% to NT$22.87 billion. Mobile phone revenue — still 41% of the company — dropped 14% sequentially and 20% year over year as rising bill-of-materials costs choked handset demand, and MediaTek expects global smartphone units to decline about 15% this year. Smart Edge Platforms, up 26% year over year and now 53% of revenue, carried the quarter.
Against that backdrop, MediaTek’s board approved a discretionary $5 billion financing budget the same day, explicitly to secure supply chain capacity and fund an expansion “from AI ASIC chips to full-scale systems and platforms.” That last phrase is the whole strategy: MediaTek is pitching pre-validated memory, I/O and connectivity subsystems, a 448G SerDes with co-packaged copper, and end-to-end 3.5D design flows spanning both TSMC CoWoS and Intel’s EMIB-T packaging.
The customer nearly everyone assumes is Google. Counterpoint Research projects that MediaTek will capture roughly 26% of global AI ASIC server compute shipments by 2028, approaching 5 million units — more than tenfold growth from the roughly 400,000 units expected in 2026 — on the back of Google’s TPU v8t “Zebrafish” and its successor, the v8e “Humufish.” That would make MediaTek the second-largest player in custom AI silicon, behind only Broadcom.
Why it matters
The MediaTek story is less about one design win than about the unbundling of the custom-silicon business model, and the reason is memory economics.
“Under the Broadcom turnkey model, the ASIC vendor handles HBM sourcing and applies a 15%-20% markup,” Counterpoint research associate David Wu said. “As HBM represents an increasingly large share of the total ASIC bill of materials, this markup becomes a high cost at scale, particularly as Google’s TPU deployment velocity continues to accelerate. By bringing chip design and HBM procurement in-house, starting with Zebrafish, Google eliminates the intermediary fee and lowers its overall cost.”
In that disaggregated arrangement, Google designs the compute die itself and MediaTek supplies the I/O die — a narrower, lower-margin role than Broadcom’s soup-to-nuts engagement, but one that scales with volume and gives the buyer leverage over its incumbent. Amazon has run a version of the same playbook by splitting Trainium work between Marvell and Alchip, and Microsoft’s Maia and Meta’s MTIA programs create similar openings that Counterpoint’s forecast does not yet assume. Bloomberg Intelligence projects the AI ASIC market will compound at 27% a year to $118 billion by 2033.
None of which makes MediaTek’s targets credible on their own. Broadcom guided to roughly $16 billion in AI semiconductor revenue for its fiscal Q3 2026 alone, and Morgan Stanley analyst Joseph Moore estimates the company will book about $120 billion in AI revenue in fiscal 2027, some $80 billion of it TPU-related. Moore expects Broadcom to hold roughly 80% of Google’s TPU business over time, calling bear-case forecasts of a 50% split or outright displacement “premature.” MediaTek participation, the firm wrote, “is real, but not disruptive.”
Gartner is blunter. “If I look at 2024 market share for MediaTek for AI accelerators in compute electronics, it was less than 1%,” VP analyst Gaurav Gupta said. “Based on recent studies, our projection is still less than 1% for MediaTek. The company has gained some traction recently in custom accelerators, like with Google. But aiming for 15-20% seems very high at this point.”
What to watch
Three things will settle it. First, the Q4 ramp: MediaTek has promised more than $2 billion of data center revenue this year from a standing start, and the first quarter of shipments will show whether yields and volumes hold. Second, packaging. Counterpoint senior analyst Ashwath Rao flags that the second-generation Humufish design is still in qualification with Intel’s EMIB-T, with mass production not targeted until late 2027 — unproven technology at Google-scale volume, and the biggest execution risk in the plan.
Third, and most consequential: whether MediaTek converts a single anchor customer into a portfolio. Tsai says the company is deepening engagements with several prospective data center customers. Until one of them becomes a named design win, MediaTek’s AI business is a bet on Google’s appetite for supplier optionality — a good bet, but a narrow one.
“The company has gained some traction recently in custom accelerators, like with Google. But, aiming for 15-20 percent seems very high at this point.”— Gaurav Gupta, VP Analyst, Gartner