Bernstein: ASIC Market Could Reach $300 Billion by 2028, MediaTek Expected to Capture a Quarter of the Share
nashnova research
Bernstein raised MediaTek's target price sharply to NT$5,500, arguing the custom AI chip market is far larger than expected — up to $300 billion by 2028 — and MediaTek's Google TPU work positions it to capture roughly a quarter of that.
How big is the custom AI chip market, really?
Bernstein lifted its 2028 datacenter ASIC/XPU — chips designed for specific AI tasks, as opposed to general-purpose GPUs — forecast from $140–160 billion to $250–300 billion, nearly doubling the prior estimate.
Waypoints: $50–70 billion in 2026, $130–150 billion in 2027 — a steep ramp.
This means → Bernstein believes the Street had drastically underpriced how much silicon AI inference will consume. This is not a tweak; it is a repricing of the entire segment.
Where does MediaTek's revenue come from, and how fast does it grow?
MediaTek ASIC revenue forecast: $2.4 billion in 2026 → $16 billion in 2027 → $44 billion in 2028 — roughly a 20× increase in three years.
ASIC's share of total revenue jumps from 11% in 2026 to 65% in 2028; its EPS contribution rises from 15% to 68%.
In plain terms = MediaTek is transforming from a smartphone-chip company into one whose main business is custom AI silicon — and the window for that shift is just three years.
Can supply keep up — where are the bottlenecks?
The $16 billion 2027 forecast assumes capacity is secured: MediaTek can reallocate wafers from other product lines, but the real chokepoint is TSMC's CoWoS advanced packaging — a process that stacks multiple chips and high-bandwidth memory together. Bernstein says Google's help has locked enough capacity.
The 2028 uplift is driven mainly by a second major project, TPU v9, expected to start volume production in late 2027 using EMIB-T packaging.
Substrate supply is another constraint — Ibiden, Shinko, and Unimicron have been mobilized, but yield progress varies. This means → across the chain — wafers, HBM, packaging, substrates — whichever link is slowest sets the delivery pace.
How far above consensus are the earnings forecasts — and why cut the valuation multiple?
Bernstein projects MediaTek 2027 EPS at NT$169.35 and 2028 at NT$369.45, roughly 22% and 38% above consensus, respectively.
Gross margin is expected to dip from 45.9% in 2026 to 42.9% in 2028, yet operating margin rises from 16.6% to 30.1%. In plain terms = each dollar earned carries slightly less gross profit, but revenue growth far outpaces cost growth — so total profit rises. That is operating leverage at work.
Yet the valuation multiple was cut from 22× to 18×. This reflects the risk that Google's 2029 "v10" project may bring in more suppliers — Broadcom, Alchip, GUC, and possibly AMD and Marvell — diluting MediaTek's incumbency advantage.
Why can the ASIC market get this big — what is the underlying logic?
Bernstein traces the expansion to Token economics in the AI-inference era: once a model's architecture stabilizes and call volumes are massive, ASICs can be deeply optimized for low-precision math, attention mechanisms, and KV Cache — a mechanism that stores previously computed results during inference to avoid redundant calculation — lifting Tokens-per-watt and Tokens-per-dollar.
Google's eighth-gen TPU splits into training-focused 8t and inference-focused 8i. The 8i uses larger on-chip SRAM, 288 GB HBM, and a dedicated collective-communication engine to optimize inference; Google claims up to 80% better inference cost-performance versus the prior generation.
This means → GPUs and ASICs are not substitutes but complementary, heterogeneous compute — frontier model training relies on GPU programmability, while high-volume inference increasingly favors purpose-built silicon. This underpins the "AI semiconductor super-cycle" thesis.
Can these forecasts actually be delivered — what is the key checkpoint?
MediaTek's 2026–2027 upside hinges on TPU v8t; 2028 hinges on TPU v9.
The core verification point is singular: whether volume production ships on node, on schedule. This reflects a simple reality — no matter how impressive the revenue projection, it must clear the capacity-and-yield gate.
Bernstein maintains an "Outperform" rating, target price raised from NT$4,380 to NT$5,500. The implied confidence: MediaTek's scale, IP portfolio, cost discipline, and project track record form a sufficient defensive moat.
市场有风险，内容仅供研究参考，不构成投资建议。