Navigate this guide · 1,631 words
The essential signal
- ByteDance is in advanced talks with MediaTek for custom AI data-center chips (ASICs), Bloomberg reported on October 15 — its second custom-silicon arrangement this year after a Qualcomm deal in May.
- US export rules bar ByteDance from NVIDIA's most advanced chips but still allow chip-design partnerships below set performance thresholds — the legal gap the custom-silicon wave is walking through.
- Custom AI ASIC shipments are projected to grow 44.6% in 2026 versus 16.1% for GPUs: as AI moves to inference at scale, purpose-built silicon is winning on economics.
The deal on the table
On October 15, 2026, Bloomberg reported that ByteDance, the Chinese technology company behind TikTok, is in advanced talks to buy custom artificial intelligence chips from Taiwan's MediaTek for its data centers [1][3][4]. The chips would be application-specific integrated circuits — ASICs — silicon engineered around a particular customer's computing workloads rather than sold as general-purpose parts [1][4]. The report, carried by multiple outlets within hours, immediately moved markets and refocused attention on the quietest arms race in technology.[1][2][3][4]
For all the noise, the deal is not done. No contract has been signed, and the volume of chips under discussion has not been disclosed [2][3]. If the talks do succeed, ByteDance would become MediaTek's second major customer for data-center ASICs after Google [2][4]. Neither ByteDance nor MediaTek has commented publicly on the negotiations, and Reuters said it could not immediately verify the report independently [4].[1][2][3][4]
What an ASIC actually is
To understand why this matters, it helps to know what an ASIC is not: it is not a graphics card you can buy off a shelf. A general-purpose GPU like NVIDIA's can run almost any parallel workload — training a chatbot today, rendering a film tomorrow. An ASIC gives up that flexibility. It is designed, down to the transistor layout, to execute a narrow set of tasks as efficiently as physically possible: chips engineered around a particular customer's computing environment [1][4].[1][4]
That trade is the whole point. When an AI model moves from research into production and starts serving requests at scale, the workload stops changing and starts repeating. At that point the flexibility of a GPU becomes overhead — transistors, power, and money spent on capability you never use. Purpose-built silicon strips that overhead out, and for a company planning to spend up to $70 billion this year on AI infrastructure, the savings compound fast [1][4].[1][4]
Why ByteDance can't just buy NVIDIA
The obvious question is why ByteDance is shopping for custom chips at all, when NVIDIA sells the world's most capable AI accelerators off the shelf. The answer is export control. US restrictions on semiconductor exports to China make it difficult for ByteDance, Alibaba, and Tencent to secure NVIDIA's most advanced AI processors [2][4]. The rules are designed to slow China's access to frontier AI compute, and at the top end of the market they are working as intended.[1][2][4]
That leaves ByteDance in an uncomfortable middle. Its AI infrastructure budget — up to $70 billion this year alone — signals compute demand on a historic scale [1][4]. With the straightforward path to top-end GPUs restricted, the company is working every alternative channel: domestic suppliers, which are not fully meeting demand, and custom silicon partnerships structured to stay within the rules [2][4].[1][2][4]
The loophole that makes it legal
Here is the crucial subtlety: the US rules do not ban all chip cooperation with Chinese companies. They set performance thresholds, and partnerships that stay below those thresholds are explicitly permitted. US rules still allow Chinese companies to work directly with chip design firms as long as the performance stays below certain thresholds [2]. Independent trackers confirm the same reading: chip design partnerships below defined performance thresholds give ByteDance and its peers a lawful route past GPU restrictions [4].[2][4]
ByteDance has already walked through this door once. In May 2026, the company agreed to buy several million ASICs for AI data centers from Qualcomm — a deal structured the same way, through the same legal gap [2][4]. The MediaTek talks are best understood as a second pass through a proven playbook: diversify suppliers, stay under the thresholds, and build AI infrastructure the export rules were not written to stop [2][4].[2][4]
ByteDance's $70 billion buildout
The chip talks are one line item in a staggering capital program. Bloomberg reports ByteDance is considering capital expenditure of up to $70 billion this year — more than double last year's total — as it builds out its AI infrastructure [1][4]. Independent tracking puts the trajectory higher still: roughly $100 billion in 2027 if business conditions hold, with the company's own guidance framing 2026 AI infrastructure spending at 160 to 200 billion yuan, roughly half of it potentially going toward processors [4].[1][4][6]
Numbers like these explain the urgency. When you are spending tens of billions on compute, every percentage point of efficiency is worth hundreds of millions — and custom silicon's entire pitch is efficiency. Owning the chip design means tuning hardware to your exact models rather than paying for generality you don't need [1][6].[1][4][6]
MediaTek's second act
For MediaTek, this is a coming-out party. The Taiwanese company is best known for smartphone processors, but it has been quietly widening its push into data-center infrastructure, building on capabilities developed in its mobile chip business [2][3]. Landing ByteDance would give it a stronger foothold in a custom-chip market where Qualcomm and Broadcom are already well established [1].[1][2][3][4]
MediaTek is not starting from zero in custom AI silicon. It is already a major vendor for Google's tensor processing units — the TPUs that power Google's AI cloud [1]. ByteDance would be its second hyperscale ASIC customer, and the validation matters as much as the revenue: winning Google and ByteDance back-to-back would announce MediaTek as a serious new force in custom AI chips [2][4].[1][2][3][4]
The rivals watching nervously
The talks are a competitive signal, and the market heard it. Qualcomm and Broadcom shares were in focus after the Bloomberg report, with retail investor sentiment turning bearish on both — and on NVIDIA — as traders priced in ByteDance shifting AI workloads to custom ASICs and reducing its reliance on GPUs [1]. For Qualcomm the sting is sharper: ByteDance was supposed to be its own custom-silicon showcase after the May deal for millions of chips [1][4].[1][4][5]
Broadcom has the most to defend. It controls an estimated 60% of the custom AI ASIC co-design market, designing Google's TPUs, Meta's MTIA accelerators, and OpenAI's planned ASIC among others [5]. A MediaTek empowered by ByteDance-scale volumes would be a new challenger in a business where Broadcom is, as market coverage puts it, already well established [1][5].[1][4][5]
Everybody's building their own
ByteDance is following a playbook the hyperscalers wrote years ago. Amazon, Google, and Microsoft all concluded that standard server hardware was designed for the average customer, not for running the world's largest platforms — so they built their own silicon, tuned precisely to their workloads, from Graviton to TPUs [1][6]. The lesson wasn't just cost. It was control: once you own the hardware stack, you can optimize more aggressively and build things competitors can't simply buy [6].[1][5][6]
Now the AI labs are running the same play. OpenAI has a planned ASIC in Broadcom's co-design pipeline, and Meta's MTIA accelerators are already in the field [5]. The logic compounds over time: a company that designs its own hardware around its own models is solving a problem its competitors haven't even started on — a head start that money alone can't shortcut [6].[1][5][6]
Why inference changed the math
The timing is not accidental. Industry projections put 2026 custom ASIC shipment growth at 44.6% against 16.1% for GPUs — the first time purpose-built silicon is outgrowing merchant GPUs [5][6]. The driver is inference. Training a model is research: architectures shift constantly, and GPUs offer the flexibility to adapt. But once a model ships and inference volumes scale, the workload becomes predictable — and predictable work is exactly what custom silicon is built to devour [5][6].[5][6]
As models move into production, spending concentrates on serving, and serving rewards chips that do one thing repeatedly without burning through energy and budget [5][6]. General-purpose chips carry overhead that purpose-built chips simply don't need: transistors and power spent on flexibility that production workloads never use. Strip that overhead at data-center scale and the economics flip decisively toward custom silicon [6].[5][6]
The concentrated market underneath
For all the talk of a gold rush, custom AI silicon is currently a concentrated business. Broadcom and Marvell together control roughly 95% of the custom AI ASIC co-design market, with Broadcom alone holding an estimated 60% share [5]. Broadcom co-designs Google's TPUs, Meta's MTIA, and OpenAI's planned ASIC; Marvell handles Amazon's Trainium and Inferentia lines and Microsoft's Maia [5].[1][4][5]
That concentration is precisely what makes the ByteDance talks threatening to incumbents. A MediaTek armed with ByteDance-scale volumes and Google's reference business would be the first credible challenger to the duopoly in years [1][4]. For ByteDance, a second supplier means leverage over pricing and supply; for the industry, it means the custom-silicon club might finally be admitting a new member [1][4].[1][4][5]
What could still go wrong
A necessary dose of skepticism: none of this has happened yet. The talks are advanced but unconfirmed, no contract exists, volumes and pricing are undisclosed, and no target shipment date appears anywhere in the reporting [2][3][4]. Reuters, notably, said it could not independently verify Bloomberg's report — a reminder that single-sourced deal reporting, even from Bloomberg, occasionally dissolves on contact with reality [4].[2][3][4][5][6]
There is also the question of what ByteDance actually gets. ASICs excel at fixed, high-volume workloads — but AI model architectures still evolve quickly, and training remains GPU-dominated precisely because GPUs offer the flexibility to adapt when architectures shift [5]. A chip frozen around today's models can age badly if the field moves, which is why even the most aggressive custom-silicon buyers keep buying GPUs too [5][6].[2][3][4][5][6]
The bottom line
Step back, and the ByteDance–MediaTek talks are a symptom of a structural shift: AI competition is becoming hardware competition. The companies winning in AI are no longer content to rent compute — they are designing the silicon it runs on, because at sufficient scale, owning the hardware is cheaper, faster, and harder for rivals to copy than renting it [1][6].[1][2][4][6]
Whether this particular deal closes matters less than what it confirms. Export controls set out to slow China's AI buildout; instead, they are accelerating a global custom-silicon industry that reduces everyone's dependence on any single supplier — including NVIDIA [2][4]. The GPU era isn't ending. But the era of AI infrastructure as someone else's general-purpose hardware is [1][6].[1][2][4][6]
Questions, answered.
Is the ByteDance–MediaTek deal confirmed?
No. Bloomberg reported advanced talks on October 15, 2026; no contract has been signed, volumes are undisclosed, and neither company has commented. Reuters said it could not independently verify the report [2][3][4].[2][3][4]
Why can't ByteDance just buy NVIDIA's AI chips?
US export restrictions bar ByteDance, Alibaba, and Tencent from buying NVIDIA's most advanced processors. Partnerships with chip designers are still permitted below set performance thresholds [2][4].[2][4]
What is an ASIC, and why does AI want them?
An application-specific integrated circuit is silicon designed for particular workloads. As AI models move into production and inference volumes scale, predictable workloads justify purpose-built chips that beat general-purpose GPUs on efficiency [1][5][6].[1][5][6]
Read the sources.
- QCOM, AVGO, NVDA Stocks In Focus As ByteDance Reportedly Closes In On A MediaTek Pact For AI Data Center ChipsStockTwits · Accessed 2026-10-23
- ByteDance in talks with MediaTek on custom AI chips as Nvidia access tightensDigital Today · Accessed 2026-10-23
- ByteDance nears deal to buy MediaTek's custom AI chips for data centersTechNode · Accessed 2026-10-23
- ByteDance in advanced talks to buy MediaTek data-center ASICsAI Weekly · Accessed 2026-10-23
- Custom AI ASIC MarketMarkets and Markets · Accessed 2026-10-23
- The Companies Winning AI Are Starting to Build Their Own ChipsMemeBurn · Accessed 2026-10-23
Contributor articles and source links do not guarantee error-free reporting. See our editorial policy.
