Kai ChenNvidia plans to put $3.5 billion into convertible bonds issued by MediaTek, which expects about $2 billion in AI-chip revenue this year. The investment exceeds the partner’s forecast annual AI-chip business—and points to what Nvidia values as inference escapes the data-center rack.
Key takeaways
- AI inference shifts competition from peak accelerator speed toward deployable systems that balance compute, memory, networking, power, software, latency and placement.
- MediaTek gives Nvidia access to high-volume OEM channels and integration expertise spanning phones, PCs, vehicles and data centers—markets Nvidia cannot reach with GPUs alone.
- Nvidia is moving its moat up the stack: NVLink, software and infrastructure can preserve its role even when customers or partners supply their own CPUs, XPUs or ASICs.
- Custom silicon gives large workload owners leverage over Nvidia for stable, high-volume inference, while general-purpose GPUs remain valuable for uncertain or variable workloads.
- Nvidia’s planned investment signals that long-term ecosystem dependence and interoperability may be more durable than profits driven by temporary GPU scarcity.
Inference makes placement part of the computation
During model training, buyers concentrated spending on scarce, general-purpose accelerators. Large, centralized workloads made cluster utilization the dominant buying decision. Inference buyers must also optimize cost per useful task, latency, memory capacity, power, privacy and physical location.
A chip can complete a kernel quickly and still lose at the system level because data movement, memory, networking, software overhead or power limits consume the gain. Improving one component exposes the next bottleneck. Inference turns balanced-system design from an engineering preference into a commercial constraint.
AI builders already act on that constraint. Perplexity split tasks between local and cloud models, keeping private data on-device and improving token efficiency. Deployment architecture is becoming part of the product rather than plumbing beneath it.
At that scale, inference redraws the AI-chip market. A workload that can run in a cloud rack, on a PC, inside a phone or in a vehicle does not create one accelerator market. It creates a routing problem across compute locations with different limits and owners. A useful system places enough compute, memory and software in the right location at an acceptable cost.
MediaTek sells the joints between the parts
MediaTek matters because it has long combined compute, connectivity, power management, software support and customer delivery in SoCs that ship through high-volume device channels. Inference carries that integration skill from phones into PCs, vehicles and data centers.
MediaTek practiced this model before the current AI cycle. In 2020, Amazon and MediaTek built the AZ1 Neural Edge processor for on-device speech recognition. The product joined specialized compute with the surrounding functions needed to deploy it.
In vehicles, MediaTek agreed to integrate Nvidia GPU chiplets and software into automotive infotainment SoCs. RTX Spark broadens the relationship through an Arm-based consumer chip family that combines Nvidia GPU technology, an Arm CPU and as much as 128GB of unified memory for local AI workloads.
MediaTek has also started using Intel Foundry’s advanced packaging alongside TSMC. By coordinating CPUs, accelerators, memory and connectivity from multiple sources, MediaTek can deliver a system rather than a collection of components.
Nvidia plans to invest $3.5 billion in MediaTek convertible bonds as MediaTek forecasts about $2 billion in AI-chip revenue this year. The scale follows the industrial logic: Nvidia gains a route into device categories, OEM relationships and custom systems where MediaTek already knows how to make heterogeneous components behave like one product.
Nvidia’s moat has to move one layer up
Nvidia already has a way to benefit when partners design some of the silicon: it can supply the software, interconnect and infrastructure that coordinate those parts.
With NVLink Fusion, MediaTek can bring customer-designed XPUs into NVLink-connected rack-scale AI factories. The broader collaboration spans cloud infrastructure, local computing and automotive systems. MediaTek supplies domain-specific integration; Nvidia supplies the coordination layer.
An OEM can pair its own CPU or XPU with Nvidia GPUs instead of buying a fully Nvidia-designed stack. Nvidia may surrender a component sale while retaining demand for its software and interconnect.
MediaTek illustrates the bargain. Its Taiwan R&D data center uses Nvidia B200 chips and DGX SuperPOD infrastructure even as the company expands its own AI-chip ambitions. Nvidia has also said MediaTek can sell the CPU developed for Project DIGITS.
Nvidia widens distribution by sharing part of the silicon and system layer with MediaTek. MediaTek gains capabilities and customer access that make it more credible beyond Nvidia’s platform. The bond aligns incentives around compatibility without deciding which company captures each layer’s economics.
Stable workloads turn customers into silicon designers
Large workload owners create pressure from the other side of the market. Once an inference workload becomes stable, high-volume and specific to one operator, that operator can optimize hardware around its models, data flow and operating environment. General-purpose flexibility loses value as control over cost and system behavior gains it.
OpenAI and Broadcom took the Jalapeño inference chip from design to manufacturing tape-out in nine months, aided by OpenAI’s models. Waymo built a custom 5nm ASIC for its robotaxis to improve driving performance and diversify away from third-party suppliers, including Nvidia.
OpenAI and Waymo sell different products, but each controls a workload or distribution channel large enough to justify silicon optimization. Their recurring inference bills make custom designs economically legible.
Custom silicon does not make GPUs obsolete. It carries design, manufacturing, software and deployment costs, while a general-purpose platform spreads those costs across customers and workloads. The more uncertain or variable the workload, the more valuable that flexibility remains.
A credible in-house design still changes bargaining power. The buyer can decide which workloads require Nvidia’s platform, which belong on an ASIC and which should move between local and cloud systems. Nvidia no longer negotiates against “no compute,” but against a workload-specific alternative controlled by the customer.
Scarcity pricing is not ecosystem dependence
Nvidia still commands scarcity pricing. The hourly rental price of a Blackwell GPU rose from $2.75 to $4.08 in two months amid rising agentic-AI demand. AWS raised reserved Nvidia GPU-capacity prices by 20% while leaving Trainium pricing unchanged.
Those price moves show that accelerator supply remains constrained and customers still value immediate access to Nvidia’s stack. MediaTek reinforces the point by building its own R&D infrastructure on B200 chips and DGX SuperPOD while expanding its AI-chip business.
When supply is tight, Nvidia collects a premium on each constrained GPU. When customers depend on its ecosystem, Nvidia captures value from the interfaces, software and architecture through which CPUs, GPUs, custom XPUs, memory and networking work together. The first advantage lasts while the component stays scarce; the second can survive a more heterogeneous market.
A cloud operator can place a stable, high-volume job on custom silicon, keep volatile workloads on Nvidia GPUs and route privacy-sensitive tasks on-device. Nvidia retains leverage only when those choices remain easier inside its architecture than outside it.
Set against MediaTek’s about $2 billion AI-chip forecast, Nvidia’s planned $3.5 billion commitment prices more than a supplier relationship: it prices routes into systems Nvidia does not build alone. The accelerator created Nvidia’s leverage; the MediaTek deal shows where Nvidia now intends to preserve it—in the terms on which the rest of the system connects.
The price of MediaTek’s reach
| As of | Measure | Amount |
|---|---|---|
| 2026-08-31 | Nvidia’s planned MediaTek convertible-bond investment | $3.5 billion |
| 2026-08-31 | MediaTek’s forecast AI-chip revenue for the year | About $2 billion |
Frequently asked questions
Why is Nvidia planning to invest $3.5 billion in MediaTek?
MediaTek already integrates compute, connectivity, power management and software into products shipped through major device channels. Nvidia is effectively buying deeper access to edge-to-cloud systems and OEM relationships it does not control alone.
How does inference change the AI-chip market?
Inference must optimize cost per task, latency, memory, power, privacy and physical location—not merely kernel speed. That splits workloads among cloud racks, PCs, phones, vehicles and custom infrastructure.
What does MediaTek contribute that Nvidia lacks?
MediaTek contributes SoC integration, advanced packaging coordination and delivery through high-volume consumer and automotive channels. It can combine components from multiple suppliers into deployable products.
Does custom silicon make Nvidia GPUs obsolete?
No. ASICs can improve economics for stable, high-volume workloads, but they bring design, manufacturing, software and deployment costs; Nvidia GPUs retain an advantage when workloads are variable or uncertain.
How can Nvidia retain leverage when customers design their own chips?
Nvidia can make its software, interconnect and infrastructure the preferred coordination layer for heterogeneous systems. NVLink Fusion, for example, allows partner-designed XPUs to participate in Nvidia-connected rack-scale infrastructure.