Jensen Huang has spent the past two years repeating a version of the same line: most custom AI chip projects never actually ship. He says it on earnings calls, he says it on stage, and he says it with the calm of someone who has watched a dozen internal silicon programs die in validation. My reaction, reading the reports of Nvidia taking a roughly $3.5 billion position in MediaTek, is that the line was always half true and half insurance policy. If custom accelerators were genuinely doomed, you would not buy equity in the company helping design them.
A caveat before analysis: I am working from reporting, not from a filing I have read line by line, and the strategic reasoning below is mine rather than anything either company has stated. Treat the dollar figure as reported and the interpretation as argument.
What money buys that a press release cannot
MediaTek is not a rival to Nvidia in any direct sense. It is a design partner — the kind of firm hyperscalers hire when they want an accelerator of their own without building a silicon organization from scratch. That places MediaTek at a specific chokepoint in the AI buildout: the path by which Google, Amazon, Microsoft, Meta, and anyone else with a capex budget converts frustration with GPU pricing into working hardware.
Owning a slice of that chokepoint is not about control. It is about visibility and optionality. An equity stake gets you a seat near the roadmap, a reason to co-engineer rather than compete, and a hedge that pays off in the scenario where custom silicon takes real inference share. Nvidia has already signaled the co-engineering direction by opening NVLink to third-party chips. A stake makes that opening look less like generosity and more like strategy.
Agent workloads are the actual battleground
This is where the story stops being about corporate finance and starts being about architecture, which is the part I care about.
Training a frontier model is a bandwidth-bound, tightly synchronized problem across thousands of accelerators. It rewards exactly what Nvidia sells: dense compute plus an interconnect fabric that behaves like one machine. Custom silicon has struggled there, and Huang’s skepticism has been mostly earned.
Agent inference is a different animal. An agent loop is not one forward pass. It is dozens or hundreds of them, interleaved with tool calls, retrieval, memory reads, and re-planning. The compute profile skews heavily toward decode rather than prefill, which means it is dominated by memory bandwidth and cache behavior rather than raw matrix throughput. Context grows across turns, so the key-value cache becomes the dominant resident structure, and the scheduling problem becomes about keeping that cache warm across unpredictable latency gaps while a tool returns.
Those characteristics are far more tractable for a fixed-function design. You do not need general-purpose flexibility to run the same transformer decode step ten thousand times a second. You need bandwidth, cheap memory capacity, and an efficient path from cache to compute. That is a specification a design house can hit, and hitting it well can beat a general-purpose GPU on cost per token by a wide margin.
So when I look at agent architectures, I see the workload that finally gives custom accelerators a credible entry point. And I suspect Nvidia sees the same thing.
Why the moat moved to the fabric
The interesting part of Nvidia’s position is that it has already conceded the chip-level fight is winnable by others, and moved the defense one layer out. Multi-agent systems that run planners, critics, retrievers, and specialized models concurrently do not fit neatly on a single accelerator. They need coherent memory across devices, low-latency interconnect, and a scheduler that can move state between models without serializing everything through host memory.
That is a systems problem, and systems problems are where Nvidia’s advantage is hardest to replicate. Networking, coherent fabric, and the software layer above it are not things a first-generation ASIC program solves. If custom chips end up plugged into Nvidia’s fabric, Nvidia keeps the margin that matters even as it loses the socket.
What I would watch next
- Fabric adoption. Does third-party silicon actually attach to Nvidia’s interconnect at scale, or do hyperscalers build their own coherent networks? This is the whole question.
- Where custom chips land. If they show up in serving fleets rather than training clusters, the agent-inference thesis holds.
- Memory system design. Watch capacity and bandwidth ratios in the next generation of custom parts. They will tell you whether designers are optimizing for long-context agent loops or for something else.
- Software gravity. CUDA remains the reason most inference stays on Nvidia. Agent frameworks that abstract the accelerator away erode that faster than any hardware release.
The honest read is that Nvidia is buying a position in the scenario where it is partially wrong. That is not weakness. It is the behavior of a company that has done the math on what agents will cost to run at scale, and decided it would rather own a piece of the alternative than argue the alternative away.
đź•’ Published: