Jensen Huang, Nvidia's chief executive, and Rick Tsai, who runs MediaTek, went on Bloomberg the morning the two companies announced an expanded partnership and a $3.5 billion Nvidia investment in MediaTek convertible bonds. Ed Ludlow took them through what NVLink Fusion lets a custom-chip customer do, why Huang says the deal is not circular financing, how the two companies now split the contents of an AI factory, and what MediaTek plans to do with the proceeds of its bond program.
👤 Guests: Jensen Huang, chief executive of Nvidia, and Rick Tsai, chief executive of MediaTek, the Taiwanese company Huang calls the world's best maker of system-on-chips, joining from Taiwan late in his evening
🎙️ Host: Ed Ludlow, who anchors Bloomberg Tech from San Francisco
📰 Published: 31 August 2026 on Bloomberg Talks
🔴 YouTube | 🟢 Spotify | 🟣 Apple Podcasts | 🔗 Episode page | ⏱️ 21 min | ✅ Time saved: 7 min
Key Takeaways
Nvidia put its chip-to-chip link inside MediaTek's system-on-chips, and a desktop machine came out of it first
"It's one petaflops." — Huang on DGX Spark, which he says runs advanced agentic AI on a desk rather than in the cloud
The same processor is the base of a machine Nvidia is building with Microsoft, "This processor is going to also be the foundation of the revolutionary new computer that's coming from Microsoft and ourselves, a reinvention of the Windows PC for the age of agents."
The deal is built so each company's wins drag the other's content into the rack
"When they win, we have an opportunity to sell a lot more. When we win, they have an opportunity to sell a lot more."
A MediaTek XPU win brings Nvidia networking with it; an Nvidia data center gives a MediaTek XPU somewhere to plug in
Huang says buying $3.5 billion of a partner's convertible bonds is not circular financing
"Well, this is not circular because obviously they do their own business and we do our own business."
"The roadmap that we're executing on right now, it's not one or two years. It's a decade long."
Custom accelerators are not a threat Nvidia is waiting for — Huang says they are already here
"I see XPUs in the market. It's not coming into the market. It's in the market."
"And so we're not intimidated by this. We're in fact welcoming it and opening our platform so that they could connect XPUs into it."
Seven types of Nvidia chip now go into an AI factory, and a rival accelerator displaces one of them
"Well, inside a gigawatt, if we were just to extrapolate, in a gigawatt, Hopper was about $18 billion of economics for Nvidia per gigawatt. Grace Blackwell is about $25 billion. And the Vera Rubin generation, it's about 40 plus billion dollars per gigawatt of total Nvidia compute economics."
"An XPU would be a replacement of one of those."
MediaTek's AI chip ramp is guided, but Tsai will not promise it comes faster
Ludlow read MediaTek's own guidance back to him: "Rick, you had said that your AI chip business would be about $2 billion this year, but would accelerate to get about 15% of an addressable market of $80 billion next year."
"We certainly expect that to happen. But it will take a little time"
What a MediaTek customer is really buying is time to market, not silicon
"And that's the key for today's world. Fast time to market."
Without the Nvidia stack, a custom accelerator has to solve NVLink switching, Ethernet switching and scale-out, scale-up, scale-across and scale-in on its own
The constraint on MediaTek is supply, and it runs past memory into substrates
"I think supply chain does present challenges, not just HBM, go down to substrate, et cetera." — Tsai
"But it's tough."
Nvidia's next HBM is custom, and it has handed the base layer to its partners
"Our next generation HBM is a custom HBM, and the HBM stacks are built on a Nvidia custom base layer."
"We've extended it to Amazon as well. And they're one of the first customers and partners building it with us."
Nvidia Wired Its Chip-to-Chip Link Into MediaTek's SoCs, and DGX Spark Came Out of It
Ludlow opened by framing the deal for the audience: "The idea is you can take custom silicon, right, XPU, and just plug it into any GPU-based architecture, share the memory, the networking." He then asked Huang to explain the technology agreement.
Huang said the two companies were already partners and were now making the relationship much larger. "Today, we're announcing a big partnership. We've already had a big partnership with MediaTek. Today, we're going to make it a lot bigger."
It began with Nvidia's NVLink chip-to-chip interface being integrated into MediaTek's parts, so that any MediaTek system-on-chip can be connected to an Nvidia GPU
The reason he gave for picking the partner is flat: "MediaTek makes the world's best SoCs."
The first product out of that work is DGX Spark, a desktop computer. "This little computer is an incredible computer. It's one petaflops."
"It allows you to run an advanced agentic AI right on your desk instead of having to run it in the cloud."
The new agreement extends that processor across multiple generations, and Huang said it will also be the foundation of a machine Nvidia is building with Microsoft: "This processor is going to also be the foundation of the revolutionary new computer that's coming from Microsoft and ourselves, a reinvention of the Windows PC for the age of agents."
The second piece is the data center. Nvidia is connecting NVLink Fusion, the NVLink system, its Spectrum-X switches and its system chassis to MediaTek, so that Nvidia networking rides along with a MediaTek accelerator win — "So when they win projects, it brings along Nvidia networking."
Because Nvidia is already in every cloud and every data center, he said, a MediaTek XPU can be connected into that estate without new plumbing
The two are also working together on robotic systems such as autonomous vehicles
The reciprocity is the whole design of the deal: "When they win, we have an opportunity to sell a lot more. When we win, they have an opportunity to sell a lot more."
MediaTek's Pitch to ASIC Customers: One Stack, and Speed to Market
Ludlow put it to Tsai that an ASIC customer builds its own accelerator and does not need to reinvent the rest of the architecture, and asked what MediaTek tells those customers about doing it with Nvidia through NVLink.
Tsai opened by thanking Nvidia and Huang "for having the confidence in MediaTek." He then described the layers the partnership is meant to cover: the top of the ASIC market at the hyperscalers, the mid-end of enterprise data centers and new clouds, and the desktop supercomputers Huang had just described.
The argument is about the pace of the market, not the silicon. "The important thing for this era of AI is it's incredibly fast changing."
What matters, he said, is enabling customers across that stack to build a data center and scale it quickly — "Well, as Jensen always like to say, speed of light."
"And we work with our customers with tremendous flexibility." That, he said, is the essence of the collaboration
The AI Chip Ramp Will Take Time, Even With Nvidia Behind It
Ludlow read MediaTek's own guidance back to Tsai: "Rick, you had said that your AI chip business would be about $2 billion this year, but would accelerate to get about 15% of an addressable market of $80 billion next year." He asked whether the Nvidia arrangement helps the business get there faster.
Tsai agreed with the direction and refused the timing. "We certainly expect that to happen. But it will take a little time"
The reason is mechanical: MediaTek is only now starting to build Huang's latest NVLink version into the chassis he had described
What the chassis sells is speed. He said it gives potential customers a vehicle for fast time to market, and repeated the phrase for emphasis: "And that's the key for today's world. Fast time to market."
Why Nvidia Bought $3.5 Billion of MediaTek Convertible Bonds
Ludlow asked the question directly: "Jensen, you're making a $3.5 billion investment in MediaTek through the purchase of convertible bonds. Why?"
Huang called it one of the largest partnerships Nvidia has. The two companies are building multiple generations of edge AI and PC chips, with Nvidia GPUs connected die-to-die to MediaTek's system-on-chips
The consumer end of it is a product line, not an experiment. He named DGX Spark and RTX Spark for PCs as a major investment, and said Nvidia has a major commitment to the entire PC industry
The investment also formalizes the AI factory side. An AI factory is no longer just the GPU, he said: it is GPUs, scale-up switches and scale-out switches, five different types of networking, and Nvidia's Vera CPUs
"And so wherever we have Nvidia in the cloud already, which we're in just about every cloud, we now have the opportunity using exactly the same AI factory architecture to plug in MediaTek XPUs."
On why it needed money attached at all, he went to the relationship. "I've known Rick for 25 years." — a relationship he called "deeply grounded on trust"
He was careful to say the target did not need the capital: "Of course, MediaTek is already immensely profitable, and they're incredibly successful already"
Huang Says Investing in a Partner Is Not Circular Financing
Ludlow noted that Huang and Colette, Nvidia's chief financial officer, had already fielded questions about circular financing on the earnings call and had been clear that they see it differently from the way the world reads it. In this case, he said, the investment goes into a partner rather than a direct customer.
"Well, this is not circular because obviously they do their own business and we do our own business." MediaTek, Huang said, is already profitable and successful on its own account
"They are the world's best SoC maker. They make more SoCs than any company in the world."
He framed the check as a signal of duration. "The roadmap that we're executing on right now, it's not one or two years. It's a decade long."
The investment, he said, is there to demonstrate the seriousness with which Nvidia takes the partnership
He also said he expects it to pay: the investment will be a phenomenal one for Nvidia and will generate incredible returns
On the location of it: "It's a great pleasure for us to invest in a technology company in Taiwan."
What MediaTek Will Do With the Bond Proceeds
Ludlow said MediaTek had announced in July that it would use its bond program — $5 billion of dollar bonds — to support and grow supply chain, and asked what the proceeds are actually for, and whether the work needs to happen with Taiwanese or US suppliers.
Tsai said that when the program was announced, "we were not able to discuss the strategic investment part of the bond"
The proceeds have several destinations, with supply chain among them, and he described that supply chain as global — wherever it is needed to enhance and enable customers
More of it goes into MediaTek's own technology. He named the company's capability to make the ASIC chip, the intellectual property for interconnects, and advanced packaging
The key thing, he said, is the set of collaboration agreements MediaTek has built and expanded with Nvidia as described in the companies' press release, and the proceeds go toward making those happen faster
XPUs Are Already in the Market, and Nvidia Says It Is Not Intimidated
Ludlow put Nvidia's own earnings-call language to Huang — that the GPU-based architecture has superior economics, said in the context of the frontier labs and in answer to a question about custom silicon — and pointed out that NVLink Fusion makes it easy to drop an XPU into the data center. He asked what had changed in a week.
Huang rejected the premise of the tense. "I see XPUs in the market. It's not coming into the market. It's in the market."
The distinction he drew is between a specialized part and a general one. "Look, an XPU is a specialized chip. Nvidia's GPU is a general purpose accelerator."
Nvidia accelerates the whole lifecycle, he said — data processing, pre-training, post-training and agentic inferencing — across closed models, open models, video and language models, and biological, physics and robotics models
"That explains why it is that we are the most fungible and the most durable and therefore the most rentable GPU, the most rentable computer infrastructure in the world."
"We're in every cloud. We're on-prem. We're at the edge."
The customers building XPUs are his customers too, and Rick's. "If they would like to put a specialized XPU into a data center, why not make it easier for them to connect it to the Nvidia infrastructure?"
"And so we're not intimidated by this. We're in fact welcoming it and opening our platform so that they could connect XPUs into it." He said Nvidia's footprint is the largest and that it is growing share in closed and open models alike
The Customers Asked for This, and Building an XPU Is Still Very Hard
Ludlow asked Tsai whether there is a case study of an ASIC customer he could walk through, and how such a customer might move in reaction to the announcement.
Tsai said MediaTek is engaging with customers on a co-go-to-market basis with Nvidia, and that the announcement itself is the point: "And I think the announcement today will further strengthen their confidence in the relationship with this strong bond between Nvidia and the MediaTek."
The market is big enough for buyers who do not need a frontier-scale cluster. "And the AI market, computer market is enormous. Enormous." Some customers need obviously smaller amounts of computing power but still want differentiation in their own offering
Huang said the deal was pulled out of Nvidia rather than pushed. "And in a lot of ways, Ed, this whole thing started because of customer requests." Customers have already committed to NVLink Fusion, he said
"Building the XPU is very, very hard, and there's no question about that." With MediaTek in the middle, he said, those customers can get help building one
"And it's incredibly hard to do when we're in our sixth generation of doing so."
Seven Types of Nvidia Chip Go Into an AI Factory; an XPU Replaces One
Ludlow asked about NVHBM, noted Huang had been clear that the bottleneck is still memory pricing and supply, and added a second question about how much of the server design on the Vera Rubin generation is now Nvidia content.
Huang answered the second part first, in per-gigawatt terms. "Well, inside a gigawatt, if we were just to extrapolate, in a gigawatt, Hopper was about $18 billion of economics for Nvidia per gigawatt. Grace Blackwell is about $25 billion. And the Vera Rubin generation, it's about 40 plus billion dollars per gigawatt of total Nvidia compute economics."
The mechanism is more chips, not a higher price on the same chip. "And the reason for that is because we've been expanding beyond the GPU to one type of switch after another type of networking." Seven different types of Nvidia chip now have to go into an AI factory
"An XPU would be a replacement of one of those."
His forecast is displacement at the margin, not exclusion. "My sense is that we'll continue to be in every single cloud, we'll continue to be in every single AI factory. Some of those racks might be augmented with XPUs."
In those racks, he said, Nvidia's networking, switching and CPUs still participate
For MediaTek the payoff is being able to sell the whole thing. "In the case of MediaTek, it allows MediaTek to offer an end-to-end solution that is complete."
The alternative is that the XPU has to solve NVLink switching, Ethernet switching and scale-out, scale-up, scale-across and scale-in on its own
MediaTek's Bottleneck Is the Supply Chain, Not Demand
Ludlow asked Tsai for the single biggest bottleneck at MediaTek right now.
"I think supply chain does present challenges, not just HBM, go down to substrate, et cetera."
Location is part of the answer. "But again, being in Taiwan, in almost like a center of the AI industry, infrastructure ecosystem really helps." MediaTek also works closely with Nvidia as a customer of its own
He did not soften the assessment: "But it's tough."
Huang added that the two supply chains now overlap. Nvidia's networking ecosystem is part of MediaTek's supply chain, he said, and MediaTek's XPU is part of Nvidia's, so both companies came out of the deal with a wider one
NVHBM Is a Custom Stack on an Nvidia Base Layer, and Amazon Has It Too
Huang then went back to the memory question he had skipped.
"Our next generation HBM is a custom HBM, and the HBM stacks are built on a Nvidia custom base layer."
That base layer has been handed outward. Nvidia has extended it to its NVLink Fusion partners, MediaTek among them
"We've extended it to Amazon as well. And they're one of the first customers and partners building it with us."
The purpose is standardization. Aggregating the supply chain into something more standard, he said, lets partners benefit from Nvidia's large install base and supply chain, and lets Nvidia benefit from theirs
Huang's bottom line is that opening NVLink Fusion to MediaTek costs Nvidia one chip out of the seven it puts into an AI factory and keeps it inside every rack a rival accelerator lands in.
Products, Companies & Tools Mentioned
Nvidia and MediaTek (The two sides of the deal: Nvidia buying $3.5 billion of MediaTek convertible bonds and opening its interconnect, MediaTek supplying the system-on-chips and custom accelerators Huang calls the world's best)
NVLink and NVLink Fusion (The interconnect at the center of the announcement — chip-to-chip into MediaTek's SoCs, and the scale-up fabric a custom XPU plugs into; Huang says Nvidia is in its sixth generation of it)
DGX Spark and RTX Spark (The first product out of the earlier partnership and its PC sibling — "It's one petaflops," running agentic AI on a desk instead of in the cloud)
Spectrum-X (The Ethernet switching Nvidia is attaching to MediaTek's chassis, so an accelerator win brings Nvidia networking with it)
Microsoft (Co-building the machine Huang calls "a reinvention of the Windows PC" for the age of agents, on the same MediaTek-connected processor)
Hopper, Grace Blackwell and Vera Rubin (The three generations Huang used to show Nvidia's per-gigawatt economics rising as it sells more of the factory; Vera CPUs are part of the current one)
NVHBM (Nvidia's next-generation custom HBM, stacked on an Nvidia custom base layer that has been extended to NVLink Fusion partners)
Amazon (Named alongside MediaTek as one of the first partners building on that custom HBM base layer)
Books & Resources Mentioned
NVIDIA and MediaTek Deepen Long-Standing Partnership to Build AI Edge to Cloud Computing Platforms (The joint press release Tsai points to for the collaboration agreements behind the investment)
Nvidia's quarterly results and earnings call (The call from the week before, where Huang and his chief financial officer answered the circular-financing questions Ludlow returns to here)
Get the latest market chatter as it happens:
If this saved you the listen, send it to one person who would rather have the time back.