NVIDIA invests $3.5 billion in MediaTek: when the GPU king buys its defense against custom chips

NVIDIA invests $3.5 billion in MediaTek: when the GPU king buys its defense against custom chips

🔎 The day NVIDIA admitted that custom chips weren't a fad

On August 31, 2026, NVIDIA invested $3.5 billion in convertible bonds in MediaTek. NVIDIA's first direct investment in a Taiwanese chip designer. The largest offshore convertible issuance in Taiwan's history ($3.9 billion in total, with Alphabet in the mix). MediaTek's stock jumped 10% at the open, according to Morningstar.

But the figure isn't the story. The story is what NVIDIA gets in exchange: MediaTek integrates NVLink Fusion into its custom chips. The ASICs built by Amazon, Google, Microsoft, and others will be able to plug directly into NVIDIA clusters. Instead of losing these customers, NVIDIA locks them into its interconnect ecosystem.

This is a major strategic pivot. NVIDIA officially announces it as a deepening of the partnership for edge-to-cloud AI platforms. The reality is more stark: if you can't kill custom chips, make sure they can't live without you.

This move is part of a dizzying financial sequence. NVIDIA is already pouring $40 billion into AI in 2026, and investing $150 billion a year in Taiwan. The $3.5 billion for MediaTek seems modest compared to these amounts. But their strategic scope far outweighs the entry price.

The Essentials

• NVIDIA is investing $3.5B in convertible bonds in MediaTek, as part of a record $3.9B issuance that also includes Alphabet.
• MediaTek integrates NVLink Fusion to allow hyperscalers and AI natives to plug their custom chips (XPU/CPU/ASIC) into the NVIDIA ecosystem.
• MediaTek's stock surged 10% following the announcement, according to Morningstar/Dow Jones.
• NVIDIA neutralizes the threat of custom chips by controlling their interconnect layer rather than fighting them head-on.
• The debate on AI's circular financing resurfaces: NVIDIA lends to OpenAI, OpenAI builds datacenters, NVIDIA invests in the chip suppliers for those datacenters.

Recommended Tools

The deal figures: $3.5B, $3.9B, 10% increase

Anatomy of the investment

NVIDIA is placing $3.5 billion in convertible bonds in MediaTek. These are not direct equity stakes, but hybrid securities that will convert into MediaTek shares at a predetermined price. According to Cryptobriefing, the conversion price is set at NT$4,513.75 per share, with a zero coupon.

The total issuance amounts to $3.9B, making it the largest offshore convertible deal ever executed in the Taiwanese market, reports BigGo Finance. The difference between NVIDIA's $3.5B and the total $3.9B is filled by Alphabet, whose exact stake has not been disclosed but is confirmed by Nikkei Asia via TechTimes.

Market reaction

MediaTek's stock surged 10% at the opening of the Taiwanese market on September 1, 2026. A massive move for a market capitalization that easily exceeds $100B. The market understood two things: MediaTek gains access to the industry's most powerful interconnect technology, and NVIDIA just validated the custom silicon model as structural, not cyclical.

According to Yahoo Finance, the deal positions MediaTek as the vendor of choice for hyperscalers looking to build custom XPUs while remaining compatible with existing NVIDIA infrastructure.

Editorial note on custom silicon revenues

The brief mentions "MediaTek targets $2B in custom silicon revenues in 2026 on an $80B addressable market by 2027." These specific figures do not appear in any of the provided sources. They are circulating within the analyst ecosystem but are not corroborated by an official press release or regulatory filing. We are flagging them without validating them.

NVLink Fusion: The strategic weapon NVIDIA offers its rivals

What NVLink Fusion is exactly

NVLink Fusion is a high-speed, low-latency interconnect technology designed to connect heterogeneous chips within the same cluster. According to the official NVIDIA page, it allows hyperscalers and AI natives to deploy custom XPUs and CPUs within the NVIDIA ecosystem.

Concretely, this means that an ASIC designed by Google for its TPU infrastructure will be able to communicate with NVIDIA GPUs via NVLink Fusion, instead of being isolated in a proprietary silo. MediaTek confirms this on its blog: NVLink Fusion becomes the design foundation for customers building custom ASICs.

Why it's a stroke of genius

NVIDIA's problem is simple: every hyperscaler is building its own AI chip. Amazon has Trainium, Google has TPU, Microsoft has Maia, Meta has MTIA. These chips are eating away at NVIDIA GPU market share.

The classic response would have been to make better and cheaper GPUs. NVIDIA chose another path: make custom chips dependent on its interconnect infrastructure. If your MediaTek-NVLink Fusion ASIC works perfectly in an NVIDIA cluster, you have no reason to build a competing proprietary interconnect. You remain within the NVIDIA fold.

Igor's Lab analyzes this strategy as an encirclement move: NVIDIA can no longer prevent hyperscalers from manufacturing custom chips, but it can ensure that these chips live within its ecosystem.

The historical parallel with Intel

In the 2000s, Intel dominated the server market with x86. Competitors tried to create alternative architectures (ARM, Power). Intel failed to stop them. But what Intel did well was control the chipsets, interconnects, and platforms.

NVIDIA is reproducing this exact pattern, but on an AI scale. The GPU becomes the center of gravity, and everything else — ASICs, custom CPUs, specialized accelerators — orbits around it via NVLink Fusion. Except that NVIDIA learned from Intel's mistake: it is not relying on a processor monopoly, but on an ecosystem monopoly.

MediaTek: from smartphone to the heart of AI datacenters

A radical transformation

MediaTek is known to the general public for its smartphone processors. Dimensity, its flagship range, equips hundreds of millions of Android phones every year. But since 2024, the Taiwanese company has been pivoting massively towards AI datacenters.

The NVLink Fusion partnership is accelerating this pivot in a spectacular way. TechPowerUp rapporte that NVIDIA is adding "new layers of collaboration" beyond simple financial investment. MediaTek is not becoming a simple subcontractor manufacturer: it is becoming NVIDIA's armed wing in the custom chip market.

MediaTek's position in the custom silicon market

The custom AI chip market is exploding. Every hyperscaler wants its own chip to reduce its dependence on NVIDIA, control its costs, and optimize its specific workloads. But designing a custom chip is an extremely complex exercise. It requires design teams, IP licenses, foundry agreements, and now, an interconnection strategy.

This is where MediaTek comes into play. By offering NVLink Fusion as an interconnection standard, MediaTek offers hyperscalers a path of least resistance: design your ASIC, MediaTek handles the NVLink integration, and you plug into existing datacenters. This dynamic is reminiscent of what Chinese players are trying to do with alternative approaches, such as la puce Zhenwu M890 d'Alibaba qui vise à détrôner NVIDIA en Chine.

Why NVIDIA chose MediaTek and not someone else

The question is valid: why not TSMC directly? Why not Broadcom, which already dominates the custom ASIC market with Google and Meta chips?

The answer comes down to two words: independence and positioning. TSMC is a neutral foundry. NVIDIA cannot tie itself exclusively to it without alienating its other partners. Broadcom is a potential competitor in interconnection. MediaTek, on the other hand, is a newcomer to the datacenter market. It needs NVIDIA. The balance of power is clear, and Reuters le confirme: the NVLink Fusion technology provides specialized connectors and memories to plug custom XPUs into the Nvidia ecosystem. MediaTek is the vehicle, not the driver.

Circular AI Financing: NVIDIA Lends to Everyone

The Mechanism Explained

The concept of "circular financing" of AI, raised by TechTimes, describes a dizzying loop. NVIDIA sells GPUs to hyperscalers. These hyperscalers generate revenue with AI. NVIDIA reinvests this revenue (directly or indirectly) into the hyperscalers' suppliers. And so on.

The most striking example precedes this deal: the $105B guaranteed by NVIDIA for OpenAI's lease in Ohio. NVIDIA lends money to OpenAI so that OpenAI can buy NVIDIA GPUs. The cycle feeds itself.

With MediaTek, the circle expands. NVIDIA invests in MediaTek. MediaTek manufactures custom chips for NVIDIA's customers. These customers also buy NVIDIA GPUs. NVIDIA wins on all fronts: GPU sales, appreciating investments, and the lock-in of the interconnection ecosystem.

The Risks of a Circular Model

This model has an obvious flaw: it relies on the perpetual growth of the AI market. If demand slows down, if current models like Gemini 3.1 Pro or GPT-5.5 no longer bring sufficient performance gains to justify infrastructure spending, the whole house of cards wobbles.

Convertible bonds add a specific risk. If MediaTek's stock falls below the conversion price of NT$4,513.75, NVIDIA is left with debt that does not convert into equity. On a $3.5B investment, the potential loss is not negligible, even for NVIDIA.

Alphabet in the Shadow of the Deal

Alphabet's participation in the $3.9B convertible is a detail that speaks volumes. Alphabet is simultaneously a customer of NVIDIA (Google Cloud buys GPUs), a competitor (Google has its TPUs), and now an investor in NVIDIA's partner for custom chips. This triangular position illustrates the complexity of alliances in AI. Nobody is purely a friend or an enemy. Everyone is both at the same time.

The Taiwanese context: why this deal is geopolitically charged

Taiwan, the epicenter of NVIDIA's strategy

NVIDIA invests 150 billion dollars per year in Taiwan, and for good reason: that is where TSMC is located, the foundry without which no NVIDIA GPU exists. But the $3.5B in MediaTek is different. It is NVIDIA's first direct investment in a Taiwanese designer, not in a foundry.

NVIDIA is therefore anchoring its operations not only in Taiwanese manufacturing, but also in Taiwanese design. This is a strong signal to Beijing, which considers Taiwan a rebel province, and to Washington, which wants to relocate the semiconductor supply chain.

The ARM factor and Vera Rubin

This deal also comes in the context of the GTC Taipei where NVIDIA unveiled Vera Rubin N1X and the era of AI agents. ARM architectures, which MediaTek masters perfectly thanks to its smartphone chips, are at the heart of NVIDIA's edge-to-cloud strategy. The partnership therefore takes on an additional technical dimension: MediaTek brings its ARM expertise, NVIDIA brings NVLink Fusion, and the two converge toward unified platforms.

The indirect response to China

While NVIDIA is locking down the custom silicon ecosystem in Taiwan, China is accelerating its own semiconductor sovereignty agenda. The investment in MediaTek indirectly strengthens Taiwan's position as an essential hub for AI chip design, making any attempt to isolate the island even more costly for Beijing.

The winners and losers of the operation

Winners

NVIDIA: Locks down the interconnect ecosystem, turning a threat (custom chips) into a recurring revenue opportunity via NVLink Fusion. The $3.5B investment is risky but the strategic positioning is priceless.

MediaTek: Gains access to the most in-demand interconnect technology on the market. Transitions from a smartphone vendor to an AI datacenter player in a single deal. The stock's +10% jump speaks volumes about the market's appreciation.

Mid-sized hyperscalers: Players that don't have the resources of Google or Amazon to build a proprietary interconnect can now order custom ASICs from MediaTek and plug them directly into NVIDIA clusters. It's a democratization of access to custom silicon.

Losers

Broadcom: The current leader in custom ASICs (Google TPU, Meta MTIA) sees a formidable competitor arriving armed with NVLink Fusion. If MediaTek can offer superior interconnect, Broadcom loses its edge.

Alternative interconnect projects: UCIe (Universal Chiplet Interconnect Express), CXL, and other open standards lose ground. Why adopt an open standard when NVLink Fusion offers better performance and access to the dominant GPU ecosystem?

Alternative foundries: Samsung Foundry, Intel Foundry, and Chinese foundries do not have access to NVLink Fusion. Custom chips manufactured there will be disadvantaged compared to those going through the MediaTek-NVIDIA-TSMC pipeline.

What this means for AI developers

Hybrid GPU/ASIC clusters

For teams training models like Claude Opus 4.7 d'Anthropic or DeepSeek V4 Pro, the NVIDIA-MediaTek deal means they will likely be working in hybrid clusters by 2027. NVIDIA GPUs for general tasks, custom ASICs for specialized workloads (embedding, high-density inference, agentic routing).

Reference benchmarks will change. Current agentic model scores — GPT-5.5 at 98.2, Gemini 3 Pro Deep Think at 95.4, Claude Opus 4.7 Adaptive at 94.3 — are measured on mostly GPU infrastructures. Hybrid clusters will introduce new performance variables.

Impact on training costs

The ultimate goal of custom chips is cost reduction. A specialized ASIC consumes less power and costs less to manufacture than a generalist GPU. If NVLink Fusion allows these ASICs to operate seamlessly within an NVIDIA cluster, training costs per model could drop significantly.

This is particularly relevant for agentic models that require millions of reasoning turns. Kimi K2.6 de Moonshot AI, which scores 88.1 in self-host mode, illustrates the value of infrastructure optimization: when you control the full stack, every compute cycle counts.

Skills to develop

AI engineers will need to understand compute heterogeneity. Knowing how to distribute a workload between GPUs and ASICs is no trivial matter. Orchestration frameworks (JAX, PyTorch, vLLM) will need to evolve to natively manage NVLink Fusion clusters. Those who master this layer will have a considerable competitive advantage.

The previous Groq: when the NVIDIA ecosystem recycles its own rejects

The irony of history

This is not the first time NVIDIA has maneuvered to bring alternative players into its orbit. Groq raised 650 million dollars and pivoted to neocloud after Nvidia snatched its soul for 20 billion. Groq's story is enlightening: a promising startup with an innovative LPU (Language Processing Unit) architecture, which found itself marginalized because the NVIDIA ecosystem was too dominant.

The deal with MediaTek is the extension of this logic, but in reverse. Instead of marginalizing alternatives, NVIDIA absorbs them. Instead of letting Groq die outside the ecosystem, NVIDIA offers future Groqs to be born inside it. It's more elegant, more profitable, and less risky from an antitrust perspective.

The lesson for AI chip startups

The message is clear: don't build outside the NVIDIA ecosystem, build inside it. NVLink Fusion is the offer you cannot refuse. You keep your custom chip, your differentiation, your IP. But the interconnect belongs to NVIDIA. And in the modern AI datacenter, the interconnect is just as important as the compute itself.

❌ Common mistakes

Mistake 1: Confusing investment and acquisition

The $3.5B is in convertible bonds, not direct shares. NVIDIA does not "own" $3.5B of MediaTek. It holds a future conversion right at NT$4,513.75. If the share price rises above this, NVIDIA realizes a gain. If it falls, NVIDIA holds debt. The distinction is important for assessing NVIDIA's actual commitment.

Mistake 2: Thinking that NVLink Fusion replaces GPUs

NVLink Fusion connects custom chips to the NVIDIA ecosystem. It does not replace GPUs. Custom ASICs will be complementary, not substitutive. A typical cluster will continue to rely heavily on NVIDIA GPUs for general-purpose computing. ASICs will handle specialized workloads.

Mistake 3: Underestimating Alphabet's role

Alphabet is participating in the $3.9B convertible alongside NVIDIA. This is not a minor detail. Alphabet is NVIDIA's customer, NVIDIA's competitor (TPU), and now a co-investor in NVIDIA's partner. Ignoring this triangular dimension means missing half the story.

Mistake 4: Extrapolating unsourced custom silicon revenues

The figure of $2B in custom silicon revenues for MediaTek in 2026 is circulating but is not corroborated by any official source in the provided documents. Treating it as a fact is a journalistic error. Mentioning it with the appropriate distance is the right approach.

❓ Frequently Asked Questions

Why did NVIDIA choose convertible bonds instead of shares?

Convertibles offer downside protection (senior claim in case of bankruptcy) while capturing the upside if the share price rises. It is a more flexible financial instrument than a direct equity purchase, with different accounting and tax implications. NVIDIA is keeping its options open.

Is NVLink Fusion an open standard?

No. NVLink Fusion is proprietary to NVIDIA. That is precisely the issue: NVIDIA is not promoting an open standard, it is imposing its own. Customers accept this because NVLink's performance is superior to open alternatives like UCIe, and because the NVIDIA GPU ecosystem is indispensable.

Does this deal threaten TSMC's position?

No, on the contrary. The custom chips designed by MediaTek with NVLink Fusion will most likely be manufactured by TSMC. The deal reinforces the TSMC → MediaTek → NVIDIA chain rather than weakening it.

Will current models be affected?

In the short term, no. Models like GPT-5.5, Gemini 3.1 Pro, or Claude Opus 4.7 run on existing infrastructure. In the medium term (2027-2028), NVLink Fusion GPU/ASIC hybrid clusters could modify training architectures and associated costs.

What is the connection to OpenAI's $105B?

The $105B guaranteed by NVIDIA for the OpenAI lease in Ohio is another example of circular financing: NVIDIA lends so that its customers can buy. The MediaTek investment follows the same logic at a different level of the stack.

✅ Conclusion

NVIDIA is not spending $3.5B to help MediaTek. It is spending it to ensure that every custom chip built by hyperscalers goes through its interconnect, its ecosystem, its standards. The GPU king is no longer fighting against ASICs: it is taming them. The rest of the market had better brace itself.