Nvidia s $5.5 Billion Bet on MediaTek and Marvell Signals Shift to AI Infrastructure Tollbooth
Nvidia has now invested $5.5 billion in five months into two companies whose primary growth businesses involve designing custom AI accelerators for the very hyperscalers that Nvidia itself supplies. On August 31, Nvidia subscribed for $3.5 billion of MediaTek's five-year, zero-coupon convertible bonds set at a conversion premium of 15% above the closing price, creating roughly 1.5% pro-rata ownership on full conversion. That follows a $2 billion investment in Marvell announced March 31. Both MediaTek and Marvell sell custom-silicon design services to AWS, Google, Meta, and Microsoft – Nvidia's largest customers.
The numbers attached to MediaTek's AI-ASIC ambitions are staggering. S&P Global's Visible Alpha consensus projects MediaTek's AI-ASIC revenue jumping from roughly NT$69.9 billion in 2026 to NT$535 billion in 2027 – a 7.7x step-up. MediaTek targets 20% of the serviceable custom accelerator market by 2027, with its first production accelerator expected in Q4 2026. DIGITIMES reports Google's TPU v10 may adopt a multi-vendor supply structure involving MediaTek alongside Broadcom. Execution risk at this scale is severe, but the commercial intent is concrete.
Nvidia's Core Business Remains Unscathed For Now
- Q2 FY2027 revenue hit $96.2 billion
- Data Center reached $89.0 billion
- Gross margin held at 75%
- Q3 guidance: $108 billion
No financial evidence yet shows custom silicon denting Nvidia's economics. Buried inside those results, Data Center networking generated $14.8 billion in Q1 FY2027 alone – up 199% year-over-year – and already equivalent to roughly 24.5% of Data Center compute revenue. Networking is one of the semiconductor industry's largest profit pools, and it belongs to Nvidia.
AWS Crossed the Line First
The strategically weightier development predates the MediaTek financing. AWS's Annapurna Labs committed to supporting NVLink Fusion with Trainium4 and became the first collaborator on Nvidia's new NVHBM memory architecture. AWS – among the most sophisticated proprietary-silicon operators on earth – chose to wrap its own accelerator in Nvidia's memory and interconnect. That decision carries more commercial significance than any press release about partnership frameworks.
NVLink Fusion packages NVLink, NVLink-C2C, NVHBM, Nvidia networking, and prevalidated rack infrastructure around third-party XPUs. A hyperscaler can swap out the Nvidia compute die for a custom chip while Nvidia retains revenue through:
- Scale-up interconnect
- NICs, DPUs, switches, CPUs
- Rack architecture
- Software
Concentration Creates Fragility
Nvidia's largest customers are also its most capable competitors:
| Metric | Detail |
|---|---|
| Single direct customer (Q2 FY2027) | 16% of revenue |
| Top 3 customers (first half) | 16%, 15%, and 13% |
| Extended payment terms | Up to one year for large investment-grade buyers |
| Supply commitments | $279 billion |
| Future equity-investment commitments | $25 billion |
| Total future obligations | ~$366 billion |
If AI infrastructure demand slows, the consequences extend well beyond lower chip shipments into inventory impairments, stranded capacity, and weakened investee valuations. Nvidia's own $4.5 billion H20 export-restriction charge in FY2026 demonstrated how fast "scarce capacity" converts into impaired inventory.
The principal architectural counterweight is UALink, whose governing consortium includes AMD, AWS, Apple, Google, Meta, and Microsoft. UALink 1.0 supports 200G-per-lane scale-up for up to 1,024 accelerators, with commercial systems under development through 2026–27. The moment UALink plus merchant Ethernet plus internal orchestration matches NVLink's deployment velocity, hyperscalers' calculus shifts toward forcing Nvidia to compete component by component.
The House Built Around Someone Else's Rooms
Every custom accelerator that lands inside NVLink Fusion infrastructure, uses NVHBM, connects through Nvidia switches, and ships in Nvidia-validated racks generates Nvidia revenue without Nvidia designing the compute die. AWS Trainium4 is already on this path. If two or three additional hyperscaler programs follow, the correct KPI for Nvidia ceases to be GPU units shipped and becomes Nvidia dollar content per rack containing a non-Nvidia accelerator.
The industry is migrating from a structure where one company owned the chip to one where profits concentrate at the interfaces: memory controllers, scale-up fabric, scale-out networking, rack integration, orchestration. Nvidia has spent $5.5 billion in five months subsidizing that migration.
A procurement team that negotiates a 20% discount on a custom accelerator die while accepting Nvidia-controlled infrastructure around it has saved money on the room and surrendered the economics of the house.
Not investment advice.