Nvidia and SK Hynix Sign Multi-Year Deal to Co-Develop Advanced AI Chips

On June 7, 2026, Nvidia and SK Hynix formalized what the semiconductor industry had long treated as an open secret: the two companies are now bound by a multi-year technology partnership to co-develop memory for the next wave of AI computing platforms [1]. The deal covers memory for Nvidia's Vera Rubin AI supercomputers, Vera CPUs, RTX Spark PCs, and Jetson Thor robotics platforms — and extends into joint work on semiconductor simulation, chip design, and autonomous factory operations [2].

The agreement arrives two days after Nvidia CEO Jensen Huang certified all three major HBM suppliers — SK Hynix, Samsung, and Micron — for the Vera Rubin platform's HBM4 memory [3]. But the formal partnership with SK Hynix goes well beyond a supply contract. It signals a deeper structural alignment between the dominant AI chip designer and the dominant high-bandwidth memory manufacturer, with implications that ripple through competitors, regulators, and entire national industrial strategies.

The Revenue Dependency

The financial relationship between Nvidia and SK Hynix has grown rapidly. In 2023, Nvidia accounted for roughly 10% of SK Hynix's total revenue. By 2024, that figure rose to 16%, representing approximately ₩10.9 trillion. In the first half of 2025, Nvidia drove an estimated 27% of SK Hynix's total revenue, according to TrendForce [4].

Source: TrendForce

Data as of Aug 18, 2025CSV

SK Hynix's full-year 2025 revenue reached 97.1 trillion won (approximately $67.9 billion), up 47% year-over-year, with HBM revenue more than doubling and reaching an estimated 40% of total DRAM revenue [5]. The company has confirmed that all DRAM, NAND, and HBM production through 2026 has been sold out, much of it committed to Nvidia [5].

This concentration cuts both ways. SK Hynix benefits from a guaranteed buyer for its most profitable product line. Nvidia benefits from preferential access to the most supply-constrained component in AI hardware. The new partnership formalizes and extends that mutual dependency, though neither company has disclosed financial terms, minimum purchase commitments, or IP ownership provisions.

What the Deal Covers: HBM4, HBM4E, and Beyond

The partnership encompasses current and next-generation memory standards. SK Hynix currently supplies HBM4 for the Vera Rubin platform and SOCAMM2 (192GB) modules for Vera Rubin data centers. HBM4E — the next evolutionary step — is under co-development for Nvidia's Rubin Ultra platform [2].

The performance targets are substantial. SK Hynix has demonstrated a 48GB HBM4E solution in a 12-layer stack using 32 Gb chips, achieving speeds up to 16 Gbps per pin and a maximum indicated bandwidth of 4 TB/s [6]. Samsung, which delivered the industry's first HBM4E samples, offers 48GB in a 12-layer configuration with pin speeds of 14 Gbps, scalable to 16 Gbps, delivering 3.6 TB/s per stack [7]. Both are targeting mass production in 2027 [8].

The co-development component extends beyond memory specifications. The companies will apply Nvidia's CUDA-X libraries and PhysicsNeMo to accelerate semiconductor simulations and TCAD (technology computer-aided design) workflows. SK Hynix will also build factory digital twins using Nvidia's Omniverse platform and cuOpt optimization tools to pursue autonomous fab operations [2].

Market Concentration and Competitive Implications

SK Hynix controls approximately 58% of the global HBM market as of Q1 2026, according to Counterpoint Research. Samsung and Micron each hold roughly 21% [9].

Source: Counterpoint Research

Data as of Apr 1, 2026CSV

Nvidia commands an estimated 84–90% of the high-end AI GPU market [10]. When the dominant GPU maker formalizes a co-development partnership with the dominant HBM supplier, the structural implications for competitors are significant.

AMD's MI300X accelerators use HBM3 from Samsung and Micron, but neither supplier matches SK Hynix's volume or yield maturity in the latest HBM generations. Samsung has positioned itself as AMD's primary HBM partner, while Micron pursues broader qualification across customers [11]. Intel, facing deeper competitive challenges, partnered with a SoftBank subsidiary in June 2025 to form Saimemory, a startup developing a power-efficient HBM alternative — an acknowledgment that the existing supply chain may not serve its needs [11].

The concern is not that SK Hynix will refuse to sell to AMD or Intel. All three HBM manufacturers supply multiple customers. The concern is about priority. In a supply-constrained market — and Huang himself publicly urged SK Hynix to produce more HBM chips at the GTC Taipei keynote on June 1, 2026 [3] — co-development partners receive first access to leading-edge products, and their architectural requirements shape the specifications that other buyers must accept.

The HBM Market Explosion

The scale of the market at stake helps explain why this partnership matters. The global HBM market has grown from roughly $2.5 billion in 2022 to an estimated $38 billion in 2025, and is projected to reach $58 billion in 2026 and $100 billion by 2028 [12].

Source: TrendForce / Industry Estimates

Data as of Jun 1, 2026CSV

This growth is driven by increasing HBM content per AI accelerator and expanding deployment volumes. Each successive Nvidia platform has demanded more memory bandwidth and capacity, and the Vera Rubin generation continues that trajectory with HBM4 as its baseline [3].

Antitrust and Export Control Considerations

Nvidia's market position has attracted regulatory attention on multiple fronts. The FTC blocked Nvidia's $40 billion acquisition of Arm in 2022 over concerns about foreclosing IP licensing to competitors [10]. In late 2024, the European Commission sent questionnaires to Nvidia's customers and rivals probing whether the company bundles GPUs with other products or imposes restrictive conditions [13].

The Nvidia–SK Hynix partnership does not involve an acquisition or exclusive supply arrangement, which limits the surface area for conventional antitrust intervention. But critics of Nvidia's market power argue that long-term co-development agreements with key suppliers function as soft lock-ins, shaping product roadmaps around one customer's architecture and making it harder for competitors to access equivalent technology on equivalent timelines [10].

On export controls, the partnership operates within a relatively permissive framework. Under the Bureau of Industry and Security's AI Diffusion Rule, South Korea is classified as a Tier 1 country — the most trusted category — alongside the U.S. and 18 allied nations, exempting organizations in those countries from AI chip procurement restrictions [14]. However, BIS has expanded the Foreign Direct Product Rule to cover semiconductor manufacturing equipment and chips, with controls extending to South Korean firms operating in China [15]. Both Samsung and SK Hynix maintain significant DRAM production facilities in China, though both companies have been gradually shifting Chinese operations toward commodity products while concentrating advanced manufacturing — including HBM — exclusively in Korean facilities [16].

The co-development of AI-specific memory between a U.S. chip designer and a Korean manufacturer does not, by itself, trigger additional licensing requirements. But any downstream transfer of co-developed technology or manufacturing know-how to restricted entities would fall under existing controls — a compliance consideration that both companies must manage as the regulatory environment continues to tighten.

Precedents: Apple-TSMC and Vertical Integration

The Nvidia–SK Hynix partnership invites comparison to Apple's relationship with TSMC, which has defined the mobile semiconductor era. Apple invested over $20 billion and grew its silicon engineering team from approximately 500 to more than 3,000 engineers between 2015 and 2024 to execute its vertical integration strategy [17]. By designing its own chips and manufacturing exclusively at TSMC, Apple gained architectural control and pricing advantages that competitors relying on off-the-shelf processors could not match [17].

The analogy has limits. Apple owns its chip designs outright and contracts manufacturing. Nvidia co-develops memory specifications but does not manufacture or hold exclusive rights to the resulting products. SK Hynix remains free to sell HBM to other customers, and Nvidia has certified all three HBM suppliers for Vera Rubin [3].

The more precise precedent may be TSMC's joint development programs with leading-edge customers, where the foundry co-optimizes process technology with one or two anchor clients before making the node available to others. The anchor clients — historically Apple and, increasingly, Nvidia itself — get access first, which translates into a time-to-market advantage that can persist for 12–18 months. The Nvidia–SK Hynix memory partnership appears to follow a similar pattern: co-development yields early access and architectural alignment, without formal exclusivity.

South Korea's Strategic Calculus

The partnership lands in the middle of South Korea's most ambitious semiconductor investment cycle. SK Hynix is building a 19 trillion won ($13 billion) advanced packaging facility (P&T7) in Cheongju, North Chungcheong Province, specifically for HBM production [18]. Beyond that, SK Group Chairman Chey Tae-won has announced plans to expand investment in the Yongin semiconductor cluster from an initial 128 trillion won to 600 trillion won ($410 billion) over 25 years through 2050 [19].

The South Korean government's Semiconductor Cluster Master Plan, announced in 2024, commits $21 billion in public infrastructure investment through 2030, including high-speed rail, industrial water supply, and electricity grid reinforcement [16]. The government expects these facilities to produce 7.7 million wafers per month by 2030 and create at least 3 million jobs over two decades [20].

New hires at SK Hynix are being assigned to production bases in Icheon, Yongin (both in Gyeonggi Province), and Cheongju, with the company recruiting high school and vocational college graduates for positions starting in mid-2026 [20]. Average bonuses have reached approximately 700 million won ($510,000) per employee before tax, making SK Hynix one of the most sought-after employers in South Korea [20].

For Seoul, the Nvidia partnership reinforces Korea's position as the indispensable node in the AI memory supply chain. But it also deepens Korea's exposure to a single customer relationship and to the broader AI capex cycle.

The Downside Risk: What Happens If AI Spending Slows?

Hyperscaler capital expenditure is projected at $635–690 billion in 2026, a 67–74% increase over 2025 [21]. Google Cloud alone reported a backlog near $460 billion in Q1 2026, nearly double the prior year [21]. These are contracted commitments, not speculative projections, which provides a demand floor for the near term.

But the gap between hyperscaler spending growth and AI revenue growth has drawn scrutiny. Federal Reserve officials have flagged the disconnect, and analysts have noted that the biggest single buyer of AI compute may be growing slower than the buildout assumes [21]. If AI adoption slows, compresses returns, or produces volatile markets, the conviction behind current capex programs will be tested.

Neither Nvidia nor SK Hynix has disclosed the contractual protections governing their partnership — minimum purchase commitments, exit provisions, or how IP ownership is allocated. In the absence of such disclosure, the risk distribution remains opaque. SK Hynix's massive capital investments in Cheongju and Yongin are calibrated to a world where HBM demand continues its exponential growth trajectory. If that trajectory flattens, the question of who bears the excess capacity risk — the memory manufacturer, the chip designer, or the hyperscaler end customers — becomes material.

Historical precedent from the DRAM industry offers a cautionary note. Memory manufacturers have repeatedly suffered severe cyclical downturns when supply overshot demand, and the HBM market's concentration in a single application (AI training and inference) makes it more vulnerable to demand shifts than diversified DRAM markets.

What This Means

The Nvidia–SK Hynix partnership codifies an existing relationship rather than creating a new one. But formalization matters. It signals to investors, competitors, and regulators that the two companies intend to align their technology roadmaps for years to come. For AMD, Intel, and other AI chip challengers, it raises the bar on memory access and co-optimization. For Samsung and Micron, it underscores the competitive cost of being the second- and third-largest HBM suppliers in a market where the largest buyer prefers the largest seller.

The deal also highlights a tension at the heart of the AI hardware boom: the technologies enabling the most consequential computing platform since the internet are being built through an increasingly narrow set of bilateral relationships — Nvidia and TSMC for logic, Nvidia and SK Hynix for memory, ASML and TSMC for lithography. Each link in this chain is commercially rational. Taken together, they describe a supply chain where concentration is the feature, not the bug — until it becomes a vulnerability.