Foxconn, NVIDIA CPO Switch Partner & More (0513)

1. Foxconn Secures Exclusive ODM for NVIDIA’s CPO Switches — Shipment Target Raised to 50,000+ Units in 2026–2027, Emerging as New Bottleneck in AI Networking Infrastructure

• Core Source

“The expected shipment volume for 2026 has been significantly raised from ‘over 10,000 units’ to ‘50,000 units or more’ for 2026–2027 combined.”

“Demand is so overwhelming that even display units are being handed over to NVIDIA, leaving the company in a ‘zero inventory’ state.”

“Foxconn is the sole ODM (design and manufacturing) partner for NVIDIA’s CPO switches.”

“CPO products carry a high average selling price (ASP) and double-digit operating margins, significantly exceeding the 5–8% margins of conventional server OEM businesses.”

“CPO business is expected to account for over 15% of Foxconn Industrial Internet (FII) revenue in 2026.”

• Expected Impact

CPO (Co-Packaged Optics) is a technology that integrates optical communication components into the same package as semiconductor chips, making it essential for transmitting vast amounts of data at ultra-high speed and low power consumption within AI clusters. As AI cluster scale expands from thousands of GPUs to tens of thousands, inter-chip communication speed and power efficiency become bottlenecks just as critical as GPU performance itself. CPO maximizes transmission efficiency while reducing energy consumption compared to conventional copper cables or separate optical module configurations.

Foxconn’s strategic value lies not merely in being a manufacturer, but in its exclusive position as NVIDIA’s sole CPO switch ODM partner. The fact that supply cannot keep pace with demand — to the point where even Computex display units are being handed over to NVIDIA — starkly illustrates the severity of the supply shortage. With operating margins at double digits, more than double the existing server business (5–8%), this is transforming Foxconn’s revenue structure from an assembly-focused contract manufacturer to a high-value-added AI infrastructure components supplier. The industry views 2026 as an inflection point for data center optical communications, and Foxconn, having already secured first-mover advantages, is well-positioned for sustained structural benefits.

2. Google and Anthropic Partner with SpaceX on Space Data Centers — AI Infrastructure Race Expands from Ground to Orbit, Starship Success Becomes the Pivotal Factor

• Core Source

“SpaceX and Google are in discussions regarding a launch contract related to Google’s space data center.”

“Google is advancing ‘Project Suncatcher,’ with plans to launch a prototype space data center satellite by next year.”

“Following last week’s computing partnership with Anthropic, Google’s move to leverage SpaceX’s space infrastructure capabilities further solidifies SpaceX’s position as a space infrastructure provider.”

“SpaceX now has two options: building space data centers for clients, or investing in space data centers to enhance Grok’s performance.”

“Falcon 9 cannot support space data center construction. Therefore, the success and increased launch frequency of the Starship program must be the top priority.”

• Expected Impact

The competitive frontier in AI data centers is expanding beyond Earth. Ground-based data centers face three physical constraints: power, cooling, and land, and space data centers are gaining attention as an alternative that can simultaneously address cooling and land limitations. In space, solar power generation and radiative cooling can be utilized, offering long-term energy cost reduction potential.

Strategically, the more significant shift is SpaceX’s business model transformation. Whereas the previous model involved SpaceX investing its own capital to build space data centers for operating Grok, the company is now positioning itself as an infrastructure provider receiving advance orders from external clients — Google and Anthropic. The fact that AI companies in direct competition with SpaceX are still relying on its infrastructure demonstrates that SpaceX’s launch capabilities represent an irreplaceable competitive moat at this point in time. However, all these scenarios are contingent on the successful commercialization of the Starship program. Since Falcon 9 cannot support space data center construction, the Starship V3 test launch result has emerged as a near-term critical catalyst.

3. NVIDIA’s Blackwell·Rubin Pipeline Surpasses $1 Trillion — GW-Based Data Center Deployment Model Drives Significant Upward Revisions to FY27–29 Revenue Estimates

• Core Source

“A pipeline of over $1 trillion has been established through 2027. Even after FY4Q26, the remaining order backlog is estimated at over $840 billion.”

“Deployment capacity outlook: 9.2GW in 2026 → 15.7GW in 2027 → 20.8GW in 2028 → 25.2GW in 2029.”

“Capacity is expected to grow by approximately 3GW per quarter. Analysts calculate approximately $25 billion in revenue generated per GW.”

“FY2027: Annual revenue of $378.9 billion, EPS of $8.45 estimated (consensus: revenue $363.4 billion, EPS $8.21).”

“FY2029: Annual revenue of $655.4 billion, EPS of $14.85 — newly introduced as a forecast.”

• Expected Impact

Wells Fargo introduced a new analytical framework that uses GW (gigawatt) of infrastructure deployment capacity — rather than conventional financial metrics — as the core revenue driver. The fundamental logic is straightforward: in a market where demand overwhelmingly exceeds supply, NVIDIA’s revenue is determined not by how much customers want to buy, but by how much NVIDIA can produce and deploy. Applying the formula of approximately $25 billion per GW, as deployment capacity expands from 9.2GW in 2026 to 25.2GW in 2029, the revenue ceiling mechanically scales upward.

Citi pointed in the same direction, projecting April-quarter revenue at $80 billion (above the consensus of $78.6 billion) and July-quarter at $89 billion. With hyperscaler capex growth projected at 83% in 2026 and 57% in 2027, NVIDIA’s FY29 revenue estimate of $655.4 billion exceeds the market consensus of $517.8 billion by 26%. This is not merely a target price upgrade — it reflects a structural conviction that NVIDIA’s results will continue to outpace consensus over the next three years. For retail investors, the key takeaway is that the successive product cycles from Blackwell to Rubin to Rubin Ultra are structured to convert demand into revenue without a cliff, in continuous succession.

4. Broadcom’s ASIC Demand Explodes — AI Revenue Projected at $115 Billion in 2027 and $180 Billion in 2028, Rising as a Formidable Rival to GPUs

• Core Source

“AI revenue is estimated to grow from approximately 49% of total revenue currently to approximately 81% by F4Q28.”

“The combined AI revenue from Google and Anthropic is modeled at approximately $80 billion in 2027, with total AI revenue reaching approximately $115 billion — exceeding the prior forecast of $100 billion — and $180 billion in 2028.”

“AVGO has four major clients — Google, Meta, Anthropic, and OpenAI — along with two undisclosed clients (believed to include ByteDance). We understand that AVGO is currently in discussions with three additional clients on custom AI chip collaboration.”

“ASIC units, starting from a lower base, are expected to grow at a 74% CAGR as various ASIC programs come online, particularly from 2027 onward.”

• Expected Impact

Broadcom is emerging as the largest beneficiary of the custom ASIC (Application-Specific Integrated Circuit) market — an alternative to NVIDIA GPUs. Broadcom leads the design and supply of Google’s TPU, Meta’s MTIA, and chips for Anthropic and OpenAI, with collaboration on three additional undisclosed clients reportedly in progress.

The reason ASICs are drawing attention over GPUs is clear. ASICs optimized for specific AI workloads offer superior power and cost efficiency compared to general-purpose GPUs, while also fulfilling hyperscalers’ strategic objectives of supply diversification and proprietary technology internalization. Citi projected the total AI accelerator TAM at $603 billion by 2028, with ASIC revenue reaching $150 billion (94% CAGR) — far outpacing GPU growth (30% CAGR). Broadcom has effectively blocked competitor entry by securing a five-year long-term agreement (LTA) with Google, and margins are improving as the revenue recognition model transitions from rack-level to chip-level sales. The expansion of Broadcom’s AI revenue share from 49% to 81% of total revenue represents a comprehensive structural transformation of its business.

5. Samsung Electronics Union Strike Imminent — DRAM Output May Fall 3–4%, Operating Profit Impact Estimated Up to 30 Trillion Won, Micron and SK Hynix Set to Benefit

• Core Source

“Samsung Electronics’ largest union, representing approximately 36,000 members, has threatened an up to 18-day strike beginning May 21, arguing that bonuses and wages are insufficient despite improved earnings driven by AI demand.”

“Some analysts project that a prolonged strike could reduce DRAM output by approximately 3–4%, with potential losses reaching up to $20 billion.”

“Assuming Samsung accepts the union’s demands (distributing 10–15% of operating profit as OPI incentives and a 5% base salary increase), additional labor costs are estimated at 17–30 trillion won compared to the April disclosed estimates.”

“2Q26 estimated contract prices are being presented at DRAM up 58–63% quarter-over-quarter and NAND up 70–75% quarter-over-quarter — figures that exceed our prior quarter-over-quarter increase estimate of 40–50%.”

• Expected Impact

The Samsung Electronics union strike is not merely a labor dispute — it is an event that could deliver a direct shock to the global memory supply chain. With the global memory market already in a supply shortage driven by surging AI server demand, a strike of up to 50,000 participants lasting 18 days could reduce DRAM output by 3–4%. JP Morgan estimated the operating profit downside risk at 6–10%, with total impact — including lost revenue opportunities — reaching up to 35 trillion won.

Paradoxically, however, the more supply disruption materializes, the greater the upward pressure on memory contract prices. With 2Q26 DRAM contract prices already projected to rise 58–63% quarter-over-quarter, additional supply disruptions could further intensify the price surge. This translates into a dual benefit for competitors Micron and SK Hynix: market share expansion and pricing upside. For retail investors, this is not a moment to view the Samsung strike solely as a risk — it is a time to simultaneously observe whether it serves as an extension of the memory up-cycle and how competitive dynamics are reshaping.

6. Ajinomoto Raises ABF Prices 30%, IC Substrate Industry Enters ‘Super Expansion Cycle’ — AI Chip Packaging Layers Expand from 3+3 to 11+11, Structural Demand Growth Underway

• Core Source

“Ajinomoto has decided to raise prices for its core ABF build-up film by 30%, with the new prices to take effect from Q3 2026.”

“Ajinomoto holds over 95% market share in the global ABF market, giving it strong pricing power across the entire supply chain.”

“As the number of layers in AI chip packaging expands from the existing 3+3 layers to 11+11 layers, and is expected to further increase to 13+13 layers after 2030, ABF demand is projected to continue its structural growth trajectory.”

“The industry forecasts that the IC substrate sector will enter what it calls a ‘Super Expansion Cycle’ over the next 2–3 years.”

“Ajinomoto has announced plans to build a third ABF plant in Gifu Prefecture, Japan, to address demand beyond 2030.”

• Expected Impact

ABF (Ajinomoto Build-up Film) is a critical material in AI chip packaging, essential for stacking circuit layers within semiconductor substrates. Ajinomoto is a de facto monopoly supplier holding over 95% market share, meaning buyers have no alternative but to absorb this 30% price increase.

The root cause of the price increase lies in the structural expansion of demand. As AI chip packaging layer counts surge from the conventional 3+3 layers to 11+11 layers now, and 13+13 layers beyond 2030, the volume of ABF film required to manufacture the same chip multiplies several times over. This is not simply a matter of more AI chips being produced — it is a structural shift in which material consumption per chip is rising. Compounding this, simultaneous supply shortages are emerging across the upstream value chain, including glass fiber, copper foil, and drill bits, leading to a prevailing industry view that supply-demand imbalances in ABF will intensify rather than ease through 2027–2028. This is a clear illustration of how the ripple effects of AI infrastructure investment are reshaping the entire materials and substrate supply chain, far beyond chips and GPUs alone.