Vera Rubin enters full production as connectivity becomes AI's next battleground
Zero One Daily Intelligence Brief — June 3, 2026
Vera Rubin enters full production as connectivity becomes AI's next battleground
Top Market Signals
Vera Rubin hits full production, yet NVIDIA says demand still outruns supply
NVIDIA's next-generation Vera Rubin platform is now in full production, but the company says global demand for AI compute already far exceeds everything the industry can build, and it is leaning on Taiwan's supply chain to scale faster.
Speaking at GTC Taipei, NVIDIA AI and data center GPU marketing director Shar Narasimhan said the platform was deliberately simplified to lift throughput: a manufacturing step that took two hours on the prior Blackwell architecture now runs about 20 times faster on Vera Rubin. He framed Taiwan's ecosystem as indispensable, crediting it with letting NVIDIA expand the new systems on the show floor.
The architecture is built for agentic AI, where every inference has to be checked. Vera Rubin pairs the Rubin GPU with a standalone 88-core Vera CPU dedicated to tool-calling and answer verification, feeding results back to the GPU in real time. Narasimhan reframed Jensen Huang's old line "the more you buy, the more you save" as "the more you buy, the more you earn," arguing the platform delivers both the lowest cost per token and the highest token throughput.
The read for Taiwan is volume, not just prestige. NVIDIA says anyone with tokens to sell is sold out, and at COMPUTEX Huang told employees the entire Taiwan ecosystem is now joining NVIDIA's CPU journey. The constraint on the AI buildout is shifting from chip design to how fast Taiwan can add capacity.
Linked stocks: NVDA US, 2330 TT, 2317 TT
Sources: 全球 AI 算力短缺，輝達：仰賴台灣生態系擴增產能, 輝達辦桌 黃仁勳：台灣生態系加入 CPU 旅程
NVIDIA ships TSMC-built CPO switches as optical interconnect goes mainstream
NVIDIA has started shipping co-packaged optics (CPO) switches built on TSMC's silicon-photonics platform, signaling that data movement between racks, not raw compute, is the next constraint on AI performance.
NVIDIA networking SVP Gilad Shainer said the company now treats data transfer as the next technical battleground. Its new Spectrum-X CPO switch, with throughput up to 400 Tb/s (terabits per second), has begun shipping to close partners, with capacity expanding in 2H26. The design integrates the optical engine directly with the switch chip rather than bolting optical transceivers on the outside, cutting power draw and transmission distance. TSMC plays the gating role through its new COUPE (Compact Universal Photonic Engine) packaging platform.
The ecosystem moved in the same direction on the same day. Silicon-photonics firms Lightmatter and Ayar Labs both joined NVIDIA's NVLink Fusion program, making their CPO products compatible with NVIDIA's optical and SerDes interfaces. Lightmatter says its approach cuts fiber and connector counts by 50% and lets a customer's semi-custom XPU link to NVIDIA switch chips across vendors. Ayar Labs is targeting the rack-scale custom-silicon case.
Shainer was blunt about limits: copper stays the default for short, in-rack links over NVLink because it is cheap, reliable and low-power. Optics earn their place only when connecting thousands of GPUs across racks. That split is exactly where the optical names and Taiwan's packaging chain now sit.
Linked stocks: NVDA US, 2330 TT, COHR US, LITE US
Sources: 輝達結盟台積電布局 CPO 交換器，下半年產能點火, TechNews
Marvell jumps 32% after Huang calls connectivity the deciding factor in AI
Marvell (MRVL US) surged about 32% to a record after Jensen Huang publicly endorsed it on stage, the clearest sign yet that investors are repricing the companies that move data between accelerators rather than the accelerators themselves.
Sharing a stage with Huang, Marvell's chief executive heard the NVIDIA founder argue that the future of AI depends not on computing power but on connectivity. The market took it as a green light: Marvell makes custom AI silicon and the high-speed interconnect that stitches large GPU and XPU clusters together, the same bottleneck NVIDIA is attacking with co-packaged optics.
The move fits a broader rotation. Commentators flagged Marvell and Broadcom (AVGO US) as the names that benefit as hyperscalers shift spend toward custom accelerators and the networking around them. The risk is valuation: the stock is now priced for flawless execution, and a single quarter that misses the new bar could unwind the move as fast as Huang's comment built it.
Linked stocks: MRVL US, NVDA US, AVGO US
Sources: Google News, Google News
Intel launches 288-core Clearwater Forest on 18A as the CPU reclaims AI relevance
Intel (INTC US) used COMPUTEX to launch its Xeon 6+ "Clearwater Forest" server CPUs, with up to 288 cores and built on its own Intel 18A process, and the stock rose more than 8% as a new narrative took hold: the CPU is becoming the bottleneck for AI agents.
Agentic workloads spend heavily on orchestration, tool-calling and verification, tasks that run on CPUs rather than GPUs. Intel leaned into that, pairing Xeon 6+ with Ethernet upgrades aimed at agentic AI. Its CFO separately pointed to firming CPU demand and a ramping 18A supply, the two pillars of the company's foundry comeback.
Not everything is clean. Intel's client chips, Panther Lake and Wildcat Lake, are reportedly in short supply, and analysts note tight laptop-chip output is testing the 18A ramp just as demand returns. Intel also confirmed a data center GPU, Crescent Island, that supports up to 480GB of LPDDR5X and skips the HBM that NVIDIA relies on, a deliberate bet on cheaper memory for inference. CEO Lip-Bu Tan, meanwhile, still calls TSMC a key partner, a reminder that the comeback is not yet self-sufficient.
Linked stocks: INTC US, NVDA US, 2330 TT
Sources: Google News, Google News
SK Hynix to double wafer capacity in five years as Huang pleads for more HBM
SK Group chairman Chey Tae-won said SK Hynix (000660 KS) will double its wafer capacity within five years, the most concrete supply commitment yet from the company that has become the lead supplier of AI memory.
The pledge landed at COMPUTEX, where Huang left a handwritten "please make more HBM" note on the SK Hynix booth, a public signal of how tight high-bandwidth memory remains. SK Hynix is widely described as the new memory leader on the strength of HBM, and the capacity plan is the supply-side answer to demand that NVIDIA says it cannot fill.
The commitment is showing up in hard spending. Air Liquide announced a new industrial-gas plant in South Korea, reported at about US$233m, dedicated to supporting SK Hynix's AI memory production. SK Group is also reported to have paused the planned sale of wafer maker SK Siltron, choosing to keep a strategic materials asset rather than raise cash, a sign that AI strategy is now overriding financial restructuring across the group.
Linked stocks: 000660 KS, NVDA US, 005930 KS
Sources: Google News, Google News
Microsoft launches in-house AI models and a quantum chip to loosen its OpenAI dependence
Microsoft (MSFT US) used its Build 2026 conference to launch its own MAI family of AI models, explicitly aimed at reducing reliance on OpenAI, and to unveil Majorana 2, a topological quantum chip it says is a thousand times more reliable than its predecessor.
The MAI models put Microsoft in direct competition with the partner it has funded for years. AI chief Mustafa Suleyman framed the goal plainly, saying he wants Microsoft to be the fourth lab that matters alongside OpenAI, Google and Anthropic. Microsoft also showed Project Solara, a chip-to-cloud platform built for "agent-first" enterprise devices designed to run AI agents rather than traditional apps.
Majorana 2 is the harder signal to read. Microsoft says the new material-stack design lifts qubit reliability by roughly a thousand times and keeps it on track to deliver a quantum computer by 2029, with development assisted by its own agentic AI research platform. On the hardware side Microsoft is still tied to NVIDIA: its new Surface Laptop Ultra and Dev Box run NVIDIA's RTX Spark chips. The throughline is a company trying to own more of its own stack, from models to qubits, while staying dependent on NVIDIA for silicon.
Linked stocks: MSFT US, NVDA US
Sources: 微軟發表新一代量子晶片Majorana 2 可靠性提升千倍, Google News
先聲 First Word — Exclusives from Chinese-Language Sources
Yuancheng Semiconductor lands AMD and MediaTek as it chases all-optical CPO
Silicon-photonics design house Yuancheng Semiconductor turned profitable in 2025 and has secured both orders and funding from AMD and MediaTek, using an asset-light, design-led model to target all-optical CPO for the 800G and 1.6T transitions. A Taiwanese fabless startup winning backing from two major chip customers shows how the optical-interconnect supply chain is forming around the same companies now joining NVIDIA's NVLink Fusion. (3 Jun 2026) Source: 左擁超微、右抱聯發科 元澄半導體直攻全光CPO設計、卡位1.6T商機
Phison CEO warns 2027 chip shortage will be worse than 2026
Phison (8299 TT) chief executive Pan Jian-cheng said the market has reached a point where "even at high prices you may not get supply," and that the 2027 supply-demand imbalance could be more severe than 2H26, as the company shifts from a pure IC component vendor to a system-level AI solutions provider. The warning matters because Phison sits at the controller layer of NAND storage, giving it an early read on memory tightness spreading down the chain. (3 Jun 2026) Source: 潘健成：2027年缺貨痛苦更甚2026 群聯啟動系統級AI轉型
China's CXMT closes the HBM gap with Samsung to about three years
CXMT's fourth-generation HBM (HBM3) is reportedly still in testing with commercialization at risk of slipping, but Korean media now put the China-Korea HBM technology gap at roughly three years, narrower than widely assumed, with Beijing said to have issued an "HBM mobilization order." For investors pricing the memory super-cycle on a long Korean and US duopoly, a faster-than-expected Chinese catch-up is the main risk to terminal pricing. (3 Jun 2026) Source: 長鑫存儲與三星HBM差距剩3年 良率低但技術已達HBM3
MiTAC to start US server-rack shipments from two new plants in 3Q26
MiTAC Holdings (3706 TT) unit Mitac Computing will bring two new US plants online in 3Q26, shipping AI server racks, with management calling full-year operating growth a near-certainty on sustained AI server demand. The detail adds to a steady drumbeat of Taiwanese ODMs building US capacity, locating AI server assembly closer to the hyperscaler customers placing the orders. (3 Jun 2026) Source: cnYES 鉅亨網
© 2026 Zero One Investment Research Pte Ltd. For informational purposes only. Not investment advice.