1. OpenAI Unveils MRC Protocol — The Opening Shot in AI Supercomputer Network Infrastructure Standards Competition
• Core Source
“AMD, Broadcom, Intel, Microsoft, and NVIDIA jointly developed and released MRC (Multipath Reliable Connection)”
“Introducing a multi-plane structure that splits 800Gb/s links into multiple 100Gb/s planes, enabling connection of over 100,000 GPUs with a two-tier Ethernet architecture”
“Instead of conventional BGP-based dynamic routing, SRv6-based static source routing is applied, enabling microsecond-level failure rerouting”
“Already deployed in Texas Abilene Stargate and Microsoft Fairwater supercomputers. In actual frontier model training environments, training continued even during link flaps and switch reboots”
“OpenAI emphasized that the AI infrastructure competition is shifting beyond simple GPU procurement toward network utilization and cluster efficiency”
• Expected Impact
As AI training clusters scale beyond tens of thousands of GPUs toward 100,000 or more, network bottlenecks have emerged as a variable just as critical as GPU performance. In conventional InfiniBand or simple Ethernet architectures, scaling large clusters increases switch layers and wiring complexity exponentially, driving up costs and failure risks in tandem. MRC resolves this by simplifying the architecture to a two-tier Ethernet structure while eliminating congestion through packet distribution across hundreds of paths, and enabling uninterrupted training operations via microsecond-level failure rerouting.
For investors, there are two key implications. First, this protocol being released as open source through OCP (Open Compute Project) makes it a strong candidate to become an industry standard. The fact that five companies — AMD, Broadcom, Intel, Microsoft, and NVIDIA — co-developed it signals an orientation toward an open ecosystem rather than vendor lock-in, and a trend may form where hyperscalers and AI startups alike build infrastructure on top of this standard. Second, the beneficiary landscape for Ethernet-based AI networking equipment becomes clear. While AI cluster networks were previously dominated by NVIDIA’s InfiniBand, MRC formalizes an Ethernet-based alternative, signaling a structural shift that distributes benefits across Broadcom (Ethernet switch chips), Arista Networks, Cisco, and the broader Ethernet ecosystem. The fact that it has already been deployed in Stargate and Fairwater — among the world’s largest AI supercomputers — confirms commercial validation is complete, clearly showing that the next phase of the AI infrastructure investment cycle is moving from ‘GPU procurement’ to ‘network efficiency optimization’.
2. TSMC Pursues Expansion of U.S. Investment to $250 Billion — Accelerating Semiconductor Supply Chain Americanization and Advanced Node Dominance
• Core Source
“Taiwanese media reported that TSMC’s investment scale in the U.S. could grow to $250 billion”
“The existing U.S. investment plan was approximately $165 billion in total”
“Plans to build 6 wafer fabs, 2 packaging factories, and 1 R&D center in the U.S. Mass production at Arizona’s first advanced packaging factory is targeted for 2028, with the second factory planned for 2029–2030”
“TSMC is converting Fab 15A in the Central Taiwan Science Park from a 28/22nm-focused line to a 4nm-capable fab. Phase 2 construction of the 1.4nm (N14) fab is also progressing faster than expected. Pilot production is possible as early as Q3 2027, with mass production targeted for the second half of 2028”
“Sources suggest TSMC may replicate the Hsinchu Science Park cluster model in Phoenix, Arizona through its expanded U.S. investment”
• Expected Impact
The prospect of TSMC’s U.S. investment scale expanding from $165 billion to $250 billion — an increase of approximately 52% — goes beyond a simple investment hike. It is a key indicator for gauging the speed and scale of global semiconductor supply chain restructuring. Combined with U.S. government support under the CHIPS Act, plans to build six wafer fabs and advanced packaging factories centered on Arizona effectively aim to transplant the Taiwan Hsinchu Science Park ecosystem to U.S. soil.
From a supply chain investor perspective, two points stand out. First, advanced packaging capacity (CoWoS, SoIC, etc.) will also be built within the U.S. One of the current bottlenecks in AI accelerators is the shortage of TSMC’s CoWoS advanced packaging capacity, and the establishment of two packaging factories in the U.S. will partially alleviate this bottleneck while reinforcing a geopolitical safety net for the U.S. domestic AI infrastructure supply chain. Second, the faster-than-expected progress on the 1.4nm (N14) process. With pilot production targeted for Q3 2027 and mass production for H2 2028, this suggests TSMC will continue to widen its process technology lead over Intel and Samsung even beyond the 2nm (N2) generation. As long as TSMC’s near-monopoly process leadership is maintained, the structural benefits for ASML (EUV equipment) and advanced packaging materials and equipment companies will persist in tandem.
3. Anthropic’s Annualized Revenue Surges 80x, Valuation Surpasses $1.2 Trillion — Reshaping the Competitive Landscape of the AI Foundation Model Market
• Core Source
“Anthropic’s implied valuation in the on-chain Pre-IPO market has surged to $1.2 trillion, surpassing OpenAI by approximately 20% for the first time”
“Anthropic’s Q1 annualized revenue surged 80x YoY, far exceeding the original target of 10x growth”
“Over 300 megawatts of capacity and more than 220,000 NVIDIA GPUs of new computing resources secured within one month, immediately utilized to raise usage limits for Claude Pro and Max subscribers”
“Anthropic’s revenue forecast for this year: $30 billion+. Over 80% of Anthropic’s revenue is B2B”
“Computing resources to be secured by next year: 7–8GW. Significantly short compared to OpenAI’s 2030 target of 30GW”
• Expected Impact
The fact that Anthropic’s annualized revenue exceeded its target by 8x while its valuation surpassed OpenAI for the first time is an important signal that the AI foundation model market is restructuring from a single-dominant-player structure into a multi-strong-player competitive regime. The fact that over 80% of revenue comes from B2B in particular means Anthropic has succeeded in differentiating its positioning in the enterprise AI market from ChatGPT.
For investors, the key focus is the structure in which demand explosions cascade through the entire computing supply chain. Anthropic urgently leased the SpaceX Colossus 1 data center (300MW+, 220,000+ GPUs) to meet demand, and has already signed computing contracts with Amazon (up to 5GW), Google-Broadcom (5GW), MS-NVIDIA ($30 billion), and FluidStack ($50 billion). The reality that the demand from just a single AI model company absorbs tens of gigawatts of infrastructure provides evidence that the combined annual CapEx guidance of $590–620 billion announced by the Big Four is by no means overinvestment. Meanwhile, Anthropic’s achievable computing resources through next year (7–8GW) fall far short of OpenAI’s target (30GW), suggesting that computing infrastructure procurement capability will become the ultimate determining variable in next-generation AI model competition.
4. ARM’s AGI CPU Demand Doubles to $2 Billion, Yet Supply Chain Bottleneck Exposed — Structural Growth and Near-Term Constraints in the AI CPU Market
• Core Source
“Arm stated in its earnings release that combined revenue from its in-house data center chip ‘AGI CPU’ for FY2027–2028 is expected to reach $2 billion. This is double the forecast provided at the time of its March launch.”
“The company disclosed that demand for the product had expanded from the original $1 billion to over $2 billion, but the short-term revenue guidance was not raised due to supply chain issues including wafers and memory”
“Foundry wafer production capacity at companies like TSMC has already reached its limits. Arm is understood to have secured over $2 billion in CPU demand across FY27–FY28, but there are limits to how much of that demand can actually be converted into revenue”
“ARM reported record quarterly and annual results with revenue up 20% YoY. Licensing revenue grew +29% and royalty revenue +11%, with data center royalties more than doubling YoY”
• Expected Impact
ARM’s latest results numerically confirmed that the role of CPUs in AI data centers is being elevated from simple computational support to the core axis of workload orchestration and agentic operations. Data center royalties more than doubled YoY, and the demand forecast for AGI CPU doubled within just a few months of launch. This is because ARM’s architecture holds advantages in power efficiency over x86, and its design direction aligns with the demand requirements of the AI agent era.
However, the structural gap between demand and supply is the core risk. With TSMC and other foundries’ advanced node wafer capacity already at its limits, it may take considerable time before ARM’s secured $2 billion in demand converts into actual revenue. This is not simply ARM’s problem alone — it reflects intensifying competition for advanced node wafer capacity among all fabless companies including NVIDIA, AMD, Apple, and Qualcomm. For investors, it is important to understand the background behind ARM’s failed near-term guidance raise, and over the medium-to-long term, the pace at which TSMC expands its advanced node capacity will be the key variable determining ARM’s revenue conversion timeline.
5. Physical Limits of CoWoS Accelerate Transition to Glass Substrate (CoPoS) — Tectonic Shift in Next-Generation AI Chip Packaging Supply Chain
• Core Source
“As AI chip die sizes continue to grow, existing CoWoS packaging is approaching its physical limits. Industry leaders led by TSMC are accelerating the transition to CoPoS technology, which uses rectangular glass substrates instead of circular silicon interposers”
“Glass substrates enable larger sizes, lower signal loss, and stronger resistance to warpage. TSMC has already begun operating a pilot production line, with large-scale mass production expected in 2028. The industrial supply chain transition has already substantively begun”
“When substrate companies order CCL, delivery lead times that were previously around 2 weeks have expanded to a maximum of 6 weeks”
“In the high-end product segment, Japan’s Nitto Boseki virtually monopolizes supply based on quality certification from major Big Tech companies”
• Expected Impact
CoWoS (Chip on Wafer on Substrate) is the core advanced packaging technology currently applied in cutting-edge AI accelerators such as NVIDIA Blackwell, but as AI chip die sizes continue to grow, the physical size limitations of the CoWoS structure using circular silicon wafers as interposers have come into view. CoPoS (Chip on Panel on Substrate) replaces this with rectangular glass substrates, offering larger area realization, lower signal loss, and superior flatness, emerging as the core technology for next-generation AI accelerator packaging.
From a supply chain perspective, this transition carries broad ripple effects. With TSMC already operating a CoPoS pilot production line and targeting mass production for 2028, the restructuring of core materials supply chains — including glass substrates, CCL (Copper Clad Laminate), and T-glass — is proceeding ahead of the curve. In particular, the low-thermal-expansion glass fiber (T-glass) required for high-performance semiconductor substrates is currently virtually monopolized in supply by Japan’s Nitto Boseki, and the fact that CCL delivery lead times have expanded from the previous 2 weeks to a maximum of 6 weeks shows that supply is already tight relative to surging demand. As the transition to glass substrates accelerates in earnest, companies with monopolistic positions in materials are expected to see simultaneous rises in bargaining power and profitability, ushering in a structural beneficiary phase.
Comment [0]