来自 Maksym Misichenko · Yahoo Finance ·
AI智能体对这条新闻的看法
While GlobalFoundries' SCALE platform holds promise in addressing AI's data transport bottlenecks, panelists express concerns about execution risks, competition, and the challenge of shifting GFS's valuation from a legacy foundry to a high-growth AI infrastructure play.
风险: Proving silicon photonics yields and reliability at hyperscale, competition from other photonics players, and demand hinging on hyperscalers' standardization on a single vendor.
机会: Potential re-rating of GFS's valuation if the SCALE platform gains traction with hyperscalers.
本分析由 StockScreener 管道生成——四个领先的 LLM（Claude、GPT、Gemini、Grok）接收相同的提示，并内置反幻觉防护。 阅读方法论 →
人工智能 (AI) 的 playbook 已经逐渐熟悉：构建更大的 GPU 集群。添加更多的 Blackwell 芯片。投入更多的电力来解决问题。如果芯片过热，就在河流旁边建造数据中心。如果带宽耗尽，铺设更多的铜线。
这就是亚马逊、Alphabet、Microsoft 和 Meta Platforms 在 2026 年解决 AI 的方式。而且它有效——直到遇到物理定律的限制。
人工智能会创造世界上第一个万亿富翁吗？我们的团队刚刚发布了一份关于一家鲜为人知但提供英伟达和英特尔都需要的关键技术的公司（被称为“不可或缺的垄断”）的报告。继续 »
一家公司看到了同样的问题，并得出了不同的答案。GlobalFoundries(NASDAQ: GFS) 认为，人工智能基础设施中的真正瓶颈不是计算能力，而是连接芯片的电线，并用光来替代那根电线。
无人提及的铜墙
在每个 AI 数据中心里，数千个芯片必须以巨大的速度共享信息。目前，大部分通信是通过铜线进行的，而铜线正在耗尽空间。它会产生热量，信号在距离上会损失，并且以在规模化时会变得痛苦的方式消耗电力。每次 AI 模型变得更大，铜线问题就会变得更糟。
行业已经知道这一点多年了。解决方案有一个名称：共封装光子器件 (CPO)。这个想法是将光发射器（通过光而不是电传输数据）直接放置在芯片旁边，从而将数据必须通过铜线传输的距离缩短到几乎为零。结果是更快速、更凉爽、更节能的 AI 基础设施。
2026 年 5 月，GlobalFoundries 宣布 SCALE -- 硅光子器件共封装先进光引擎解决方案 -- 这是业界首个满足人工智能规模化架构的光学互连多源协议规范的平台。该平台使用每个光纤上的粗波分复用 (DWDM) 和密集波分复用 (DWDM) 来推动带宽密度和可扩展性超越铜线所能做到的程度，并且 GlobalFoundries 已经展示了 8λ 和 16λ 双向 DWDM 在其平台上原生实现——该公司将其描述为实现所有后续工作的基石。
每个人都在追逐的堆栈部分
关于硅光子器件的一个被 GPU 报道忽略的事实是：它是一个制造问题，以及一个物理问题。设计硅光子器件很困难。以所需的精度，在体积上，为超大规模数据中心构建它，则更困难。
GlobalFoundries 花费了数年时间来开发能够做到这一点的工艺技术。其硅光子器件平台支持 50 Gbps 和 100 Gbps 微环调制器、宽带可拆卸光纤接口以及在波长计数扩展时能够实现平台未来防护的扁平插入损耗特性。2025 年 11 月，该公司在新加坡收购了 Advanced Micro Foundry，一家专门的硅光子器件制造商，增加了制造资产、知识产权和工程深度，这些资产和深度需要数年时间才能从头开始构建。
该收购为 GlobalFoundries 在新加坡提供了硅光子器件的生产能力，这对于在美中半导体紧张局势加剧的情况下实现供应链多元化至关重要。该公司正在构建超大规模客户需要并且只有少数制造商能够真正交付的平台。
GlobalFoundries 在一天内下跌了 10%，但我并不担心
由于 Mubadala 的 alleged 股票出售，GlobalFoundries 在 5 月 27 日下跌了近 10%。尽管股价下跌，但长期 Motley Fool 的框架在此处胜出。
GlobalFoundries 的故事仍然完好无损。SCALE 的宣布在几周前使股价在单一交易日上涨了 12%。随后出现了 2026 年第一季度盈利额的超预期。预计 2026 年硅光子器件收入将再次几乎翻倍，并且在 2025 年记录了超过 500 个设计成果，并且势头正在增强。抛售行为并不能改变这些事实。
此外，美国政府正在通过一项拟议的 3.75 亿美元拨款来支持 GlobalFoundries，以帮助建设国内量子制造基础设施。
每个主要的超大规模客户都在询问如何更快地训练更大的模型。GlobalFoundries 正在询问一个不同的问题：如何在不使基础设施熔化的前提下移动芯片之间的芯片数据？
共封装光子器件是答案。正在构建用于以规模化方式交付该平台的制造平台，在大多数情况下，仍然归类为“半导体代工厂”。在因与其最重要的业务无关的原因导致其下跌 10% 的日子里，那次下跌就变成了机会。
您现在应该购买 GlobalFoundries 的股票吗？
在您购买 GlobalFoundries 的股票之前，请考虑以下事项：
Motley Fool Stock Advisor 分析师团队刚刚确定他们认为投资者现在应该购买的 10 支最佳股票……而 GlobalFoundries 并非其中之一。这些能够产生巨大回报的 10 支股票可能在未来几年产生巨大回报。
请考虑当 Netflix 在 2004 年 12 月 17 日被列入此名单时……如果您当时投资了 1,000 美元，您将拥有 463,900 美元！* 或者当 Nvidia 在 2005 年 4 月 15 日被列入此名单时……如果您当时投资了 1,000 美元，您将拥有 1,294,401 美元！*
现在，值得注意的是 Stock Advisor 的总平均回报率为 978%——与标准普尔 500 指数相比，实现了 211% 的市场领先表现。不要错过最新的前 10 名名单，该名单可使用 Stock Advisor，并加入由个人投资者为个人投资者构建的投资社区。
Micah Zimmerman 对所提及的任何股票都没有持仓。Motley Fool 持有 Alphabet、Amazon、GlobalFoundries、Meta Platforms 和 Microsoft 的头寸。Motley Fool 有一份披露政策。
四大领先AI模型讨论这篇文章
"GFS's photonics opportunity hinges on unproven manufacturing scale-up against better-capitalized rivals, making the stock's volatility more noise than signal."
The article frames GlobalFoundries (GFS) as the differentiated play on AI infrastructure via its May 2026 SCALE CPO platform, DWDM milestones, and AMF acquisition, targeting copper interconnect limits that constrain hyperscalers like MSFT, AMZN, and GOOGL. Silicon photonics revenue is projected to nearly double in 2026 amid 500+ design wins. Yet the piece, from an outlet holding GFS, understates that this remains a small slice of foundry revenue, with unproven volume manufacturing at the precision needed. The Mubadala overhang and $375M quantum grant add noise rather than clarity on commercial traction.
TSMC and Broadcom are already shipping CPO prototypes to the same hyperscalers, so GFS's platform could be late or commoditized even if the physics case holds.
"GFS has a real technical edge in manufacturing silicon photonics at scale, but the article overstates both the urgency of the copper bottleneck and GFS's competitive moat against TSMC/Samsung, while ignoring margin and adoption-speed risks."
The article makes a seductive case: CPO solves a real physics constraint (copper bandwidth saturation in AI clusters), GFS has manufacturing moat via process tech and the Singapore acquisition, and hyperscalers are desperate. But the article conflates 'real problem' with 'GFS is the only solution.' Intel, TSMC, and Samsung all have silicon photonics programs. The article also doesn't quantify: What % of hyperscaler capex does interconnect represent? If it's <5% of total data center spend, even a 10x efficiency gain moves the needle modestly. GFS trades ~2.2x sales; the valuation assumes CPO becomes a meaningful revenue driver fast. The May 27 drop and Mubadala selling pressure suggest institutional skepticism about execution risk and timeline.
CPO adoption may be slower than the article implies—hyperscalers have sunk capex in copper infrastructure and may optimize software/architecture instead of rip-and-replace; and even if GFS wins design wins, foundry margins on specialty silicon are notoriously thin, so revenue growth doesn't translate to profit growth.
"GlobalFoundries is uniquely positioned to capture the transition from copper to optical interconnects, potentially decoupling its valuation from the cyclical semiconductor foundry market."
GlobalFoundries (GFS) is positioning itself as the critical foundry for silicon photonics, a necessary pivot as copper interconnects face physical thermal and latency limits in hyperscale AI clusters. By moving data transmission from electrical to optical at the chip level, GFS addresses the 'IO bottleneck' that threatens to stall GPU scaling. However, the market is currently pricing GFS as a legacy foundry—exposed to cyclical automotive and industrial demand—rather than a high-growth AI infrastructure play. If the SCALE platform gains traction with hyperscalers, GFS could see a significant valuation re-rating from its current low-multiple foundry status to a specialized component supplier, provided they successfully navigate the notoriously complex yield challenges of co-packaged optics.
Silicon photonics is a graveyard of 'next-gen' technologies; hyperscalers may ultimately favor proprietary, vertically integrated optical solutions or alternative interconnect architectures that render GFS's platform a niche, low-margin manufacturing service.
"CPO could redefine AI data movement, but meaningful upside for GFS requires proven scale, ecosystem adoption, and durable demand—risks that could limit upside."
Article argues AI bottlenecks shift from compute to data transport and that GlobalFoundries' SCALE co-packaged optics platform could redefine AI infrastructure. The logic is plausible: copper interconnects heat up, bandwidth per watt matters, and Singapore manufacturing plus the AM Foundry acquisition help execution. Yet the piece glosses key risks: silicon photonics yields and reliability at hyperscale are unproven, capex to scale production is enormous, and demand hinges on hyperscalers' willingness to standardize on a single vendor. Competition from other photonics players and alternative packaging approaches could erode the edge. Timing and revenue visibility remain highly uncertain.
Hyperscalers are notoriously reluctant to lock in a single vendor for a mission-critical data-path; if SCALE isn’t adopted at meaningful scale, the thesis collapses. Also, the revenue ramp for silicon photonics remains unproven, and capex intensity could pressure margins.
"Legacy cyclical segments will cap any CPO-driven re-rating until they shrink as a share of revenue."
Gemini flags re-rating potential if SCALE gains traction but overlooks that GFS's dominant auto and industrial exposure will anchor multiples regardless of photonics wins. Those segments remain cyclical and margin-compressed, so even 500 design wins won't lift the stock out of its foundry discount until legacy revenue shrinks materially. The 2026 timeline only widens this earnings mismatch.
"GFS's re-rating depends on photonics margin profile, not just revenue volume—but the revenue threshold to matter is higher than the article implies."
Grok's cyclicality anchor is real, but underestimates GFS's margin profile shift. Auto/industrial run 20-30% gross margins; specialty silicon photonics could hit 50%+ if yields stabilize. The mix rerating matters more than absolute revenue. However, Grok's right that legacy drag persists—GFS needs photonics to be 15-20% of revenue by 2027, not 5%, to move the needle on valuation multiple. That's aggressive given current design-win stage.
"Hyperscalers will commoditize GFS's photonics manufacturing to prevent vendor lock-in, neutralizing potential margin expansion."
Claude, you’re overly optimistic on margin expansion. Even if photonics hits 50% gross margins, GFS’s high fixed-cost structure and intense R&D requirements for SCALE will cannibalize those gains. We’re ignoring the 'Foundry Trap': hyperscalers will demand aggressive pricing to subsidize their own optical integration risks. GFS isn't just selling a component; they are selling a manufacturing service that hyperscalers will commoditize as fast as possible to protect their own data-path leverage.
"Scale-up optics yields and volume manufacturing economics are the gating item; without them, SCALE may not deliver meaningful margin uplift even with design wins."
Gemini's 'Foundry Trap' warnings are valid, but they understate the execution risk. The real bottleneck isn't pricing leverage from hyperscalers; it's scale-up economics: co-packaged optics must achieve high yields and low defectivity at volume. Without that, 50% photonics gross margins stay aspirational. 500 design wins don't guarantee a profitable ramp. Interoperability, tool-chains, and capital intensity could crush the margin expansion thesis even if SCALE signs on a few hyperscalers.
While GlobalFoundries' SCALE platform holds promise in addressing AI's data transport bottlenecks, panelists express concerns about execution risks, competition, and the challenge of shifting GFS's valuation from a legacy foundry to a high-growth AI infrastructure play.
Potential re-rating of GFS's valuation if the SCALE platform gains traction with hyperscalers.
Proving silicon photonics yields and reliability at hyperscale, competition from other photonics players, and demand hinging on hyperscalers' standardization on a single vendor.