저자: 미상 · 외국인
요약 보기
AI 성장 병목(메모리·광통신·맞춤 칩) 때문에 하드웨어 기업들이 가격 결정력을 가진다고 주장. 마이크론 매출 411억 달러·매출총이익률 85%, 코히런트 데이터센터 매출 16.2억 달러, 마벨 데이터센터 매출 46% 성장 등 실적을 근거로 제시. 세 기업(MU, COHR, MRVL)을 AI 병목 승자로 꼽고 저평가 PER을 강조.
다듬은 글
If you still believe the AI cycle has peaked or that the boom is heading for a quick crash, pay close attention to BlackRock CEO Larry Fink. With more than $10 trillion in assets under management , Fink doesn't track short-term market sentiment. He observes where institutional capital is running into physical barriers, warning that AI infrastructure is directly abutting structural bottlenecks that are constraining supply and pushing real demand far beyond current limits.
I think that delays in the development of AI will make this area the domain of only large firms. We have a shortage of cheap energy in many countries. We have a shortage of chips and memory. And while this deficit continues, developing AI is very expensive. And because of this cost, it becomes the domain of large companies.
And the faster we can scale up , the faster we can democratize AI and make it accessible to everyone. What the market perceives as a slowdown in growth is actually a massive shift in pricing power to specialized hardware monopolies controlling these very bottlenecks. The overall market for high-speed optical transceivers for AI alone is growing to $26 billion this year.
Meanwhile, the transition to next- generation internetwork structures with multi-terabit bandwidth is forcing hyperscalers to enter into multi-year supply contracts. Far from a cyclical slowdown, these physical infrastructure constraints have created unprecedented operational queues for critical equipment suppliers.
This bottleneck crisis has created an extraordinary structural barrier for the three key market leaders who control the physical plumbing of AI. The first is a high-speed memory titan that has completely sold out all of its high-performance production capacity by 2026. Backing up its product offerings with a massive multi-year order backlog of $100 billion .
The second is a leader in optical interconnects, generating over $1.6 billion in quarterly data center revenue, scaling advanced 1.6 terabit optical modules to address thermal and speed limitations. The third is an architect of custom silicon and high-speed electro-optics, which forecasts revenue of about $18 billion amid explosive demand for custom AI chips.
Instead of betting on application-level software, these three “ barrier companies” wield absolute pricing power while hyperscalers race to overcome the physical barrier of AI. First on the list is Micron Technology, ticker MU, a cornerstone of American semiconductor memory and a true gatekeeper in the ongoing expansion of artificial intelligence.
While casual investors have historically viewed memory as a volatile commodity, AI workloads have fundamentally rewritten Micron's financial profile. High-performance data centers require massive amounts of memory to power power-hungry GPUs , turning the underlying hardware into a scarce, high-value asset.
Micron's latest financial trajectory makes this structural shift abundantly clear: Revenue for the third fiscal quarter rose to $41.46 billion , up significantly from $23.86 billion in the second quarter and $9.30 billion in the same period last year. Gross margin has seen monumental growth, rising from 38% a year ago to 85% in the third quarter, with management forecasting fourth-quarter revenue of around $50 billion and gross margin of around 86%.
What makes this momentum particularly compelling is the structural transparency behind these numbers. worth approximately $22 billion with multi-year supply commitments. Importantly, these take- or-pay agreements include fixed or minimum pricing structures that protect profitability even under conservative scenarios, providing businesses with long-term revenue predictability that is rare in memory cycles.
In addition to its high-speed memory offerings, Micron recently introduced the industry's first ultra-dense 512GB DDR5 module, providing a crucial first-mover advantage for energy-efficient data center architectures. This single- module innovation reduces operating power consumption by over 60% compared to traditional configurations, directly addressing the severe energy constraints faced by today's hyperscale data centers.
Demand remains extremely high. Industry analysis indicates a prolonged global memory shortage that will last for years as the repurposing of capacity to produce advanced high-speed memory constrains the supply of standard DRAM and NAND modules. Hyperscalers are aggressively increasing capital expenditures, leading to annual revenue projections of around $129 billion .
Even with a notable acceleration in revenue growth, the stock's forward valuation remains low relative to the overall growth profile, trading at a low single-digit price- to-earnings ratio, highlighting the clear gap between market valuation and actual cash generation . With diversified strengths in computing, mobile, cloud, and automotive embedded systems, Micron controls a critical bottleneck in the global AI hardware ecosystem.
Moving from the limitations of physical memory to the speed of light inside a server rack, we turn to Coherent Corp—key architects solving the physical, optical, and thermal challenges of next- generation artificial intelligence. As copper wiring inside huge data centers reaches severe physical limits in terms of distance, heating, and signal degradation, AI computing is scaled not over copper, but over high-speed optical connectivity.
Coherent is central to this transformation: its data center and communications segment generated $1.62 billion, representing 79% of total quarterly revenue, and annual sales reached $7.12 billion. The latest financial results show impressive operational acceleration: fourth-quarter revenue rose 34% year- over-year to a record $2.05 billion, while adjusted gross margin expanded to 40.2%.
The real strength of Coherent's competitive advantage lies in its deep vertical integration of complex semiconductors and thermal management systems. The company is actively implementing a large-scale production transition from legacy 3-inch to 6-inch indium phosphide wafers at facilities in Texas and Europe, doubling its internal laser output while achieving four times the product yield at about half the price.
To overcome the enormous thermal constraints of megawatt- scale AI racks, Coherent introduced advanced 300mm silicon carbide substrates with high thermal conductivity and proprietary Thermadite cooling solutions that improve heat dissipation by up to 20%. Additionally, its integrated photonics platform, spanning the full optical signal chain for next-generation 1.6 terabit architectures, positions the company as an indispensable system- level partner, rather than simply a supplier of off-the-shelf components.
This fundamental market position is backed by solid institutional validation, as evidenced by multi-billion dollar strategic partnerships and $2 billion in equity investments from leading GPU architects to secure long-term optical capabilities . Management expects further acceleration in sales, forecasting first- quarter revenue of over $3 billion by the end of fiscal 2027, with a gross margin trend of 45%.
Following a recent valuation revision, Coherent shares trade at a discounted forward price-to- earnings multiple of around 28x and an extremely low PEG ratio of 0.55, a significant discount to expected earnings growth. By controlling optical interconnects, high-voltage distribution, and advanced thermal cooling, Coherent holds the critical hardware keys to unlocking the next generation of AI at scale .
Third on the list, moving from laser transmission to specialized computing structures, is Marvell Technology , ticker symbol M R V L. A top architect of communications and specialized chips, laying out high-speed optical backbones for global AI data centers. While retail investors' attention is often focused solely on CPUs, Marvell dominates the complex infrastructure that connects these processors together across racks and remote facilities.
Recent financial results underscore this dominance: data center revenue in the second quarter grew 46% year- over-year to $2.17 billion, accounting for 79% of the company's total sales. Management subsequently raised its long-term outlook, expecting annual revenue to reach approximately $12 billion in 2027 and $18 billion in fiscal 2028, supported by non-GAAP operating margin targets of around 40%.
Marvell's competitive advantage extends to both specialized chips and advanced optical interconnects. The company has over 20 projects using specialized computing and acceleration systems at all four major US hyperscalers, backed by multi-year commercial agreements with key platform partners such as Google and Amazon.
To circumvent the significant costs of high-bandwidth memory and the physical limitations of servers, Marvell has certified its StrataX memory expansion controllers with next-generation processor platforms , enabling seamless memory aggregation across AI clusters. Additionally, as large-scale grid bottlenecks force operators to split physical data centers across regions, Marvell's leadership in 800- gigabit and 1.6- terabit coherent optical bridges to connect distributed facilities into unified AI factories.
In addition to standard optical modules, Marvell has revolutionary intellectual property in co- packaged optics through its photonic structure technology, which enables massive scalability of chiplet-to-chiplet interconnects at speeds of 16 terabits per second directly for future computing systems. Management forecasts that the specialized chip order backlog alone will reach $11 billion by fiscal 2029, while the scalable switching business continues to rapidly double .
Even as major semiconductor players enter the custom chip space, the massive expansion of optical networks and data center interconnections is creating a growing market. By controlling the critical silicon infrastructure that combines compute, memory, and switching, Marvell remains uniquely positioned as the master "controller" for the entire AI development.
While the hype around software changes daily, the physical infrastructure limitations that Larry Fink pointed out are very real. Micron , Coherent, and Marvell Technology have absolute pricing power over memory, optics, and specialized chips. Build your portfolio around hardware " controllers" that break through the barriers for more institutional analytics, see you in the next video.
Comments 0