저자: Oshima · 외국인
요약 보기
AI 인프라 투자 구조를 다룬 영상. 뉴스상 200조원 규모는 중복 합산된 상한치로 실제 집행 규모는 수십조원 수준이라 주장. 벤더 파이낸싱은 밀스톤·지분 연동·다중 벤더 경쟁 등으로 순환거래와 구분되며 건전하다고 평가. NVIDIA LOI, AMD-OpenAI 6GW 성과연동 구조, 액침/DLC 냉각 등 기술·재무 설계를 설명하고, 브로드컴 커스텀 칩 협업은 NVIDIA·AMD와 병행 성장하는 구도로 해석.
다듬은 글
Good evening. This is Oshima from the American Master Oshima Fund Garage. Uh, the previous video was Nvidia x OpenAI x AMD. Is this a dangerous cycle? We received many questions about this. Why is the liquid example not involved? How is vendor finance different from circular trading? So, this time, we'll organize things around a single axis: a system that doesn't stop, and gently and quietly examine the AI factory from both the perspective of funding flow and on-site implementation .
Okay, then let's begin the main part. The boom in generative AI, such as ChatGPT, has moved from the era of creating and demonstrating AI to the next stage of continuous operation behind the scenes of daily life and business. What matters here is n't a single impressive performance or a flashy announcement.
Overcoming daily troubles, peak loads, maintenance, and failures without interruption. This is the essence of Phase 2, or the second act of AI and the revolution. In the world of technology, we often refer to availability, but what's important now is not the amount of time spent in motion, but the concept of permanence and sustainability, which assumes continuous motion.
Consistent growth does n't come from accidental inventions or one-off KPIs; it's the result of layers upon layers of design that ensure continuous improvement . The design can be easily understood by dividing it into three main layers . Firstly, the capital layer—the funding design, or contract structure, milestones, and risk sharing—for accelerating the preparation of equipment and recovering costs through operation, serves as the fuel tank for continuous operation.
Secondly, we will integrate the physical layer, power cooling, network storage, and interest operations into a single AI factory, creating a system that can stably and continuously output computations. Thirdly, the layers of regulations, such as safety regulations, data governance , service level agreements (SLs), maintenance contracts, and accounting disclosure rules, play the role of road signs that support operations that do not stop from a societal perspective .
These oxygen atoms do not exist independently in a vertical manner; they only become strong when they interlock with each other. If the capital design is weak , investment will be interrupted; if the physical design is weak, operations will stop; and if the institutional design is flawed, expansion will be hindered .
Conversely, the better the oxygen supply, the more practical KPIs such as operating rate, water usage, and computational power per unit of electricity gradually improve, ultimately leading to a steady increase in corporate value. This time, we will unravel the second act of the AI revolution by taking a bird's-eye view of oxygen, like a map, and ensuring that it does not stop .
The Nikkei Shimbun reported on October 7th that OpenAir is investing 200 trillion yen in infrastructure with AMD . The figure of 200 trillion yen has taken on a life of its own, but this is a simplified, rough estimate based on three major projects as the upper limit . Specifically, the first is the Oracle, SoftBank, and OpenA Stargate project, which involves the construction of a $ 500 billion GW- sized facility over four years.
The second is the OpenAir x Nvidia GWLOI and Letter of Intent, which means NVIDIA will be investing up to $100 billion in stages . The third figure is the 6GW investment by OpenAI x AMD , meaning OpenAI will invest up to 10% in AMD and introduce GPUs. This figure was created by adding these three entities together without considering any overlap.
From a practical standpoint, the structure is such that NVD and AMD GPUs are mounted inside a Stargate box, so subtracting the three elements results in the 2030 series, which is just a place name. Therefore, 200 trillion yen is a fictitious value that does not exist in accounting or in terms of fund execution, and can be understood as a symbolic value that has been expanded by the framework and the phased execution stages.
Next, let's talk about the unit price. News reports often quote the catchy unit price of 1 GW = 50 billion yen . However, this is a headline unit price based on multiple upper limit assumptions, and the actual range will vary greatly depending on the design stage, such as the degree of liquid adoption, the use of rack mist networks, storage configuration, and the difficulty of power supply .
If we only consider the portion of the cost closest to that of a GPU, a reasonable estimate would be around $15 billion to $30 billion per GBW. Simply multiplying this by $50 billion to estimate the total would exceed our actual cash flow. So what should we investors look at? It's the design of the money flow, rather than the monetary value.
Here, we will examine this by dividing it into three points. First, eliminate duplicates. The Stargate is a box. The NDIA 's LO and upper limit transition AMD main contract drug implementation does not involve creating a separate layer. The mechanism by which headlines become larger is precisely this overlap.
Secondly, separate the upper limit from the execution phase. NDIA's Letter of Intent (LOI) indicates an intention to invest up to $100 billion in stages, and does not promise immediate cash flow . On the other hand, AMD and 6GW are implementing a performance-linked ballast that OpenAir will purchase, and it is only 1,000, with the rights being determined in stages using a milestone method.
Thirdly, to what extent have the process progressed in the execution phase, specifically regarding the designated land, power construction equipment, and power supply? In other words, it means separating and interpreting the phases in which news and funding actually move within a given framework on a map.
In short, 200 trillion yen is a symbolic value indicating direction, a breakdown of temporary information, that is, recalculating the finance within the framework of the LOI for Stargate's infrastructure investment NvVidea, and the finance of the phased execution of the main contract with AMD. This reinterpretation reveals that the more realistic range in the implementation phase, compared to the simple summation in the headlines, falls to tens of trillions of yen .
Therefore, instead of being swayed by the impact of the numbers, understanding who takes on risk at what stage and how it will be recovered through KPIs—reading this blueprint of the financial cycle—is the core of urban decision-making in the age of AI . Here, we will organize key points to consider to avoid confusing vendor finance with circular transactions, drawing on the telecommunications bubble of the 2000s as a cautionary tale .
In conclusion, what we are seeing now is not the old-fashioned, high-pressure sales method of lending and then forcing a purchase. The three-part mechanism—the clear purpose, the risk-sharing phase of implementation, the gradual allocation of funds, and the multi-vendor competition— supports the prospect of investment recovery .
First, there's the investment. Rather than simply accounts receivable or deferred payments, the fee assumes a portion of the upside in case of success and the downside in case of failure through balances and equity. This is the exact opposite of a circular transaction that prioritizes sales above all else .
Sharing the burden through capital means that the incentive to simply sell now is weakened, and attention is paid to the soundness of the deal . This is the first difference. Next is the execution phase: power generation, construction equipment cooling, and power supply. Training and cashiering are coordinated at each stage of this process.
In other words, funds are released in accordance with the actual progress, so there is little room to prematurely record sales or profits while the project is still incomplete . This structure makes it less likely for accounting practices to become reliant on formalities, such as recording sales simply because there is a contract, as was common in the 2000s.
What's important here is milestones. As the process progresses with funding, the risk decreases and the cost of capital also decreases. It's about ensuring that the project doesn't stall financially. The third is multi-vendor competition. Multiple vehicles can be bid on and substitute for each layer of GPU network cooling storage construction .
This acts as a brake on price intelligence and the internal pricing system that relies solely on those involved to determine the price . The more a company becomes locked into a particular business model, the more likely it is to create a sound pattern that resembles circular trading . Conversely, the very existence of an open competitive environment functions as a safety valve that guarantees the soundness of capital.
When these three points are combined, the key lies in three safety valves: price independence, and the forced provision of milestone funding. Independence means avoiding scientific decision-making solely by the relevant parties by utilizing third-party planning, comparative estimates, and index linkages.
Milestones are about marking tangible progress and creating cash milestones. A competitive environment must always ensure the possibility of switching to another company. As long as these factors are being considered simultaneously, the cycle of funds and sales will not be one where they simply go back and forth on paper.
So why can this be called infrastructure finance? The answer lies in the nature of cash flow . Capacity reservations are primarily driven by operational cash flow, such as recurring revenue based on SLAs and uptime backed by long-term contracts . Construction risk, supply risk, performance risk, demand risk.
The design, which involves allocating these resources to the appropriate entities and having them undertaken in stages, is similar to project finance for power generators and communication backbones. In other words, it's not a cycle designed to make headlines show high sales quickly , but rather a financial framework based on long-term operation that allows for recouping investments .
Just in case, I'll add a warning about the danger . These contracts are merely a way to inflate the resale value of training programs that lack any real substance. A transaction in which funds move back and forth at the same price and daily price, making third-party comparison impossible . If you see these signs, you should suspect circular trading.
However, the current mainstream trend is actually the opposite, moving towards supporting the implementation of operations that do not stop at the three points of the emperor's price competition . That's why, as an investor, instead of focusing on the sheer size of the offering, it's more important to analyze the contract structure, the training conditions, and whether the competitive environment is properly designed .
This, I believe, is the correct perspective on vendor finance in the AI era, based on the lessons learned from the 2000s . Here, we will re-examine the Letter of Intent ( LOI), as discussed in the context of the relationship between NVIDIA and OpenAI, as a form of financial framework . While a Letter of Intent (LOI) does not have the same legal binding force as a final contract , there is a reason for this weakness .
The goal isn't to finalize everything right away, but to maintain flexibility and move forward as needed . For example, it enables the gradual securing and commencement of land, electricity, power distribution, cooling equipment , and operational personnel for construction companies . National-level projects inevitably encounter bottlenecks in terms of regulations, grid connection coordination, materials, and personnel.
The Letter of Intent (LOI) is the cover page of the blueprint that signals the issue and places all stakeholders on the same map. NVD's position in leading this framework is not merely that of a supplier, or seller of reaction chips. We define upstream requirements such as required power, rack density, cooling methods, network configuration, and operational SLAs, and then bring together suppliers and various players involved in construction and operation—essentially acting as the architect of social infrastructure .
Your company doesn't need to manufacture everything in-house . Rather, providing standards and specifications and indirectly controlling the entire ecosystem allows for a better balance of speed and scale . This is the concept behind the articular control indirect architecture. And importantly, it's crucial to adopt the perspective of incorporating the market share expansion of others as part of your own market expansion.
Even if AMD's contract changes the GPU ratio at a specific site, the data center structure remains the same, and the requirements for power, cooling, network storage, and operation are consistent. In other words, if the number of computers increases, the importance curve for the entire ecosystem, including software, middleware, network cooling, power, infrastructure services, and human resources, will shift to the right.
NVIDIA can accelerate the expansion of its overall scope precisely because it writes the container first using Letters of Intent (LOI). As a result, AMD's advancement will lead to a pi-type competition, supplementing the value proposition of the entire ecosystem . In short, LOIs are not weak because they are uncertain; rather, they are a capital structure designed to move forward amidst uncertainty.
The milestones leading up to the final agreement mark out funding and risk, allowing room for development in response to changes in policy, technology, and supply and demand . NVD is using this framework to its advantage, designing standard supply based on actual conditions, and leading the implementation to keep the AI factory running .
We view the framework of AMD and OpenAI as the finance of implementation. There are two . With a serious 6GBW plan, the supply will be implemented in stages, and a portion of the payment will be deferred in 1000 balants, thus avoiding large upfront cash accumulation and transferring value only as much is achieved.
That's why, as the plan moves forward, the supply chain progresses and the risks to fulfilling it are minimized. First, the rationale behind open-air design. By exercising a balance in accordance with the overall operation and power supply, without relying on large upfront payments or significant discounts, you can receive substantial benefits.
In other words, the order is to purchase the equipment first, and then offer performance-based discounts . This reduces the immediate cash burden and allows us to maintain agility during the growth phase. Next, let's look at the advantages of AMD. This is important. By not lowering prices in the short term, and instead offering the equivalent of a discount in the form of shares only upon future success, upward price reduction pressure can be avoided.
In other words, instead of bringing the price down in advance, this is a deferred payment linked to the growth. With this format, the final price is a value transfer through shares, and the accounting sales price can be set straightforwardly, making it easier to explain to the supply chain. As a result, it becomes easier to justify the advance investment to banks and parts suppliers while maintaining price independence .
As the process progresses, inaccuracies decrease and capital costs are reduced . Furthermore, this design does not disrupt the competitive environment or price independence. The 6GW milestone is marked by a specific condition, and if the performance, unit price, and delivery time don't fit, it won't be the next step.
This approach is fundamentally different from pushy lending/buying schemes or circular transactions where only numbers are exchanged, as it leaves room for relative external comparisons with other vendors . This is where the soundness of implemented finance lies. OpenAI, having swept the load, immediately receives deferred discounts based on correct answers while avoiding cache stagnation.
AMD, in turn, returns short-term gains to maintain its ASP average sales price, rewarding only successes with standardized dilutions. The logic of both parties aligns as much as they move forward, so the AI factory never stops. This is a competitive advantage as a capital design that goes beyond scientific negotiation, and it is a mechanism that provides constant support from all sides .
In the AI factory setting, the direct-tip (DC) liquid solution, which is the Blackwell generation, has become the de facto standard . When power consumption exceeds 100kW per unit, air cooling reaches its limits in terms of airflow, overall noise, hot spots, and underfloor wiring control. The concept is similar to the steep-ended plumbing in a house.
A thin stream of cold liquid is directed directly to the CPU and GPU chips, where it absorbs heat and the warmed liquid is returned. The returned liquid is cooled again in the radiator box in the machine room and then circulated. The system branches out to each server from the piping infection , and uses a one -touch type that is easy to plug and unplug, minimizing maintenance power outages .
The cooling unit, which contains a circulation pump and heat exchanger, is the heart of the system, automatically adjusting the flow rate and temperature . This significantly increases the computing power that can be handled per unit of floor space for the same floor area, resulting in improved costs related to the operation of the equipment using electricity .
The important thing here is not to simply dismiss cooling as a matter of device resolution . The extreme example assumes an integrated design where power design, spatial planning, and maintenance and operation are all considered together . For example, the allocation of power receiving capacity , redundant routes, piping systems and valve placement, aging water rights, and work routes during replacement must all be incorporated into the design phase to ensure that operations do not stop.
While the liquid core system is strong in specific requirements, the current mainstream is D2C, which offers high compatibility with existing facilities and is easier to expand in stages. The consolidation of the vertical assembly process is handled by Integra, not by individual parts vendors. For example, Supermicro is pursuing high density by optimizing rack awareness to include not only servers and racks but also infected piping and mariholds .
Vertive reduces the time required to integrate cooling units (CDUs) and chiller piping power supplies into a single module . Schneider Electronics provides integrated visibility and management of power cooling monitoring, simultaneously improving uptime and ease of maintenance. In short, the key is not which pump to buy, but who will design the whole system, take responsibility for its construction, and oversee its operation .
The effects can be divided into three categories. Firstly, the computational load per floor area is increased by multiplying density by the fire resistance ratio, and secondly, the computational load per unit of electricity is increased by improving energy efficiency. Thirdly, by leveling out maintenance, planned downtime will be shortened and operational efficiency will be boosted.
Cooling is not a cost center, but a revenue-generating design lever. Companies that can properly integrate these groundbreaking examples are the true landowners and growing entities of the AI era. Up to this point, we've looked at the foundation of a system that doesn't stop at capital design and cooling implementation.
Now, let's consider what the latest developments mean. There have been reports that OpenAI and Broadgom have reached an agreement to jointly develop custom reaction mechanisms . Whenever this kind of story comes up, the question that inevitably arises is whether the NBA is alright. To put it simply, it's not a matter of one-size-fits-all replacement, but rather a situation where you need to choose the right one for the job .
If you try to understand the roles as being divided into three categories, it's easy to get confused . First, NVIDIA will provide a full-stack foundation, from support and learning to training and networking. Its significance as a starting point for launching a new model remains considerable. On the other hand, as demand expands, the constraints on supply prices and electricity become apparent when relying solely on NVID.
Therefore, AMD will step in as the second supplier, increasing cost and delivery flexibility. A key point is that the system is designed to move forward while avoiding immediate collapse by utilizing performance-based deferred payment . As a result, OpenAI can secure backup tanks to prevent shutdowns .
Broadcom will enhance the mass production and communication capabilities of data centers with application-specific custom ethics and large-scale Ethernet . This is a different area from NVD and AMD, with a mouse-based approach that is lower cost and lower power consumption than specific workloads . In terms of the timeline, it seems natural to have an initial deployment in the latter half of 2026, followed by a phased expansion .
There are simply three ways to interpret this from an investor's perspective. The first scenario is a zero-sum/positive-sum game, meaning that in a situation where demand is expanding along with the volume of goods, NVD's foundation plus AMD 's lead and Broadcom's A run in parallel . The second reason is decentralization to avoid bottlenecks ; considering the constraints of HBM, advanced packaging, power, and network, a multi- vendor approach is actually more rational.
And the third point is that the way it appears in numbers varies from case to case . In other words, we will focus on different aspects of performance in the second stage, such as the strong integration of NVIDIA systems and software, the progress of phased contracts for AMD, and the accumulation of custom thick and network projects for Broadcom.
Of course, NBD's market share and scientific negotiating power are not universal . The more second and third options there are, the more fluid the allocation becomes. However, this doesn't mean that the reliability of the underlying infrastructure will be relatively lower; rather , it means that the optimal combination to prevent downtime will be selected for each project.
In general, this will work to increase the overall processing load of the ecosystem . NVDAMD Broadco, with its distinct role, is in a parallel storage relationship that will accelerate the second act of the AI revolution . It's about storage, not competition, and it's about plausibility, not zero-sum.
I believe that looking at it from this perspective allows us to interpret the meaning of the news without losing focus . So, that's all for today . If you enjoyed this video, please click the like button and subscribe to the channel. Please feel free to leave comments as well . I hope that everything will go smoothly and that I can use the lessons learned to improve future events . Thank you for watching until the end .
댓글 0