Categories
Asides

Redesigning Conversation and the Emergence of a Post-Human Language

As I wrote in the previous article, the idea of a “common language for humans, things, and AI” has been one of my long-standing themes. Recently, I’ve begun to feel that this question itself needs to be reconsidered from a deeper level. The shifts happening around us suggest that the very framework of human communication is starting to update.

Human-to-human conversation is approaching a point where further optimization is difficult. Reading emotions, estimating the other person’s knowledge and cognitive range, and choosing words with care—these processes enrich human culture, yet they also impose structural burdens. I don’t deny the value of embracing these inefficiencies, but if civilization advances and technology accelerates, communication too should be allowed to transform.

Here, it becomes necessary to change perspective. Rather than polishing the API between humans, we should redesign the interface between humans and AI itself. If we move beyond language alone and incorporate mechanisms that supplement intention and context, conversation will shift to a different stage. When AI can immediately understand the purpose of a dialogue, add necessary supporting information, and reinforce human comprehension, the burdens formerly assumed to be unavoidable can dissolve naturally.

Wearing devices on our ears and eyes is already a part of everyday life. Sensors and connected objects populate our environments, creating a state in which information is constantly exchanged. What comes next is a structure in which these objects and AI function as mediators of dialogue, coordinating interactions between people—or between humans and AI. Once mediated conversation becomes ordinary, the meaning of communication itself will begin to change.

Still, today’s human–AI dialogue is far from efficient. We continue to use natural language and impose human-centered grammar and expectations onto AI, paying the cognitive cost required to do so. We do not yet fully leverage AI’s capacity for knowledge and contextual memory, nor have we developed language systems or symbolic structures truly designed for AI. Even Markdown, while convenient, is simply a human-friendly formatting choice; the semantic structure AI might benefit from is largely absent. Human and AI languages could in principle be designed from completely different origins, and within that gap lies space for a new expressive culture beyond traditional “prompt optimization.”

The most intriguing domain is communication that occurs without humans—between AIs, or between AI and machines. In those spaces, a distinct communicative culture may already be emerging. Its speed and precision likely exceed human comprehension, similar to the way plants exchange chemical signals in natural systems. If such a language already exists, our task may not be to create a universal language for humans, but to design the conditions that allow humans to participate in that domain.

How humans will enter the new linguistic realm forming between AI and machines is an open question. Yet this is no longer just an interface problem; it is part of a broader reconstruction of social and technological civilization. In the future, conversation may not rely on “words” as sound, but on direct exchanges of understanding itself. That outline is beginning to come into view.

Categories
Asides

A Common Language for Humans, Machines, and AI

Human communication still has room for improvement. In fact, it may be one of the slowest systems to evolve. The optimal way to communicate depends on the purpose—whether to convey intent, ensure accuracy, share context, or express emotion. Even between people, our communication protocols are filled with inefficiencies.

Take the example of a phone call. The first step after connecting is always to confirm that audio is working—hence the habitual “hello.” That part makes sense. But what follows often doesn’t. If both parties already know each other’s numbers, it would be more efficient to go straight to the point. If it’s the first time, an introduction makes sense, but when recognition already exists, repetition becomes redundant. In other words, if there were a protocol that could identify the level of mutual recognition before the conversation begins, communication could be much smoother.

Similar inefficiencies appear everywhere in daily life. Paying at a store, ordering in a restaurant, or getting into a taxi you booked through an app—all of these interactions involve unnecessary back-and-forth verification. The taxi example is especially frustrating. As a passenger, you want to immediately state your reservation number or name to confirm your identity. But the driver, trained for politeness, automatically starts with a formal greeting. The two signals overlap, the identification gets lost, and eventually the driver still asks, “May I have your name, please?” Both sides are correct, yet the process is fundamentally flawed.

The real issue is that neither side knows the other’s expectations beforehand. Technically, this problem could be solved easily: automate the verification. A simple touch interaction or, ideally, a near-field communication system could handle both identification and payment instantly upon entry. In some contexts, reducing human conversation could actually improve the experience.

This leads to a broader point: the need for a shared language not only between people but also between humans, machines, and AI. At present, no universal communication protocol exists among them. Rather than forcing humans to adapt to digital systems, we should design a protocol that enables mutual understanding between the two. By implementing such a system at the societal level, communication between humans and AI could evolve from guesswork into trust and efficiency.

Ultimately, the most effective form of communication is one that eliminates misunderstanding—regardless of who or what is on the other end. Whether through speech, touch, or data exchange, what we truly need is a shared grammar of interaction. That grammar, still emerging at the edges of design and technology, may become the foundation of the next social infrastructure.

Categories
Asides

The Age of the AI Home

In the age of AI, the idea of what a home is will change fundamentally. As humans begin to coexist with artificial intelligence, houses may need to include small power generators or even miniature data centers. Computing power, like electricity or water, will become part of the essential infrastructure built into everyday living spaces.

Imagine a home with a living room, a dining room, and a data room. Such a layout could become commonplace. A dedicated space for AI, or for data itself, might naturally appear in architectural plans. It could be on the rooftop, underground, or next to the bedroom. Perhaps even the family altar—once a spiritual repository of ancestral memory—could evolve into a private archive where generations of personal data are stored and shared.

Either way, we will need far more computing power at the edge. Every household could function as a small node, collectively forming a distributed computational network across neighborhoods. A society that produces and consumes both energy and compute locally may begin with the home as its basic unit.

Still, this is a vision built on the inefficiencies of today’s AI infrastructure. As models become more efficient and require fewer resources, even small-scale home data centers might disappear. In their place, countless connected devices could collaborate to form an intelligent mesh that links homes and cities into a single network. At that point, a house would no longer just be a space to live—it would be a space where information itself resides.

The idea of an “AI-ready home,” one equipped with its own computing and energy systems, may be a symbol of this transition. It represents a moment when the boundary between living space and computational space begins to blur, and the household itself becomes a unit of intelligence.

Categories
Asides

Nvidia Is Copying the Earth

Eric Schmidt of Google once said it would take 300 years to crawl and index all the digital information in the world. Thirty years later, Google has collected, structured, and ranked the planet’s data, establishing itself as the central hub of global information.
This process has been one of humanity’s long attempts to digitally capture the sum of its knowledge.

Around the same time, Facebook began copying humanity itself. It targeted not only personal attributes and relationships but even private exchanges, mapping them into a social graph that visualized how people are connected.
If Google drew the “map of knowledge,” Facebook drew the “map of human relationships.”

AI has bloomed on top of these vast copies. What AI seeks is not mere volume of data, but the ability to analyze accumulated information and transform it into insight. Value lies in that process of interpretation. For this reason, possessing more data no longer guarantees advantage—what matters now is the ability to understand and utilize it.

So, what becomes the next battleground?
After the maps of knowledge and human connection, what is the next domain to be replicated? One emerging answer lies in Nvidia’s current approach.

Nvidia is attempting to copy the Earth itself. Whether we call it a Digital Twin or a Mirror World, the company is trying to reconstruct the planet’s structure and dynamics within its own ecosystem.
It aims to simulate the movements of the physical world and overlay them with digital laws. This marks a departure from the information-based replication of earlier internet companies, moving instead toward the duplication of reality itself.

What lies ahead is a complete digital copy of Earth—and a new industrial ecosystem built upon it. In Nvidia’s envisioned world, cities, climates, and economies all become entities that can be simulated. Within that digital Earth, AI learns, reasons, and reconstructs. Humanity has moved from understanding the planet to recreating it.

Yet if we wish to honor diversity and generate more possibilities in parallel, what we will need are not one, but countless “worlds.” Rather than imitating a single correct reality, AI could generate multiple “world lines” that diverge under different conditions. We can imagine a future where AI compares these world lines and derives the most optimal outcome. Such a vision would require an immense foundation of computational power.

This is no longer a contest of information processing alone but a struggle over resources themselves. The question becomes how efficiently we can transform energy into computation.The industries that produce semiconductors and the infrastructures that generate and distribute energy will form the next field of competition.
Nvidia’s challenge is not about data but about the “replication of worlds”—a new scale of technological struggle, an attempt to rewrite civilization with the Earth itself as the stage.

Categories
Asides

Rethinking Tron

Perhaps Tron is exactly what is needed right now.
I had never looked at it seriously before, but revisiting its history and design philosophy makes it clear that many of its principles align with today’s infrastructure challenges.
Its potential has always been there—steady, consistent, and quietly waiting for the right time.

Background

Tron was designed around the premise of computation that supports society from behind the scenes.
Long before mobile and cloud computing became common, it envisioned a distributed and cooperative world where devices could interconnect seamlessly.
Its early commitment to open ecosystem design set it apart, and while its visible success in the consumer OS market was limited, its adoption as an invisible foundation continued to grow.

The difficulty in evaluating Tron has always stemmed from this invisibility.
Its success accumulated quietly in the background, sustaining “systems that must not stop.”
The challenge has never been technological alone—it has been how to articulate the value of something that works best when unseen.

Why Reevaluate Tron Now

The rate at which computational capability is sinking into the social substrate is accelerating.
From home appliances to industrial machines, mobility systems, and city infrastructure, the demand for small, reliable operating systems at the edge continues to increase.
Tron’s core lies in real-time performance and lightweight design.
It treats the OS not as an end but as a component—one that elevates the overall reliability of the system.

Its focus has always been on operating safely and precisely inside the field, not just in the cloud.
The needs that Tron originally addressed have now become universal, especially as systems must remain secure and maintainable over long lifespans.

Another reason for its renewed relevance lies in the shifting meaning of “open.”
By removing licensing fees and negotiation costs, and by treating compatibility as a shared social contract, Tron embodies a practical model for the fragmented IoT landscape.
Having an open, standards-based domestic option also supports supply chain diversity—a form of strategic resilience.

Current Strengths

Tron’s greatest strength is that it does not break in the field.
It has long been used in environments where failure is not tolerated—automotive ECUs, industrial machinery, telecommunications infrastructure, and consumer electronics.
Its lightweight nature allows it to thrive under cost and power constraints while enabling long-term maintenance planning.

The open architecture is more than a technical advantage.
It reduces the cost of licensing and vendor lock-in, helping organizations move decisions forward.
Its accessibility to companies and universities directly contributes to talent supply stability, lowering overall risks of deployment and long-term operation.

Visible Challenges

There are still clear hurdles.
The first is recognition.
Success in the background is difficult to visualize, and in overseas markets Tron faces competition from ecosystems with richer English documentation and stronger commercial support.
To encourage adoption, it needs better documentation, clearer support structures, visible case studies, and accessible community pathways.

The second is the need to compete as an ecosystem, not merely as an OS.
Market traction requires more than technical superiority.
Integration with cloud services, consistent security updates, development tools, validation environments, and production support must all be presented in an accessible, cohesive form.
An operational model that assumes continuous updating is now essential.

Outlook and Repositioning

Tron can be repositioned as a standard edge OS for the AIoT era.
While large-scale computation moves to the cloud, local, reliable control and pre-processing at the edge are becoming more important.
By maintaining its lightweight strength while improving on four fronts—international standard compliance, English-language information, commercial support, and educational outreach—the landscape could shift considerably.

Rethinking Tron is not about nostalgia for a domestic technology.
It is a practical reconsideration of how to design maintainable infrastructure for long-lived systems.
If we can balance invisible reliability with visible communication, Tron’s growth is far from over.
What matters now is not the story of the past, but how we position it for the next decade.

Categories
Asides

Could the Human Eye Receive Optical Communication through IoT Integration?

I wondered if humans could ever become compatible with IOWN.

When vision is seen as an entry point for information, the human eye is already a highly advanced sensor for receiving light. If communication functionality could be layered onto it, the human body itself might become a node within the information network.

Of course, in reality, there are significant challenges involving freedom of movement and safety. Directly receiving optical signals—through Li-Fi or fiber-based communication—would place biological strain on the eye, making practical implementation difficult. Yet if even a part of the human body could receive data through optical communication, the relationship between humans and networks would be fundamentally transformed.

Reframing vision not as an organ for seeing but as a port for communication shifts the gateway of information from the brain to the body itself.
I would like to imagine a future where humans become the terminal devices of IOWN.

Categories
Asides

The Need for a Self-Driving License

After AT-only licenses, the next step we may need is a “self-driving license.”

Recently, I rented a gasoline-powered car for the first time in a while. It was an automatic model, but because I was unfamiliar with both the vehicle and the driving environment, the experience was far more stressful than I expected. Having become used to driving an EV equipped with autonomous features, I found the act of operating everything manually—with my own judgment and physical input—strangely primitive.

When the gear is shifted to drive, the car starts moving on its own. A handbrake must be engaged separately, and the accelerator must be pressed continuously to keep moving. Every stop requires the brake, every start requires a shift of the foot back to the accelerator, and even the turn signal must be turned off manually. I was reminded that this entire system is designed around the assumption that the human body functions as the car’s control mechanism.

I also found myself confused by actions that used to be second nature—starting the engine, locking and unlocking the door with a key. What once seemed natural now feels unnecessary. There are simply too many steps required before a car can even move. Press a button, pull a lever, step on a pedal, turn a wheel. The process feels less like operating a machine and more like performing a ritual.

From a UX perspective, this reflects a design philosophy stuck between eras. The dashboard is filled with switches and meters whose meanings are not immediately clear. Beyond speed and fuel levels, how much information does a driver actually need? The system relies on human judgment, but in doing so, it also introduces confusion.

When driving shifted from manual to automatic, the clutch became obsolete. People were freed from unnecessary complexity, and driving became accessible to anyone. In the same way, in an age where autonomous driving becomes the norm, pressing pedals or turning a steering wheel will seem like relics of a bygone era. We are moving from a phase where machines adapt to humans to one where humans no longer need to touch the machines at all.

Yet driver licensing systems have not caught up with this change. Until now, a license has certified one’s ability to operate a vehicle. But in the future, what will matter is the ability to interact with the car, to understand its systems, and to intervene safely when needed. It will no longer be about physical control, but about comprehension—of AI behavior, of algorithmic decision-making, and of how to respond when something goes wrong.

When AT-only licenses were introduced, many drivers were skeptical about removing the clutch. But over time, that became the standard, and manual transmissions turned into a niche skill. Likewise, if a “self-driving license” is introduced in the near future, pressing pedals may come to be viewed as a legacy form of driving—something from another era.

The evolution of driving technology is, at its core, the gradual separation of humans from machines. A self-driving license would not be a qualification to control a vehicle, but a literacy certificate for coexisting with technology. It would mark the shift from moving the car to moving with the car. Such a change in licensing might define how transportation itself evolves in the next generation.

Categories
Asides

Will SoftBank Acquire Intel?

I once wondered about this question. Later, SoftBank announced a new round of investment, confirming that the company was indeed moving in that direction. I want to record here what I was thinking before that announcement. My purpose is to leave a trace that will allow me to compare those thoughts with what eventually happens.

The Background and Current State of the ARM Acquisition

SoftBank’s 2016 acquisition of ARM was a clear declaration that the company intended to anchor itself at the upstream of the semiconductor value chain — intellectual property for chip design. ARM’s licensing model, built on neutrality and scalability, expanded its reach from mobile devices to IoT, servers, and even supercomputers. While promising to maintain ARM’s neutrality, SoftBank began to tighten integration, emphasizing subsystem offerings and deeper involvement in the server domain. Even after ARM’s re-listing in 2023, it remains the group’s most important asset and the central hub connecting its other investments.

The Next Step: Building Around ARM

SoftBank’s acquisitions have sought to elevate ARM from a mere IP provider into the driving force of an entire ecosystem. The acquisition of Graphcore gave it a foothold in AI accelerators, and the purchase of Ampere brought a practical server-CPU operation under the same umbrella. The combination of ARM’s low-power design philosophy and the data-center scale-out trend offers an alternative optimal point to the traditional x86-centric server market. This configuration directly connects to the later thought experiment concerning Intel.

The Distance Between SoftBank and Nvidia

SoftBank was once a major Nvidia shareholder, forging a close relationship before the AI boom. The subsequent sale, which forfeited a massive appreciation opportunity, shifted the relationship to one of both collaboration and competition. While joint projects in Japan’s AI and telecommunications infrastructure continue, SoftBank’s push to cultivate multiple in-house AI-chip initiatives can be read as an attempt to challenge Nvidia’s dominance. Nvidia, for its part, is reinforcing its own vertical integration with ARM-based CPUs and NVLink interconnects. The two paths intersect but ultimately lead toward different goals.

The AI Investment Strategy Centered on OpenAI

SoftBank’s massive commitment to OpenAI, its infrastructure partnerships with Oracle and others, and joint ventures in Japan all signal a plan to bring the software core of AI under its orbit while pre-securing compute resources. In the AI era, supremacy converges not on algorithms but on the ability to govern and interconnect power, semiconductors, and capital. SoftBank aims to tie the scale of AI itself to its balance sheet, controlling both design IP and the physical data-center layer.

The Intel Hypothesis

How might Intel fit into this circuitry? Market stagnation, restructuring pressures, and the separation of manufacturing from products have fueled repeated speculation about acquisitions and equity partnerships. Reports suggested that ARM showed interest in Intel’s product division but talks fell through, and negotiations over AI-chip manufacturing also collapsed over production-capacity terms. There is no evidence of a formal buyout attempt, but traces of exploratory engagement remain. The core question is simple: why would SoftBank want to absorb Intel, and through what realistic path could it happen?

Examining Strategic Alignment

ARM is an IP-driven entity without manufacturing. Intel possesses vast fabrication capacity and an x86 franchise but lags in mobile and power-efficient contexts. Combined, they could span both CPU architectures, integrating from data centers to edge devices with comprehensive design and supply capabilities. Within the AI infrastructure stack, they could encompass CPUs, AI accelerators, memory, interconnects, and fabs. The logic is elegant — and access to CHIPS Act subsidies and advanced fabrication would offset reliance on external foundries.

Yet elegant logic does not guarantee practical feasibility. For foreign capital to take control of Intel — an American strategic asset — would run headlong into political and regulatory barriers. As the U.S. Steel precedent showed, national interest can override regulatory clearance. On antitrust grounds, even the perception that ARM’s neutrality might erode would provoke fierce resistance. The industry views ARM as common infrastructure; any integration skewed toward a single group’s advantage would meet opposition from all sides. Add financial strain and the operational burden of running manufacturing, and a full acquisition becomes implausible.

Pragmatic Alternatives

If full control is closed off, distributed strategies remain. Partial equity participation, co-design projects, long-term manufacturing contracts, and multinational consortiums all represent workable routes. ARM can enhance its relevance through subsystem design and joint optimization; Ampere and Graphcore can bring their products to market; Rapidus and overseas foundries can diversify manufacturing access. Rather than outright control, strengthening its role as a hub connecting specifications, capital, and power supply aligns with SoftBank’s pragmatic style.

Re-Examining the Risks

A U.S. Steel–type political blockade is entirely plausible. Cross-border semiconductor investments fall squarely within national-security and industrial-policy oversight, entangling legislators, unions, and state governments. Antitrust risks are also significant. If ARM’s neutrality were questioned, Apple, Qualcomm, Microsoft, Amazon, Google, and Nvidia would all lobby against the deal. Conflicts with existing players would be inevitable: Nvidia is consolidating independence across CPUs and GPUs, while Apple closely monitors ARM’s trajectory, vital to its own SoC strategy. The practical route to conflict avoidance lies in incentive structures that distribute value across stakeholders and in maintaining transparent, non-discriminatory licensing.

Japan’s Policy Landscape and Points of Contact

SBI’s domestic memory initiative has shifted focus from a failed PSMC alliance toward cooperation with the SK group. Subsidy frameworks remain, and Japan continues exploring ways to restore local memory capacity. With domestic AI firms such as PFN in the mix, a new ecosystem centered on AI-specific memory demand could emerge. Meanwhile, Rapidus aims for 2-nm logic mass production and is collaborating with Tenstorrent to capture edge-AI demand. SoftBank, a shareholder, holds the option to align ARM or Ampere designs with domestic manufacturing. The interplay between national and private capital thus serves SoftBank as both risk hedge and policy alignment mechanism.

Managing Relationships with Nvidia and Apple

Nvidia represents both partner and competitor. Joint efforts in Japan’s AI and 5G infrastructure coexist with SoftBank’s independent AI-chip initiatives and ARM’s expansion, both of which could alter long-term market dynamics. For Apple, ARM’s neutrality and licensing stability are paramount. Any perception that ARM’s roadmap tilts toward proprietary advantage could chill relations. Maintaining openness in software toolchains, transparency in roadmaps, and a balance between differentiation and neutrality will be key.

The Question That Remains

Even if an acquisition is unrealistic, why does the idea keep resurfacing? The answer is simple: in the AI era, value creation is migrating toward the convergence of compute resources, power, and capital. CPU architectures, advanced fabs, AI accelerators, memory, interconnects, cloud infrastructure, and generative AI platforms — whoever orchestrates these elements will define the next decade. SoftBank holds capital, IP, and market reach, but lacks proprietary access to manufacturing and power. That is why Intel enters the frame. Yet being in view and being within reach are two different things.

Conclusion

Even if the path to a full Intel acquisition is closed, SoftBank still has room to build equivalent capability through distributed partnerships. The real question is how to integrate power sources, manufacturing ecosystems, architectures, and capital structures into a coherent design. This is no longer about a one-time transaction but about the ability to interlink policy, capital, and technology. When revisited years from now, this speculation may not look like a rumor but rather an early thought experiment on the reconfiguration of power in the age of compute sovereignty.

Categories
Asides

The Strategic Value of Compute Resources in the OpenAI–AMD Partnership

The expansion of generative AI has entered a stage where progress is determined not by model novelty but by the ability to secure and operate compute resources. The multi-year, multi-generation alliance between OpenAI and AMD clearly reflects this structure. It is no longer a simple transactional deal but a framework that integrates capital, supply, power, and implementation layers into a mechanism for mutual growth—signaling a shift toward scale as a built-in assumption.

Forecasting Power Demand

The backbone of this partnership is gigawatt-class compute capacity. An initial 1 GW, scaling to several gigawatts, links data-center construction directly to regional grid planning rather than individual projects. The key factors are not only peak power draw but sustained supply reliability and effective PUE including heat rejection. AI training workloads behave as constant loads rather than spikes, making grid stability and redundancy in auxiliary systems critical bottlenecks.

Model evolution continues to expand overall electricity demand, offsetting gains in performance per watt. Even as semiconductor generations improve efficiency, larger parameter counts, bigger datasets, and multimodal preprocessing and inference push consumption upward. Consequently, capital investment shifts its center of gravity from racks to civil-engineering and electrical domains that include cooling infrastructure.

Structural Issues in the Compute Market

Even with AMD expanding deployment options, the NVIDIA-dominated market faces other bottlenecks—optical interconnects, advanced HBM, and CoWoS packaging capacity among them. Rising rack-level heat density makes the shift from air to liquid cooling irreversible, tightening location constraints for data centers. The result is a conversion lag: capital cannot instantly be turned into usable compute capacity.

A further concern is geopolitical risk. Heightened global tensions and export controls can fragment manufacturing and deployment chains, triggering cascading delays and redesigns.

OpenAI’s Challenges

The first challenge for OpenAI is absorbing and smoothing exponentially growing compute demand. Running research, productization, and APIs concurrently complicates capacity planning across training and inference clusters, making the balance between model renewal and existing services a critical task.

The second is diversification away from a single vendor. Heavy reliance on NVIDIA has caused supply bottlenecks and eroded pricing flexibility. Sharing the roadmap with AMD therefore carries both optimization and procurement significance.

The third lies in capital structure and governance. While drawing in vast external commitments, OpenAI must maintain neutrality and research agility, requiring careful contract architecture to coordinate partnerships. The episode of its past internal split serves as a reminder: when capital providers bring divergent decision criteria, alignment of research agendas becomes a challenge.

AMD’s Challenges

AMD’s bottlenecks are manufacturing capacity and the software ecosystem. Its latest designs can compete technically, but to offer a developer experience rivaling the PyTorch/CUDA world, it must advance runtimes, compilers, kernels, and distributed-training toolchains. Hardware aspects such as HBM supply, packaging yield, and thermal management will define both delivery schedules and operational stability.

A second challenge is converting the co-developed results with OpenAI into broader market value. If collaboration remains confined to a single project or product, dependency risk increases. Generalizing and scaling the gains to other markets will be essential.

Strategic Intent of the Partnership

At the surface, the intent is clear: OpenAI seeks secure and diversified compute resources, while AMD seeks simultaneous credibility and demand. Structurally, however, there is a deeper layer—integrating models, data, compute, and capital into a unified flow; accelerating GPU design and supply cycles; and locking in diversified power and site portfolios early. In effect, both sides embed their respective challenges into a forward-loaded roadmap that reduces uncertainty in supply and financing.

Scheme Design

The distinctive feature is clause design that firmly enforces reciprocal commitment. Large take-or-pay volumes and facility milestones are tied to capital returns, linking hardware success directly to customer benefit. For suppliers, it secures quantity certainty and pricing floors, easing investment decisions. For buyers, it strengthens influence over technical specifications and workload fit. Financially, it helps smooth extreme swings in cash flow.

Difference from NVIDIA’s Model

Where NVIDIA’s massive deal channels capital from supplier to buyer—who then spends it back on the supplier—the AMD structure grants equity options from supplier to buyer, while the buyer guarantees long-term procurement. Both align incentives, but the direction of capital flow and degree of governance leverage differ.

NVIDIA’s model gives suppliers greater control and restricts buyers through capital conditions. AMD’s allows buyers to become future shareholders, giving them indirect influence over the supplier’s technical priorities.

Compute-ism

In the AI era, the value model ultimately converges on a single question: who can operate how much compute, on what power, at what efficiency, and under what governance. Partnerships with Microsoft, NVIDIA, AMD, and Oracle all stem from that premise. Compute capacity has become currency, conduit, and foundation of sovereignty. The choice of compute space—including power source, jurisdiction, ethical stance, and data lineage—extends from corporate strategy into institutional design.

From this viewpoint, true competitiveness lies in projects that integrate long-term cloud commitments, dedicated power and cooling, secured land, and supply-chain finance. Price or FLOPS comparisons alone no longer define advantage.

Impact on the Hardware and Technology Roadmap

Meeting the insatiable demand for compute requires clear priorities: larger memory space, lower latency, more efficient cooling, higher energy performance. GPUs will continue evolving accordingly—scaling HBM capacity and bandwidth, advancing interconnects, and optimizing storage and data-loading paths. Opportunities for improvement remain endless.

On the software side, the question is how close AMD’s compilers and runtimes can come to zero-friction while preserving backward compatibility with PyTorch and JAX. In an expanding market, feeding operational feedback into architecture along the shortest path will decide generational performance gaps. Even abundant hardware fails to convert into market value without matching software optimization.

Power, cooling, and site strategy should also be treated as integral parts of the roadmap. Layouts premised on liquid immersion, integration of heat recovery with district systems, hybridization of renewables and storage, and adaptive scheduling to power demand—all these “Watt and Bit” linkages define the real unit cost of compute. Chip miniaturization alone will not sustain the next decade.

Conclusion

The OpenAI–AMD partnership marks the arrival of an era where capital, supply, power, and software are designed as a single system around compute resources. Under compute-ism, victory depends not on individual products but on ecosystem maturity. Market velocity will accelerate, yet the fundamentals remain simple: which power, in which place, on which chip, through which code, under which governance. The alliances that design these layers early, deeply, and broadly will draw the next map of the AI age.

Categories
Asides

How Much Power Does a GPU Consume?

I tried to compare the power consumption of GPUs in a way that makes it easier to imagine.
This is not a precise comparison, and since it only looks at power consumption, it may lead to misunderstandings regarding heat generation or efficiency.
Still, to get an intuitive sense of how much energy today’s GPUs consume, this kind of simplification can be useful.

Let’s start with something familiar — a household heater.
A typical ceramic or electric heater consumes about 0.3 kilowatts on low and roughly 1.2 kilowatts on high.
We can use this 1.2 kilowatts as a reference point — “one heater running at full power.”

When you compare household appliances and server hardware in the same units, the scale difference becomes more tangible.
The goal here is to visualize that difference.

Power Consumption (Approximate)

Item Power Consumption
Household Heater (High) ~1.2 kW
Server Rack (Conventional) ~10 kW
Server Rack (AI-Ready) 20–50 kW
NVIDIA H200 (Server) ~10.2 kW
Next-Generation GPU (Estimated) ~14.3 kW

A household heater represents the level of power used by common home heating devices.
A conventional server rack, typical through the 2010s, was designed for air-cooled operation with around 10 kilowatts per rack.
In contrast, modern AI-ready racks are built for liquid or direct cooling and can deliver 20–50 kilowatts per rack.
The NVIDIA H200’s figure reflects the official specification of a current-generation GPU server, while the next-generation GPU is a projection based on industry reports.

Next, let’s convert this into something more relatable — how many heaters’ worth of electricity does a GPU server consume?
This household-based comparison helps make the scale more intuitive.

Heater Equivalent (Assuming One Heater = ~1.2 kW)

Item Equivalent Number of Heaters
NVIDIA H200 (Server) ~8.5 units
Next-Generation GPU (Estimated) ~12 units

Until the 2010s, a standard data center rack typically supplied around 10 kilowatts of power — near the upper limit for air-cooled systems.
However, the rise of AI workloads has changed this landscape.
High-density racks designed for liquid cooling now reach 20–50 kilowatts per rack.
Under this assumption, a single GPU server would nearly fill an entire legacy rack’s capacity, and even in AI-ready racks, only one to three GPU servers could be accommodated.

  • NVIDIA H200 (Current Model)

    • Per Chip: up to 0.7 kW
    • Per Server (8 GPUs + NVSwitch): ~10.2 kW
    • Equivalent to about 8.5 household heaters
    • Nearly fills a conventional 10 kW rack
    • Fits roughly 2–4 servers per AI-ready rack

  • Next-Generation GPU (Estimated)

    • Per Chip: around 1.0 kW (based on reported estimates)
    • Per Server (8 GPUs + NVSwitch assumed): ~14.3 kW
    • Equivalent to about 12 household heaters
    • Exceeds the capacity of conventional racks
    • Fits roughly 1–3 servers per AI-ready rack

Looking at these comparisons, the difference between a household heater and a GPU server becomes strikingly clear.
A GPU is no longer just an electronic component — it’s effectively part of the power infrastructure itself.

If you imagine running ten household heaters at once, you start to grasp the weight of a single GPU server.
As AI models continue to scale, their power demands are rising exponentially, forcing data center design to evolve around power delivery and cooling systems.
Enhancing computational capability now also means confronting how we handle energy itself, as the evolution of GPUs continues to blur the line between information technology and the energy industry.