The AI Bubble May Only Be Beginning, Three Accelerators, One Bottleneck — and Why We Have Become More Bullish on SoftBank Group

In our previous post, we explained why we had resumed investing in SoftBank Group.

Our earlier caution was not driven by pessimism about AI itself. The principal reasons we had stayed on the sidelines were geopolitical risk and interest rates.

What mattered most in bringing us back was a qualitative change in AI demand.

With the rise of agents, "AI labor" is beginning to break free from the physical constraint of human working hours. A person can work only 24 hours a day. But that same person can run multiple agents in parallel and commission far more than 24 hours of AI work.

The ceiling on demand is beginning to disappear.

The supply side, however, remains constrained by physics.

In our previous article, we argued that the most important bottleneck was becoming power.

Since then, our thinking has developed one step further.

Finance is beginning to enter the scarcity created by that power constraint.

And finance may become the third accelerator that transforms today's genuine compute shortage into a much larger investment cycle — and eventually, a bubble.

The Three Accelerators of the AI Revolution

We now think about the AI industrial cycle through three accelerators.

The first is:

GPU = Compute Accelerator

GPUs accelerate the enormous amount of computation required by AI.

The second is:

AI = Innovation Accelerator

AI accelerates software development, semiconductor design, scientific research, drug discovery, materials science and increasingly intellectual work itself.

The third is now emerging:

Finance = Capital Accelerator

AI infrastructure requires enormous amounts of capital.

But long-term compute contracts, customer prepayments, GPU- and data-center-backed lending, project finance and private credit can allow companies to build infrastructure far beyond what their own equity could support.

In simple terms:

GPUs accelerate computation.

AI accelerates innovation.

Finance accelerates the construction of AI infrastructure.

These three accelerators are beginning to reinforce one another.

Yet all three ultimately collide with one physical bottleneck:

Energy.

Power Scarcity Can Turn Compute Into a Premium Asset

This is the central idea.

At first glance, power shortages appear to be a brake on the AI investment boom.

We increasingly believe they may have the opposite effect.

Power scarcity can prolong the shortage of compute and transform compute into a premium asset.

AI demand can grow extremely quickly.

GPU production can also expand with sufficient investment, although with a lag.

Power plants, transmission lines, substations and grid interconnections cannot expand nearly as quickly.

The result is a persistent imbalance:

AI demand > available compute

That imbalance creates a scarcity premium.

And that premium changes the economics of investment.

A data-center project that would not meet its hurdle rate under normalized compute prices may become economically attractive when compute commands a substantial scarcity premium.

The cycle then becomes:

Power shortage→ Compute shortage→ Scarcity premium→ Higher expected returns on compute assets→ Lower investment hurdle→ More financial capital enters→ More AI infrastructure gets built

Real demand creates scarcity.

Scarcity creates a premium.

The premium attracts finance.

Finance accelerates the creation of more compute.

We believe this feedback loop may become the core mechanism through which the next phase of the AI bubble is formed.

A Bubble Built on Something Real

This is not a bubble built on imaginary demand.

The demand is real.

Agents are increasing inference workloads.

Power constraints are preventing supply from responding quickly enough.

Compute is genuinely scarce.

And that is precisely why financial capital can enter with confidence.

The sequence is therefore:

Real demand creates a scarcity premium, and finance amplifies that premium into an investment boom.

The bubble is not AI itself.

It is the scarcity premium created by constrained compute supply — and the amount of financial capital eventually deployed on the assumption that this premium will persist.

Compute Is Structurally Deflationary

There is an important distinction here.

AI compute is fundamentally a rapidly deflationary industry.

Better GPUs, higher performance per watt, inference optimization and more efficient models continuously reduce the cost of performing a given amount of AI work.

Compute prices should decline over time.

That is not, by itself, a bearish signal.

What matters is whether usage grows faster than the cost of compute falls.

The current dynamic looks more like:

Compute cost ↓→ AI usage ↑↑→ Total inference demand ↑

This is the Jevons effect applied to intelligence.

So a decline in dollars per token, or even in compute pricing more broadly, should not automatically be interpreted as the end of the cycle.

The critical variable is not the absolute price of compute.

It is the scarcity premium embedded in that price.

The Premium Creates the Bubble — and Its Disappearance Ends It

This also tells us how the bubble may eventually end.

We do not believe it requires AI demand to collapse.

The key is power supply eventually beginning to catch up with AI infrastructure demand.

As generation, transmission, grid interconnection and data-center construction expand — while GPU performance per watt continues to improve — effective compute supply will rise rapidly.

Consider a simplified example.

The market begins with:

Demand = 100Supply = 60

Compute is extremely scarce.

Several years later:

Demand = 200Supply = 180

AI demand has doubled.

Demand still exceeds supply.

There is still no absolute compute glut.

But the extreme scarcity premium that existed when supply was only 60 can no longer be sustained.

The cycle then reverses:

Power constraint eases→ Scarcity premium compresses→ Expected returns on compute assets decline→ Investment hurdle rates rise→ Collateral values weaken→ Financing conditions tighten→ New infrastructure investment slows

AI can continue succeeding.

AI usage can continue growing.

Compute can continue becoming cheaper through technological progress.

And the financial bubble can still burst.

Power scarcity creates the premium.

The premium helps create the bubble.

Power supply catching up removes the premium.

And the disappearance of that premium reverses the financial cycle.

Crucially, financial markets will not wait until compute supply actually exceeds demand.

The turning point may come while compute is still scarce — once investors become convinced that the physical bottleneck is beginning to disappear.

What If the Financial Bubble Is Still in Its Earliest Stage?

This is where our view has changed most.

The AI boom itself is obviously not new.

GPU spending is already enormous.

Data-center investment is already enormous.

But the third accelerator — finance — has only recently begun to enter the compute buildout in a much more meaningful way.

Long-term compute contracts can support financing.

GPUs and data centers can support asset-backed lending.

Customers can prepay for capacity.

Private credit can fund infrastructure.

Project-finance structures can reduce the amount of equity required to build each new unit of compute.

Compute is beginning to evolve from equipment into a financeable asset class.

That leads us to an important distinction:

The AI boom is already well underway.

But the financial bubble around AI compute may still be in the earliest stages of formation.

Technology alone did not create the Industrial Revolution.

Finance accelerated the construction of factories, railways and ships beyond what individual entrepreneurs could have funded themselves.

The same mechanism may now be coming to AI.

Technology creates the opportunity.

Finance accelerates industrialization.

And eventually, finance can create overshoot.

Open-Weight Models May Increase Compute Demand

The rapid improvement of open-weight models is sometimes viewed as a threat to AI infrastructure demand.

We see another possibility.

As models become cheaper and more accessible, the cost of an individual inference falls.

But if lower prices cause usage to grow even faster, total inference demand rises.

There is another consequence.

In a world of Llama, Qwen, DeepSeek, Kimi and many other competing models, a cloud provider cannot know with certainty which workload will dominate six or twelve months from now.

That uncertainty increases the value of:

compute that can run almost anything.

This favors flexible, general-purpose AI infrastructure.

It is one reason we have become more convinced of the durability of NVIDIA-based compute.

A specialized accelerator may deliver extraordinary performance on a particular workload.

But a broadly programmable NVIDIA cluster offers optionality across models, customers and workloads.

That can support utilization, residual value and ultimately financeability.

The cycle may therefore be:

Lower model costs→ Broader AI adoption→ More inference→ Greater demand for flexible compute

Open-weight AI may accelerate compute demand rather than destroy it.

Why SoftBank Group?

This framework makes SoftBank Group particularly interesting to us.

Through OpenAI, it has exposure to Intelligence.

Through Arm, it has exposure to Compute Architecture from the cloud to the edge.

Through initiatives such as SB Neo, it is moving into Compute itself.

Through large-scale AI and energy infrastructure development, it is addressing the industry's most important physical bottleneck: Energy.

And SoftBank Group brings another core capability:

Finance and Capital Allocation.

In other words, it is attempting to position itself across:

Intelligence × Compute × Energy × Finance

As AI moves from technological invention toward industrial deployment at hundreds of megawatts, gigawatts and potentially tens of gigawatts, the ability to recognize technological change, assemble capital and scale infrastructure becomes increasingly valuable.

These are areas in which SoftBank Group has historically been strong.

SoftBank Group's Weakness: Innovation Creation

There is, however, an important weakness that should not be ignored.

NVIDIA created CUDA.

SpaceX created reusable orbital rockets and Starship.

OpenAI created frontier AI models.

SoftBank Group's historical strength has been different.

It has demonstrated an unusual ability to recognize technological transitions, allocate capital aggressively and scale what it believes will matter next.

But its record of creating foundational technological innovation internally is less established.

Innovation Recognition is not the same thing as Innovation Creation.

This is why we continue to watch Izanagi closely.

As discussed in our previous article, the assets and capabilities assembled or pursued around Arm, Ampere, Graphcore and other semiconductor targets suggest an ambition extending beyond passive investment.

If Izanagi ultimately produces a genuinely differentiated AI compute architecture or AI-factory platform, it could represent something new for SoftBank Group:

not merely identifying the next technology, but creating it.

Whether Izanagi achieves that remains uncertain.

But if it does, it could address one of the most important weaknesses in the SoftBank Group investment thesis.

Beyond NAV

In our previous article, one reason for resuming our SoftBank Group position was its substantial discount to NAV.

That remains relevant.

But the compute thesis introduces another possibility.

If businesses such as SB Neo begin producing:

Contracted GW→ Revenue→ EBITDA→ Free cash flow

then SoftBank Group begins to acquire operating value that cannot be captured simply by adding up the market value of its holdings.

And if it can repeatedly use long-term compute contracts and external finance to build infrastructure, recycle capital and fund the next project, it becomes more than a holding company.

It becomes a capital-allocation engine capable of creating new NAV and new cash flows.

Masayoshi Son himself has used the metaphor of the "goose that lays the golden eggs," arguing that investors should value not only the eggs but the goose that continues to create them.

AI compute may provide a way for that argument to become visible in operating numbers.

The potential re-rating would then come not only from rising NAV, but from the market assigning value to the process that creates future NAV.

Beyond that lies another potential growth cycle.

Large-scale reasoning will remain concentrated in the cloud, while automobiles, robots, industrial equipment, PCs and other devices increasingly run AI at the edge.

That creates another growth vector for Arm.

And as Physical AI develops, cloud compute and edge compute should reinforce one another.

Why We Are More Bullish Than Before

In our previous article, we argued that the nature of AI demand had changed and that power was becoming the binding physical constraint.

We still believe that.

What has changed is that we now see another mechanism forming on top of those forces:

Finance = Capital Accelerator

Our current framework is therefore:

Compute Accelerator × Innovation Accelerator × Capital Accelerator

confronting one major physical bottleneck:

Energy.

Power scarcity creates compute scarcity.

Compute scarcity creates a premium.

The premium attracts financial capital.

Financial capital accelerates infrastructure construction.

Eventually, power supply catches up.

The scarcity premium begins to disappear.

And the financial cycle reverses.

Real demand creates the premium.The premium creates the bubble.The disappearance of the premium ends the bubble.

But the third accelerator — finance — has only recently begun to operate at scale.

That is why we increasingly view the AI investment cycle not as a bubble already approaching its end, but as a financial and industrial cycle that may still be near the beginning of its most powerful phase.

And if this framework is correct, SoftBank Group — with exposure to Intelligence, Compute, Energy and Finance — occupies one of the most interesting positions in that cycle.

That is why our conviction has increased since our previous article.

We remain bullish.

But precisely because the thesis is becoming stronger, we intend to keep looking for evidence that could disprove it.

The variable we will watch most closely is not simply compute prices, which should decline structurally with technological progress.

It is whether power scarcity itself is beginning to disappear — and with it, the scarcity premium that supports the economics and financing of AI compute.

That, in our view, is where the eventual turning point will be.

This article reflects the views of Fin Inc. and the author as of the date of publication and describes positions we hold. It is provided for informational purposes only and does not constitute investment advice, a research report, or a recommendation or solicitation to buy or sell any security. Fin Inc. and/or its affiliates, including the author, hold a position in SoftBank Group and may buy or sell it at any time without notice. Statements regarding Izanagi, future AI infrastructure development, compute financialization and market cycles include interpretations and forward-looking inferences based on public information and may prove incorrect. Figures and industry observations should be independently verified. Investing involves risk, including possible loss of principal. Please make investment decisions at your own responsibility and consult a licensed financial adviser where appropriate.