The Shift from Software to Compute

01 · The capital signal

Concentration is not the same as conviction.

The Q3 2026 PitchBook-NVCA Venture Monitor reports that AI captured 82.7% of U.S. venture deal value through September. That supports the claim that more than 80% of venture dollars have concentrated in AI, but only with the geography, period and denominator attached. [1]

It does not mean 82.7% of all startups were AI companies. It does not describe every venture market worldwide. And it does not mean that 82.7% went directly into chips or data centers. Funding a model developer is not the same transaction as financing a power plant.

This distinction matters because a few enormous financings can dominate a dollar-weighted statistic. Capital concentration tells us where investors are placing large bets; it does not establish that those bets will earn attractive returns. A great technology, a great business and a great investment are three separate propositions.

The more useful reading is that AI has drawn attention toward the resources required to deliver intelligence at scale. KPMG’s Q3 review describes investment spanning models, infrastructure, autonomous systems and industry-specific applications. This is a broader ecosystem, not a single trade. [2]

My interpretation: the investment map is expanding downward into physical constraints even as entrepreneurs continue building upward into software experiences. The layers are becoming more interdependent. The statistic is a starting point for diligence, not a substitute for it.

02 · The physical stack

The cloud has a street address.

A digital product can hide a long physical dependency chain. Accelerators need memory and fast connections. Servers need cooling and electrical equipment. Buildings need land and permits. Power commitments need a practical path to delivery. One constrained layer can leave expensive assets elsewhere waiting.

Compute is therefore not simply a count of chips. What matters economically is usable computation: the right hardware and software, available at the right time, with adequate reliability, networking and cost. An idle accelerator is capacity on paper, not productive output.

This stack is a conceptual map, not a claim that every application must own the layers beneath it. Renting infrastructure can be rational. Owning it can be rational. The question is where a company needs control and where flexibility is worth more than ownership.

03 · Power and time

Electricity is part of the product roadmap.

The International Energy Agency’s 2026 update projects global data-center electricity use rising from 485 TWh in 2025 to about 950 TWh in 2030. This is an outlook, not a guaranteed outcome, and it includes non-AI computing. It makes the industrial scale of the opportunity visible. [3]

For an investor, the important distinction is between an attractive location and an executable project. Land is not energized capacity. A power agreement is not necessarily an operational connection. A queue position is not a completed substation. Dates, obligations and dependencies deserve as much attention as the headline megawatts.

The commercial clock and the construction clock also differ. Customers may change their requirements while a facility is being built. Financing costs can accrue before revenue begins. A delay can affect the value of a project even when the underlying demand story remains persuasive.

Nor is the only answer to build more. More efficient models, better scheduling, improved utilization and appropriate task routing can make existing capacity more productive. Lower cost per task may also encourage more use. Neither efficiency nor demand growth should be treated as a one-variable forecast.

The useful question is not simply who has announced the most capacity. It is who can deliver reliable, economically useful capacity on a schedule that customers need and contracts can support.

04 · The economics of intelligence

Software is not dead. Its economics need inspection.

Software was never weightless; the cloud merely allowed many companies to rent its physical requirements. AI makes those requirements harder to ignore when a product’s usage carries meaningful model-serving costs. That is not true to the same degree for every company, but it changes what investors should measure.

Training builds or adapts a model. Inference uses it. The two have different spending patterns, and a company may pay for either directly or through a provider. More users can bring more revenue while also bringing additional computation, tool calls, storage, monitoring and human review.

A subscription price therefore tells us very little on its own. The analytical task is to connect the value delivered to the resources consumed. A product that saves a customer substantial time may support strong pricing. A product that attracts heavy usage without equivalent willingness to pay can create impressive engagement and disappointing cash economics.

My preferred question is what it costs to produce an accepted outcome at the required quality and latency. A cheap answer that needs repeated correction may be expensive in practice. A stronger model may cost more per request while completing the task in fewer attempts. Verification and rework belong in the comparison.

Applications can still be excellent businesses. Distribution, workflow integration, trusted data access and customer relationships may be more defensible than a temporary model advantage. Infrastructure enables value; software often determines whether someone actually receives it. The opportunity is the connection between the two, not a contest in which only one can survive.

05 · Match the capital

Not every AI asset belongs in a venture fund.

This is an illustrative financing map, not a claim that every project fits neatly into one category. The distinction is important: venture dollars raised by a startup, corporate capital expenditure and debt raised against infrastructure are different pools of capital. Adding them together without care can exaggerate the scale of investment or count the same economic activity twice.

The financing structure should reflect the risk. Venture equity can absorb uncertainty in pursuit of exceptional upside. An asset financed with fixed obligations needs a credible path to cash generation. Hardware may age economically faster than a building, while customer commitments may have different terms from the debt financing both.

Before underwriting the AI narrative, I would examine delivery dates, contract enforceability, customer concentration, utilization, maintenance requirements, refresh spending and downside liquidity. A forecast that works only at full capacity and favorable refinancing terms is not made robust by attaching AI to its description.

06 · Judgment and responsibility

Scarcity is a thesis. Durability is a question.

A bottleneck can support pricing power. It can also attract the supply that eventually weakens it. The essential question is whether an advantage rests on a temporary shortage, a difficult-to-replicate capability, a durable customer relationship, or some combination of the three. Those are different investment cases.

I would look for evidence in five places: useful performance per dollar; credible access to power and capacity; recurring customer value; resilience under lower prices or slower demand; and a capital structure that can survive delays. The right entry valuation remains part of the analysis. Strong demand does not make every price sensible.

I would also separate commercial independence from a circular financing story. When investors, suppliers and customers overlap, the relationships may be strategically useful. But diligence should trace who ultimately pays, whether obligations are binding, and how much demand survives without continued external financing.

The social questions are equally practical. Who pays for grid upgrades? How are water and land constraints handled? Which communities receive benefits, and which bear costs? Who can access affordable compute? These are questions for project design and public negotiation, not obstacles to be dismissed as resistance to progress.

There is real promise here: better tools for research, more capable small businesses, more useful services and new forms of creation. But an infrastructure boom is not the same thing as a social dividend. The latter depends on what gets built on top, how widely benefits spread, and whether customers and communities are better off.

The shift from software to compute is ultimately a shift in attention. The screen is no longer enough. We have to see the industrial systems behind it, the financial obligations beneath them, and the human purposes they are meant to serve.

The next great advantage may be knowing which parts of the future must be built—and which are merely being priced as inevitable.

Sources & reading notes

What the evidence says. What this essay argues.