The Return of the Real World: When AI Strategy Becomes Industrial Strategy

A formula from Jensen Huang started this series:

Input: electrons. Output: tokens. In the middle: Nvidia.

I first read it as a claim about technology. The previous three essays followed that claim from chips into systems, from systems into institutional knowledge, and from there into the relationships and trade-offs required to keep a complex production network working.

By the end of the summer, the boundary had moved again.

Nvidia’s DSX architecture now reaches across chips, networking, software, power, cooling and facility operations. In its July filing, the company reported $279 billion of supply and capacity commitments, primarily around memory and manufacturing capacity. In August, it announced financing partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR intended to mobilise more than $500 billion of third-party capital for AI infrastructure. A week later, Nvidia agreed to provide credit support around land, power and shell for the first 4.25 gigawatts of a data-centre campus in Ohio, with OpenAI as the expected customer¹.

This is still recognisably the same strategy Huang described in April, but the object being coordinated is much larger than a GPU.

That makes the final question of the series fairly simple: how far can Nvidia extend the system it optimises without having to own it?

From the cloud back to the ground

One of the great achievements of cloud computing was to hide the physical machine.

A software company did not have to know where its servers were, how their electricity was contracted, or which transformer served the building. AWS and its peers turned an industrial asset into something that felt like a variable input. Capacity appeared behind an API.

This changed the kinds of companies the software era produced. A small team could reach millions of users without financing factories or distribution networks. Product cycles shortened. Marginal distribution costs approached zero. The physical infrastructure remained enormous, but from the application developer’s point of view it mostly disappeared.

Frontier AI brings part of it back into view.

A large AI cluster needs advanced logic, HBM, packaging, networking, optical components, power delivery, cooling, land and a grid connection. These inputs come from different industries with very different clocks. A model can change in months. Semiconductor capacity takes years to plan. Electricity infrastructure can take longer again.

The IEA captures the mismatch well. Its base case has global data-centre electricity demand roughly doubling to 945 TWh by 2030, but the more interesting observation is about timing: a data centre can often be operational within two or three years, while the energy infrastructure around it generally takes longer to plan and build². We worked through the same mismatch from the energy side in Jevons Meets Jensen, the closing part of our AI × Energy series.

The result is not that power has permanently replaced compute as the bottleneck. It is that the bottleneck moves.

Bottleneck migration

The first obvious shortage of the generative-AI boom was GPUs. Then advanced packaging and HBM became harder to ignore. Larger clusters put more pressure on networking. Higher rack densities pulled liquid cooling and power delivery into the design problem. In several markets, available megawatts and grid interconnections now limit how quickly installed compute can grow.

None of these constraints arrives in a clean sequence, and none disappears when another becomes more visible. A project can secure the chips and wait for a transformer. Another can have power and struggle to finance the equipment. A cloud operator can build capacity and discover that utilisation is the scarce thing.

If the output of the system depends on several complements arriving together, the value of improving any one component depends on the state of the others. A faster accelerator does little if memory starves it; another rack produces no tokens if the site cannot power it; an available gigawatt creates no revenue if the machines cannot be financed or the customers do not arrive. The strategic question shifts accordingly: not only what is scarce today, but where the next constraint is likely to appear and whether the system can reorganise around it.

Nvidia has repeatedly moved toward those interfaces. Mellanox put networking inside the strategic perimeter before large distributed clusters made that decision look obvious. NVLink and NVSwitch did the same within the rack. CUDA ties software performance to the hardware roadmap. DSX now extends the design surface into the facility itself.

Nvidia describes DSX as a framework spanning chips, systems, software, facilities and partner technologies, with software that optimises token performance per megawatt³. The language is self-serving, as product language usually is, but it tells us what Nvidia thinks the product has become: less the chip, or even the rack, than the functioning AI factory.

Integration without ownership

The natural response to a coordination problem is vertical integration: if two activities need to be tightly synchronised, put them inside the same firm. Nvidia has mostly chosen another route. It does not own TSMC, the HBM manufacturers, most of the ODMs assembling its systems, the utility or the data-centre developer. The new financing platforms are explicitly independent rather than Nvidia-owned lending vehicles. Instead, Nvidia tries to set enough of the architecture, interfaces and economic incentives that these independently owned pieces arrive in a compatible form and on a compatible schedule.

It is also a useful way to reread the first three essays. CUDA is not only software lock-in but a common programming target across changing hardware. NVLink does more than move data faster; it helps define how accelerators behave as a larger machine. Reference designs reduce uncertainty for manufacturers and operators, while long-dated purchase commitments make it easier for suppliers to build capacity against future demand. DSX carries the same logic into power, cooling and operations. The financing initiative carries it into the cost and availability of capital.

Ronald Coase’s old question about the boundary of the firm is relevant here, but Nvidia’s answer is not simply to make the firm larger⁴. It is trying to make coordination across firms cheaper without absorbing all of those firms onto its own balance sheet.

You could call this architectural integration without ownership.

It is a very attractive position when it works. Nvidia keeps the high-margin layer and leaves most of the physical assets elsewhere, while still influencing how the surrounding system is designed.

The catch is that influence increasingly comes with commitments.

How asset-light is asset-light?

Nvidia remains fabless, and its conventional capex is tiny compared with the fabrication and data-centre capacity its products rely on. The description is accurate, but it captures less of the economic perimeter than it used to.

As of July 26, 2026, Nvidia disclosed $366 billion of future commitments across several categories. The largest was $279 billion of supply and capacity commitments, mostly memory and manufacturing capacity. It also disclosed $29 billion of cloud-service agreements, $25 billion of data-centre leases not yet commenced, $25 billion of equity-investment commitments and $8 billion of conventional capex commitments⁵.

The categories should not be collapsed into one pseudo-capex number. Nvidia states that some supply agreements can be cancelled, rescheduled or adjusted, and the accounting treatment differs significantly across them.

What the filing does show is how much future industrial activity Nvidia now helps organise. Its fabless model works because foundries, memory makers, assemblers, data-centre developers and customers are willing to build around expectations of Nvidia demand. As the sums grow, that coordination requires more than a good product roadmap: it requires credible volume commitments, long-term customer demand and, increasingly, financing structures able to bridge the time between construction and utilisation. The distinction between operator and investor starts to blur at exactly this point.

Capital becomes part of the deployment stack

In August, Nvidia announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to create independent financing platforms for AI compute infrastructure. The stated ambition is to mobilise more than $500 billion of third-party capital over time⁶.

The announcement uses the language of a new asset class. Nvidia wants its compute to be treated as durable, fungible infrastructure capable of supporting long-duration financing.

There is a clear industrial precedent for this move. Productive assets do not always have to be owned by the company using them. Aircraft, telecom towers, renewable-energy projects and industrial equipment are routinely financed through structures that separate use from ownership, and AI infrastructure is beginning to borrow from that playbook. What is less conventional is the role of the technology supplier itself: Nvidia is helping to build the financing market because access to capital can determine whether its systems are deployed in the first place.

The Ohio project announced with SB Energy makes that explicit. Nvidia agreed to provide credit support around land, power and shell for an initial 4.25 IT-GW of capacity at PORTS-Pike, with an option over another 3.75 IT-GW. SB Energy is expected to build, own and operate the data centre under a 20-year lease to OpenAI, while Nvidia supplies the compute architecture and has invested $1.5 billion in SB Energy⁷.

The legal and economic exposure is limited compared with owning the whole site, but the direction is notable. Nvidia is helping remove a downstream constraint so that more Nvidia systems can be deployed.

Amazon is exploring a different structure for the same underlying problem. Reuters reported in October that Amazon had considered placing roughly $8 billion of Nvidia Grace Blackwell hardware in a special-purpose vehicle funded by outside investors and leasing the chips back⁸. Broadcom has gone further in another part of the market, offering financing connected to Anthropic’s infrastructure commitments⁹.

Once structures like these become material, interest rates, lease terms and residual-value assumptions sit surprisingly close to the cost of a token.

The awkward question of asset life

This is where the industrial analogy becomes less comfortable.

A lender financing an aircraft has decades of evidence about how the asset depreciates, how it can be redeployed and what a used engine is worth. A GPU is being financed into a market where the supplier is simultaneously promising a much better generation every year or two.

Nvidia argues that high-end compute can remain productive for many years. Credit investors are more cautious. Reuters reported on October 1 that banks and private-credit investors commonly underwrite GPUs on depreciation schedules closer to three or four years, and have been reluctant to rely on long-dated residual values without strong customer contracts, guarantees or other forms of credit support¹⁰.

The disagreement is not mainly about accounting. It is about the rate of technical progress.

If Rubin materially lowers the cost of inference relative to Blackwell, an older Blackwell system can remain physically functional while becoming economically less competitive. Strong demand can still absorb it at a lower tier. Software improvements can extend its useful life. CUDA compatibility can make redeployment easier. None of that produces a stable residual-value curve yet.

Nvidia is therefore trying to create a financeable infrastructure asset while continuing to improve that asset class unusually quickly. The faster the frontier moves, the more compelling the next generation becomes and the harder it may be to finance the previous one on infrastructure-style durations. Wall Street’s hesitation does not invalidate the strategy; it shows that the asset class Nvidia wants to create does not yet have the operating history that makes infrastructure finance boring.

Where the moat stops and the exposure starts

Parts 1 to 3 mostly looked at the upside of coordination: system design can produce gains that are invisible in a chip benchmark, institutional knowledge compounds across repeated engineering cycles, and long relationships can preserve more value than maximum extraction in a single transaction. The same logic also creates exposure.

A company that coordinates more layers becomes sensitive to more ways those layers can fail. Nvidia’s own July filing now names the availability of land, power, shell and capital as factors that can delay customers’ deployments and affect Nvidia’s revenue timing¹¹.

That list would have sounded odd in an Nvidia filing a decade ago.

There are several ways this can go wrong without AI itself disappointing. Power can arrive later than the compute; customers can sign infrastructure commitments faster than they create economic demand; hyperscaler ASICs can become good enough for a larger share of workloads; lenders can decide that GPUs depreciate too quickly to support the expected leverage; or large clusters can run below the utilisation assumptions used to justify them. Suppliers can also build capacity against forecasts that later move, while regulators may become more concerned about one architecture becoming embedded across critical infrastructure. At the limit, the bottleneck could migrate out of infrastructure altogether: if compute, power and capital become abundant before valuable applications catch up, demand becomes the scarce input.

This is the part of the Nvidia story that a pure “moat” analysis misses. The architecture can become more defensible at the same time as the company becomes coupled to a wider set of external assumptions.

The question is no longer only how hard Nvidia is to displace. It is also how much of the surrounding system has to remain economically coherent for Nvidia’s commitments to pay off.

Five questions I would keep

The original series was supposed to end with a map of the AI stack. I think a set of questions is more useful.

1. Where is the bottleneck moving?

Not which resource is scarce in the abstract, but which resource limits the next unit of useful deployment: logic, HBM, packaging, networking, power, cooling, land, capital or demand.

2. Is Nvidia still improving system economics, not only component performance?

The relevant evidence increasingly sits in utilisation, token cost, time to production and performance inside a fixed power envelope. Vendor benchmarks are a starting point, not the answer.

3. Can Nvidia keep coordinating complements it does not own?

Suppliers have to invest, data-centre developers have to build, customers have to commit, financiers have to lend and developers have to stay inside the software ecosystem. Nvidia does not need to own those actors, but its economics increasingly depend on their plans lining up.

4. Does each extension of the perimeter strengthen the moat more than it adds risk?

Mellanox looks, in retrospect, like a clean extension of the architecture. Credit support for gigawatt-scale data centres is a different kind of move. At some point solving a customer’s deployment constraint can shade into financing the demand for your own product. The boundary will have to be judged deal by deal.

5. What would make the systems thesis less convincing?

A few developments would matter.

Hyperscaler ASICs could reach comparable system economics across a much broader range of workloads.

Rack-level architectures could standardise enough that accelerators become easier to substitute.

CUDA could lose its position as the default development target.

Financing Nvidia systems could require progressively stronger guarantees because lenders refuse to take technology risk.

Or infrastructure could simply cease to be scarce before the installed base earns an adequate return.

Those are the conditions I would use to test the argument.

Redrawing the stack

Huang’s stack in the April interview was:

chips → systems → cloud → models → applications

The unusual choice was to separate chips from systems. That distinction set up much of this series, starting with Part 1.

The map now needs another layer underneath it:

energy + land + capital
↓
AI factory: power + cooling + networking + compute
↓
models
↓
applications and workflows
↓
useful output

Even that diagram is too clean. Capital does not sit neatly at the bottom. Software reaches all the way down into the operation of the data centre. Feedback from applications changes which models get built, which changes what compute is valuable. Supply commitments alter the incentives of suppliers years before a rack is switched on.

The useful thing about drawing the stack is therefore not the taxonomy itself. It is noticing which interfaces the taxonomy hides.

That is where Nvidia has spent much of the last decade.

It began with a programmable accelerator and the software around it. It connected accelerators together, then racks, then clusters. It bought networking. It worked further into the supply chain. It is now publishing reference architectures for the facility and helping organise the capital required to build it.

Calling Nvidia a semiconductor company is still technically correct. Calling it a software company, as Huang prefers, captures another part of the economics.

After following the company through these four essays, “systems company” seems closer to the strategic description.

Its advantage is not that it owns every component. It is that it has repeatedly identified which interface is preventing the overall system from producing more useful output, then moved close enough to that interface to influence it.

That also gives the series its limit.

The system cannot expand forever without changing the nature of the company. A supplier that begins guaranteeing customers, supporting infrastructure finance and taking equity stakes is doing more than selling a better chip. Some of those moves can strengthen the architecture. Others may eventually tell us that the architecture needs financial support to keep growing at the expected rate.

That is what I would watch now.

The formula remains a good one: electrons in, tokens out.

The interesting part was always the middle.