(Jensen Huang, Founder and CEO of NVIDIA)
Welcome to Global Data Center Hub. Join investors, operators, and innovators reading to stay ahead of the latest trends in the data center sector in developed and emerging markets globally.
Jensen Huang has a habit of calling the next phase of AI infrastructure before the market prices it in.
In a recent long-form interview on an investor podcast, he laid out where the build-out stands heading into 2027: what is scarce, what is getting more expensive, and who is actually paying for it.
The conversation spent little time on chatbots or model launches.
It went to the physical and financial layers: the sites, the power, the cost per gigawatt, and the money that turns chips into running capacity.
The message was clear: the chips are no longer the hard part. Power, sites, and financing are.
One caveat governs everything that follows: Huang runs the company that sells most of the world’s AI chips, and the hosts are Nvidia shareholders. Every claim about demand, market share, and durability comes from people with a direct stake in the answer.
Most of his argument holds up. Three parts do not, or at least not yet.
Here are the ten takeaways where he is right, and the three where investors should push back.
The Bottleneck Has Moved
1. Chips are no longer the binding constraint.
Huang named land, power, and shell (buildings ready to receive equipment) as one of the most important obstacles to next year’s build-out. The host relayed a street estimate of roughly 50 GW of announced chips against 35 to 40 GW of sites that may be ready. The exact gap is an estimate, but the direction matches what the market is showing: chips can be made faster than sites can be energized. The scarce asset is now the ready-to-power site, not the GPU.
2. Power scarcity rewards the most efficient chip.
The common view is that a power crunch hurts chipmakers by capping how many chips can be installed. Huang argues the opposite. When every watt is limited, buyers who still need to hit revenue targets choose the hardware that produces the most output per watt. He says this is why Nvidia’s share is rising even at its current scale. The logic is sound: when one input is scarce, whatever uses it best becomes more valuable.
3. Supply will follow ready sites, not borders.
Because demand exceeds supply, Huang said Nvidia can redirect chips to whichever customers and regions have sites ready. He cited early build-outs in Africa, Central Asia, Southeast Asia, Latin America, and the Middle East, and said he does not like that the U.S. may not keep pace. For markets that can deliver power and permits quickly, this is a direct opening: the world’s largest supplier is telling buyers it will go wherever the megawatts are.
The Cost Curve Has Turned
4. The end of Moore’s law makes spending rise on its own.
For decades, twice the computing power cost the same a year later. Huang says that era is over: twice the transistors now cost twice as much, however long you wait. That means capex climbs even if demand for compute stays flat. He put pre-AI computing spend at about $0.5 trillion a year and sees compute alone reaching about $1.5 trillion by 2030, inside $3 to 4 trillion of total annual infrastructure spending. Most AI spending debates leave this effect out.
5. Each gigawatt now carries more than twice the capital.
Huang said Nvidia’s content per gigawatt rose from about $18 billion to more than $40 billion as the company added networking, CPUs, storage, and security to what it sells. He expects the number to keep rising. He also contrasted roughly $5 billion per gigawatt for general-purpose computing with about $50 billion for the newest systems. For anyone sizing debt or equity per megawatt, older cost assumptions understate the capital each campus now carries.
The Demand Base Is Wider Than It Looks
6. Two AI labs are not carrying the build-out.
The bear case asks whether two frontier labs can earn enough to justify trillions in spending. Huang split Nvidia’s demand into rough thirds: everyday computing such as data processing and simulation, five or six AI labs rather than two, and thousands of startups renting compute through neoclouds. He said about 80 percent of those startups use open models to build new products rather than compete with the labs. The exact split is his, but the broader point is right: demand is more spread out than the headlines suggest.
7. AI labs will build a copy of the internet.
Huang expects labs to build digital twins of the internet, so agents can learn to use websites and tools without acting on the live web. It is the same approach robotics already uses, where machines train in simulated factories before entering real ones. Few people are modeling this, but it would add a new, compute-heavy layer of demand: simulation for agents, on top of training and inference.
8. Breakthroughs are now search problems.
The host cited a frontier lab solving the Navier-Stokes problem with about 10,000 agents and an estimated $4 millions of compute. Huang’s point was that the result rested on decades of human work, and the AI’s role was to search every path from there. He believes many scientific fields sit at a similar edge. If he is right, science becomes a large and recurring buyer of compute, and the cost of each breakthrough falls as compute gets cheaper.
How Compute Gets Bought and Deployed
9. Orders are placed before the product exists.
Huang said buyers now commit $50 to 100 billion to systems that have not been built. A system at that scale cannot be tested in advance, because testing it would take a giant data center and years of work. Trust in the supplier’s track record replaces testing, the same way Nvidia places all its leading-edge chip production with one foundry it can plan with five years out. This changes how delivery and counterparty risk should be judged on large compute contracts.
10. Safety will be engineered, not aligned.
Huang argued that perfect alignment is impossible, because goals like “be successful” can be read countless ways. His answer mirrors how companies treat new employees: start with zero trust and minimum access, monitor from outside the system so the AI cannot adjust its behavior, and widen limits as trust is earned. This matters for infrastructure because enterprise adoption, the next large wave of demand, depends on agents being safe to deploy inside corporate networks.
Where He Is Not
Three claims deserve more scrutiny than the interview gave them.
1. “Flexible” does not mean “durable.”
Huang’s financing argument runs as a chain: general-purpose chips can be re-rented to any customer, for any model, in any region, which makes them durable, easier to finance, and cheaper to own. The first link holds. The second does not follow. He notes Nvidia still updates software on chips sold eight years ago, but software support is not economic life. Each new generation needs more power per rack and better cooling, so older chips and the buildings around them lose value even while they still work. Lenders should price GPUs on their economic life, not their physical one.
2. The chipmaker as lender is a risk, not just an upside.
Huang presented two new business lines as pure growth: software licensing at a few thousand dollars per GPU per year, and revenue share with neoclouds that Nvidia helps finance, supply, and fill with customers. The second deserves a harder look. When one company is supplier, investor, and source of demand, the risks are linked. If demand softens, the neocloud, its lenders, and Nvidia’s revenue share all weaken at once. Operators taking these terms should know exactly what they give up and how dependent they become.
3. The tenfold jump in value per gigawatt is asserted, not proven.
Huang’s own numbers set the test: about $5 billion per gigawatt for general-purpose computing versus about $50 billion for the newest systems. For that spending to pay off, each gigawatt must produce roughly ten times more economic value. He argues it will, and a wider base of buyers helps. But spreading the bill does not shrink it. Someone still has to earn ten times more per gigawatt, and the interview offered conviction rather than evidence.
Strategic Implications for Digital Infrastructure
Strip out the book-talking and three implications survive for anyone investing in or operating digital infrastructure.
First, the energized megawatt is the asset. If chips can be made faster than sites can be powered, value moves to capacity that has cleared every stage: announced, permitted, financed, and energized. Announced capacity should be discounted until it shows the later stages.
Second, cost models need updating. With Nvidia content per gigawatt more than doubling, per-megawatt capex, debt sizing, and equity checks built on older assumptions understate the capital each campus needs. Models should carry both the higher cost and a shorter economic life for the hardware inside.
Third, emerging markets with ready power gain negotiating strength. The largest supplier says it will send chips wherever sites are ready. Markets that can deliver power and permits on a credible timeline can attract capacity that U.S. sites cannot absorb.
The trade-off is dependence: vendor-financed, single-anchor builds should be priced for that concentration.