At the G20 Innovation Ministerial in Chapel Hill, North Carolina, U.S. Commerce Secretary Howard Lutnick asked Jensen Huang a simple question on behalf of every government in the room. If a country wants to participate in the AI revolution, what should it actually build? AI, he said, is a five layer cake. Most people think AI is only the models. It is not. The models sit in the middle of a stack, and every layer under and above them has to exist for the thing to work at all. I posted the layers on X earlier this week, so here is the longer version: * What each layer does * Which companies sit in it * What the research is currently saying about the money flowing through the whole structure Layer 1: Energy Jensen started at the bottom, and the ordering is the point. Nothing in AI happens without electricity. His description of what a data center really does is the cleanest one I have heard: it transforms electricity into mathematics. That makes power generation the true base of the stack, not an afterthought. Names investors look at here include * Vistra ($VST) runs a large fleet of gas and nuclear plants selling into competitive power markets, exactly the position you want when data center demand tightens supply * Bloom Energy ($BE) builds fuel cells that generate power on site, letting a data center come online without waiting years for a grid connection * GE Vernova ($GEV) makes the gas turbines, grid equipment and transformers that everyone needs at once, with an order book booked out well ahead * Constellation Energy ($CEG) is the largest nuclear operator in the United States and has been signing long term supply deals directly with hyperscalers Compute capacity is now being planned around where electricity is available rather than the other way round. This layer also has a property the others do not: demand here is real regardless of which model wins. Whether the leading model in 2029 comes from a lab in California or Shenzhen, it still has to be plugged into something. Layer 2: Chips Jensen described this as the world he lives in, and the concentration is extreme. Four companies sit at different points of the same chokepoint. * Nvidia ($NVDA) designs the accelerators the industry trains and runs models on, and its CUDA software layer is the reason switching away is so hard * Broadcom ($AVGO) co designs the custom chips hyperscalers build to reduce their Nvidia dependency, and supplies the networking silicon that ties racks together * TSMC ($TSM) manufactures the leading edge chips for almost everyone, which makes it the single point every design has to pass through * ASML ($ASML) is the only company making the extreme ultraviolet machines those chips are printed with, a monopoly at the bottom of the bottleneck The bottleneck is also shifting. Nvidia is now committing enormous sums to lock up memory supply, which tells you the constraint is no longer only logic chips. Capacity here gets bought years ahead of the revenue it is supposed to serve, which makes this the layer where a demand miss would show up first. Layer 3: Infrastructure This is the layer people underestimate. Jesnen broke it down as land, power, and the shell that holds the computers, with energy plugged in at one end and money coming out the other. These are the companies renting out compute rather than owning the models that run on it. * Nebius ($NBIS) is the European AI cloud that came out of the Yandex breakup, now building GPU capacity for labs and enterprises * CoreWeave ($CRWV) is the pure play GPU cloud that grew out of crypto mining and now leases capacity to the largest AI labs under multi year contracts * IREN ($IREN) converted bitcoin mining sites into AI data centers, which gave it something scarce: land with power already connected * Oracle ($ORCL) turned itself into a serious AI cloud provider by winning enormous compute contracts, funded with a lot of debt It is also the most financially engineered part of the stack. Sites get built with private credit and private equity money, secured against a promise that a hyperscaler will lease the finished building. Contracted revenue years out looks like safety while demand climbs, and like concentration risk the moment a single anchor tenant changes its plans. Layer 4: Models The layer most people mean when they say “AI”, and the one Jensen spent the least time on. All four fund the layers beneath them out of businesses that have nothing to do with AI. * Alphabet ($GOOGL) owns the full stack, from its own TPU chips to its own models to the distribution that puts them in front of billions of users * Microsoft ($MSFT) turned its OpenAI relationship and Azure into the enterprise route to market, selling AI into software companies already pay for * Meta ($META) builds models to improve its own advertising machine, and is spending accordingly * Amazon ($AMZN) combines AWS, its own Trainium chips and its stake in Anthropic, giving it a position at both the model and infrastructure layer Huang made a broader point here that is easy to skip past. AI is not only language. Anything with structure can be learned and predicted, and biology, chemistry and physics all have structure. That is why he expects different regions to lead in different model domains rather than everyone chasing one general chatbot. Layer 5: Data and applications Jensen called this the most important part of the cake, which is a striking thing for a chip CEO to say. This is the software that turns model output into something a business actually uses. * Palantir ($PLTR) connects messy organisational data to decisions, the unglamorous work that has to happen before any model is useful inside a government or a large company. I just wrote a deep dive recently: * ServiceNow ($NOW) owns the workflow layer enterprises already run on, so AI agents get dropped into processes rather than bolted on beside them. I also wrote a piece on now: * Snowflake ($SNOW) stores and organises the data models are trained and queried against, and models are only as good as what they are pointed at * CrowdStrike ($CRWD) secures the endpoints and now the AI systems themselves, a layer that grows precisely because AI expands the attack surface His advice to the assembled countries was not “build your own foundation model”. It was to push AI diffusion, meaning adoption into existing industries: education, healthcare, manufacturing, science, everything in between. No country has to win every layer. Pick the one you can actually compete in. That translates almost directly to a portfolio. You do not need exposure to all five. You need to know which one you own and why. Who is paying for the cake A five layer map tells you where the money goes. It does not tell you whether the money comes back. Some of the research crossing my desk this weekend is worth putting next to Jensen’s framework, because it describes the same structure from the financing side. The commitments are enormous and increasingly invisible. * Nvidia has disclosed roughly 279 billion dollars in supplier commitments, more than double the 119 billion figure from a quarter earlier * The four big hyperscalers carry around 248 billion in lease liabilities and 356 billion in long term debt on their balance sheets * On top of that sit roughly 904 billion in leases that have not started yet and about 1.52 trillion in future purchase commitments * Those last two do not show up as liabilities today The financing is getting circular. Some of the money customers use to buy AI chips is coming from the chipmakers themselves, in the form of investments, guarantees and financing deals. Revenue that arrives this way looks like demand on the income statement, but part of it started out as the seller’s own cash. This happened before. In the late 1990s, the companies selling telecom equipment lent their customers the money to buy it. Sales looked strong right up until the customers ran out of road, and then the sales disappeared along with them. One way to watch for stress is the price of insuring these companies’ debt against default, known as a credit default swap. That cost has been climbing since late 2025, which means bond investors see slightly more risk than they did. It is still nowhere near the level that signals real trouble, so read it as a thermometer rather than an alarm History is mixed, not doomed. RBC looked at eleven major investment booms going back to canal mania, and eight produced a bubble that burst * Most ran for decades before resolving * Three ended without tears: post war European reconstruction, the mainframe and PC era, and the pre AI cloud buildout * Booms that fill a real shortage usually end better than booms created by cheap money and loose rules. In the first case there is a genuine gap to fill and a rough sense of how big it is. In the second, the demand only exists while the policy does. Technology booms are the hard middle case, because demand is real but the right amount of capacity is a guess. Even the bad ones leave something behind. * The fibre laid in 1999 was a disaster for the people who financed it and a gift to everyone who used the internet afterwards * If this cycle overshoots, the power plants and data centers do not disappear How I read it Jensen is telling countries what to build. But the credit market is asking who pays if the demand shows up smaller or later than all those promises assume. Both are describing the same cake from opposite ends. Three simple things I take from it. * The layers are not equally risky. Energy and chips are limited by physical reality, and you cannot fake a turbine or a fab. Infrastructure is where the borrowed money sits. Applications are where the revenue has to eventually appear to justify everything underneath * A promise you cannot see is still a promise. A lease that starts in 2029 was signed in 2026, and it has to be paid whether or not AI demand met expectations * Time in the trade matters more tha
