The Missing Ingredient

Jensen Huang's five-layer AI cake is missing something and it isn't technology.

Photo by Alexander Van Steenberge on Unsplash

Jensen Huang has a gift for making the complex simple. The NVIDIA CEO’s description of AI as a five-layer cake, energy, chips, cloud infrastructure, AI models, and applications, has given executives, investors and policymakers a shared map of the territory.

It’s a good map, but to paraphrase Alfred Korzybski, maps leave things out, and what this one omits is the ingredient that determines whether the whole cake rises or not.

Not a sixth layer sitting on top. Something more like baking powder: invisible in the finished product, present through every layer, and entirely responsible for whether you end up with something edible or a flat, expensive, indigestible mess.

That ingredient is human infrastructure.

The cake, layer by layer

Huang’s framework starts at the bottom with energy: the power that makes everything else possible. Then chips: the specialised silicon that runs AI workloads. Above that, cloud infrastructure: the data centres and networks that host the compute. Then the models themselves: the large language models and multimodal systems that are the visible face of AI. And finally, applications: the products, services and workflows where economic value is actually created.

The logic is clean, with each layer dependent on the one below. The application layer is where Huang says the real benefit lands, where AI stops being a technology and starts being a capability.

But a capability for whom? Deployed by whom? Integrated, adapted, questioned, and directed by whom?

The answer is humans. And the question of whether those humans have the adaptive capacity, Learning Power and relational intelligence to work effectively alongside AI is simply not on the map.

Scoring the scorecard

Run the five layers as a quick geopolitical audit and the picture becomes uncomfortable, particularly for the UK.

The United States leads on models and capital. The race to Artificial General Intelligence is being run at a pace that even its own engineers admit outstrips any governance architecture. But Gillian Tett in The Financial Times has been persistent on a point that the AI headlines tend to obscure: American AI ambitions could be undone by the humdrum issue of power. Data centres demand extraordinary quantities of electricity. The grid is under strain. The energy layer, the foundation of Huang’s cake, is becoming a genuine constraint on US AI dominance. And the relentless focus on faster models means the human capability to use them wisely is an afterthought, not a design principle.

China is taking a different approach entirely. Practical over theoretical, open over closed, long-term over quarterly. Chinese AI investment is concentrated in health, education, agriculture and infrastructure, domains where human capability still matters, where the interface between person and system is the point, not an inconvenience. China is winning on the energy layer, particularly in green technology. And it has made a strategic choice about its human infrastructure that the West is only beginning to understand.

ByteDance (the company behind TikTok) built two versions of the same app. The version available inside China called Douyin has strict time limits for under-18s, no late-night access, and an algorithm that actively surfaces educational content. The version exported to the rest of the world was optimised for engagement and addiction. They knew exactly what the algorithm did and they chose not to do it to their own children.

The UK and Europe are, as Huang’s framework makes clear, laggards at almost every visible layer. Behind on energy transition. Dependent on US cloud providers. Marginal in frontier model development. On chips, the story is more complicated, and more revealing. Britain produced arguably the most important chip architecture in the world. ARM, headquartered in Cambridge, is the invisible foundation of virtually every smartphone on the planet and, increasingly, of AI data centres globally. Half of all compute shipped to the world’s top cloud providers in 2025 ran on ARM architecture. NVIDIA licenses it. Apple builds on it. Google, Amazon, Microsoft - all of them. And yet ARM is majority-owned by a Japanese conglomerate and listed in New York. This is a pattern - Britain created it and didn’t hold it. At the application layer, where economic benefit is supposed to land, the questions are whether there are businesses capable of building on these foundations, and workers capable of using them.

The bill has arrived

Research published by the Resolution Foundation, recently highlighted by John Burns-Murdoch in The Financial Times, found that Britain has the highest proportion of young people not in education, employment or training (NEETs) in the developed world. Nearly a million under-25s. Sixty percent have never held a paid job. A quarter cite health, including mental health, as the reason.

This is not simply a welfare statistic, it is a human infrastructure statistic. It describes the condition of the generation that will have to work alongside AI, the people at the application layer of Huang’s cake. And it tells us that layer is severely compromised.

The harvest of a decade of unregulated social media exposure to young people is now being counted in disengagement, mental health crisis and structural economic exclusion. Britain imported the addictive version of the algorithm and applied it to an entire generation. China protected its own.

The government’s response? The Home Office has launched a review of visas to attract top talent from abroad. Which is to say: the answer to a depleted domestic human infrastructure is to import someone else’s. A workaround, not a solution. It depletes another country’s sixth layer to compensate for the failure to invest in our own.

Starting my career at Procter & Gamble was instructive on this point. P&G has a longstanding principle that shapes everything from hiring to leadership development: build and promote from within. The logic is not sentimentality; it is that the capability you develop internally compounds over time in ways that imported capability never can. You can hire brilliant individuals but you cannot hire an organisational learning culture.

The grid that can’t be built

Nowhere is this more visible, or more consequential, than in Britain’s Great Grid Upgrade.

The £60 billion programme to overhaul the UK’s electricity transmission network is, adjusted for inflation, costing twenty times more than the construction of the original grid and supergrid combined. It is Critical National Infrastructure - the energy foundation that every other layer of Huang’s cake depends on. Without it, Britain’s AI ambitions are literally unpowered.

And it is struggling. Not primarily because of planning delays or pylon protests or the complexity of connecting Scottish wind to southern demand, though all of those are real. It is struggling because of a skills shortage that the energy union Prospect described in terms that bear repeating: “dire shortages,” “chronic understaffing,” “woefully under-resourced.” In their most recent survey, 82% of energy workers said staffing levels were too low. 69% reported tangible skills gaps in their organisation.

One specialist recruiter put the argument plainly: the biggest constraint on grid delivery is not generation capacity. It is “the growing shortage of specialist skills required to upgrade and operate the UK’s electricity grid. This is where delivery risk is quietly building.”

National Grid has recognised this. It launched what it calls a “pioneering enterprise model” - a Great Grid Partnership with supply chain partners - explicitly as a response to skills shortage. The physical infrastructure problem and the human infrastructure problem are, in their own framing, the same problem.

Technology infrastructure deploys in quarters. Human infrastructure develops over years. Planning both on the same timeline means you are structurally late on the variable that takes longest to mature. Britain is discovering this in real time, in the programme that most needs to succeed.

Not a layer, an ingredient.

So is human infrastructure the sixth layer of the cake, or something else?

The sixth layer framing is useful as an entry point. It says: you have left something out of your model, and that omission explains a great deal of what isn’t working.

But it’s deeper than that: human infrastructure isn’t additive, it runs through every layer. Energy systems need humans who can plan, build and adapt complex delivery programmes under conditions of radical uncertainty. Chip fabrication requires engineering cultures that sustain deep expertise across generations. Cloud infrastructure depends on organisations capable of operating and evolving systems that were never fully designed. Models need humans who can interrogate outputs rather than simply accept them. And applications, where the value is supposed to land, are only as good as the human capability to integrate AI into real work, real decisions, real relationships.

For the UK this is not a new vulnerability. We discovered North Sea oil in the same decade Norway did. Around twenty years later, in 1990, Norway built the Government Pension Fund. Now worth over $1.7 trillion, Norway developed the institutional capability to think in generations rather than quarters whilst Britain extracted and spent. The human infrastructure to make a different choice - the policy capability, the long-term governance instinct, the wisdom to resist short-term extraction - was the difference. ARM tells the same story in a different decade. The capability to create is not the same as the capability to retain.

Before the crisis, or after

When I published an earlier piece arguing that AI is the wrong unit of analysis, it drew hundreds of responses. One comment stood out: Sener Cem Irmak, founder of Koru Impact, an evidence-first adaptive systems consultancy working with C-suites and private equity, offered a diagnosis of the governance problem at the heart of this argument.

“We’ve seen this pattern before,” he wrote. “Risk management spent decades in this liminal state: too technical for the board, too strategic for operations, too cross-cutting for any function. Everybody’s concern, nobody’s discipline. Then Basel II, COSO, and 2008 forced it into its own governance architecture: dedicated reporting lines, board-level accountability, measurable KRIs.

Human infrastructure is at that inflection point, but the question isn’t which function should own it. The question is whether organisations build a dedicated governance architecture for adaptive capacity before a crisis forces it, or after.”

The risk management precedent is illuminating. Before 2008, risk was structurally orphaned in most organisations: acknowledged in principle, unowned in practice, measured inconsistently if at all. The financial crisis didn’t create the risk management discipline, but it forced it into governance architecture that boards could no longer ignore. Chief Risk Officers acquired real authority. Measurable frameworks became mandatory. The function that had been everyone’s concern and nobody’s discipline became a defined professional field with its own reporting lines, its own metrics, its own seat at the table.

Human infrastructure is in exactly that pre-2008 position today. Acknowledged everywhere, owned nowhere. Every organisation depends on the adaptive capacity of its people. Almost none measures it, governs it, or funds it with the seriousness the dependency deserves.

Sener added a sequencing point that belongs at the centre of every infrastructure investment conversation: technology deploys in quarters, human infrastructure develops over years. Planning both on the same timeline means you are structurally late on the variable that takes longest to mature. The Great Grid Upgrade is discovering this. Britain’s AI ambitions are about to.

The measurability work already exists. For nearly three decades Professor Ruth Crick has been leading research into “Learning Power”: not skills, not qualifications, but the adaptive capacity that allows individuals and organisations to keep learning as conditions change. The disposition to engage with complexity rather than retreat from it. The relational intelligence to work across difference. The capacity to hold uncertainty without collapsing into either false certainty or paralysis.

Professor Crick’s Learning Power framework has been doing precisely what Sener identified as the precondition for governance: making the unmeasured measurable, making the unfunded fundable. You cannot govern what you cannot measure. And you cannot fund what you cannot govern.

The only question, as Sener framed it, is whether organisations and nations build the governance architecture before the crisis, or after. History suggests most will wait. But those that don’t will have a structural advantage that compounds, quietly, over years, in exactly the way that human infrastructure does.

What the AI billionaires can’t see

There is another relevant thread in recent commentary worth naming. The men who built the current wave of AI, the founders, the funders, the stage-fillers, are exhibiting signs of a particular kind of anxiety. Not about Artificial General Intelligence in the abstract. About something more personal: the cognitive style that made them successful is now being replicated by their own product. Linear pattern recognition. The ability to process information faster than others, spot trends earlier, optimise systems. The intelligence that gets you into Stanford, through Y Combinator, onto a stage in front of a quarter million people. That intelligence is now a commodity. They know it. They can feel it. And they’re projecting their anxiety on to us.

Jensen Huang himself put it plainly when asked who was the most intelligent person he’d ever met. He declined to name anyone, because the question itself revealed the problem. “The definition of smart is someone who solves problems, technical,” he said. “But I find that’s a commodity, and we’re about to prove that Artificial Intelligence is able to handle that part easiest.” Software programming, he noted, was supposed to be the ultimate marker of intelligence and it was the first thing AI solved. So the definition of smart, he concluded, needs to change entirely.

And yet, when the people building AI are asked not what the technology can do, but what their own children will need, something shifts. The Wall Street Journal recently asked several leading AI figures exactly that question. Ethan Mollick pointed toward flexibility and broad education: the things you’d invest in “in any time of uncertainty.” Daniela Amodei, President of Anthropic, was more direct: “When I think about what my kids will need as they get older, it’s human qualities: the ability to relate, to empathise and be around other humans. What’s not going to be replaceable is how you treat other people, how well you communicate with them, how kind you are.”

They are describing, for their children, exactly what they are failing to build for their organisations and their nations. They can see the missing ingredient, they just haven’t yet worked out how to bake with it.

What they cannot feel, because they have never had to develop it, because it was never what got them on the stage, is the capability that AI cannot replicate: genuine curiosity, relational depth, embodied judgement, the capacity to learn in conditions of ambiguity, the ability to ask the question that doesn’t yet have an answer.

That is what human infrastructure names. Not soft skills. Not culture. The human substrate - the measurable, developable adaptive capacity of people and organisations to engage with what they don’t yet know.

Jensen Huang is right about the layers. He’s right that the application layer is where value lands. He’s right that energy is the foundation.

But the nation most likely to win the AI era is not the one with the best chips or the biggest models. It is the one that has invested, seriously, measurably, at scale, in the human capability to use all of the above wisely.

Britain is not currently that nation, but it could be.

The teaspoon of baking powder in every cake never gets the credit, but take it out and see what you’re left with.

—

A note on sources

Jensen Huang’s five-layer framework was articulated at Davos and expanded in conversation with BlackRock CEO Larry Fink (blog on NVIDIA’s website here). His remarks on the redefinition of intelligence appeared on the A Bit Personal Podcast with Jodi Shelton.

Alfred Korzybski was a Polish-American scientist and philosopher working in the 1930s who coined the phrase “the map is not the territory”. He reminds us that our mental models of the world are not the same as the world itself - maps help us simplify complexity. Excellent read on mental models by Shane Parrish here.

The Wall Street Journal’s interviews with Ethan Mollick and Daniela Amodei on AI and the future of education were shared widely, including by Dr Will Van Reyk Deputy Headmaster of St Paul’s School.

Link to The Resolution Foundation’s NEET research here - thanks to John Burns-Murdoch for reporting on it in The Financial Times.

The Great Grid Upgrade skills shortage data comes from Prospect union workforce surveys and specialist energy sector commentary.

The Douyin/TikTok comparison is documented across multiple independent investigations into ByteDance’s domestic and international product strategies.

The observation that the AI billionaires’ anxiety stems from the obsolescence of linear pattern recognition was made with characteristic precision by Abi Awomosu on Substack.

The Learning Power framework is the result of twenty-six years of research led by Professor Ruth Crick (originally at the University of Bristol) founder of WILD Learning. Crick’s research has been validated across more than 100,000 individuals and 190 organisations using Structural Equation Modelling (the rigorous statistical method) to confirm the robustness of its 8 dimensions measurement framework. WILD stands for Work Integrated Learning Design and is the framework that scaffolds “learning journeys”.