The $7.6 Trillion AI Cake: Five Layers, Two Chip Giants, and My Map

Wall Street has priced about $7.6 trillion of AI spending over the next five years, and that single number is Exhibit A for why AI stocks keep beating the market. The video's claim is that even that number will prove too small, with the record earnings of Nvidia and Broadcom as evidence. The host, Alex, an MIT-trained engineer, maps where the money actually flows and says his own buying follows the same map. The promised tour is clear: the true size of the spend, the five layers absorbing it, the duel between Nvidia GPUs and Broadcom custom silicon, and the stocks he buys as a result.

Nvidia posted $96.2 billion in quarterly revenue, up 106%, with data center sales at $89 billion, up 117%. Broadcom grew 86% to $29.6 billion while its AI revenue rose 221%. The striking part is not the growth but its acceleration: both companies report faster AI revenue growth nearly every quarter. The mix rhymes too: about 73% of Broadcom's AI revenue comes from custom chips for six giants, 27% from networking, while Nvidia splits roughly 80% compute and 20% networking gear.

Still, rates and dollars must not be confused; as the video stresses, percentages clock the pace while dollars weigh the bulk. Broadcom added $11.5 billion of quarterly AI revenue over the past year; Nvidia added $48 billion, more than four times as much. Next-quarter signals point the same way: roughly 94% AI growth implied at Nvidia, 236% at Broadcom. That does not mean Broadcom is winning the race; growing fast off a small base and growing a huge base are different feats.

The gap between institutional forecasts and field signals is the core thesis. Big-bank charts pencil in global AI spending growth of 32% next year, 21% in 2028, and just 4% by 2031. Yet Nvidia sees the top five cloud firms alone spending over $800 billion this year and $1.3 trillion next, meaning five companies outspend the whole-planet estimate by about 30%. Google and Amazon raised capex budgets mid-year, some more than once, because every added unit of capacity sells out instantly. The gap could run to hundreds of billions of dollars a year.

The map comes from a Jensen Huang blog post describing AI as a five-layer cake. The frame: AI is not an app or a single model but a utility like power or the internet; if electricity is metered in watts and the internet in bits, AI is metered in tokens. The catch is that intelligence cannot be cached: every answer must be produced live, so all five layers had to be built from scratch. The layers are energy, chips, infrastructure, models, and applications, each standing on the one below, so the money trail is best walked bottom-up.

Energy sits at the bottom because every token is moving electrons, managed heat, and power converted into computation. Per company presentations, a gigawatt of AI capacity maps to about $18 billion on Hopper, $25 billion on Blackwell, and $40 billion on Vera Rubin, with Broadcom's mix averaging $20-30 billion. That is why the whole buildout is priced in gigawatts. The US expects roughly 90 gigawatts of new data center demand by 2030, about 18 gigawatts of always-on load a year. The grid is adding a record 86 gigawatts this year, yet over 90% is solar, wind, and batteries, which count at roughly a quarter of their nameplate toward always-on supply. Retirements of coal and gas plants mean firm supply is shrinking while paper capacity grows, and the median grid-connection wait runs past five years. The only clearing mechanism is price: electricity goes up. Hence the layer's winners are those who already own power: Constellation, restarting a sound Three Mile Island reactor under a 20-year Microsoft deal; Vistra, with 20-year nuclear deals signed with Amazon and Meta; and GE Vernova, whose gas turbines are effectively sold out for years.

Chips are where gigawatts become tokens. Nvidia builds a universal platform anyone can buy; Broadcom co-designs single-customer circuits, ASICs, mostly for inference. When power is the constraint, every watt counts, and predictable high-volume workloads justify trading flexibility for efficiency: search queries, social feeds, billions of chatbot prompts. Vera Rubin is in full production, set to be the fastest ramp in company history, about 20% of next quarter's data center revenue, with a first-ever full-year guide pointing to 70% growth. Broadcom counts six major custom-silicon customers: Google's TPUs, Meta's inference chips entering production this quarter, Anthropic scaling from 1 gigawatt toward 10 by 2028, and OpenAI's jalapeno chip now shipping. Management's visibility claim runs to $115 billion of AI revenue next year and $230 billion the year after, a quadrupling in two years. Yet the video's verdict is that the two architectures barely compete: the giants buy both at gigawatt scale, GPUs for speed and flexibility, ASICs for cost at billions of requests. Both sides are sold out, so the true winners supply them both: TSMC, with two-thirds of revenue from AI and high-performance compute, plus Micron and SK Hynix, whose high-bandwidth memory sales more than tripled.

The infrastructure layer makes thousands of chips behave as one machine, and networking splits in two. Scale-up wiring fuses chips inside the rack: Nvidia's NVLink lets every GPU read its neighbors' memory almost as its own, today over two miles of copper per rack, collapsing into a coffee-table-sized board with Rubin Ultra. Scale-out does the opposite across racks over fiber, with laser transceivers converting electrons to light and back. Nvidia's Ethernet revenue grew about 160% year over year with new switches riding the Rubin platform; Broadcom's AI networking grew nearly as fast, its flagship Tomahawk switches serve all six custom-chip customers, a first 200-terabit switch is taped out, and Tomahawk Ultra takes Ethernet inside the rack against NVLink. Around them: Arista, expecting over $3.6 billion in AI revenue this year; Vertiv, riding power and cooling dollars per rack with guidance raised across the board; Lumentum and Coherent, each backed by $2 billion Nvidia investments plus purchase commitments, with Lumentum's revenue doubling in a year. Then the names Wall Street barely covers: Fabrinet, assembling optical modules for Nvidia, Amazon, and Cisco with nine analysts on the stock, and Powell, sub-$7 billion builder of industrial switchgear with backlog up 69% and four analysts. The video's counter to co-packaged-optics fears is worth logging: optical port additions are outrunning laser consolidation, so lasers keep selling.

Models and applications are the top of the cake, and concentration is the story. Models rank among the costliest products ever made; labs need tens of billions just to start, and the largest still burn cash. The scale is hard to picture: Google alone processes 3.2 quadrillion tokens a month, roughly a billion books a day, eight times everything humanity has written. The big labs are still private; OpenAI and Anthropic filed to go public, Anthropic possibly as soon as next month. The only directly investable frontier-model builders today are Google and Meta; Nvidia builds models too but mostly gives them away to sell chips. Microsoft's OpenAI stake and Amazon's Anthropic stake count as well, with Amazon booking over $50 billion in paper gains last quarter. The real money sits at the very top: Meta's ad engine, arguably the largest AI application on earth, grew revenue 27% on AI targeting with 12% higher pricing per ad, off a base serving nearly half the planet. Google's AI search reaches billions monthly, while drug discovery, humanoids, and autonomy fleets remain early, risky bets.

The investing plan is one line: own two or three winners per layer by market share, average in over time, ignore short-term noise. The logic is that funds must stay conservative for fear of a single bad quarter while individuals need not. The list: Constellation, Vistra, and GE Vernova for energy; Nvidia and Broadcom plus TSMC for chips; Coherent, Lumentum, and Fabrinet plus a new Powell position for infrastructure; Google and Meta at the top. Google condenses the pitch: up 45% in a year, more than double the index, defying the big-companies-can't-grow reflex. The closing verdict: Wall Street prices $7.6 trillion, the earnings point higher, the map is known, and the job is to own the winners.