The Five-Layer Cake
In January, on a stage at Davos, the man who sells the world its AI chips told the audience that chips are only a fifth of the story. Jensen Huang's frame: AI is a five-layer cake. Energy at the bottom. Then chips. Then infrastructure - the data centers and clouds. Then models. Then, on top, applications.
"AI is... essential infrastructure, like electricity and the internet... This is becoming the largest infrastructure buildout in human history." - NVIDIA's own March 2026 essay laying out the five-layer stack
The frame is a useful machine for a question nobody on that stage asked out loud: of the seven companies that dominate the US market, who actually owns the layers - and who is renting them from a rival?
Own all five and you compound on your own terms: your power, your silicon, your data centers, your model, your distribution. Rent even one, and your moat has a landlord.
So we used Barebone to score every Magnificent 7 company across all five layers, using only deals and products publicly committed as of April 28. Dominant scores 2, Has scores 1, Partial scores 0.5, None scores 0 - a maximum of 10.
The result surprised us twice. Alphabet sweeps the board with 8.0. And Nvidia - the company synonymous with the AI trade - finishes dead last with 3.0. Here are the receipts.
The Scorecard
| # |
Company |
Energy |
Chips |
Infra |
Models |
Apps |
Score |
| 1 |
Alphabet |
Has |
Dominant |
Has |
Dominant |
Dominant |
8.0 |
| 2 |
Meta |
Has |
Has |
Has |
Has |
Dominant |
6.0 |
| 3 |
Microsoft |
Has |
Partial |
Dominant |
Rents |
Dominant |
6.0 |
| 4 |
Amazon |
Has |
Has |
Dominant |
Partial |
Has |
5.5 |
| 5 |
Tesla |
Partial |
Has |
Partial |
Has |
Has |
4.0 |
| 6 |
Apple |
None |
Has |
Partial |
Rents |
Dominant |
3.5 |
| 7 |
Nvidia |
None |
Dominant |
Partial |
Partial |
None |
3.0 |
Scoring notes: "Rents" counts 0.5 for Microsoft (a 27% OpenAI stake plus model IP rights through 2032) and 0 for Apple (a pure license - no equity, no IP). Tesla is scored standalone: xAI is Elon Musk's private company, not Tesla's. We break the Meta - Microsoft tie below.
Read the shape of the bars, not just the totals. Meta is the only company besides Alphabet with every layer at Has or better - which is why it edges Microsoft despite the identical score: Microsoft's thinnest layer is models, and models are the layer the rest of the cake exists to produce.
Two placements need defending: the bottom and the top. Start at the bottom.
The Bottom of the Table
Nvidia: dominant in one layer, absent from three
Nvidia at #7 is not a short thesis. The framework measures breadth of ownership, not value capture - and Nvidia owns the chips layer outright, the one layer every other company on this list still has to bleed through to build anything.
But look at the rest of its row. No energy position. No consumer application. Its model work is real but peripheral, and its infrastructure presence runs mostly through other people's clouds and a web of strategic stakes. The most famous of those: last September, Nvidia agreed to put $5 billion into Intel at $23.28 a share. On April 24, after a first quarter that demolished estimates, Intel closed at $82.57 - the stake is up roughly 3.5x in seven months.
Now read Nvidia's chips column upside down, as a risk map. Google just signed Broadcom to build its TPUs through 2031. Meta committed more than a gigawatt of its own custom silicon. Amazon trains Anthropic's Claude on Trainium. Tesla handed Samsung $16.5 billion for its next chip. Four of Nvidia's six peers are spending billions to need the vendor less.
Apple: the best apps business, renting the brain
Apple owns arguably the single best application-layer asset on earth - the iPhone installed base - and real silicon underneath it. That's where the ownership ends. On January 12, Apple and Google confirmed what Bloomberg first reported in November: Siri's next brain is a 1.2-trillion-parameter custom Gemini model, leased for roughly $1 billion a year - about eight times the size of the 150-billion-parameter model Apple Intelligence had been running. It runs inside Apple's Private Cloud Compute, but it is Google's model. No proprietary frontier model, no AI-scale energy deals, no hyperscale infrastructure.
Apple has run this playbook forever: pick the best supplier, squeeze it, ship the experience. It built the most profitable hardware franchise in history renting modems, memory, and screens.
The difference is that nobody's modem decided what the product could think.
Tesla: four layers, if you count the ones next door
Tesla scores better than most people expect - for less comfortable reasons. Its AI5/AI6 chip program is anchored by a $16.5 billion Samsung foundry deal running through 2033 at the Taylor, Texas fab. FSD is a genuine proprietary foundation model, trained on fleet data no one else has - it just speaks driving, not language. The car is the app. But infrastructure is now Partial: Dojo, the in-house training supercomputer, was shut down in August 2025, with Musk calling it "an evolutionary dead end." Training runs on bought silicon.
The bull case quietly assumes a sixth layer: xAI, with its frontier models and Colossus data centers. But xAI belongs to Musk, not Tesla - and at November's annual meeting, the proposal to authorize a Tesla investment in xAI failed to win approval: 1.06 billion shares for, 916 million against, an unusually large 473 million abstaining.
The crossover everyone prices in is an org chart, not a balance sheet.
The Middle: One Empty Floor Each
Amazon: everything but the model
Amazon owns the biggest infrastructure layer in the business - AWS is still the largest cloud - plus a real chips layer in Trainium and a real energy layer: a 960-megawatt data center campus beside the Susquehanna nuclear plant, an expanded power agreement signed last June, and X-energy small modular reactors in development. Anthropic's April 6 compute announcement spelled out Amazon's role in writing: "Amazon remains our primary cloud provider and training partner."
That sentence is also the problem. Amazon's frontier-model exposure is a contract with Anthropic, not a subsidiary, and its in-house models haven't cracked the frontier tier. The company guiding to roughly $200 billion of 2026 capex - the largest budget in the group - is building the world's best kitchen and buying the recipe.
Microsoft: the most sophisticated rental agreement in tech
Microsoft signed the deal that defines the energy layer's revival - the 20-year power purchase agreement behind Constellation's $1.6 billion restart of Three Mile Island Unit 1, targeted for 2027. Azure is a Dominant infrastructure layer, now carrying OpenAI's $250 billion Azure spending commitment. The applications layer - Office, Windows, GitHub - is as Dominant as distribution gets.
The models layer is where the framework bites. GPT is rented. It's the most sophisticated lease in tech - roughly 27% of OpenAI, a stake valued around $135 billion, plus model IP rights through 2032 - renting with an option on the landlord. But October's restructuring also ended Microsoft's right of first refusal on OpenAI's new cloud workloads.
The tightest partnership in AI got looser the moment OpenAI could afford it. That's the weak-link clause in action, at the very top of the market.
The Top Two
Meta's last four months read like someone working through the cake checklist. Energy: on January 9 it announced nuclear agreements with Oklo, Vistra, and TerraPower for up to 6.6 gigawatts - more than New Hampshire draws - on top of last June's 20-year Clinton deal, with the Prometheus supercluster waiting at the other end. Chips: on April 14, Broadcom announced an expanded MTIA partnership - an initial commitment above one gigawatt of custom accelerators, billed as the industry's first 2-nanometer AI chip. Models: on April 8, Meta Superintelligence Labs shipped Muse Spark, its first post-Llama frontier model - closed weights, a clean break from the open-source era, with a Contemplating mode that Meta says scores 58% on Humanity's Last Exam. Applications were never the question: billions of daily users across Facebook, Instagram, and WhatsApp.
The honest read: Muse Spark is three weeks old, and as CNBC put it this morning, the model "has promise" while Wall Street still wants the strategy. Meta is paying $115 - 135 billion this year to find out.
Alphabet: the sweep
Alphabet is the only company that owns every layer and dominates three. Energy: a power purchase agreement underwriting NextEra's restart of the Duane Arnold nuclear plant (targeted for early 2029), small modular reactors with Kairos Power and the TVA, and a fusion offtake with Commonwealth Fusion Systems. Chips: the TPU is the only non-Nvidia AI silicon that other frontier labs queue up to buy - Anthropic contracted for up to one million TPUs last October, more than a gigawatt arriving this year, then expanded by multiple gigawatts on April 6; Broadcom will co-build future generations through 2031. Infrastructure: the third-largest cloud, self-built end to end. And the models layer carries the best third-party validation in the industry: Apple - the most ruthless component shopper alive - evaluated the field and chose to pay Google a billion dollars a year for Gemini.
Then distribution: Search, Android, YouTube, Workspace. Not just reach - training data the others can't license at any price.
Alphabet could lose any single layer and still hold four. Nobody else could lose one and keep a full hand.
The Price of Owning the Stack
Owning layers is a balance-sheet decision, and 2026 is when it gets expensive: Amazon guided to roughly $200 billion of capex, Alphabet to $175 - 185 billion, Meta to $115 - 135 billion, and street estimates put Microsoft at $110 - 120 billion. Call it ~$620 billion across four companies, up about 64% from $378 billion in 2025.
Notice who isn't on that chart. Apple, Nvidia, and Tesla keep their income statements clean precisely because they rent layers - or, in Nvidia's case, because they are the layer. Owning the cake means converting today's cash into depreciating concrete and silicon. Genius if token demand keeps compounding. A write-down generator if it doesn't.
Where the Framework Breaks
Before anyone treats the scorecard as a buy list, here is April's actual tape, ordered by our ranking:
The market half-agrees. The four stack owners ripped - Amazon +24.7%, Alphabet +21.6%, Meta +17.3%, Microsoft +16.0% from March 31 through April 28 - while the two biggest renters, Tesla (+1.1%) and Apple (+6.7%), sat out the rally. But the framework's last-place company gained +22.2%, nearly leading the group. Nvidia closed April 28 at $213.17; Alphabet at $349.78; Meta at $671.34.
Three honest problems with the cake:
It measures breadth, not value capture. Profit pools concentrate where supply is scarcest, and for three years that has been the chips layer - exactly why the framework's worst-ranked company has been the defining trade of the decade. Owning five layers adequately can be worth less than owning one layer absolutely.
Ownership cuts both ways. Apple spent fifteen years as the most profitable company on earth while renting components, and vertical empires have died of their own capex before. If models commoditize, the companies that sank hundreds of billions into the model layer own the most expensive depreciation schedule in corporate history - and the renters get to walk away.
The scores are judgment calls. Muse Spark is three weeks old. Microsoft's 27% of OpenAI could plausibly score a full point, which would tie it with Meta. Tesla's row changes completely if xAI ever folds in. A framework that fits in one table is a lens, not a verdict.
What This Means
The cake is not a stock-picking machine. It's a map of where moats are being dug - and it suggests five things worth watching from here.
Third-party silicon adoption. The chips layer stops being a Nvidia monopoly the day other accelerators win external customers at scale. Anthropic buying TPUs by the gigawatt was the first real crack. Watch who signs next, and for whose chip.
Whether Siri's new brain ships - and what Apple does after. The Gemini deal is either Apple's bridge to its own frontier model or the moment it became a permanent tenant.
Meta's cadence. One closed model doesn't make a model layer. Whether Muse Spark's successors arrive on a Gemini-like rhythm is the test of whether $115 - 135 billion is buying a stack or a science project.
Capex revisions. Most of this list reports earnings within days of this writing. Every raised number is a vote that the cake is worth owning; the first meaningful cut is the signal the market will actually trade on.
Where the giants put equity. When these companies buy stakes instead of products - Nvidia naming Intel as its bottleneck seven months before Intel's blowout quarter - they're telling you which layer worries them. That disclosure is free.
The Magnificent 7 spent a decade being lumped together as "Big Tech." The five-layer cake un-lumps them: one company owns the whole stack, two own most of it, and four - including the most beloved name in the group - are structurally dependent on a rival's layer. The question the next year will answer is whether the market starts pricing the whole cake, or keeps paying for the one layer everyone still has to buy.
Data: Barebone | Sources: NVIDIA five-layer stack essay (March 2026) and Davos 2026 remarks, Meta newsroom (January 9 and April 8, 2026), Broadcom - Meta and Broadcom - Google announcements (April 2026), Anthropic compute announcement (April 6, 2026), Apple - Google joint statement (January 12, 2026), Microsoft - OpenAI agreement (October 28, 2025), Tesla 2025 annual meeting results, Constellation, NextEra and Talen announcements, Q4 2025 earnings calls | Prices as of April 28, 2026 close
