Anthropic is on pace to be the size of Meta. Not eventually — on current growth trajectories, within a year or two. That single fact is a reasonable proxy for how distorted the AI value chain has become, and it's the right place to start before working through where the money in this ecosystem actually sits.
The AI industry breaks into four layers: the hyperscalers, the chip makers, the frontier AI labs, and the fibre/networking companies that wire it all together. Ranked by current revenue, the hyperscalers are the largest, followed by the chip makers. The frontier labs are still small in absolute terms — but their growth rates are so steep that a name like Anthropic, sitting at roughly $65 billion in annualized revenue today, could plausibly be running at 3-4x that within two years. That would put it in Meta's revenue neighbourhood, a company it has no comparable balance sheet, headcount, or operating history against.
Profitability tells a different, more sobering story. The hyperscalers and chip makers enjoy very high margins — Amazon is the one exception among the hyperscalers, running a comparatively thin 9% net margin against Microsoft's 36%, Alphabet's 29%, and Meta's 28%. Nvidia sits at the top of the entire chain: a 63% net margin, a level of profitability more common in software monopolies than industrial hardware businesses.
The Thesis: Capex Is Everyone Else's Revenue
A report from S&P suggests that capex spending by Alphabet, Amazon, Microsoft, and Meta will exceed $1.0 trillion in 2027, versus close to $700 billion in 2026 and $350 billion in 2025. The capex is expected to continue at more than $1.1 trillion annually in 2028 and 2029. The overwhelming majority of it is earmarked for AI buildout. That figure is the single most important number in this entire ecosystem, because it isn't really "cost" from a system-wide view — it's revenue in transit.
Every dollar of hyperscaler capex is revenue in transit.
Every dollar Microsoft, Amazon, Alphabet, and Meta spend on GPUs, data centre shells, networking gear, and cooling systems becomes booked revenue for Nvidia, for Corning, for Vertiv, for Eaton. The hyperscalers are, in effect, the funding mechanism for the rest of the chain. That's the frame for everything that follows: which layer captures the most value per dollar of that $1.7 trillion capex spend across 2026 and 2027, which layer carries the balance-sheet risk if the spending ever slows, and which layer is most exposed to a demand shock that traces back to a single customer's capex guidance.
Hyperscalers — The Funders
Microsoft, Amazon, Alphabet, and Meta are the largest revenue base in the ecosystem and, historically, among the most profitable businesses in corporate history. What's changed is capital intensity. Capex-to-revenue ratios that used to sit in the 15-30% range for this group are now running 45-57% at some of these names — a level of spending typically associated with utilities or telecoms, not software companies. Alphabet posted its first negative free cash flow quarter in company history in Q2 2026. Alphabet, Amazon and Meta have funded their capex through debt in 2025 and H1 2026, while Microsoft has not raised any debt in this period. Debt issuance across the group hit roughly $80 billion in 2025, and has raced to more than $140 billion in the first half of 2026, as capex funding shifts from pure cash generation toward the balance sheet.
This is the layer carrying much of the capital-allocation risk. Suppliers book the revenue as the buildout occurs; the hyperscalers still have to earn an adequate return on the assets they are building.
Chip Makers — The Highest-Margin Layer
This is where the ecosystem's economic rent currently appears to concentrate. Nvidia converts roughly two-thirds of every revenue dollar to net profit and runs a 114% return on equity with minimal leverage — a level of capital efficiency with few precedents in industrial history. Micron's memory business has been almost as dramatic: revenue nearly quintupled over several quarters in fiscal 2026 as HBM went from a niche product to a supply-constrained bottleneck, and net margin now sits around 56%. TSMC, as the sole advanced-node foundry serving both Nvidia and AMD, effectively taxes the entire GPU supply chain regardless of which chip designer wins any given generation.
Broadcom and AMD sit a tier below on margin (38% and 23% respectively) but are still comfortably ahead of anything downstream in networking or cooling.
Value capture runs almost perfectly inverse to distance from the chip.
Frontier AI Labs — Smallest Today, Least Understood Risk
Anthropic and OpenAI are the smallest segment by current revenue and the only one where the underlying financials are self-disclosed rather than audited. Anthropic's run-rate has moved from $9 billion at the end of 2025 to roughly $47 billion by May 2026 and $65 billion by July 2026 — a trajectory with almost no precedent in software history. OpenAI's run-rate, by contrast, has been roughly flat near $25 billion since February 2026 after a much faster ramp the year before. Their profitability is not disclosed and remains to be seen whether it is sustainable.
Both companies filed confidential S-1s in June 2026. This is the layer to watch most closely in the future, both because the growth math is the most extreme in the ecosystem and because it's the layer where the numbers are least independently verified.
Fibre & Networking — The Picks and Shovels
This is the segment furthest from the chip and it shows in the margins — Cisco's 21% is the high end, and several names (Nokia, Prysmian, Ciena) sit in the mid-single to high-single digits. Applied Optoelectronics is running a net loss despite explosive revenue growth, still scaling toward profitability on its 800G/1.6T transceiver ramp. But this is also the segment with the clearest forward demand signal: data-centre fibre demand grew roughly 76% year-on-year in 2025 and is projected to reach 30% of total global fibre demand by 2027, up from under 5% in 2024. Corning's multi-billion-dollar anchor-tenant deals with Nvidia and Meta are a direct read on how this layer is starting to trade order-book visibility for a seat at the table — accepting thinner margins in exchange for multi-year, take-or-pay-style demand certainty.
Where the Opportunities and Risks Sit
The opportunity is straightforward: $1.7 trillion over 2026 and 2027 in hyperscaler capex is real, committed, and flowing disproportionately toward the chip layer, with the networking and cooling layers picking up a smaller but fast-growing share. The risk is concentration. A meaningful slowdown in hyperscaler capex guidance — even a deceleration in the rate of increase, not an outright cut — would ripple through every layer below it, and the layer most exposed isn't the one with the thinnest margins, it's the one trading forward earnings multiples on the assumption that this capex cycle continues uninterrupted. That describes most of the networking, cooling, and frontier-lab names discussed here.
There are three key bets, and the future performance depends on how they play out:
- AI demand continues to grow.
- Hyperscaler capex continues to grow.
- The company you invest in captures enough of that spending at attractive margins to justify its valuation.
In the articles that follow this piece, we'll go deeper into each layer individually — growth and profitability over the past decade where the history exists, and where the opportunities and risks concentrate within each segment specifically.