The Layer That Grows

Everyone is building agent infrastructure. Almost nobody is putting agents to work in an actual business. That gap is the reason this company exists, and this piece is the sourced version of the picture on the front page: what happened the last time the money went into the bottom of a stack, and what it means for the person reading this.

Every number below was opened at its source before it was printed. The sources are at the end, and so is a list of what these numbers do not say.

The bottom first

The internet was built from the bottom up, and the money knew it.

Between the first quarter of 1996 and the last quarter of 2000, annual US investment in communications equipment grew from about $62 billion to over $135 billion, in constant 1996 dollars.¹ In five years, companies laid nearly 40 million miles of fibre-optic cable across the United States, spending about $90 billion.²

At the peak, in March 2000, the most valuable company on earth was a router maker. Cisco passed Microsoft at $555.5 billion against $541.6 billion.³ It got there selling under $19 billion of networking gear a year;⁴ Microsoft, the platform of that era, booked $22.96 billion the same year.⁵ The infrastructure layer was worth more than the platform layer. It was not earning more.

Then it broke. Roughly $700 billion of stock-market value left listed S&P telecom firms between 2000 and 2002.⁶ By early 2002, a Merrill Lynch estimate put less than 9 per cent of the fibre in the world in use.⁷ Most of what had been paid for sat dark.

The top grows

The application layer did not wait for the dust to settle. In 2004 the whole of Google turned over $3.2 billion, up 118 per cent in a year,⁸ while most of the fibre it would eventually run over was still unlit.

What changed next was the price of standing on that infrastructure. Marc Andreessen’s own estimate is that running a basic internet application cost about $150,000 a month in 2000 and about $1,500 a month in 2011, on Amazon’s cloud.⁹ The over-built bottom became something you could rent by the hour.

By 2015 the money had moved up the stack. Alphabet earned $75.0 billion.¹⁰ Facebook, which owned no cable and no data-centre business of its own, earned $17.9 billion, growing 44 per cent a year.¹¹ Amazon’s consolidated sales were $107.0 billion, of which $7.9 billion was its cloud platform; the rest was the application business running on top.¹² Each of those single application companies earned multiples of what Cisco earned at the height of the infrastructure boom.

By market value the picture was more mixed: in November 2014 the most valuable company was Apple, at $701.7 billion, a device-and-platform company.¹³ The venture framing of the era, written in 2016, was that the shared plumbing stayed thin and the applications built on top got fat.¹⁴ Read only that description here; the same essay’s prediction about what would happen next has aged badly.

That is the shape of the internet wave: infrastructure first, worth more than it earned; then the layer that touched real work expanded until it held the value.

Where AI is

AI is at the plumbing stage, far larger.

The five biggest US cloud companies put roughly $380 billion into AI infrastructure in 2025 and have committed $660 to $690 billion for 2026.¹⁵ Their combined quarterly capital expenditure went from $37.6 billion at the start of 2022 to $140.6 billion in the last quarter of 2025,¹⁶ compounding at about 72 per cent a year since GPT-4 shipped.¹⁷ NVIDIA’s data-centre business alone booked $193.7 billion in the fiscal year to January 2026, up from $115.2 billion the year before,¹⁸ and on 29 October 2025 NVIDIA became the first company ever worth $5 trillion.¹⁹ As of September 2026 it still holds the top slot, at about $5.09 trillion.²⁰

In June 2024 Sequoia put a name to the gap: the buildout needed $600 billion a year of revenue to pay for itself, and it did not have it.²¹ Two years on, the same author’s figure is $1.5 trillion a year.²²

The top of the stack is still small. By Menlo Ventures’ count, enterprises bought $19 billion of AI applications in 2025 — the software they actually use to do work — and $12.5 billion of foundation-model APIs;²³ applications were a little more than 6 per cent of the whole software market.²⁴ The model layer is becoming visible: annualised revenue run rates of $65 billion at Anthropic and over $40 billion at OpenAI by mid-2026.²⁵ And half of businesses now pay for AI at all — 50.4 per cent in March 2026, on Ramp’s platform, up from 35 per cent a year before²⁶ — but the median firm spends $11.95 per employee per month.²⁷

Adoption is wide. The cheque is tiny. That is what the top layer looks like at the plumbing stage.

Why this is yours

Here is the part the pyramid on the front page cannot say by itself.

The infrastructure layer of this wave is chips, data centres and the companies that can spend $690 billion in a year. Nobody reading this is going to compete there, and nobody needs to.

The application layer is different in kind. It is not made of silicon. It is made of the work a business already does: the mail that has to be answered, the invoices that have to be chased, the files that have to be kept current, the appointments, the books, the follow-ups. When the internet’s top layer expanded, it expanded into ordinary activities — buying things, finding things, talking to people — and the companies that held it were the ones that took those activities seriously as software.

So the layer that grows in this wave is going to be built out of businesses like yours. The question is only whether it is built by you, under rules you can read, with a person you can name checking the work, or built at you, by whoever gets there first.

That is the whole reason we do what we do. We do not build the plumbing. We put AI that does real work into a business, and we keep it answerable to the person whose name is on the door. The front page says what that looks like day to day. This piece is just the evidence that the top layer is where the next ten years of value goes — and that it is, for once, the layer within reach of a small firm.

What these numbers do not say

The picture is honest only if its edges are drawn, so here they are.

• Nothing here measures a whole layer. Every figure is one company or one survey standing inside a layer. Alphabet’s $75.0 billion is not “the application layer”; Cisco’s $18.9 billion is not “the infrastructure layer”. The band heights on the front page are words — most of it, growing, small, almost none — and never percentages, because the sources support the words and not the percentages.
• The three layers do not add up and are not slices of anything. Hyperscaler capital spending partly buys NVIDIA’s hardware; much of NVIDIA’s data-centre revenue is those same purchases seen from the other side.
• Capital spending, revenue and stock-market value are three different claims. “The most valuable company in the world” is not “the layer that earns the most”, and in 2000 and 2015 the two readings disagree.
• Which layer a company sits in is our call. Microsoft and Apple straddle platform and hardware; Amazon straddles retail and cloud; NVIDIA’s software makes it partly a platform. Amazon’s 2015 total already contains AWS, so the two must never be added. AWS in 2015 is filed as infrastructure; foundation-model APIs are filed as platform.
• A run rate is not revenue. The Anthropic and OpenAI figures annualise a short recent period, are unaudited, and reach us second-hand through Bloomberg’s reporting.
• The 2026 capital spending is committed and projected, not spent. Do not read it in the past tense.
• The 2025 enterprise spend and the 6 per cent share come from one venture firm’s survey of enterprise buyers, not a global total. Ramp’s figures cover businesses on Ramp’s own platform — US small and mid-sized firms, in a sample Ramp says skews towards technology. Both are directional, and both carry the thesis, which is why they are attributed on the front page and not in a footnote.
• Fiscal years do not line up with the slider. NVIDIA’s fiscal 2026 ended in January 2026 and maps roughly to calendar 2025; Cisco’s and Microsoft’s fiscal 2000 ended in July and June of that year.
• The $700 billion is a fall in market capitalisation, not money burned. The “less than 9 per cent of fibre lit” figure is a 2001 estimate reported in 2002, and the same article carries the rebuttal that unlit fibre was laid deliberately ahead of demand.
• Cisco’s peak is reported at $555.5 billion on a basic-share basis and higher elsewhere on a diluted basis; call the moment March 2000. Apple’s $701.7 billion is an intraday figure from 25 November 2014, so that stop is late 2014 rather than 2015. NVIDIA’s $5.09 trillion comes from a ranking that updates in real time and covers US-listed companies only; it was read on 2 September 2026 and will drift.
• Several widely repeated numbers are absent because they could not be sourced, and nothing was filled in by inference.
• The internet-to-AI parallel is our analogy. None of these sources draws it. The two eras rhyme in shape; nobody here is claiming they rhyme in outcome.

Sources