The Shovel Sellers' Paradox: What the AI Infrastructure Trade Forgot to Audit
Cloudflare does not make money. It trades at somewhere between two hundred and forty-six and three hundred times forward earnings, and last week a fund manager went on national television and told people to buy it anyway — not despite the fact that it had already climbed past the price analysts assigned it, but because of it. Buried inside that recommendation is a sentence the market has trained itself not to hear. I want to spend this piece listening to it.
The segment itself was ordinary. A CNBC contributor named Jeff Kilburg ran through three names — Arista Networks, Fortinet, Cloudflare — under the familiar banner of "AI capital expenditure." Two of them, Fortinet and Cloudflare, were trading above their analyst targets. Arista sat roughly twenty percent below. Kilburg already owns CrowdStrike and Palo Alto Networks. He said so, on air. He was, in the plainest sense, talking his book.
That is not the scandal. The scandal is how few people noticed that the book he was talking was not an AI book at all.
The thesis here is not about intelligence. It is not about models, or alignment, or the strange new thing that happens when a machine learns to finish your sentence. It is about who gets paid while the machine is being built. Kilburg's strategy has a name among people who prefer their metaphors biblical: sell the shovels. Do not own the miners. Own the companies that sell them the picks.
So: do not own Microsoft, Meta, Google, Amazon — the hyperscalers who carry the capital expenditure risk, who burn cash on data centers and hope the return arrives before the patience does. Own the layer beneath them. Arista sells the Ethernet switches that stitch AI clusters together. Fortinet sells the firewalls that guard the perimeter. Cloudflare sells the edge that carries the traffic. When the miners dig, the shovel sellers bill.
This is a coherent story. It is also a story with a single load-bearing wall, and I have spent enough of my life in open source to recognize the sound a structure makes just before it fails.
The backdrop matters more than the picks. The ten-year Treasury yield has touched a level not seen since 2002. FactSet projects S&P 500 earnings growth near twenty-nine percent — the third consecutive reading above twenty-five percent, against a decade average closer to eight. And Kevin Gordon of Schwab, asked to comment, offered the only honest sentence in the entire segment: one hyperscaler's capital expenditure coming in below expectations could disrupt the AI-driven market.
Read that again. Not "could pressure valuations." Not "might trigger a correction." Could disrupt the market. The whole trade, all three names, hangs from one hook, and the hook is a quarterly capex line item at four companies.
I have watched this shape before. In 2020 I was a junior developer advocate at Aragon, running governance workshops, and I remember the first time I understood that a treasury vote and a capital expenditure decision are the same act of faith. Both concentrate power in the hands of whoever shows up. Both dress that concentration in the language of growth. And both collapse the moment the belief that sustains them is withdrawn.
Here is what the segment did not say, and it is the part that would have made the recommendation honest.
AI training clusters are not computers. They are conversations. Thousands of GPUs passing gradients back and forth, and the volume of that east-west traffic grows faster than the number of chips — superlinearly, because every node can talk to every other node. At a certain scale, the bottleneck stops being the GPU and becomes the wire between the GPUs. This is why eight-hundred-gigabit and 1.6-terabit Ethernet suddenly matter. This is why Arista's operating margin sits near fifty percent. The company is not really selling switches. It is selling the right to stand between the chips while they argue.
But the segment never mentioned the fight that will decide whether Arista keeps that margin. NVIDIA — the company whose GPUs created the bottleneck — sells its own networking. Spectrum-X on the Ethernet side, InfiniBand on the proprietary side, bundled with the silicon everyone already has to buy. When you control the compute, you can afford to give away the network. Arista's moat is real, but it is being approached by a company that owns the ground the moat was dug into.
Listen to what the repository refuses to say. The article names the beneficiaries and stays silent on the threat. That silence is not an oversight. It is the shape of the recommendation.
When a stock trades above the price analysts say it is worth, the market is telling you it has stopped trusting the analysts and started trusting a longer story. Kilburg was doing exactly that — making a two-to-three-year bet and letting the twelve-month target absorb the blame. That is a legitimate move. It is also a move that requires you to believe you can see further than the people whose job is to see.
Then there is Cloudflare, the piece of the story the host, at least, was willing to interrogate. A company that does not turn a profit, valued at two hundred and forty-six times forward earnings by one count and three hundred by another. Evans, the anchor, compared it to early Netflix — the kind of valuation you cannot defend with arithmetic, only with a story about what happens later. Kilburg's reply was that the earnings growth supports it.
The void between tokens holds the true value, and the void between earnings and price is where this trade actually lives. Cloudflare is not being bought for what it earns. It is being bought for what it might earn if the edge becomes the new center — if inference migrates away from the hyperscalers' data centers and out toward the periphery of the network. That is a genuine thesis. It is also a thesis with no timetable, and a company without profits cannot wait indefinitely for the future to arrive.
The most dangerous thing about the segment, though, was not any single pick. It was the word "diversified."
Kilburg called Fortinet a diversification — from CrowdStrike, from Palo Alto, from the security names he already owned. Three security companies is not diversification. It is one bet placed three times. And the deeper problem runs through all three names: Arista, Fortinet, and Cloudflare are all levered to the same single variable, the same capex line, the same belief. They will rise together, and on the morning the belief breaks, they will fall together. The portfolio is not a portfolio. It is a position.
I keep returning to the liquidity mining years for the same reason. In 2021, every protocol could show you a chart of total value locked that went up and to the right, and every chart was an advertisement for the incentives that produced it. Turn off the emissions and the TVL evaporated, because the users had never been users. They had been yield. Growth without belonging is just noise, and the AI infrastructure trade is measuring belonging with the same instrument. The shovels sell because the miners buy, and the miners buy because the shovels sell, and somewhere inside that loop is a capex guidance number that has not yet been revised downward.
And then there is the constraint the entire segment ignored: power.
The bottleneck in AI infrastructure is migrating from the chip to the substation. Data centers are becoming power plants with server racks attached, and the utilities, the grids, the cooling systems are the real limit on how fast clusters can grow. If the electricity is not there, the switches are not deployed, and Arista's order book becomes a function of interconnection queues it does not control. None of this appeared in the recommendation. The story was told as though infrastructure were frictionless, as though the only thing standing between the capex and the profit were the courage to buy. Infrastructure is never frictionless. It is concrete, copper, and cooling, and it is built by people who need permits.
There is a second hidden fragility, quieter and more structural. The hyperscalers are not passive customers. Meta designs its own switches. Google designs its own silicon. Microsoft builds its own security. The shovel sellers are, in the long run, training their largest customers to become their competitors. Every dollar of capex that flows to a supplier is also a dollar spent learning how to replace that supplier. The segment treated the suppliers as permanent beneficiaries. In truth they are temporary contractors on a construction site the client intends to own.
This is where the AI story and the story I actually work on converge, and I want to be precise, because the easy version of the argument is wrong.
The easy version says: decentralization fixes this. Put the infrastructure on-chain, distribute the compute, and no single capex line can sink the market. That is a slogan, not an architecture. I have spent the last year leading a team building Veritas — an open-source framework for verifying AI-generated content on-chain — and I can tell you that decentralized infrastructure has its own single load-bearing wall. Ours is liquidity. Theirs is capital expenditure. The same fragility, wearing different clothes.
What actually connects the two is the question neither the segment nor most of my own industry is willing to ask: who audits the ledger? When AI generates a claim, a face, a voice, a line of code, where is the record that says so? The watermarking standards I spent six months negotiating with five AI labs — the ones we folded into the Ethereum protocol — exist because the alternative is a world in which the infrastructure of truth is owned by whoever owns the models. The market is busy pricing the shovels. Nobody is pricing the pen that writes the record.
We do not write code; we weave conviction. And the conviction underneath the AI infrastructure trade is a conviction about continuity — that the capex keeps flowing, that the rate stays kind, that the customers never learn to build. Every one of those is a belief about the future, held by people who call it analysis.
Here is the contrarian turn, and it cuts against my own side as much as it cuts against Kilburg's.
The AI infrastructure trade and the crypto infrastructure trade are not opposites. They are the same trade in different costumes, and the decentralization community is not immune to it. Both are bets on a layer that sits beneath the visible action. Both reward the sellers of shovels — the validators, the bridge operators, the rollup providers — while the miners absorb the risk. Both measure health with a single number that is easy to subsidize and hard to earn. And both, when the number turns, discover that their diversification was always a single position.
My own industry spent three years telling itself that cross-chain interoperability would fix the user experience. The bridges got cheaper. The experience did not. Dencun lowered the cost of moving value between rollups, and it is still an order of magnitude more painful than pressing withdraw on a centralized exchange. We solved the price and called it the problem. The AI trade is doing the same thing: it lowered the cost of a bet and called it an investment. The OP Stack and the ZK Stack do not differ in their cryptography so much as in who convinces more projects to deploy first — a competition of adoption, not of truth. The shovels win by being chosen, not by being right.
The blind spot, then, is not that Kilburg is talking his book. Everyone talks their book. The blind spot is that the audience believes there is a book to read — that somewhere in the capex line, the P/E ratio, the analyst target, there is a fact that settles the question. There is not. There is only a covenant: a promise that the future will resemble the present, signed by people who cannot enforce it. Open source is not a license; it is a covenant. So is a capex forecast. The difference is that open source publishes its terms, and the market hides them in a target price the stock has already passed.
So what do we watch? The obvious things: hyperscaler capital expenditure guidance, every quarter, without exception. The ten-year yield, and whether it holds above the level that starves the unprofitable. NVIDIA's share of the Ethernet market, which is the slow erosion of Arista's moat. And the day Cloudflare reports its first real profit, if it ever does.
But the signal that matters most is quieter. Watch whether anyone — in AI or in crypto — builds infrastructure that does not need to be believed in. The shovels are expensive because the miners are hopeful. The ledger is silent about the hope, and the silence in the ledger speaks louder than code. Faith in the fork, hope in the merge. The question is not whether the capex holds. The question is who holds the record when it does not.