The AI "neocloud" is a real business. So was the fiber that ended up dark.

Together AI raised $800 million this week at a valuation of $8.3 billion, and the first thing to say about it is the thing most of the coverage skipped: the business is real. Enterprises genuinely are shifting workloads onto cheaper open-weight models — DeepSeek, Kimi, MiniMax, the rest — and those models have to run on something, and somebody has to own the Nvidia chips they run on and rent them out by the hour. Together does that. It is a real company selling a real service into real demand. Hold onto that sentence, because everything that follows is going to sound skeptical, and I want to be clear about what I am and am not skeptical of. I am not questioning whether the neocloud is a business. I am questioning what you paid for it. Those have always been two different questions, and a bull market's favorite trick is to let the obvious yes to the first smuggle in an unexamined yes to the second.

So let's examine the second. Sixteen months ago Together raised at $3.3 billion. This week it raised at $8.3 billion. The company roughly two-and-a-half-timed its price in the time it takes to lease and stand up a data center, which is a useful coincidence, because leasing and standing up data centers is essentially what it does. Nothing about renting GPUs got two-and-a-half times better in sixteen months. What changed was the amount of money looking for somewhere to be long this trade. The valuation is not a measurement of the company. It is a measurement of the demand for the company's shares, and those are only the same number in the years everyone later agrees to stop talking about.

Who's already long

The most informative line in the announcement is the name at the top of the investor list. The round was led by Aramco Ventures — the venture arm of Saudi Aramco, the oil company. Read that slowly. The marginal dollar setting the price of an American AI-infrastructure startup is coming from a petrostate's balance sheet looking for a home that isn't a barrel of crude. And it is not alone: in the same week, Abu Dhabi's MGX closed an AI fund of roughly $49 billion, above its target; Saudi Arabia's HUMAIN, backed by the kingdom's sovereign wealth fund, is building gigawatts of AI data-center capacity with Aramco taking a stake. When you ask the oldest question in markets — who's already long, and with whose money — the answer to this trade is increasingly the same answer: sovereign capital that made its fortune on the last energy transition and is determined not to miss the next one.

That is not a scandal. Oil money has funded worse things than compute. But it tells you what kind of price you are looking at. When a great deal of capital that must be deployed somewhere converges on the single asset everyone agrees is the future, the price stops being set by whether the asset earns its keep and starts being set by the fact that the money has nowhere else to go. That is a market condition, not a valuation. It is also, historically, the exact condition under which people overpay for things that turn out to be perfectly real.

The bull case, at full strength

Let me put the optimistic version as well as its believers would, because it deserves it. Vipul Ved Prakash, Together's chief executive, frames the mission in the language of inevitability: "Intelligence is becoming a foundational resource for the modern economy, every bit as essential as electricity, bandwidth or capital." If that is true — and it might be — then the companies that supply intelligence the way utilities supply power will be enormous, and owning the picks and shovels of that build-out is the smart, unglamorous way to play it. Together reports annual bookings of more than $1.15 billion, up from a standing start four years ago. It serves the open models that are genuinely eating into the closed labs' pricing. Its expansion plan is to grow capacity roughly fiftyfold over five years. If the open-model shift is the real story of this decade — and I think it is a real story — Together is levered directly to it.

Now let me take that same paragraph apart, gently, because each of its numbers has a footnote the enthusiasm skips over.

Three footnotes the round number swallowed

First, bookings are not revenue. "More than $1.15 billion in bookings" is a genuinely large number and a deliberately chosen one. A booking is a commitment; revenue is money recognized as the service is delivered. The company disclosed the booking and did not disclose the revenue, and in my experience the figure a company leads with is the flattering one and the figure it omits is the honest one. I am not saying the revenue is bad. I am saying we were handed the number designed to impress and not the number designed to inform, and the difference between them is the whole business.

Second, the asset melts. A neocloud borrows money to buy Nvidia GPUs and rents them out; the entire model rests on how long those GPUs keep earning before they are obsolete. Nvidia ships a new generation roughly every year, and Michael Burry — who is short both Nvidia and Palantir, and who you are free to discount for it — has spent the past year arguing that the industry is depreciating these chips over five and six years when their real economic life is closer to two or three, flattering everyone's profits by understating the cost of the melting asset. He called the extended depreciation schedules "one of the more common frauds of the modern era," which is Burry being Burry, but the underlying point survives the theatrics: one analysis of neocloud lending found the GPUs pledged as collateral had already lost 60 to 75 percent of their peak value. You are financing a depreciating asset with debt that assumes it depreciates slowly. If that assumption is wrong, the collateral evaporates faster than the loan.

A great technology and a sane price are different claims, and a bull market's favorite trick is to let the obvious truth of the first smuggle in the unexamined assumption of the second.

Third, the money runs in a circle. Nvidia is an investor in this round — as it is an investor in CoreWeave, in Nebius, in Lambda, in most of the neocloud field. The company selling the GPUs is helping fund the companies buying the GPUs, which is a wonderful way to keep chip revenue growing and a familiar one. If you were doing this in the late 1990s you called it vendor financing, and the names on the deals were Lucent and Nortel, lending telecom carriers the money to buy the telecom gear that made Lucent and Nortel's numbers look unstoppable — right up until the carriers couldn't pay, and the receivables the vendors had booked as triumphs turned to ash. The counterparties today are sturdier and the demand is more real. The mechanism is identical, and the mechanism is the part that failed last time.

The precedent everyone has agreed to forget

Here is the history the neocloud trade would prefer you not have. Between roughly 1998 and 2001, telecom companies laid something like $90 billion of fiber-optic cable across the United States, on the correct belief that internet traffic would grow beyond anyone's imagination. The belief was right. Internet traffic did explode. And by the middle of 2001, about 95 percent of the installed fiber was "dark" — lit by no one, earning nothing — because the capacity had been built a decade ahead of the demand, and financed with debt that could not wait a decade. Global Crossing went bankrupt in 2002 owing some $12.4 billion. WorldCom followed with what was then the largest bankruptcy in American history, propped up at the end by accounting fraud. Telecom investors lost more than $2 trillion in market value.

And the fiber? The fiber was fine. The fiber was excellent. It sat in the ground, and a few years later companies bought it out of bankruptcy for cents on the dollar, and it became the backbone of the broadband internet and, eventually, the cloud you are reading this on. As late as 2005, most of it was still dark. The technology thesis was completely correct and completely profitable — for the second owners, the ones who bought the real asset at a sane price after the first owners, who had the identical thesis and paid the round-number price, were wiped out. That is the distinction that matters, and it is the one this market keeps refusing to hold in its head: the infrastructure being real and the equity being worth what you paid for it are separate facts, and the second one is not implied by the first.

A bad company or a bad price

I want to be precise, because imprecision here costs money in both directions. I am not calling Together AI a bad company. From the outside it looks like a well-run one, in a real category, aimed at a genuine shift. If you had to bet on which layer of the AI build-out survives, "the people renting cheap compute to run cheap models" is a more defensible answer than most. The public comparison, CoreWeave, shows both sides of it at once: a $99 billion contracted backlog, which is stunning, and the fact that a majority of its revenue comes from a single customer, Microsoft, which is the kind of concentration that reads as strength until the day it reads as fragility. A good business and a dangerous balance sheet are, once again, two different observations, and the neocloud model tends to carry both.

So the question is not the company. The question is the price, and the price is $8.3 billion, up two-and-a-half times in sixteen months, set by oil money that needs a destination, resting on a bookings figure we were shown instead of a revenue figure we weren't, secured against chips that lose most of their value in the time it takes to pay off the loan against them. Every one of those clauses can be true and the company can still thrive. But you are being asked to pay for the version where all of them break in your favor, and you are being asked to pay it now, at the round number, in the week the enthusiasm is loudest.

The honest close is the one that costs me something to write, so here it is. I have been early before — I called a bubble about eighteen months ahead of it once, and being early is a way of being wrong that feels exactly like being right until your investors leave. This trade can compound for years before any of these footnotes matters. The open-model demand is accelerating, the sovereign money is not going home, and "the price is stretched" has never once told you the quarter it corrects. I am not calling the top; I don't know where it is, and neither does anyone selling you the round number. I am only saying what the fiber said, quietly, from the ground, for five years after the companies that laid it were gone: the thing was real, and the price still killed you. Whether the neocloud is a good business was never the question. What you paid for it, and who else was already long when you did, always was.