Fiber Was Real Too: What the AI Capital Wave Is Actually Pricing

Consider what crossed the wire in a single week of July. Switch, a Las Vegas data-center operator, hired Goldman Sachs and JPMorgan¹ for an IPO that could raise $10 billion and value the company near $80 billion including debt. CoreWeave, the AI cloud provider, was reported to be exploring put options² to insure itself against a fall in memory-chip prices. Flex, a three-year-old AI banking startup, doubled its valuation to $1.2 billion¹⁰ in roughly six months. In the background, OpenAI is wrestling over whether to list at $1 trillion, and San Francisco's median home price sits at a record $1.7 million⁹ on the strength of AI paychecks. Five stories, one question: is this capital wave standing on durable revenue, or are we watching the late innings of a mania?

I think the honest answer is both, and that the combination has a historical name. The right analogy for this moment is not 1999, when companies with no revenue went public on eyeball counts. It is 2000-era telecom, when the fiber-optic buildout was real, the demand forecasts eventually came true, and the equity holders got wiped out anyway because the infrastructure was financed with debt at prices that assumed no cycle would ever arrive. This is not a story about fake demand. It is a story about what real demand costs when everyone rushes to finance it at once.

Start with Switch, because it is the cleanest test. Data-center leasing is a genuinely good business; nobody serious disputes that. Equinix, the largest public benchmark, reported $9.2 billion of 2025 revenue and $4.5 billion of adjusted EBITDA¹², and just raised 2026 guidance to a 51 percent margin on AI-driven demand. But Equinix earned its market value, in the $80-90 billion range, over two decades of audited quarterly filings. Switch was taken private for $11 billion in 2022¹, discussed a private raise at a valuation of at least $40 billion this spring, and is now floating a number near $80 billion, all without a single public financial statement in four years. The valuation has roughly septupled since 2022 entirely in private marks. Maybe the S-1 will justify it with contracted, powered capacity at fat yields. Until then, the honest description of that number is momentum, and the fact that its owners want to convert it to public currency in the fourth quarter, at the strongest IPO pace since 2021, tells you something about what sellers think of the window.

CoreWeave is the harder case, because the bulls have real ammunition. The company's first quarter was extraordinary by any topline measure: revenue of $2.08 billion, up 112 percent, and a $99.4 billion revenue backlog⁴, including a $21 billion commitment from Meta and ten customers pledging at least $1 billion each. Those are signed contracts, not vibes, and they distinguish this cycle from anything in 1999. Here is what happened in the same quarter, though: the GAAP net loss widened to $740 million from $315 million, interest expense doubled to $536 million, and, per the company's own presentation⁵, adjusted operating income collapsed to $21 million from $163 million, a margin of 1 percent, down from 17. CoreWeave ended the quarter with nearly $25 billion of debt⁶ and spent $7.7 billion on capital equipment in three months. If contracted demand truly de-risked the leverage, unit economics should be improving with scale. They are moving the other way. And the backlog itself is a concentrated bet: 62 percent of 2024 revenue came from Microsoft alone⁴, so the headline number is largely a wager on a handful of hyperscalers never pausing their capex.

The memory hedge makes the fragility legible. To secure high-bandwidth memory and storage for its GPU fleet, CoreWeave signed long-term supply deals with Micron and SanDisk that guarantee the chipmakers price floors². If memory prices fall when new fabrication capacity ramps around 2028, CoreWeave is stuck paying above market. So executives have discussed buying put options, likely on memory-maker stocks, since no liquid market for the chips themselves exists³. Defenders compare this to airlines hedging jet fuel, and the comparison is fair as far as it goes. But airlines hedge fuel from profitable operations. CoreWeave would be buying imperfect proxy insurance against contracts it had to sign to get supply at all, while losing money on a GAAP basis and carrying $25 billion of debt. Prudent, yes. Also a portrait of a company whose margin for error is thin enough that it needs Wall Street to insure it against its own procurement.

Then there is OpenAI, where the timing pressures are least subtle. The company filed confidentially for an IPO in June; its advisers offered a choice between listing sooner at a lower price or waiting for the full trillion, and Sam Altman reportedly called any sub-$1 trillion valuation a non-starter⁷, pushing the debut toward 2027. His own CFO, Sarah Friar, has cited disclosure readiness and $600 billion in infrastructure commitments as reasons to wait⁸. The numbers explain the caution: OpenAI posted a $38.5 billion net loss on $13.07 billion of 2025 revenue⁷. Revenue is now running near $2 billion a month, which is genuinely remarkable, but the sharpest deadline in the process belongs to a shareholder: SoftBank carries a $40 billion bridge loan due March 2027⁸ that a public listing would help resolve. When an IPO calendar is being negotiated between a symbolic valuation and a lender's repayment schedule, the issuance is being timed for sellers, not buyers.

The steady rejoinder to all of this, and it deserves a real hearing, is that scale and cash flow make this cycle different. JPMorgan's strategists note that two-thirds of the 25 largest IPOs in history were followed by positive S&P 500 returns¹³ over the next year, and estimate $1.5 trillion of 2026 buybacks can absorb the roughly $260 billion of expected equity issuance. Today's listers have revenue; even little Flex backs its $1.2 billion mark with a nine-figure run rate growing fourfold¹¹. All true, and it is why I do not expect a 2000-style Nasdaq collapse on any particular date. But the telecom precedent answers this directly. Bandwidth demand grew relentlessly through the crash; the companies that built the fiber still failed, because contracted growth financed with leverage at perfect-world prices cannot survive a single demand pause or price cycle. The question was never whether compute demand is real. It is whether these specific capital structures survive the first bad year.

San Francisco housing shows what happens when the private marks leak into the real economy. Redfin's March data had the metro median up 14.4 percent, condos up 24.4 percent, homes selling 8.9 percent over list with 1.8 months of supply, and an agent reporting that "A lot of 22-year-olds are getting $500,000 signing bonuses from AI companies"⁹. That is real income, not sentiment. But much of it is compensation and secondary-sale liquidity priced off valuations that have never cleared a public market. Housing is where paper marks become thirty-year mortgages, which makes it the most leveraged expression of this cycle's confidence and the place a repricing would hurt ordinary people most.

So my reading is this: the wave is financing real, contracted, cash-producing assets, and pricing them as if the cycle has been abolished. The signals I trust least are precisely the ones that live outside audited disclosure: Switch's private markup, OpenAI's trillion-dollar floor, the condo bids in Pacific Heights. Schroders' Duncan Lamont recently flagged that analysts' expected long-term S&P 500 earnings growth has hit 20.2 percent, above the 18.6 percent record set at the 2000 peak¹⁴, calling it "yet another way in which 2026 is looking like 1999." I would amend him by one year. The fiber was real then too, and the people who financed it at the top still lost nearly everything.