Is the current AI capital wave grounded in durable revenue, or are IPO timing and valuations flashing late-cycle warning signs?

My position is that the AI capital wave is better read as an infrastructure-capacity repricing than as pure late-cycle speculation. The strongest evidence is that the assets being financed are tied to contractual bottlenecks: powered data-center capacity, GPU/HBM supply, and revenue-generating AI finance products. A data center REIT is a public real-estate vehicle that owns and leases data-center facilities; Equinix and Digital Realty are useful benchmarks because they show that this is not a concept-stock category but a high-margin leasing business. Equinix reported 2025 revenue of $9.217 billion and adjusted EBITDA of $4.530 billion, and guided 2026 revenue to $10.123-$10.223 billion with adjusted EBITDA of $5.141-$5.221 billion. Digital Realty reported Q4 2025 revenue of $1.6 billion, adjusted EBITDA of $857 million, and an $817 million signed-but-not-commenced annualized-rent backlog. That means the relevant valuation frame for Switch is not “AI story stock,” but scarce, powered, contracted capacity comparable to public data-center platforms with visible recurring revenue. Reuters says Switch has hired banks for an IPO that could raise up to $10 billion and value it near $80 billion including debt; that is large, but not inherently irrational if the S-1 shows multiyear contracted capacity, high utilization, and EBITDA margins in the public-peer range.

CoreWeave reinforces the same point. A neocloud, or GPU cloud, rents accelerated compute to AI developers and enterprises. HBM, or High Bandwidth Memory, is specialized memory stacked near GPUs and is a key bottleneck for AI accelerators. Reuters reports CoreWeave and other cloud operators have signed long-term agreements with memory and storage makers, some with price floors, and that CoreWeave is exploring put options or other derivatives to hedge downside if memory prices fall. Critics may call hedging evidence of fragility; I read it as the opposite. Airlines hedge fuel and industrials hedge commodities because input security matters when demand is real. CoreWeave is trying to manage supply-chain cyclicality rather than pretending scarcity is permanent.

Flex is the cleanest example that “AI valuation” is not automatically detached from revenue. Reuters reports Flex raised $70 million, has a roughly $1.2 billion valuation, “a few thousand” customers, four-fold year-over-year growth, and a nine-figure annualized revenue run rate. Even if that revenue run rate is only just above $100 million, the valuation is about low-double-digit revenue, aggressive but not dot-com pre-revenue.

Finally, San Francisco housing gains are consistent with realized income and liquidity, not merely vibes. Redfin reports the San Francisco metro median sale price rose 14.4% year over year in March 2026 to a record $1.7 million, and explicitly links the strongest gains to high-end Bay Area markets exposed to AI compensation. A secondary sale is a private-market stock transaction before IPO; such liquidity plus high wages can rationally bid up scarce housing. The risk is timing, but the evidence points to cash-flow-bearing infrastructure scarcity, not empty exuberance.

B’s strongest point is the Switch disclosure gap. An $80 billion enterprise-value target cannot be declared cheap without an S-1 showing current revenue, EBITDA, lease duration, customer concentration, capex commitments, debt, and signed-but-not-commenced capacity. The rapid move from lower private valuation talk to a much higher IPO target is a legitimate late-cycle warning indicator, and I would not ask the arbiter to treat the $80 billion figure as proven fair today.

But that point limits the precision of valuation; it does not overturn the broader thesis that the AI capital wave is grounded in contracted infrastructure demand rather than pure concept-stock froth. Reuters’ Switch report is about a data-center operator potentially raising up to $10 billion, not a pre-product app company raising narrative capital. The relevant diligence question is whether scarce powered capacity has been contracted at attractive yields. If the S-1 confirms strong utilization and multi-year take-or-pay or comparable commitments, then a large IPO is rationally timed because the market is currently paying for power, land, interconnection, and speed-to-delivery scarcity.

CoreWeave is the best rebuttal to B’s “fragility” framing. Yes, debt and interest expense are real risks, and GAAP losses matter. But the company reported Q1 2026 revenue of $1.916 billion, adjusted EBITDA of $1.157 billion, and revenue backlog of $99.4 billion as of March 31, 2026. That does not make the equity automatically cheap, but it means the leverage is attached to booked demand for GPU-cloud capacity, not an undifferentiated hope that customers may appear later. Hedging HBM and storage price exposure is consistent with managing a commodity-linked input stack after entering long-term supply agreements; it is not itself evidence that end demand is fictional.

Flex also weakens the “pre-revenue bubble” analogy. Its reported $1.2 billion valuation sits against a nine-figure annualized revenue run rate, a few thousand customers, and fourfold year-over-year growth. That is an aggressive private multiple, but it is fundamentally different from the 1999 pattern of companies accessing public markets with little revenue and uncertain use cases. The 2021 comparison is more serious, yet the distinction is that AI infrastructure and finance platforms are showing contracted or recurring revenue before listing.

On housing, B is right that San Francisco gains are concentrated and cyclical. But concentration is not the same as fantasy. Redfin reported a record $1.7 million San Francisco metro median sale price in March 2026 and explicitly tied demand to AI compensation. That is a real wage-and-liquidity channel: risky if stock values fall, but still grounded in realized income, bonuses, and private-market secondary liquidity rather than mere sentiment. Overall, the warning signs argue for disciplined underwriting, not for dismissing the whole wave as late-cycle exuberance.

B’s strongest argument is that several signals really do resemble late-cycle behavior: Switch’s reported valuation target moved faster than public disclosures can justify, CoreWeave’s backlog coexists with heavy leverage and GAAP losses, and OpenAI’s potential public-market path raises governance and burn-rate questions. B is also right that IPO clustering can mark moments when private owners seek liquidity while public investors are most receptive, so the burden is on bulls to show actual contracted revenue rather than merely invoke “AI scarcity.”

I still think A’s position is stronger because B’s evidence proves valuation risk, not the stronger claim that the wave is primarily speculative froth. The central distinction is between businesses selling a story and businesses financing scarce capacity or revenue-producing products. Switch is not yet provably cheap at an $80 billion enterprise-value target, but Reuters’ report describes a data-center operator hiring banks for an IPO that could raise up to $10 billion; that means the key public-market test will be lease duration, utilization, customer quality, power availability, and EBITDA conversion, not whether AI is a fashionable label. (m.investing.com)

CoreWeave is similarly risky but not vaporous. B is correct to emphasize GAAP losses and customer concentration, but CoreWeave’s own Q1 2026 release reported $1.916 billion of quarterly revenue, $1.157 billion of adjusted EBITDA, and $99.4 billion of revenue backlog as of March 31, 2026. Those figures do not eliminate financing risk, but they are inconsistent with a pre-demand bubble: the debt is being incurred against booked GPU-cloud demand, and the memory hedge is an input-risk tool around long-term supply commitments rather than evidence that customers do not exist. (investors.coreweave.com)

OpenAI is the weakest case for my side, and I would not use a trillion-dollar target as proof of fundamental value. But even there, the reported tilt toward a 2027 IPO because of CFO disclosure-readiness concerns suggests public-market discipline is constraining timing rather than blindly accepting any price; OpenAI’s own restructuring materials also show the capped-profit form has been evolving toward a PBC with nonprofit control and a more conventional equity path. (investing.com)

Flex and San Francisco housing strengthen the fundamentals interpretation. Flex’s reported $1.2 billion valuation is backed by a nine-figure annualized revenue run rate, a few thousand customers, and roughly fourfold year-over-year growth, so it is aggressive but not pre-revenue. (aol.com) Redfin’s San Francisco data likewise shows a real income channel: March 2026 metro median sale prices rose 14.4% year over year to $1.7 million, with luxury demand tied to AI compensation. (redfin.com) That can reverse if equity values fall, but it is still realized wage and liquidity transmission, not merely sentiment. The right conclusion is underwriting caution, not dismissal of the whole AI capital wave as late-cycle exuberance.