Chapter 2 · Does the AI Bubble Exist?
The strongest arguments that the AI boom is a genuine earnings-driven supercycle — not a mania — drawn from the best steelman sources in the corpus.
Start with the difference that defines everything: the dot-com index was led by story stocks with no profits trading at 60x+ sales. Today's concentration is in companies that mint actual cash. Nvidia produces roughly $49B of free cash flow per quarter. Microsoft, Alphabet, Amazon and Meta together generate half a trillion dollars of operating cash flow annually. Even after the record capex, these remain investment-grade giants self-funding the majority of their buildout. Tom Lee's framing — that 2027 US AI spend will exceed the entire Department of Defense budget — cuts both ways, but his deeper point stands: railroads-on-steroids comparisons usually precede decades of productivity gains, not immediate collapse.
Revenue, meanwhile, is exploding off a real base. Anthropic's run-rate went from $9B at end-2025 to $65B by July 2026 — seven-fold growth in eight months — and the company posted its first positive adjusted operating income in Q2 2026. OpenAI runs at ~$40B annualized. These are not Pets.com sock puppets; they are the fastest-revenue-scaling enterprises ever measured. The demand signals are physical, not narrative: SK Hynix earned 76% operating margins selling memory; power interconnection queues stretch years; DDR memory prices multiplied. Scarcity like this does not accompany fake demand.
Nasdaq trades near 24-25x forward vs ~60x for the composite in March 2000; Nvidia around 24-30x with $49B+ quarterly free cash flow; Mag7 profits are real and growing. In 2000 many index leaders had no profits at all.
Anthropic hit a $65B annualized run-rate by July 2026 (+~7x YoY) with its first positive adjusted operating income in Q2 2026; OpenAI ~$40B run-rate; Microsoft, Google Cloud and Nvidia book genuine AI revenue today. Total AI token/subscription revenue more than tripled in a year.
Mag7 generate hundreds of billions in operating cash flow ($500B+/yr combined); even after record capex they hold large cash piles; Nvidia alone has ~$49B/qtr FCF plus an $80B buyback authorization. Dot-com startups financed with convertible bonds; today's core is investment-grade self-funding.
BlackRock's Fink calls AI infrastructure a new 'financeable asset class'; the $500B Apollo-led vehicle recreates the MBS playbook that funded past buildouts for years. Bears have been calling this crash since ChatGPT launched; the market melted up anyway.
ChatGPT-class tools reached hundreds of millions of users faster than any consumer product in history; Anthropic's run-rate went $9B → $65B inside eight months; API token volumes grew multiples in a year; Ramp data shows median business AI spend rising steadily though concentrated in power users.
Multi-year interconnection queues for data-center power; SK Hynix Q2 2026 profit >13x YoY at 76% operating margins; DRAM/memory prices up multiples (DDR3 +600%); Cisco CEO: supply 'massively short' on power, capacity, compute, memory. Scarcity economics supports current asset returns (SiliconANGLE scarcity→surplus framework argues the turn comes only when scarcity breaks).
Futures still price Fed easing into 2027 despite hawkish noise; Ned Davis research cited by CNBC: all five major bubbles of the past century saw rising yields INTO their peaks — rates pop bubbles only 'eventually', often years later than bears expect.
Railways and electricity went through identical blowoffs then reshaped everything; Tom Lee: 2027 US AI spend will exceed the DoD budget, railroads-on-steroids framing; even ECB frames it as 'rational enthusiasm or bubble?' — both can be partially true. Nikkei editor's choice argues dot-com comparison overdone because earnings are real this time.
$16.1T of billionaire wealth (vs $4.4T in 2007) needs assets; Europe/Japan/China offer lower growth or capital controls; gold at record highs reflects the search for alternatives, not availability of them. TINA flows have funded every 'final top' call since 2023.
Commerce discussions on federal backstops for AI compute; OpenAI offered the US government a 5% stake structure; Stargate framed as national-champion project; defense-adjacent procurement growing. Bulls argue Washington cannot allow a disorderly AI collapse mid-AI-race with China.
Scroll.in: a crash exposing AI deployment as slower/costlier than hyped would push enterprises back to 'dependable, cost-effective outsourcing India has long provided' — data engineering, cloud migration, compliance work. Nifty IT bottomed and bounced hard in 2026 as results stayed solid (Infosys raised FY27 guidance; TCS margins ~24%).
Nifty diversification: top-3 stocks ≈25% weight vs Samsung+SK Hynix >50% of KOSPI; SEBI blocks single-stock leveraged ETFs of the kind that amplified Korea's crash; domestic SIP inflows (₹25k+ crore/month industry-wide) provide a steady bid; India is an AI USER with no frontier-lab exposure — cheap post-bust AI is a net tailwind for adoption.
>2,100 Global Capability Centres employing 2.36M people generating $98B+/yr (Scroll.in); Aadhaar/UPI/BHASHINI stack built without frontier-lab dependence (The Week: India could emerge a relative winner from an AI bust); Microsoft $17.5B (2026-29), AWS up to $35B (→2030), Google-Adani Vizag gigawatt campus committed INTO India.
On valuation — the bears' strongest ground — bulls counter with forward multiples that are elevated but nowhere near 1999: Nasdaq near 24-25x forward earnings vs ~60x at the 2000 peak; Nvidia around 24-30x despite hypergrowth. The Nikkei editor's-choice argument crystallizes it: unlike 1999, when the infrastructure (fiber) was laid by bankrupt startups and monetized a decade later by others, today's builders capture their own traffic — every token sold through Azure, AWS or Google Cloud bills the same companies doing the spending.
Finally, the escape hatches. Wall Street securitization can extend the runway for years — BlackRock's Larry Fink calls AI infrastructure a new financeable asset class, and history says credit expansion defers reckonings far longer than bears expect. The US government is structurally invested in AI winning against China, making disorderly collapse politically unacceptable. And if inference costs keep falling the way they have for three straight years, today's capex gets validated on schedule. As one corpus source put it: the cleanest falsifiable test of the whole bear case is the token-price curve.