The Numbers Are Real. The Valuations May Not Be.
KazuThe five largest technology companies reported earnings last week. The results were, by any conventional measure, extraordinary. Microsoft posted $82.9 billion in quarterly revenue, up 18% year over year, with Azure growing at rates that caused analysts to revise their full-year estimates upward mid-call. Amazon's net sales reached $181.5 billion, up 17%, with AWS expanding 28%. Google Cloud grew 63%. Meta's revenue surged 33% to produce $26.8 billion in net profit in a single quarter, a figure larger than the annual revenue of most Fortune 500 companies. Apple, navigating tariff headwinds, came in at $111.2 billion in sales, beating estimates.
Across five companies, combined 2026 AI capital expenditure is now on track to exceed $650 billion. Microsoft alone has guided for $190 billion in capital spending this year.
Alongside the earnings, Anthropic (the AI safety company founded by former OpenAI researchers) reported $30 billion in annualized revenue and is in discussions with investors about a new funding round that would value it at $900 billion, surpassing OpenAI's most recent valuation of $850 billion. A private AI company with no public market pricing, no disclosed path to profitability, and a founding mission centered on preventing AI from destroying humanity is now being valued, by serious institutional investors, at roughly the GDP of the Netherlands.
The chart attached to the question this moment keeps raising is a BofA Global Research graphic that has been circulating in investor presentations for months. It shows concentration levels in the S&P 500 over the past two centuries: the railroad bubble, the Nifty Fifty, Japan at its 1989 peak, the dot-com era. At each of those peaks, the largest companies in the relevant index reached approximately 40 to 44 percent of total market weight before the correction arrived. The AI Big 10 (the Magnificent Seven plus Broadcom, Micron, and AMD) currently represent approximately 40% of the S&P 500.
The chart makes the pattern visible. What it cannot answer is whether the pattern applies.
What the Earnings Week Actually Showed
The question hanging over last week's calls, stated explicitly on the Meta earnings call and implicit in every analyst question to Microsoft and Amazon, was whether the AI investment supercycle was producing commensurate returns, or whether the hyperscalers were building infrastructure faster than demand could absorb it.
The Q1 numbers suggested, provisionally, that demand is holding. Key figures from the week:
- Microsoft: Revenue $82.9 billion (+18%), operating income $38.4 billion (+20%), net income $31.8 billion (+23%). Azure growth accelerated, driven by AI workloads
- Amazon: Net sales $181.5 billion (+17%). AWS revenue grew 28%, reaching an annualized run rate that would make it a top-20 company in the S&P 500 on its own
- Meta: Revenue up 33%, net profit $26.8 billion. Zuckerberg announced the first model release from Meta Superintelligence Labs and raised full-year capex guidance to $125-145 billion, sending shares down 6% after hours
- Apple: Revenue $111.2 billion (+16.6%), beating estimates despite significant tariff exposure in its China supply chain
- Alphabet: Google Cloud grew 63%, contributing to overall revenue growth that exceeded expectations
The Anthropic Number
The $900 billion valuation for Anthropic requires examination independent of the earnings week.
The company's $30 billion in annualized revenue represents genuine commercial traction: enterprise API contracts, Claude deployments in corporate workflows, and AWS integration that has given it distribution at scale. Its model releases (Claude 3, Claude 3.5, and subsequent iterations) have competed directly with OpenAI on benchmark performance and won enterprise customers on reliability and safety positioning.
The valuation implies a revenue multiple of 30x on annualized figures for a company that has not disclosed a path to profitability, operates in a market where its primary competitor is backed by Microsoft, and whose compute costs scale with usage in ways that make margin expansion structurally difficult at current model architectures.
Those are not disqualifying conditions for a private valuation in a category where investors are pricing for potential rather than present earnings. They are worth naming when the number being discussed is $900 billion.
Why This Might Be Different
The BofA concentration chart is compelling precisely because the pattern is so consistent. Railroads, Nifty Fifty, Japan, dot-com: each reached approximately 40% concentration before collapsing. The argument for why this time might diverge rests on a fundamental difference between the current AI leaders and the companies that led prior bubbles.
At the peak of the 2000 bubble, the largest technology companies were valued at extraordinary multiples on revenues that were themselves largely speculative. Pets.com, Webvan, and comparable names had achieved enormous market caps on the premise of business models that had not demonstrated profitability at any scale. The infrastructure companies (Cisco, Sun Microsystems, Lucent) were priced on the assumption of indefinite network equipment demand that evaporated when enterprise buying slowed.
The current AI Big 10 generate profits that are large by any historical comparison. Microsoft's trailing twelve-month free cash flow exceeds $70 billion. Amazon's AWS segment generates operating income in excess of $30 billion annually. Meta's advertising business, which funds its AI investment, produced $26.8 billion in net profit in a single quarter. These are not companies valued on the premise of future business models. They are companies valued at high multiples on demonstrated, growing, and in several cases accelerating earnings. The marginal cost of delivering AI capability also declines as models improve, meaning the capital being deployed today builds capacity whose per-unit cost decreases over time. That is a structurally different proposition than building railroad track into territory that will never generate sufficient freight revenue.
The Case for Concern
The $650 billion in annual AI capital expenditure across five hyperscalers represents a bet not just on demand growth, but on demand growth sufficient to generate returns on assets that depreciate rapidly. Data centers built for 2024 model architectures may be partially obsolete within five years as model efficiency improves. The compute cluster optimized for one generation of training runs is not automatically optimized for the next.
The concentration figure itself carries information that earnings results cannot resolve. When 40% of the S&P 500's market capitalization is held by ten companies, the index's performance becomes tightly coupled to the fortunes of those companies. Passive investing flows (now representing a majority of total equity investment) amplify this effect automatically: every dollar entering the index must allocate 40 cents to the same ten names, pushing valuations higher independent of fundamental developments at those companies.
Meta's after-hours decline on earnings night illustrates the tension precisely. The company beat revenue estimates and raised guidance. The market's concern was whether $125-145 billion in annual capital expenditure, committed to AI infrastructure whose monetization timeline extends well beyond the current quarter, represents a rational allocation of capital or an escalating bet that all five hyperscalers are making simultaneously and that none of them can unilaterally walk away from.
Large, profitable companies making enormous, simultaneous, strategically necessary bets on infrastructure whose returns are uncertain and long-dated: that is not a bubble in the traditional sense. It may be something without a precise historical analogue.
What the Pattern Cannot Resolve
The BofA chart documents what happened at 40% concentration in prior cycles. It does not explain what caused those corrections, only that they occurred:
- The railroad collapse was driven by overcapacity and rate regulation
- The Nifty Fifty correction followed the 1973 oil shock and a decade of inflation
- Japan's 1989 peak ended when the Bank of Japan raised rates aggressively to cool a real estate and equity bubble that had become untethered from underlying economic activity
- The dot-com crash was a capital structure failure in companies whose revenues were insufficient to service their debt
The earnings results from last week are substantial, specific, and verifiable. What remains genuinely uncertain is whether the valuations placed on those results (Anthropic at $900 billion on $30 billion in annualized revenue, the AI Big 10 at 40% of the S&P 500) represent a fair price for the future being built, or whether, as in prior cycles, that future is further away and more complicated than the present momentum suggests.