This is BMInsider PRO content
You're reading a Deep Dive — our most in-depth research. Unlock full access to all analyses, Smart Money breakdowns & more.
Unlock BMInsider PRO →The Day the Story Cracked
For roughly two and a half years, the AI infrastructure trade has rested on a single, unspoken assumption: that the demand side — anchored by OpenAI — would always grow faster than the supply side could build for. Every Mag7 capex announcement, every Nvidia earnings beat, every Oracle cloud-revenue surprise traced back to that one premise. On Monday evening, the Wall Street Journal published a report that put a small but visible crack in it.
According to the WSJ, OpenAI has missed several of its internal targets: a goal of one billion weekly active users by the end of 2025, an annual revenue target for ChatGPT, and multiple monthly revenue milestones in early 2026 — particularly as Anthropic and Google's Gemini have gained share in enterprise markets. The most consequential detail, picked up by every wire service the next morning: CFO Sarah Friar reportedly told leadership she was concerned about OpenAI's ability to honor its compute contracts in future years if the top line continues to underperform.
The market reaction was sharp and unambiguous. CoreWeave fell more than 5% pre-market on Tuesday. Oracle declined approximately 5.5%. Nvidia became the worst-performing Magnificent Seven name on the day. AMD and Broadcom both dropped around 4%. Each of these names is a primary beneficiary, in different ways, of the OpenAI compute footprint. The reaction was not panic — most of these stocks have rallied for years and remain near highs — but it was the first time in a long time the market treated AI infrastructure exposure as a two-sided risk.
What the WSJ Actually Reported
The article is worth reading carefully because the framing matters. The WSJ did not claim OpenAI is failing. It reported that the company has fallen short of its own internal projections, which is a meaningfully different statement. Internal targets at hyper-growth companies are typically aggressive by design — they exist to drive the organization, not to be hit. But the cumulative pattern is what got investors' attention.
The targets the WSJ identified as missed include: the one-billion weekly active user threshold, which OpenAI had publicly suggested it could reach by year-end 2025; the company's full-year 2025 ChatGPT revenue target; and several monthly revenue milestones in early 2026 that were tied to enterprise wins. The narrative around enterprise weakness is what lent the report credibility. Anthropic's Claude has gained meaningful share in coding and analytical workloads over the past nine months, and Google's Gemini 2.5 deployment within Workspace has shifted some prosumer demand. The WSJ's CFO quote — that Friar privately worried about compute-contract obligations — is the line that triggered the share price reaction.
OpenAI's response was measured. A spokesperson told Bloomberg the company is "firing on all cylinders," citing strong demand from enterprise customers and growing interest in advertising. Sam Altman sent an internal memo the same week acknowledging that "a lot of the things that we do that look weird — buying huge amounts of compute while our revenue is relatively small" — would eventually make sense. That memo did not have the calming effect Altman likely intended.
The Numbers Behind the Hype
To understand whether the WSJ report is a meaningful signal or a passing news cycle, the actual revenue trajectory needs to be examined honestly. OpenAI's growth, even if slower than internal targets, remains historically extraordinary. CFO Sarah Friar publicly confirmed in January 2026 that the company recorded $20 billion in revenue for full-year 2025, against $6 billion in 2024 and $2 billion in 2023 — a 3x year-over-year growth rate sustained over multiple years.
Sacra estimates that annualized revenue reached approximately $25 billion in February 2026, up from $20 billion at year-end 2025. Weekly active users were reported at 910 million as of early 2026, up from 800 million in October 2025 and 700 million in July. Paying business users surpassed 9 million by February, up from 5 million in August. By any normal corporate yardstick, this is hypergrowth. The issue is not the absolute numbers. The issue is that OpenAI's compute commitments, capital structure, and valuation now imply a future revenue scale that even hypergrowth may struggle to deliver.
Consider the cost structure. OpenAI's gross margin in 2025 was 33%, constrained by inference costs that reached $8.4 billion that year and are projected to rise to $14.1 billion in 2026. The company projected cash burn of approximately $9 billion in 2025 and $17 billion in 2026, with cash-flow positivity not expected until 2030. OpenAI also agreed in October 2025 to pay Microsoft 20% of total revenue through 2032 under a renegotiated partnership — a structural drag on free cash flow generation that many retail investors do not fully appreciate. The internal target of $85 billion in revenue by 2030 leaves very little margin for execution slippage.
The Domino Effect: Oracle, Nvidia, Hyperscalers
Why does any of this matter for stocks? Because OpenAI is not just one customer among many. It is the central pillar of a complex ecosystem of compute commitments, equity stakes, and forward revenue assumptions that span most of the AI-exposed names in the S&P 500.
Start with Microsoft, which has the largest direct exposure. Following the October 2025 recapitalization, Microsoft held an approximately 27% stake in OpenAI Group PBC on an as-converted diluted basis, valued at roughly $135 billion. Microsoft is also OpenAI's primary cloud provider, the largest single beneficiary of OpenAI's compute spend, and a 20%-of-revenue partner through 2032. If OpenAI's revenue trajectory disappoints, Microsoft is hit on three vectors simultaneously: equity value mark-down, cloud-revenue growth deceleration, and partnership distribution decline.
Oracle is a more recent but more concentrated bet. The company has signed multi-year compute agreements with OpenAI worth tens of billions of dollars, anchored by the Stargate joint venture announced in early 2025 — the $500 billion AI infrastructure project involving Trump, Oracle, OpenAI, and SoftBank. Oracle's stock performance over the past 18 months has substantially reflected expectations of these contracts converting into recognized revenue. A failure of OpenAI to fulfill its commitments would directly threaten that revenue conversion. CoreWeave's situation is similar but at smaller scale and with weaker counterparty diversification.
Nvidia is the most interesting case because the exposure is indirect but enormous. Nvidia does not sell directly to OpenAI in meaningful volume; it sells to the hyperscalers and infrastructure providers who serve OpenAI. But OpenAI's compute demand is the single largest driver of hyperscaler GPU procurement. CEO Jensen Huang stated repeatedly during Nvidia's Q2 2026 earnings call that the top four hyperscalers — Amazon, Google, Microsoft, and Meta — would spend roughly $600 billion annually on capex, much of it directed at AI infrastructure to serve workloads anchored by OpenAI and similar foundation model companies. If that anchor demand softens, GPU procurement decisions get re-evaluated. Nvidia's revenue concentration in hyperscaler customers is not a secret, but the second-order dependency on OpenAI specifically is.
Capex 2024–2026: The Trillion-Dollar Bet
The aggregate capex numbers tell their own story. In 2025, the four primary hyperscalers — Microsoft, Alphabet, Amazon, and Meta — spent approximately $381 billion on capital expenditure, the vast majority directed at AI infrastructure. For 2026, current guidance ranges between $635 billion and $700 billion, depending on where in the guidance ranges each company lands. Specifically, Alphabet has guided $175–185 billion, Amazon roughly $200 billion, Meta $115–135 billion, and Microsoft is tracking toward $120 billion or more on a fiscal-year basis. Oracle adds another roughly $50 billion. The combined number sits at around $660–690 billion.
This is the largest concentrated capital deployment in technology history. To put it in context: combined Q1 2026 AI-related capex from the Magnificent Seven is estimated at $78 billion — a 45% year-over-year increase. The four hyperscalers are expected to issue more than $400 billion in new debt in 2026 to finance the buildout, more than double the $165 billion raised in 2025, according to Morgan Stanley. Total issuance of investment-grade U.S. corporate bonds is expected to reach a record $2.25 trillion this year, with roughly 18% of that coming from hyperscaler AI financing.
The free cash flow implications are starting to matter. Barclays models a near-90% drop in Meta's 2026 free cash flow. Morgan Stanley projects negative free cash flow of approximately $17 billion at Amazon in 2026; Bank of America's estimate is closer to negative $28 billion. Microsoft is expected to see a 28% decline in free cash flow before recovery in 2027. None of this is fatal. All four companies have balance sheets and operating cash flow that can absorb the spend. But the quality of the spend now depends on the quality of the demand, and the WSJ report has put exactly that question on the table.
Echoes of Bubbles Past — and Differences
The natural reference point for any capex-driven equity story is the 1999–2001 telecom and dotcom buildout. The parallels are not subtle. In 1999, telecom carriers and dotcom-era infrastructure providers committed to fiber, switching, and data center investments based on demand projections that turned out to be roughly five years ahead of reality. The aggregate over-investment was estimated at over $500 billion. The unwind crushed equity valuations across telecom, networking equipment, and component suppliers, with companies like Lucent, Nortel, JDSU, and Global Crossing losing 90% or more of their market value.
What is different this time is the financing structure. The dotcom-era capex was substantially debt-financed by carriers and project-financed by special purpose vehicles, with very little equity cushion at the buildout-company level. The 2026 hyperscaler capex is being funded by some of the most cash-generative businesses in human history. Microsoft, Google, Meta, and Amazon collectively generated over $400 billion in operating cash flow in 2025. Even with the dramatic free cash flow compressions analysts are modelling, none of these companies face a solvency question. This is not a Lucent setup.
What is similar is the concentration risk. The dotcom build was not justified solely by demand from one customer; it was justified by demand from one category — internet adoption — that grew slower than expected. The 2026 build is justified by AI workload growth, with OpenAI as the most visible single demand driver. If AI workload growth disappoints, the failure mode is closer to slow utilization ramp than to outright bankruptcy — but the equity multiple compression on names like Oracle, CoreWeave, and the second-tier infrastructure providers could still be severe.
The Bear Case
The bear case does not require believing OpenAI fails. It requires believing that OpenAI grows somewhat slower than expected, that competitive pressure from Anthropic and Google compresses pricing, and that enterprise AI adoption converts to revenue more slowly than current capex implies. In this scenario, hyperscaler utilization rates on 2026-built infrastructure underwhelm in 2027. Cloud-revenue acceleration disappoints. Multiple compressions hit the names with the most concentrated AI exposure first — Oracle, CoreWeave, AMD's data center segment — before spreading to Nvidia and the broader semiconductor complex.
Importantly, this scenario does not require a recession or a credit event. It just requires the demand-supply gap to close in the wrong direction for one to two years. The historical analog is the 2001-2003 telecom-equipment-multiple compression, which happened against a backdrop of moderate economic growth, not financial crisis. Equity multiples on capex-heavy AI names could halve from current levels in this scenario, even with the underlying companies remaining profitable.
The Bull Case
The bull case is structurally simpler and currently better priced into stocks. Enterprise AI adoption is genuinely accelerating. Inference demand — the actual production usage of AI models, as opposed to training — is growing faster than most analysts modeled even a year ago. OpenAI's GPT-5.5 deployment, expanded Codex availability, and emerging advertising business are all real revenue drivers. Microsoft's Azure AI revenue and Google Cloud's AI-attached revenue have both accelerated in recent quarters. If these trends continue, the 2026 capex eventually pays off through 2028 utilization, and free cash flow recovers in 2027 as analysts already model.
The bull case also accommodates the WSJ report. OpenAI growing at "only" 3x annually, falling short of its own aggressive internal targets, is fully consistent with the bull thesis on aggregate AI infrastructure demand. Anthropic and Google taking share from OpenAI is bearish for OpenAI specifically, but largely neutral for total compute demand. The infrastructure thesis was never about OpenAI winning monopoly share. It was about total AI workloads growing fast enough to justify the buildout.
What Retail Investors Should Watch
Five concrete metrics are worth tracking over the next three quarters. First, Mag7 capex guidance language at the upcoming earnings — specifically, whether companies defend their numbers with customer-commitment language or with model-confidence language. Second, Nvidia's data center segment revenue growth and customer concentration disclosures in the next earnings cycle. Third, Oracle's remaining performance obligations (RPO) figure, which is the cleanest single indicator of contracted compute demand. Fourth, any disclosure from OpenAI, formal or informal, about updated revenue trajectory or compute deployment pace. Fifth, Microsoft Azure AI revenue growth disclosed during earnings — Microsoft is the most visible window into the OpenAI demand picture.
For position sizing, the dependency mapping matters. Direct OpenAI-exposure names — Oracle, CoreWeave — carry the highest concentration risk. Hyperscaler names with AI exposure but diversified customer bases — Microsoft, Alphabet, Meta, Amazon — carry moderate risk. Semiconductor names sit in the middle, with Nvidia carrying more concentration risk than AMD or Broadcom because of its hyperscaler customer base. Diversification across the dependency chain, rather than across "AI stocks" generically, is the relevant risk control.
Bottom Line
The WSJ report did not break the AI infrastructure thesis. It reminded the market that the thesis has dependencies. Two years of one-way capex announcements created a market psychology in which any AI-adjacent name could be bought without examining its specific revenue conversion path. That psychology is now incrementally less robust. The companies with strong fundamentals and diversified demand will be fine. The names that were priced for an OpenAI monopoly outcome will not. As always, the difference is in the details — and the details are about to be tested in real time during the next two earnings cycles.